2017-08-08 16:36:23 -07:00
# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements. See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership. The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License. You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied. See the License for the
# specific language governing permissions and limitations
# under the License.
2018-01-30 10:45:25 -08:00
from __future__ import print_function
2017-04-03 15:18:41 -07:00
import sys
2016-03-19 23:45:52 -07:00
import os
2017-06-26 22:37:11 -07:00
import time
2018-01-30 10:45:25 -08:00
import multiprocessing as mp
2017-08-28 20:09:41 -07:00
import unittest
2017-06-26 22:37:11 -07:00
import mxnet as mx
import numpy as np
2017-10-14 19:44:32 -07:00
import unittest
2018-04-03 10:33:56 -07:00
from nose . tools import assert_raises
2017-07-28 16:26:28 -07:00
from mxnet . test_utils import check_consistency , set_default_context , assert_almost_equal
2018-04-03 10:33:56 -07:00
from mxnet . base import MXNetError
2018-04-09 14:43:53 -07:00
from mxnet import autograd
2017-06-26 22:37:11 -07:00
from numpy . testing import assert_allclose
2016-03-19 23:45:52 -07:00
curr_path = os . path . dirname ( os . path . abspath ( os . path . expanduser ( __file__ ) ) )
sys . path . insert ( 0 , os . path . join ( curr_path , ' ../unittest ' ) )
2018-02-18 03:11:58 -08:00
from common import setup_module , with_seed
2015-10-24 15:57:42 -07:00
from test_operator import *
2017-02-07 13:37:43 +08:00
from test_optimizer import *
2017-04-27 12:14:37 -07:00
from test_random import *
2017-08-06 23:20:59 -07:00
from test_gluon import *
2017-08-24 12:02:41 -07:00
from test_loss import *
2018-02-13 11:13:04 -08:00
from test_exc_handling import *
2017-06-26 22:37:11 -07:00
#from test_rnn import *
2017-07-10 10:55:46 -07:00
from test_gluon_rnn import *
2017-10-06 14:51:26 -07:00
from test_sparse_ndarray import test_create_csr , test_create_row_sparse , test_sparse_nd_slice
2017-10-24 21:09:05 -07:00
from test_sparse_ndarray import test_create_sparse_nd_empty , test_create_sparse_nd_from_sparse
from test_sparse_ndarray import test_create_sparse_nd_from_dense , test_create_sparse_nd_infer_shape
2017-12-02 02:47:22 +08:00
from test_sparse_ndarray import test_sparse_nd_check_format , test_sparse_nd_copy
2018-01-04 07:56:18 +08:00
from test_sparse_ndarray import test_sparse_nd_setitem , test_sparse_nd_binary_scalar_op
2017-08-30 23:12:06 -07:00
from test_sparse_operator import *
from test_ndarray import *
2016-03-19 23:45:52 -07:00
2016-10-19 00:06:32 -07:00
set_default_context ( mx . gpu ( 0 ) )
del test_support_vector_machine_l1_svm
del test_support_vector_machine_l2_svm
2017-04-03 15:18:41 -07:00
def check_countsketch ( in_dim , out_dim , n ) :
2017-08-29 10:34:56 -07:00
sym = mx . sym . contrib . count_sketch ( name = ' countsketch ' , out_dim = out_dim )
2017-04-03 15:18:41 -07:00
shape = [ ( n , in_dim ) , ( 1 , in_dim ) , ( 1 , in_dim ) ] #shape of input x, hash h and hash s
arr = [ mx . nd . empty ( shape [ i ] ) for i in range ( 3 ) ]
arr_grad = [ mx . nd . empty ( shape [ i ] ) for i in range ( 3 ) ]
x = np . random . uniform ( - 10 , 10 , shape [ 0 ] )
arr [ 0 ] [ : ] = x #input x
h = np . random . randint ( 0 , out_dim , shape [ 1 ] )
arr [ 1 ] [ : ] = h #hash h
s = np . random . randint ( 0 , 2 , shape [ 2 ] ) * 2 - np . ones ( shape [ 2 ] )
arr [ 2 ] [ : ] = s #hash s
# forward
exe_list = [ sym . bind ( mx . gpu ( 0 ) , arr , arr_grad ) ]
for exe in exe_list :
exe . forward ( is_train = True )
out1 = [ exe . outputs [ 0 ] . asnumpy ( ) for exe in exe_list ]
2017-04-27 12:14:37 -07:00
2017-04-03 15:18:41 -07:00
a = np . zeros ( ( n , out_dim ) )
temp = np . multiply ( x , s )
for num_sample in np . arange ( 0 , n ) :
for idx in np . arange ( 0 , in_dim ) :
a [ num_sample ] [ h [ 0 ] [ idx ] ] + = temp [ num_sample ] [ idx ]
assert_almost_equal ( a , out1 [ 0 ] , rtol = 1e-3 , atol = 1e-12 )
2017-04-27 12:14:37 -07:00
2017-04-03 15:18:41 -07:00
# backward
out_grad = mx . nd . empty ( ( n , out_dim ) )
out_grad [ : ] = np . random . normal ( - 3 , 3 , ( n , out_dim ) )
for exe in exe_list :
2017-04-27 12:14:37 -07:00
exe . backward ( [ out_grad ] )
2017-04-03 15:18:41 -07:00
a = np . zeros ( ( n , in_dim ) )
for j in np . arange ( 0 , n ) :
for i in np . arange ( 0 , in_dim ) :
a [ j , i ] = out_grad . asnumpy ( ) [ j , h [ 0 , i ] ] * s [ 0 , i ]
assert_almost_equal ( a , arr_grad [ 0 ] . asnumpy ( ) , rtol = 1e-3 , atol = 1e-12 )
2017-04-27 12:14:37 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( 0 )
2017-04-03 15:18:41 -07:00
def test_countsketch ( ) :
nrepeat = 2
minindim = 40
maxindim = 100
minoutdim = 5
maxoutdim = 30
maxn = 200
for repeat in range ( nrepeat ) :
in_dim = np . random . randint ( minindim , maxindim )
out_dim = np . random . randint ( minoutdim , maxoutdim )
n = np . random . randint ( 1 , maxn )
check_countsketch ( in_dim , out_dim , n )
2018-02-18 03:11:58 -08:00
2017-04-03 15:18:41 -07:00
def check_ifft ( shape ) :
shape_old = shape
if len ( shape ) == 2 :
if shape [ 1 ] % 2 != 0 :
lst = list ( shape )
lst [ 1 ] = lst [ 1 ] * 2
shape = tuple ( lst )
shape_old = shape
shape = ( shape [ 0 ] , shape [ 1 ] * 2 )
if len ( shape ) == 4 :
if shape [ 3 ] % 2 != 0 :
lst = list ( shape )
lst [ 3 ] = lst [ 3 ] * 2
shape = tuple ( lst )
shape_old = shape
shape = ( shape [ 0 ] , shape [ 1 ] , shape [ 2 ] , shape [ 3 ] * 2 )
2017-08-29 10:34:56 -07:00
sym = mx . sym . contrib . ifft ( name = ' ifft ' , compute_size = 128 )
2017-04-03 15:18:41 -07:00
init = [ np . random . normal ( size = shape , scale = 1.0 ) ]
arr_grad = [ mx . nd . empty ( shape ) ]
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' ifft_data ' : shape , ' type_dict ' : { ' ifft_data ' : np . float32 } } ]
exe_list = [ sym . simple_bind ( args_grad = arr_grad , * * ctx ) for ctx in ctx_list ]
2017-04-27 12:14:37 -07:00
2017-04-03 15:18:41 -07:00
for exe in exe_list :
for arr , iarr in zip ( exe . arg_arrays , init ) :
arr [ : ] = iarr . astype ( arr . dtype )
# forward
for exe in exe_list :
exe . forward ( is_train = True )
out1 = [ exe . outputs [ 0 ] . asnumpy ( ) for exe in exe_list ]
2017-04-27 12:14:37 -07:00
2017-04-03 15:18:41 -07:00
if len ( shape ) == 2 :
init_complex = np . zeros ( shape_old , dtype = np . complex64 )
for i in range ( 0 , shape_old [ 1 ] ) :
init_complex . real [ : , i ] = init [ 0 ] [ : , 2 * i ]
init_complex . imag [ : , i ] = init [ 0 ] [ : , 2 * i + 1 ]
a = np . fft . ifft ( init_complex , n = None , axis = - 1 , norm = None )
assert_almost_equal ( a . real , out1 [ 0 ] / shape_old [ 1 ] , rtol = 1e-3 , atol = 1e-12 )
2017-04-27 12:14:37 -07:00
2017-04-03 15:18:41 -07:00
if len ( shape ) == 4 :
init_complex = np . zeros ( shape_old , dtype = np . complex64 )
for i in range ( 0 , shape_old [ 3 ] ) :
init_complex . real [ : , : , : , i ] = init [ 0 ] [ : , : , : , 2 * i ]
init_complex . imag [ : , : , : , i ] = init [ 0 ] [ : , : , : , 2 * i + 1 ]
a = np . fft . ifft ( init_complex , n = None , axis = - 1 , norm = None )
assert_almost_equal ( a . real , out1 [ 0 ] / shape_old [ 3 ] , rtol = 1e-3 , atol = 1e-12 )
# backward
if len ( shape ) == 2 :
out_grad = mx . nd . empty ( shape_old )
out_grad [ : ] = np . random . normal ( - 3 , 3 , shape_old )
for exe in exe_list :
exe . backward ( [ out_grad ] )
temp = exe . grad_arrays [ 0 ] . asnumpy ( )
temp = np . zeros ( shape_old )
for i in range ( shape_old [ 1 ] ) :
temp [ : , i ] = exe . grad_arrays [ 0 ] . asnumpy ( ) [ : , 2 * i ]
2017-04-27 12:14:37 -07:00
2017-04-03 15:18:41 -07:00
a = np . fft . fft ( out_grad . asnumpy ( ) , n = None , axis = - 1 , norm = None )
assert_almost_equal ( a . real , temp , rtol = 1e-3 , atol = 1e-12 )
if len ( shape ) == 4 :
out_grad = mx . nd . empty ( shape_old )
out_grad [ : ] = np . random . normal ( - 3 , 3 , shape_old )
for exe in exe_list :
exe . backward ( [ out_grad ] )
temp = exe . grad_arrays [ 0 ] . asnumpy ( )
temp = np . zeros ( shape_old )
for i in range ( shape_old [ 3 ] ) :
temp [ : , : , : , i ] = exe . grad_arrays [ 0 ] . asnumpy ( ) [ : , : , : , 2 * i ]
2017-04-27 12:14:37 -07:00
2017-04-03 15:18:41 -07:00
a = np . fft . fft ( out_grad . asnumpy ( ) , n = None , axis = - 1 , norm = None )
assert_almost_equal ( a . real , temp , rtol = 1e-3 , atol = 1e-12 )
2017-04-27 12:14:37 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( 0 )
2017-04-03 15:18:41 -07:00
def test_ifft ( ) :
nrepeat = 2
maxdim = 10
for repeat in range ( nrepeat ) :
for order in [ 2 , 4 ] :
shape = tuple ( np . random . randint ( 1 , maxdim , size = order ) )
check_ifft ( shape )
2018-02-18 03:11:58 -08:00
2017-04-03 15:18:41 -07:00
def check_fft ( shape ) :
2017-08-29 10:34:56 -07:00
sym = mx . sym . contrib . fft ( name = ' fft ' , compute_size = 128 )
2017-04-03 15:18:41 -07:00
if len ( shape ) == 2 :
if shape [ 1 ] % 2 != 0 :
lst = list ( shape )
lst [ 1 ] = lst [ 1 ] * 2
shape = tuple ( lst )
shape_old = shape
if len ( shape ) == 4 :
if shape [ 3 ] % 2 != 0 :
lst = list ( shape )
lst [ 3 ] = lst [ 3 ] * 2
shape = tuple ( lst )
shape_old = shape
init = [ np . random . normal ( size = shape , scale = 1.0 ) ]
arr_grad = [ mx . nd . empty ( shape ) ]
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' fft_data ' : shape , ' type_dict ' : { ' fft_data ' : np . float32 } } ]
exe_list = [ sym . simple_bind ( args_grad = arr_grad , * * ctx ) for ctx in ctx_list ]
2017-04-27 12:14:37 -07:00
2017-04-03 15:18:41 -07:00
for exe in exe_list :
for arr , iarr in zip ( exe . arg_arrays , init ) :
arr [ : ] = iarr . astype ( arr . dtype )
#forward
for exe in exe_list :
exe . forward ( is_train = True )
out1 = [ exe . outputs [ 0 ] . asnumpy ( ) for exe in exe_list ]
out = np . fft . fft ( init , n = None , axis = - 1 , norm = None )
if len ( shape ) == 2 :
out = np . reshape ( out , ( out . shape [ 1 ] , out . shape [ 2 ] ) )
out2 = np . append ( out . real , out . imag , axis = 1 )
a = np . zeros ( out1 [ 0 ] . shape )
p = 0
for i in range ( out2 . shape [ 1 ] / / 2 ) :
a [ : , p ] = out2 [ : , i ]
a [ : , p + 1 ] = out2 [ : , i + out2 . shape [ 1 ] / / 2 ]
p = p + 2
2017-04-27 12:14:37 -07:00
2017-04-03 15:18:41 -07:00
if len ( shape ) == 4 :
out = np . reshape ( out , ( out . shape [ 1 ] , out . shape [ 2 ] , out . shape [ 3 ] , out . shape [ 4 ] ) )
out2 = np . append ( out . real , out . imag , axis = 1 )
a = np . zeros ( out1 [ 0 ] . shape )
for i in range ( out1 [ 0 ] . shape [ 0 ] ) :
for j in range ( out1 [ 0 ] . shape [ 1 ] ) :
p = 0
for k in range ( out2 . shape [ 3 ] ) :
a [ i , j , : , p ] = out2 [ i , j , : , k ]
a [ i , j , : , p + 1 ] = out2 [ i , j + out1 [ 0 ] . shape [ 1 ] , : , k ]
p = p + 2
2017-04-27 12:14:37 -07:00
2017-04-03 15:18:41 -07:00
assert_almost_equal ( a , out1 [ 0 ] , rtol = 1e-3 , atol = 1e-6 )
2017-04-27 12:14:37 -07:00
2017-04-03 15:18:41 -07:00
# backward
if len ( shape ) == 2 :
out_grad = mx . nd . empty ( ( shape [ 0 ] , 2 * shape [ 1 ] ) )
out_grad [ : ] = np . random . normal ( - 3 , 3 , ( shape [ 0 ] , 2 * shape [ 1 ] ) )
# out_grad_to_complex
out_grad_complex = np . zeros ( shape , dtype = np . complex64 )
for i in range ( 0 , shape [ 1 ] ) :
out_grad_complex . real [ : , i ] = out_grad . asnumpy ( ) [ : , 2 * i ]
out_grad_complex . imag [ : , i ] = out_grad . asnumpy ( ) [ : , 2 * i + 1 ]
for exe in exe_list :
2017-04-27 12:14:37 -07:00
exe . backward ( [ out_grad ] )
2017-04-03 15:18:41 -07:00
a = np . fft . ifft ( out_grad_complex , n = None , axis = - 1 , norm = None )
assert_almost_equal ( a . real , exe . grad_arrays [ 0 ] . asnumpy ( ) / shape [ 1 ] , rtol = 1e-3 , atol = 1e-8 )
2017-04-27 12:14:37 -07:00
2017-04-03 15:18:41 -07:00
if len ( shape ) == 4 :
out_grad = mx . nd . empty ( out1 [ 0 ] . shape )
out_grad [ : ] = np . random . normal ( - 3 , 3 , out1 [ 0 ] . shape )
# out_grad_to_complex
out_grad_complex = np . zeros ( shape , dtype = np . complex64 )
for i in range ( 0 , shape [ 3 ] ) :
out_grad_complex . real [ : , : , : , i ] = out_grad . asnumpy ( ) [ : , : , : , 2 * i ]
out_grad_complex . imag [ : , : , : , i ] = out_grad . asnumpy ( ) [ : , : , : , 2 * i + 1 ]
for exe in exe_list :
2017-04-27 12:14:37 -07:00
exe . backward ( [ out_grad ] )
2017-04-03 15:18:41 -07:00
a = np . fft . ifft ( out_grad_complex , n = None , axis = - 1 , norm = None )
assert_almost_equal ( a . real , exe . grad_arrays [ 0 ] . asnumpy ( ) / shape [ 3 ] , rtol = 1e-3 , atol = 1e-6 )
2018-02-18 03:11:58 -08:00
@with_seed ( 0 )
2017-04-03 15:18:41 -07:00
def test_fft ( ) :
nrepeat = 2
maxdim = 10
for repeat in range ( nrepeat ) :
for order in [ 2 , 4 ] :
shape = tuple ( np . random . randint ( 1 , maxdim , size = order ) )
check_fft ( shape )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2016-07-05 11:29:40 -07:00
def test_batchnorm_with_type ( ) :
Batch Norm rewrite without mshadow, 1D, 2D, 3D, float16, float32, float64 as well as operator gtest framework (#5936)
* Batch Norm rewrite without mshadow as well as operator gtest framework
* performance testing
* lint fixes
* use CUDNN for this test
* remove superfluous omp define
* Fix file names in comments
* build, run, clean gtest works (although a test is failing)
* CR comments
* Adjust timing tests for more strenuous sample
* Remove temp resource allocation
* DeviceTensor3 added, forEachFast not yet converted
* DeviceTensor3 version working
* DeviceTensor3 working
* .
* Fix for use_global_stats
* fixed bug with testing suite for double (Float64)
* python unit tests working for batchnorm
* python unit tests
* Update documentation for mxnet.initializer.Mixed (#5937)
* Update documentation for SVMOutput. (#5931)
* Update documentation for SVMOutput.
* Update doc for SVMOutput - fix formatting.
* Adding install instruction for Ubuntu-CPU-Python (#5885)
* edit ndarray API docs (#5806)
* edit docs in broadcast_reduce_op
* edit docs in broadcast_reduce_op
* minor change
* lint fix
* fix
* mx.nd.ones
* mx.nd.repeat
* mx.nd.reverse
* add example in repeat
* optimizer update
* fix nanprod
* fix optimizer_op api doc
* fix reduce_op api doc
* fix nd.ones api doc
* mx.nd.repeat doc change
* Update broadcast_reduce_op.h
* Symbol docs fixes (#5930)
* symbol docs minor formatting changes
* deepcopy, infer_shape, infer_shape_partial docs modified
* Few more small fixes
* arithmetic functions fixes
* some more modifications
* changes after review
* small change
* grad function note added
* More API Doc Edits (#5886)
* edit activation doc
* doc l2_normalization
* edit MakeLoss doc
* edit blockgrad doc
* blockgrad fileline fix
* edit MakeLoss doc cont.
* doc change 'tensor' to 'multidimensional array'
* l2normalization doc improve
* makeloss doc improve, blockgrad doc improve
* fix doc in activation, l2_normalization, make_loss
* fix minor grammar
* use .describe to avoid build failure.
* Update documentation for mxnet.image.imdecode (#5957)
* Update documentation for mxnet.image.imdecode
* Update documentation for mxnet.image.imdecode (clarify that we need OpenCV and not the CV2 Python library)
* Fix script by adding path to Dockerfile (#5958)
* Clean install script
* Add test for pip installations
* Remove debug statements & comments
* Make test runnable as script and from framework
* Fix path to Dockerfiles
* Putting failing cases at the end
* Update doc for Custom operator. (#5875)
* Update doc for Custom operator.
* Update doc for Custom operator.
* Fix formating in doc for Custom operator.
* Fix formating in doc for Custom operator.
* Minor change to ndarray.Custom documentation.
* Minor edit in doc for Custom operator.
* Minor change to doc for Custom operator. Data is 'NDArray-or-Symbol'.
* Minor formatting change for Custom operator documentation.
* For Custom operator doc, move example into ndarray_doc.py.
* Minor change in Custom operator documentation
* Improve the doc of pick + Update dmlc-core (#5946)
* Add PickParam to fix the docstring and the initial value for axis
* Update dmlc-core
* Update dmlc-core
* Image docs modified (#5973)
* imageIter doc modified
* edited imageiter
* ADD missing Libri_sample.json, FIX minor bugs in speech_recognition example (#5962)
* [KVStore] Add support for other data types (#5818)
* Fix kvstore type
* Fix lint
* Parse inputs to DataDesc
* Make module support dtype
* Fix lint
* Add default dtype in Comm
* Fix lint
* Revert rename
* [cpp-package] Add C++ basic tutorial and build instruction (#5971)
* Add C++ basic tutorial and build instruction
* Remove binaries
* Fix lint
* Avoid sign-compare
* Update documentation for mxnet.metric.np (#5977)
* Getting rid of identity (#5935)
* Activation ops (#5938)
* [Ops] Add op: 'relu'
* Add op: 'sigmoid'
* Introduce 'kernel_launch_op'
* Add tests and describe; move it to elemwise_unary_op
* Fix GPU version
* Convert caffe AbsVal to mx.symbol.abs in caffe converter (#5984)
* Correction to LSTMCell docstring (#5986)
* [Module] fix input_grads order (#5980)
* fix input_grads order + update dmlc-core
* set label to be optional
* update env_var doc (#5964)
* Adjusting make, Callback removed
* batch norm gpu testing
* Batch Norm rewrite without mshadow as well as operator gtest framework
* performance testing
* lint fixes
* use CUDNN for this test
* remove superfluous omp define
* Fix file names in comments
* build, run, clean gtest works (although a test is failing)
* CR comments
* Adjust timing tests for more strenuous sample
* Remove temp resource allocation
* rearrange source into cc and cu files
* lint fixes
* Trigger build
* Use latest mshadow
* temporarily revert channel position parameter field
* Add more tests for batchnorm
* Add more tests for batchnorm
* test_operator_gpu working for all types
* Compiles after AccReal
* Compiles after AccReal
* All tests working
* All tests working
* build, run, clean gtest works (although a test is failing)
* vc++ requires explicit int type for omp for loop
* Repair cpp-package
* signed/unsigned fixed in cuda file
* lint fixes in tests and cpp-package directories
* more lint
* use IsWriting() helper
* Fall-through for unsupported MKL shapes/types
* Fall-through for unsupported MKL shapes/types
* cleaner mkl_off approach
* Warning only whem MKL is requested
* Warning only whem MKL is requested
* lint
* ..
* python problem fixed
* python problem fixed
* Merge branch 'batchnorm' into batchnorm_pr
# Conflicts:
# src/operator/batch_norm.cc
# src/operator/batch_norm.cu
# tests/cpp/operator/batchnorm_test.cc
* lint fix
* lint fix
* lint fix
* lint fix
* lint fix
* Fix visual c++ compile problem
* .
* .
* All unit tests pass again
* lint fix
* fix strange compile errors in CUDNN batchnorm header
* FInish using flags instead of bools
* lint
* Fix timing pass count for forward pass
* Fix R script install roxygen problem
* code formatting, addition of doc strings is causing IDE to add spaces before the calls
* removed commented
* cr comments
* Change back to compilable code
* For CPU mode, store as invstd
* move testing code around a little
* lint fix
* Use AccReal in some places to avoid fp16 problems
* Fix minor invstd problem in cuda version
* remove unused scale param
* add permutation unit test, handle cudnn doesn't like 3D
* .
* lint
* .
* Remove mkl_off
* lint fix and time cudnn when enabled
2017-05-15 20:27:28 -07:00
ctx_list_v1_2D = [
{ ' ctx ' : mx . cpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float32 } } ,
]
ctx_list_v2_2D = [
{ ' ctx ' : mx . cpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float32 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float16 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float64 } } ,
]
ctx_list_v2_1D = [
{ ' ctx ' : mx . cpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float32 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float16 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' norm_data ' : ( 10 , 2 , 10 ) , ' type_dict ' : { ' norm_data ' : np . float64 } } ,
]
ctx_list_v2_3D = [
{ ' ctx ' : mx . cpu ( 0 ) , ' norm_data ' : ( 4 , 2 , 3 , 5 , 5 ) , ' type_dict ' : { ' norm_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' norm_data ' : ( 4 , 2 , 3 , 5 , 5 ) , ' type_dict ' : { ' norm_data ' : np . float32 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' norm_data ' : ( 4 , 2 , 3 , 5 , 5 ) , ' type_dict ' : { ' norm_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' norm_data ' : ( 4 , 2 , 3 , 5 , 5 ) , ' type_dict ' : { ' norm_data ' : np . float16 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' norm_data ' : ( 4 , 2 , 3 , 5 , 5 ) , ' type_dict ' : { ' norm_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' norm_data ' : ( 4 , 2 , 3 , 5 , 5 ) , ' type_dict ' : { ' norm_data ' : np . float64 } }
]
# V1, 2D
sym = mx . sym . BatchNorm_v1 ( name = ' norm ' , fix_gamma = False )
check_consistency ( sym , ctx_list_v1_2D )
sym = mx . sym . BatchNorm_v1 ( name = ' norm ' , fix_gamma = True )
check_consistency ( sym , ctx_list_v1_2D )
# V2, 2D
sym = mx . sym . BatchNorm ( name = ' norm ' , fix_gamma = False , cudnn_off = True )
check_consistency ( sym , ctx_list_v2_2D )
sym = mx . sym . BatchNorm ( name = ' norm ' , fix_gamma = False , cudnn_off = True )
check_consistency ( sym , ctx_list_v2_2D )
sym = mx . sym . BatchNorm ( name = ' norm ' , fix_gamma = True , cudnn_off = True )
check_consistency ( sym , ctx_list_v2_2D )
sym = mx . sym . BatchNorm ( name = ' norm ' , fix_gamma = True , cudnn_off = True )
check_consistency ( sym , ctx_list_v2_2D )
# V2, 1D
sym = mx . sym . BatchNorm ( name = ' norm ' , fix_gamma = False , cudnn_off = True )
check_consistency ( sym , ctx_list_v2_1D )
sym = mx . sym . BatchNorm ( name = ' norm ' , fix_gamma = False , cudnn_off = True )
check_consistency ( sym , ctx_list_v2_1D )
sym = mx . sym . BatchNorm ( name = ' norm ' , fix_gamma = True , cudnn_off = True )
check_consistency ( sym , ctx_list_v2_1D )
sym = mx . sym . BatchNorm ( name = ' norm ' , fix_gamma = True , cudnn_off = True )
check_consistency ( sym , ctx_list_v2_1D )
#
# # V2, 3D
sym = mx . sym . BatchNorm ( name = ' norm ' , fix_gamma = False , cudnn_off = True )
check_consistency ( sym , ctx_list_v2_3D )
sym = mx . sym . BatchNorm ( name = ' norm ' , fix_gamma = True , cudnn_off = True )
check_consistency ( sym , ctx_list_v2_3D )
2018-02-18 03:11:58 -08:00
@with_seed ( )
Batch Norm rewrite without mshadow, 1D, 2D, 3D, float16, float32, float64 as well as operator gtest framework (#5936)
* Batch Norm rewrite without mshadow as well as operator gtest framework
* performance testing
* lint fixes
* use CUDNN for this test
* remove superfluous omp define
* Fix file names in comments
* build, run, clean gtest works (although a test is failing)
* CR comments
* Adjust timing tests for more strenuous sample
* Remove temp resource allocation
* DeviceTensor3 added, forEachFast not yet converted
* DeviceTensor3 version working
* DeviceTensor3 working
* .
* Fix for use_global_stats
* fixed bug with testing suite for double (Float64)
* python unit tests working for batchnorm
* python unit tests
* Update documentation for mxnet.initializer.Mixed (#5937)
* Update documentation for SVMOutput. (#5931)
* Update documentation for SVMOutput.
* Update doc for SVMOutput - fix formatting.
* Adding install instruction for Ubuntu-CPU-Python (#5885)
* edit ndarray API docs (#5806)
* edit docs in broadcast_reduce_op
* edit docs in broadcast_reduce_op
* minor change
* lint fix
* fix
* mx.nd.ones
* mx.nd.repeat
* mx.nd.reverse
* add example in repeat
* optimizer update
* fix nanprod
* fix optimizer_op api doc
* fix reduce_op api doc
* fix nd.ones api doc
* mx.nd.repeat doc change
* Update broadcast_reduce_op.h
* Symbol docs fixes (#5930)
* symbol docs minor formatting changes
* deepcopy, infer_shape, infer_shape_partial docs modified
* Few more small fixes
* arithmetic functions fixes
* some more modifications
* changes after review
* small change
* grad function note added
* More API Doc Edits (#5886)
* edit activation doc
* doc l2_normalization
* edit MakeLoss doc
* edit blockgrad doc
* blockgrad fileline fix
* edit MakeLoss doc cont.
* doc change 'tensor' to 'multidimensional array'
* l2normalization doc improve
* makeloss doc improve, blockgrad doc improve
* fix doc in activation, l2_normalization, make_loss
* fix minor grammar
* use .describe to avoid build failure.
* Update documentation for mxnet.image.imdecode (#5957)
* Update documentation for mxnet.image.imdecode
* Update documentation for mxnet.image.imdecode (clarify that we need OpenCV and not the CV2 Python library)
* Fix script by adding path to Dockerfile (#5958)
* Clean install script
* Add test for pip installations
* Remove debug statements & comments
* Make test runnable as script and from framework
* Fix path to Dockerfiles
* Putting failing cases at the end
* Update doc for Custom operator. (#5875)
* Update doc for Custom operator.
* Update doc for Custom operator.
* Fix formating in doc for Custom operator.
* Fix formating in doc for Custom operator.
* Minor change to ndarray.Custom documentation.
* Minor edit in doc for Custom operator.
* Minor change to doc for Custom operator. Data is 'NDArray-or-Symbol'.
* Minor formatting change for Custom operator documentation.
* For Custom operator doc, move example into ndarray_doc.py.
* Minor change in Custom operator documentation
* Improve the doc of pick + Update dmlc-core (#5946)
* Add PickParam to fix the docstring and the initial value for axis
* Update dmlc-core
* Update dmlc-core
* Image docs modified (#5973)
* imageIter doc modified
* edited imageiter
* ADD missing Libri_sample.json, FIX minor bugs in speech_recognition example (#5962)
* [KVStore] Add support for other data types (#5818)
* Fix kvstore type
* Fix lint
* Parse inputs to DataDesc
* Make module support dtype
* Fix lint
* Add default dtype in Comm
* Fix lint
* Revert rename
* [cpp-package] Add C++ basic tutorial and build instruction (#5971)
* Add C++ basic tutorial and build instruction
* Remove binaries
* Fix lint
* Avoid sign-compare
* Update documentation for mxnet.metric.np (#5977)
* Getting rid of identity (#5935)
* Activation ops (#5938)
* [Ops] Add op: 'relu'
* Add op: 'sigmoid'
* Introduce 'kernel_launch_op'
* Add tests and describe; move it to elemwise_unary_op
* Fix GPU version
* Convert caffe AbsVal to mx.symbol.abs in caffe converter (#5984)
* Correction to LSTMCell docstring (#5986)
* [Module] fix input_grads order (#5980)
* fix input_grads order + update dmlc-core
* set label to be optional
* update env_var doc (#5964)
* Adjusting make, Callback removed
* batch norm gpu testing
* Batch Norm rewrite without mshadow as well as operator gtest framework
* performance testing
* lint fixes
* use CUDNN for this test
* remove superfluous omp define
* Fix file names in comments
* build, run, clean gtest works (although a test is failing)
* CR comments
* Adjust timing tests for more strenuous sample
* Remove temp resource allocation
* rearrange source into cc and cu files
* lint fixes
* Trigger build
* Use latest mshadow
* temporarily revert channel position parameter field
* Add more tests for batchnorm
* Add more tests for batchnorm
* test_operator_gpu working for all types
* Compiles after AccReal
* Compiles after AccReal
* All tests working
* All tests working
* build, run, clean gtest works (although a test is failing)
* vc++ requires explicit int type for omp for loop
* Repair cpp-package
* signed/unsigned fixed in cuda file
* lint fixes in tests and cpp-package directories
* more lint
* use IsWriting() helper
* Fall-through for unsupported MKL shapes/types
* Fall-through for unsupported MKL shapes/types
* cleaner mkl_off approach
* Warning only whem MKL is requested
* Warning only whem MKL is requested
* lint
* ..
* python problem fixed
* python problem fixed
* Merge branch 'batchnorm' into batchnorm_pr
# Conflicts:
# src/operator/batch_norm.cc
# src/operator/batch_norm.cu
# tests/cpp/operator/batchnorm_test.cc
* lint fix
* lint fix
* lint fix
* lint fix
* lint fix
* Fix visual c++ compile problem
* .
* .
* All unit tests pass again
* lint fix
* fix strange compile errors in CUDNN batchnorm header
* FInish using flags instead of bools
* lint
* Fix timing pass count for forward pass
* Fix R script install roxygen problem
* code formatting, addition of doc strings is causing IDE to add spaces before the calls
* removed commented
* cr comments
* Change back to compilable code
* For CPU mode, store as invstd
* move testing code around a little
* lint fix
* Use AccReal in some places to avoid fp16 problems
* Fix minor invstd problem in cuda version
* remove unused scale param
* add permutation unit test, handle cudnn doesn't like 3D
* .
* lint
* .
* Remove mkl_off
* lint fix and time cudnn when enabled
2017-05-15 20:27:28 -07:00
def test_batchnorm_versions ( ) :
def test_batchnorm_versions_helper ( batchnorm_op_list , data , fix_gamma , use_global_stats ) :
ctx_list = [ ]
sym_list = [ ]
# BatchNormV1 cpu
if ' batchnorm_v1_cpu ' in batchnorm_op_list :
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' batchnorm_data ' : data , ' type_dict ' : { ' batchnorm_data ' : np . float32 } } )
sym_list . append ( mx . sym . BatchNorm_v1 ( fix_gamma = fix_gamma ,
use_global_stats = use_global_stats ,
name = ' batchnorm ' ) )
# BatchNormV1 gpu (organic)
if ' batchnorm_v1_gpu ' in batchnorm_op_list :
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' batchnorm_data ' : data , ' type_dict ' : { ' batchnorm_data ' : np . float32 } } )
sym_list . append ( mx . sym . BatchNorm_v1 ( fix_gamma = fix_gamma ,
use_global_stats = use_global_stats ,
name = ' batchnorm ' ) )
# BatchNorm cpu
if ' batchnorm_cpu ' in batchnorm_op_list :
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' batchnorm_data ' : data , ' type_dict ' : { ' batchnorm_data ' : np . float32 } } )
sym_list . append ( mx . sym . BatchNorm ( fix_gamma = fix_gamma ,
use_global_stats = use_global_stats ,
name = ' batchnorm ' ) )
# BatchNorm gpu (organic)
if ' batchnorm_gpu ' in batchnorm_op_list :
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' batchnorm_data ' : data , ' type_dict ' : { ' batchnorm_data ' : np . float32 } } )
sym_list . append ( mx . sym . BatchNorm ( fix_gamma = fix_gamma ,
use_global_stats = use_global_stats ,
name = ' batchnorm ' , cudnn_off = True ) )
# BatchNorm gpu cudnn (if cudnn is enabled)
if ' batchnorm_cudnn ' in batchnorm_op_list :
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' batchnorm_data ' : data , ' type_dict ' : { ' batchnorm_data ' : np . float32 } } )
sym_list . append ( mx . sym . BatchNorm ( fix_gamma = fix_gamma ,
use_global_stats = use_global_stats ,
name = ' batchnorm ' , cudnn_off = False ) )
check_consistency ( sym_list , ctx_list )
def test_1d_batchnorm ( fix_gamma , use_global_stats ) :
data = ( 2 , 3 , 20 )
test_batchnorm_versions_helper ( batchnorm_op_list = [ ' batchnorm_cpu ' ,
' batchnorm_gpu ' , ' batchnorm_cudnn ' ] ,
data = data ,
fix_gamma = fix_gamma , use_global_stats = use_global_stats )
def test_2d_batchnorm ( fix_gamma , use_global_stats ) :
data = ( 2 , 3 , 10 , 10 )
test_batchnorm_versions_helper ( batchnorm_op_list = [ ' batchnorm_v1_cpu ' , ' batchnorm_v1_gpu ' ,
' batchnorm_cpu ' ,
' batchnorm_gpu ' , ' batchnorm_cudnn ' ] ,
data = data ,
fix_gamma = fix_gamma , use_global_stats = use_global_stats )
def test_3d_batchnorm ( fix_gamma , use_global_stats ) :
data = ( 2 , 3 , 3 , 5 , 5 )
test_batchnorm_versions_helper ( batchnorm_op_list = [ ' batchnorm_cpu ' ,
' batchnorm_gpu ' ] ,
data = data ,
fix_gamma = fix_gamma , use_global_stats = use_global_stats )
test_1d_batchnorm ( True , False )
test_1d_batchnorm ( False , False )
test_1d_batchnorm ( False , True )
test_1d_batchnorm ( True , True )
test_2d_batchnorm ( True , False )
test_2d_batchnorm ( False , False )
test_2d_batchnorm ( False , True )
test_2d_batchnorm ( True , True )
test_3d_batchnorm ( True , False )
test_3d_batchnorm ( False , False )
test_3d_batchnorm ( False , True )
test_3d_batchnorm ( True , True )
2016-07-05 11:29:40 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( 1234 )
2016-03-19 23:45:52 -07:00
def test_convolution_with_type ( ) :
2016-12-23 23:55:49 -08:00
sym1 = mx . sym . Convolution ( num_filter = 3 , kernel = ( 3 , 3 ) , name = ' conv ' )
data = mx . sym . Variable ( ' conv_data ' )
w = mx . sym . Variable ( ' conv_weight ' )
b = mx . sym . Variable ( ' conv_bias ' )
w = mx . sym . transpose ( w , axes = ( 0 , 2 , 3 , 1 ) )
sym2 = mx . sym . transpose ( data , axes = ( 0 , 2 , 3 , 1 ) )
sym2 = mx . sym . Convolution ( sym2 , w , b , layout = ' NHWC ' , num_filter = 3 , kernel = ( 3 , 3 ) )
sym2 = mx . sym . transpose ( sym2 , axes = ( 0 , 3 , 1 , 2 ) , name = ' conv ' )
sym = [ sym1 , sym1 , sym1 , sym1 , sym1 , sym2 , sym2 ]
2016-03-19 23:45:52 -07:00
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' conv_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' conv_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' conv_data ' : np . float64 } } ,
2016-12-23 23:55:49 -08:00
{ ' ctx ' : mx . cpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ,
# NHWC
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 10 , 10 ) , ' conv_weight ' : ( 3 , 2 , 3 , 3 ) ,
' type_dict ' : { ' conv_data ' : np . float32 , ' conv_weight ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 10 , 10 ) , ' conv_weight ' : ( 3 , 2 , 3 , 3 ) ,
' type_dict ' : { ' conv_data ' : np . float16 , ' conv_weight ' : np . float16 } }
]
2017-04-18 22:00:04 -07:00
# wider tolerance needed for true-fp16 NCHW test above
tol = { np . dtype ( np . float16 ) : 0.5 ,
np . dtype ( np . float32 ) : 1e-3 ,
np . dtype ( np . float64 ) : 1e-5 ,
np . dtype ( np . uint8 ) : 0 ,
np . dtype ( np . int32 ) : 0 }
check_consistency ( sym , ctx_list , tol = tol )
2017-04-21 21:15:00 -07:00
# test ability to turn off training on bias
check_consistency ( sym , ctx_list , grad_req = { ' conv_data ' : ' write ' , ' conv_weight ' : ' write ' , ' conv_bias ' : ' null ' } , tol = tol )
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
2017-04-18 22:00:04 -07:00
# Apply N symbols against each of M contexts, checking that all NxM combinations match.
def check_consistency_NxM ( sym_list , ctx_list ) :
# e.g. if sym_list=[sym1, sym2] and ctx_list=[ctx1, ctx2, ctx3], then resulting lists are:
# sym_list=[sym1, sym1, sym1, sym2, sym2, sym2] and ctx_list=[ctx1, ctx2, ctx3, ctx1, ctx2, ctx3]
check_consistency ( np . repeat ( sym_list , len ( ctx_list ) ) , ctx_list * len ( sym_list ) )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2016-12-23 23:55:49 -08:00
def test_convolution_options ( ) :
2017-03-17 12:42:11 -07:00
# 1D convolution
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ]
2017-04-18 22:00:04 -07:00
# Pad > 0
2018-01-02 10:47:41 -08:00
sym = mx . sym . Convolution ( layout = ' NCW ' , num_filter = 3 , kernel = ( 3 , ) , pad = ( 1 , ) , name = ' conv ' )
2017-04-18 22:00:04 -07:00
sym_no_cudnn = mx . sym . Convolution ( num_filter = 3 , kernel = ( 3 , ) , pad = ( 1 , ) , cudnn_off = True , name = ' conv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
# Stride > 1
2018-01-02 10:47:41 -08:00
sym = mx . sym . Convolution ( layout = ' NCW ' , num_filter = 3 , kernel = ( 3 , ) , stride = ( 2 , ) , name = ' conv ' )
2017-04-18 22:00:04 -07:00
sym_no_cudnn = mx . sym . Convolution ( num_filter = 3 , kernel = ( 3 , ) , stride = ( 2 , ) , cudnn_off = True , name = ' conv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
# Dilate > 1
2018-01-02 10:47:41 -08:00
sym = mx . sym . Convolution ( layout = ' NCW ' , num_filter = 3 , kernel = ( 3 , ) , dilate = ( 2 , ) , name = ' conv ' )
2017-04-18 22:00:04 -07:00
sym_no_cudnn = mx . sym . Convolution ( num_filter = 3 , kernel = ( 3 , ) , dilate = ( 2 , ) , cudnn_off = True , name = ' conv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
1x1 convolution acceleration (#7613)
* 1x1 convolution acceleration
* GEMM directly without im2col or col2im in 1x1 convolution(stride=1,pad=0). The 1x1 convolution is used very common in modern CNN networks such as Googlenet/Inception/Resnet/Mobilenet etc.
* cpplint
* fix linalg_impl (#7611)
* fix linalg_impl
* fix
* fix
* fix
* set build status to success only after job ends (#7628)
Earlier code marks status as success initially. So any new PR shows jenkins status as success if we see the check mark on github. On opening the full build status, we see that builds haven't even started or are running.
If something fails, variable changes to failure then. So even without this merge, a red mark on github indicates that build has failed correctly. That behavior is unchanged.
* Fix build status of a test (#7629)
installs bc required by sh2ju.sh and changes the regex match to capital alphabet as it clashes with a warning thrown by opencv driver
* entire codebase build with mshadow_use_clas=0 (#7625)
* Update README.md (#7630)
* unit test for csv iter, doc update for libsvmiter (#7623)
* add unit test for csv iter
* fix lint
* add libsvm to mxnet.io doc
* update libsvm doc
* gpu access of ndarray (#7496)
* gpu access of ndarray
* gpu access from C++ api
* gpu access fix
* Update c_api.cc
* Update c_api.cc
* refactor cudnn algo reg to no use string (#7561)
* refactor cudnn algo reg to no use string
* refactor ctx list
* fix
* refactor save_inputs
* Update io.md (#7634)
* fix tests (#7633)
* [build] explicitly install JDK8 (#7574)
* explicitly install openjdk8
* handle earlier version of ubuntu
* install software-properties-common
* update -y
* update commands
* Indents correction
* Add script to build doc files for all versions (#7636)
* Add script to build doc files for all versions
* Fix
* Use add versipn script of each different version
* add fashion mnist and move mnists to s3 (#7635)
* add fashion mnist and move mnists to s3
* refactor
* add doc for dataset (#7644)
* Change apache package URL to https (#7622)
* Pip installer for CoreML Converter: mxnet-to-coreml (#7624)
* Fixing CoreML converter's README: typos/grammar/etc.
* CoreML converter README update: Talk about layers first and then about models.
* Providing examples on converting various standard models; calling out issues with InceptionV3.
* Fixing CoreML converter's README: typos/grammar/etc.
* CoreML converter README update: Talk about layers first and then about models.
* Providing examples on converting various standard models; calling out issues with InceptionV3.
* Pip installer for converter: mxnet-coreml-converter.
Runs only on MacOS and python 2.7. Once inside the directory pip_package, user needs
to run:
python setup.py bdist_wheel
twine upload dist/*
Once uploaded it'll look like this:
https://testpypi.python.org/pypi/mxnet-coreml-converter
Also updated the README for converter to reflect this.
Note that we are going with a package per tool for the time being. Please leave feedback if you think it is better to adopt the policy of all the tools in one single package.
Unit tests continue to pass.
* More informative pypi package documentation.
* Updating MacOS in release notes to 10.11 after testing on it.
* Changing the name to mxnet-to-coreml and version to 0.1.0.
* Added license to setup.py
* Updating readme files with the correct pip package name.
* Parallelize windows unit tests of python 2 and 3 in jenkins (#7646)
* parallelize python windows tests
* reordered for clarity
* Removed asset loaded insecurely and added the asset to be loaded from the origin securely (#7649)
* skip failing test temporarily (#7648)
* lower really high threshold to fix test failure (#7650)
* Doc updates for install and doc generation (#7647)
* fluent (#7584)
* add 1x1 convolution to tests
* indent
* Refactor random linalg contrib namespaces (#7604)
* Refactor namespaces contrib, linalg, random, and sparse for op registration
Change examples in documentation
Change namespace usage in examples
Fix pylint
Remove unused import
Switch name and alias in linalg and random
Change stype comparison from string to int for functions used internally
Change documentation to use the right namespace
Register ops under ndarray/op.py and symbol/op.py
Remove unused import
Change .cu op names
* Add __all__ to ndarray and symbol modules
* Revert "Add __all__ to ndarray and symbol modules"
This reverts commit 8bc5de77bfdb40ff48dc570e2c6c49ec5d43ea64.
* Add __all__ to ndarray and symbol modules
* fix gluon fasionmnist dataset (#7655)
fix gluon fasionmnist dataset
* Parallelize Python 2 and 3 unit test cases in Jenkins CI. (#7658)
* Parallelize Python 2 and 3 unit test cases.
* Parallelize python 2 and 3 unit tests cases in jenkins
* Parallelize python 2 and 3 unit tests cases in jenkins
* Change namespace and make logging functionality changes (#7627)
* Change namespace and make logging functionality changes
* Help comment changes
* update mklml and mkl mac support (#7587)
* 1x1 convolution acceleration
* GEMM directly without im2col or col2im in 1x1 convolution(stride=1,pad=0). The 1x1 convolution is used very common in modern CNN networks such as Googlenet/Inception/Resnet/Mobilenet etc.
* cpplint
* Indents correction
* add 1x1 convolution to tests
* indent
* 1x1 convolution acceleration
* GEMM directly without im2col or col2im in 1x1 convolution(stride=1,pad=0). The 1x1 convolution is used very common in modern CNN networks such as Googlenet/Inception/Resnet/Mobilenet etc.
* cpplint
* Indents correction
* add 1x1 convolution to tests
* indent
* cpplint
* indent
2017-08-31 02:18:47 +08:00
# 1x1 convolution
2018-01-02 10:47:41 -08:00
sym = mx . sym . Convolution ( layout = ' NCW ' , num_filter = 3 , kernel = ( 1 , ) , pad = ( 0 , ) , name = ' conv ' )
1x1 convolution acceleration (#7613)
* 1x1 convolution acceleration
* GEMM directly without im2col or col2im in 1x1 convolution(stride=1,pad=0). The 1x1 convolution is used very common in modern CNN networks such as Googlenet/Inception/Resnet/Mobilenet etc.
* cpplint
* fix linalg_impl (#7611)
* fix linalg_impl
* fix
* fix
* fix
* set build status to success only after job ends (#7628)
Earlier code marks status as success initially. So any new PR shows jenkins status as success if we see the check mark on github. On opening the full build status, we see that builds haven't even started or are running.
If something fails, variable changes to failure then. So even without this merge, a red mark on github indicates that build has failed correctly. That behavior is unchanged.
* Fix build status of a test (#7629)
installs bc required by sh2ju.sh and changes the regex match to capital alphabet as it clashes with a warning thrown by opencv driver
* entire codebase build with mshadow_use_clas=0 (#7625)
* Update README.md (#7630)
* unit test for csv iter, doc update for libsvmiter (#7623)
* add unit test for csv iter
* fix lint
* add libsvm to mxnet.io doc
* update libsvm doc
* gpu access of ndarray (#7496)
* gpu access of ndarray
* gpu access from C++ api
* gpu access fix
* Update c_api.cc
* Update c_api.cc
* refactor cudnn algo reg to no use string (#7561)
* refactor cudnn algo reg to no use string
* refactor ctx list
* fix
* refactor save_inputs
* Update io.md (#7634)
* fix tests (#7633)
* [build] explicitly install JDK8 (#7574)
* explicitly install openjdk8
* handle earlier version of ubuntu
* install software-properties-common
* update -y
* update commands
* Indents correction
* Add script to build doc files for all versions (#7636)
* Add script to build doc files for all versions
* Fix
* Use add versipn script of each different version
* add fashion mnist and move mnists to s3 (#7635)
* add fashion mnist and move mnists to s3
* refactor
* add doc for dataset (#7644)
* Change apache package URL to https (#7622)
* Pip installer for CoreML Converter: mxnet-to-coreml (#7624)
* Fixing CoreML converter's README: typos/grammar/etc.
* CoreML converter README update: Talk about layers first and then about models.
* Providing examples on converting various standard models; calling out issues with InceptionV3.
* Fixing CoreML converter's README: typos/grammar/etc.
* CoreML converter README update: Talk about layers first and then about models.
* Providing examples on converting various standard models; calling out issues with InceptionV3.
* Pip installer for converter: mxnet-coreml-converter.
Runs only on MacOS and python 2.7. Once inside the directory pip_package, user needs
to run:
python setup.py bdist_wheel
twine upload dist/*
Once uploaded it'll look like this:
https://testpypi.python.org/pypi/mxnet-coreml-converter
Also updated the README for converter to reflect this.
Note that we are going with a package per tool for the time being. Please leave feedback if you think it is better to adopt the policy of all the tools in one single package.
Unit tests continue to pass.
* More informative pypi package documentation.
* Updating MacOS in release notes to 10.11 after testing on it.
* Changing the name to mxnet-to-coreml and version to 0.1.0.
* Added license to setup.py
* Updating readme files with the correct pip package name.
* Parallelize windows unit tests of python 2 and 3 in jenkins (#7646)
* parallelize python windows tests
* reordered for clarity
* Removed asset loaded insecurely and added the asset to be loaded from the origin securely (#7649)
* skip failing test temporarily (#7648)
* lower really high threshold to fix test failure (#7650)
* Doc updates for install and doc generation (#7647)
* fluent (#7584)
* add 1x1 convolution to tests
* indent
* Refactor random linalg contrib namespaces (#7604)
* Refactor namespaces contrib, linalg, random, and sparse for op registration
Change examples in documentation
Change namespace usage in examples
Fix pylint
Remove unused import
Switch name and alias in linalg and random
Change stype comparison from string to int for functions used internally
Change documentation to use the right namespace
Register ops under ndarray/op.py and symbol/op.py
Remove unused import
Change .cu op names
* Add __all__ to ndarray and symbol modules
* Revert "Add __all__ to ndarray and symbol modules"
This reverts commit 8bc5de77bfdb40ff48dc570e2c6c49ec5d43ea64.
* Add __all__ to ndarray and symbol modules
* fix gluon fasionmnist dataset (#7655)
fix gluon fasionmnist dataset
* Parallelize Python 2 and 3 unit test cases in Jenkins CI. (#7658)
* Parallelize Python 2 and 3 unit test cases.
* Parallelize python 2 and 3 unit tests cases in jenkins
* Parallelize python 2 and 3 unit tests cases in jenkins
* Change namespace and make logging functionality changes (#7627)
* Change namespace and make logging functionality changes
* Help comment changes
* update mklml and mkl mac support (#7587)
* 1x1 convolution acceleration
* GEMM directly without im2col or col2im in 1x1 convolution(stride=1,pad=0). The 1x1 convolution is used very common in modern CNN networks such as Googlenet/Inception/Resnet/Mobilenet etc.
* cpplint
* Indents correction
* add 1x1 convolution to tests
* indent
* 1x1 convolution acceleration
* GEMM directly without im2col or col2im in 1x1 convolution(stride=1,pad=0). The 1x1 convolution is used very common in modern CNN networks such as Googlenet/Inception/Resnet/Mobilenet etc.
* cpplint
* Indents correction
* add 1x1 convolution to tests
* indent
* cpplint
* indent
2017-08-31 02:18:47 +08:00
sym_no_cudnn = mx . sym . Convolution ( num_filter = 3 , kernel = ( 1 , ) , pad = ( 0 , ) , cudnn_off = True , name = ' conv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
2017-03-17 12:42:11 -07:00
# 2D convolution
2016-12-23 23:55:49 -08:00
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ]
2017-04-18 22:00:04 -07:00
# Pad > 0
2016-12-23 23:55:49 -08:00
sym = mx . sym . Convolution ( num_filter = 3 , kernel = ( 3 , 3 ) , pad = ( 1 , 1 ) , name = ' conv ' )
2017-04-18 22:00:04 -07:00
sym_no_cudnn = mx . sym . Convolution ( num_filter = 3 , kernel = ( 3 , 3 ) , pad = ( 1 , 1 ) , cudnn_off = True , name = ' conv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
# Stride > 1
2016-12-23 23:55:49 -08:00
sym = mx . sym . Convolution ( num_filter = 3 , kernel = ( 3 , 3 ) , stride = ( 2 , 2 ) , name = ' conv ' )
2017-04-18 22:00:04 -07:00
sym_no_cudnn = mx . sym . Convolution ( num_filter = 3 , kernel = ( 3 , 3 ) , stride = ( 2 , 2 ) , cudnn_off = True , name = ' conv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
# Dilate > 1
2016-12-23 23:55:49 -08:00
sym = mx . sym . Convolution ( num_filter = 3 , kernel = ( 3 , 3 ) , dilate = ( 2 , 2 ) , name = ' conv ' )
2017-04-18 22:00:04 -07:00
sym_no_cudnn = mx . sym . Convolution ( num_filter = 3 , kernel = ( 3 , 3 ) , dilate = ( 2 , 2 ) , cudnn_off = True , name = ' conv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
1x1 convolution acceleration (#7613)
* 1x1 convolution acceleration
* GEMM directly without im2col or col2im in 1x1 convolution(stride=1,pad=0). The 1x1 convolution is used very common in modern CNN networks such as Googlenet/Inception/Resnet/Mobilenet etc.
* cpplint
* fix linalg_impl (#7611)
* fix linalg_impl
* fix
* fix
* fix
* set build status to success only after job ends (#7628)
Earlier code marks status as success initially. So any new PR shows jenkins status as success if we see the check mark on github. On opening the full build status, we see that builds haven't even started or are running.
If something fails, variable changes to failure then. So even without this merge, a red mark on github indicates that build has failed correctly. That behavior is unchanged.
* Fix build status of a test (#7629)
installs bc required by sh2ju.sh and changes the regex match to capital alphabet as it clashes with a warning thrown by opencv driver
* entire codebase build with mshadow_use_clas=0 (#7625)
* Update README.md (#7630)
* unit test for csv iter, doc update for libsvmiter (#7623)
* add unit test for csv iter
* fix lint
* add libsvm to mxnet.io doc
* update libsvm doc
* gpu access of ndarray (#7496)
* gpu access of ndarray
* gpu access from C++ api
* gpu access fix
* Update c_api.cc
* Update c_api.cc
* refactor cudnn algo reg to no use string (#7561)
* refactor cudnn algo reg to no use string
* refactor ctx list
* fix
* refactor save_inputs
* Update io.md (#7634)
* fix tests (#7633)
* [build] explicitly install JDK8 (#7574)
* explicitly install openjdk8
* handle earlier version of ubuntu
* install software-properties-common
* update -y
* update commands
* Indents correction
* Add script to build doc files for all versions (#7636)
* Add script to build doc files for all versions
* Fix
* Use add versipn script of each different version
* add fashion mnist and move mnists to s3 (#7635)
* add fashion mnist and move mnists to s3
* refactor
* add doc for dataset (#7644)
* Change apache package URL to https (#7622)
* Pip installer for CoreML Converter: mxnet-to-coreml (#7624)
* Fixing CoreML converter's README: typos/grammar/etc.
* CoreML converter README update: Talk about layers first and then about models.
* Providing examples on converting various standard models; calling out issues with InceptionV3.
* Fixing CoreML converter's README: typos/grammar/etc.
* CoreML converter README update: Talk about layers first and then about models.
* Providing examples on converting various standard models; calling out issues with InceptionV3.
* Pip installer for converter: mxnet-coreml-converter.
Runs only on MacOS and python 2.7. Once inside the directory pip_package, user needs
to run:
python setup.py bdist_wheel
twine upload dist/*
Once uploaded it'll look like this:
https://testpypi.python.org/pypi/mxnet-coreml-converter
Also updated the README for converter to reflect this.
Note that we are going with a package per tool for the time being. Please leave feedback if you think it is better to adopt the policy of all the tools in one single package.
Unit tests continue to pass.
* More informative pypi package documentation.
* Updating MacOS in release notes to 10.11 after testing on it.
* Changing the name to mxnet-to-coreml and version to 0.1.0.
* Added license to setup.py
* Updating readme files with the correct pip package name.
* Parallelize windows unit tests of python 2 and 3 in jenkins (#7646)
* parallelize python windows tests
* reordered for clarity
* Removed asset loaded insecurely and added the asset to be loaded from the origin securely (#7649)
* skip failing test temporarily (#7648)
* lower really high threshold to fix test failure (#7650)
* Doc updates for install and doc generation (#7647)
* fluent (#7584)
* add 1x1 convolution to tests
* indent
* Refactor random linalg contrib namespaces (#7604)
* Refactor namespaces contrib, linalg, random, and sparse for op registration
Change examples in documentation
Change namespace usage in examples
Fix pylint
Remove unused import
Switch name and alias in linalg and random
Change stype comparison from string to int for functions used internally
Change documentation to use the right namespace
Register ops under ndarray/op.py and symbol/op.py
Remove unused import
Change .cu op names
* Add __all__ to ndarray and symbol modules
* Revert "Add __all__ to ndarray and symbol modules"
This reverts commit 8bc5de77bfdb40ff48dc570e2c6c49ec5d43ea64.
* Add __all__ to ndarray and symbol modules
* fix gluon fasionmnist dataset (#7655)
fix gluon fasionmnist dataset
* Parallelize Python 2 and 3 unit test cases in Jenkins CI. (#7658)
* Parallelize Python 2 and 3 unit test cases.
* Parallelize python 2 and 3 unit tests cases in jenkins
* Parallelize python 2 and 3 unit tests cases in jenkins
* Change namespace and make logging functionality changes (#7627)
* Change namespace and make logging functionality changes
* Help comment changes
* update mklml and mkl mac support (#7587)
* 1x1 convolution acceleration
* GEMM directly without im2col or col2im in 1x1 convolution(stride=1,pad=0). The 1x1 convolution is used very common in modern CNN networks such as Googlenet/Inception/Resnet/Mobilenet etc.
* cpplint
* Indents correction
* add 1x1 convolution to tests
* indent
* 1x1 convolution acceleration
* GEMM directly without im2col or col2im in 1x1 convolution(stride=1,pad=0). The 1x1 convolution is used very common in modern CNN networks such as Googlenet/Inception/Resnet/Mobilenet etc.
* cpplint
* Indents correction
* add 1x1 convolution to tests
* indent
* cpplint
* indent
2017-08-31 02:18:47 +08:00
# 1x1 convolution
sym = mx . sym . Convolution ( num_filter = 3 , kernel = ( 1 , 1 ) , pad = ( 0 , 0 ) , name = ' conv ' )
sym_no_cudnn = mx . sym . Convolution ( num_filter = 3 , kernel = ( 1 , 1 ) , pad = ( 0 , 0 ) , cudnn_off = True , name = ' conv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
2016-12-23 23:55:49 -08:00
2017-03-17 12:42:11 -07:00
# 3D convolution
ctx_list = [ { ' ctx ' : mx . cpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 5 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 5 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 5 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float64 } } ,
2016-12-23 23:55:49 -08:00
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 5 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ]
2017-04-18 22:00:04 -07:00
# Pad > 0
2016-12-23 23:55:49 -08:00
sym = mx . sym . Convolution ( num_filter = 3 , kernel = ( 2 , 3 , 3 ) , pad = ( 1 , 1 , 1 ) , name = ' conv ' )
2017-04-18 22:00:04 -07:00
sym_no_cudnn = mx . sym . Convolution ( num_filter = 3 , kernel = ( 2 , 3 , 3 ) , pad = ( 1 , 1 , 1 ) , cudnn_off = True , name = ' conv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
# Stride > 1
2016-12-23 23:55:49 -08:00
sym = mx . sym . Convolution ( num_filter = 3 , kernel = ( 2 , 3 , 3 ) , stride = ( 2 , 2 , 2 ) , name = ' conv ' )
2017-04-18 22:00:04 -07:00
sym_no_cudnn = mx . sym . Convolution ( num_filter = 3 , kernel = ( 2 , 3 , 3 ) , stride = ( 2 , 2 , 2 ) , cudnn_off = True , name = ' conv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
1x1 convolution acceleration (#7613)
* 1x1 convolution acceleration
* GEMM directly without im2col or col2im in 1x1 convolution(stride=1,pad=0). The 1x1 convolution is used very common in modern CNN networks such as Googlenet/Inception/Resnet/Mobilenet etc.
* cpplint
* fix linalg_impl (#7611)
* fix linalg_impl
* fix
* fix
* fix
* set build status to success only after job ends (#7628)
Earlier code marks status as success initially. So any new PR shows jenkins status as success if we see the check mark on github. On opening the full build status, we see that builds haven't even started or are running.
If something fails, variable changes to failure then. So even without this merge, a red mark on github indicates that build has failed correctly. That behavior is unchanged.
* Fix build status of a test (#7629)
installs bc required by sh2ju.sh and changes the regex match to capital alphabet as it clashes with a warning thrown by opencv driver
* entire codebase build with mshadow_use_clas=0 (#7625)
* Update README.md (#7630)
* unit test for csv iter, doc update for libsvmiter (#7623)
* add unit test for csv iter
* fix lint
* add libsvm to mxnet.io doc
* update libsvm doc
* gpu access of ndarray (#7496)
* gpu access of ndarray
* gpu access from C++ api
* gpu access fix
* Update c_api.cc
* Update c_api.cc
* refactor cudnn algo reg to no use string (#7561)
* refactor cudnn algo reg to no use string
* refactor ctx list
* fix
* refactor save_inputs
* Update io.md (#7634)
* fix tests (#7633)
* [build] explicitly install JDK8 (#7574)
* explicitly install openjdk8
* handle earlier version of ubuntu
* install software-properties-common
* update -y
* update commands
* Indents correction
* Add script to build doc files for all versions (#7636)
* Add script to build doc files for all versions
* Fix
* Use add versipn script of each different version
* add fashion mnist and move mnists to s3 (#7635)
* add fashion mnist and move mnists to s3
* refactor
* add doc for dataset (#7644)
* Change apache package URL to https (#7622)
* Pip installer for CoreML Converter: mxnet-to-coreml (#7624)
* Fixing CoreML converter's README: typos/grammar/etc.
* CoreML converter README update: Talk about layers first and then about models.
* Providing examples on converting various standard models; calling out issues with InceptionV3.
* Fixing CoreML converter's README: typos/grammar/etc.
* CoreML converter README update: Talk about layers first and then about models.
* Providing examples on converting various standard models; calling out issues with InceptionV3.
* Pip installer for converter: mxnet-coreml-converter.
Runs only on MacOS and python 2.7. Once inside the directory pip_package, user needs
to run:
python setup.py bdist_wheel
twine upload dist/*
Once uploaded it'll look like this:
https://testpypi.python.org/pypi/mxnet-coreml-converter
Also updated the README for converter to reflect this.
Note that we are going with a package per tool for the time being. Please leave feedback if you think it is better to adopt the policy of all the tools in one single package.
Unit tests continue to pass.
* More informative pypi package documentation.
* Updating MacOS in release notes to 10.11 after testing on it.
* Changing the name to mxnet-to-coreml and version to 0.1.0.
* Added license to setup.py
* Updating readme files with the correct pip package name.
* Parallelize windows unit tests of python 2 and 3 in jenkins (#7646)
* parallelize python windows tests
* reordered for clarity
* Removed asset loaded insecurely and added the asset to be loaded from the origin securely (#7649)
* skip failing test temporarily (#7648)
* lower really high threshold to fix test failure (#7650)
* Doc updates for install and doc generation (#7647)
* fluent (#7584)
* add 1x1 convolution to tests
* indent
* Refactor random linalg contrib namespaces (#7604)
* Refactor namespaces contrib, linalg, random, and sparse for op registration
Change examples in documentation
Change namespace usage in examples
Fix pylint
Remove unused import
Switch name and alias in linalg and random
Change stype comparison from string to int for functions used internally
Change documentation to use the right namespace
Register ops under ndarray/op.py and symbol/op.py
Remove unused import
Change .cu op names
* Add __all__ to ndarray and symbol modules
* Revert "Add __all__ to ndarray and symbol modules"
This reverts commit 8bc5de77bfdb40ff48dc570e2c6c49ec5d43ea64.
* Add __all__ to ndarray and symbol modules
* fix gluon fasionmnist dataset (#7655)
fix gluon fasionmnist dataset
* Parallelize Python 2 and 3 unit test cases in Jenkins CI. (#7658)
* Parallelize Python 2 and 3 unit test cases.
* Parallelize python 2 and 3 unit tests cases in jenkins
* Parallelize python 2 and 3 unit tests cases in jenkins
* Change namespace and make logging functionality changes (#7627)
* Change namespace and make logging functionality changes
* Help comment changes
* update mklml and mkl mac support (#7587)
* 1x1 convolution acceleration
* GEMM directly without im2col or col2im in 1x1 convolution(stride=1,pad=0). The 1x1 convolution is used very common in modern CNN networks such as Googlenet/Inception/Resnet/Mobilenet etc.
* cpplint
* Indents correction
* add 1x1 convolution to tests
* indent
* 1x1 convolution acceleration
* GEMM directly without im2col or col2im in 1x1 convolution(stride=1,pad=0). The 1x1 convolution is used very common in modern CNN networks such as Googlenet/Inception/Resnet/Mobilenet etc.
* cpplint
* Indents correction
* add 1x1 convolution to tests
* indent
* cpplint
* indent
2017-08-31 02:18:47 +08:00
# 1x1 convolution
sym = mx . sym . Convolution ( num_filter = 3 , kernel = ( 1 , 1 , 1 ) , pad = ( 0 , 0 , 0 ) , name = ' conv ' )
sym_no_cudnn = mx . sym . Convolution ( num_filter = 3 , kernel = ( 1 , 1 , 1 ) , pad = ( 0 , 0 , 0 ) , cudnn_off = True , name = ' conv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-03-17 12:42:11 -07:00
def test_convolution_versions ( ) :
# 2D convolution NCHW
ctx_list = [ { ' ctx ' : mx . cpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ]
conv_v1_cpu = mx . sym . Convolution_v1 ( num_filter = 3 , kernel = ( 3 , 3 ) , pad = ( 1 , 1 ) , name = ' conv ' )
conv_v1_gpu = mx . sym . Convolution_v1 ( num_filter = 3 , kernel = ( 3 , 3 ) , pad = ( 1 , 1 ) , cudnn_off = True , name = ' conv ' )
conv_cudnn = mx . sym . Convolution ( num_filter = 3 , kernel = ( 3 , 3 ) , pad = ( 1 , 1 ) , name = ' conv ' )
conv_cpu = mx . sym . Convolution ( num_filter = 3 , kernel = ( 3 , 3 ) , pad = ( 1 , 1 ) , name = ' conv ' )
conv_gpu = mx . sym . Convolution ( num_filter = 3 , kernel = ( 3 , 3 ) , pad = ( 1 , 1 ) , cudnn_off = True , name = ' conv ' )
syms = [ conv_v1_cpu , conv_v1_gpu , conv_cudnn , conv_cpu , conv_gpu ]
check_consistency ( syms , ctx_list )
# 3D convolution NCDHW
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 5 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 5 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' conv_data ' : ( 2 , 2 , 5 , 7 , 7 ) , ' type_dict ' : { ' conv_data ' : np . float32 } } ]
conv_cudnn = mx . sym . Convolution ( num_filter = 3 , kernel = ( 2 , 3 , 3 ) , pad = ( 1 , 1 , 1 ) , name = ' conv ' )
conv_cpu = mx . sym . Convolution ( num_filter = 3 , kernel = ( 2 , 3 , 3 ) , pad = ( 1 , 1 , 1 ) , name = ' conv ' )
conv_gpu = mx . sym . Convolution ( num_filter = 3 , kernel = ( 2 , 3 , 3 ) , pad = ( 1 , 1 , 1 ) , cudnn_off = True , name = ' conv ' )
syms = [ conv_cudnn , conv_cpu , conv_gpu ]
check_consistency ( syms , ctx_list )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2016-12-23 23:55:49 -08:00
def test_pooling_with_type ( ) :
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' pool_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' pool_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' pool_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' pool_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' pool_data ' : np . float32 } } ]
sym = mx . sym . Pooling ( kernel = ( 3 , 3 ) , pool_type = ' max ' , pooling_convention = ' valid ' , name = ' pool ' )
check_consistency ( sym , ctx_list )
sym = mx . sym . Pooling ( kernel = ( 3 , 3 ) , pool_type = ' max ' , pooling_convention = ' full ' , name = ' pool ' )
check_consistency ( sym , ctx_list )
sym = mx . sym . Pooling ( kernel = ( 300 , 300 ) , pool_type = ' max ' , global_pool = True , name = ' pool ' )
2016-05-27 23:27:19 -07:00
check_consistency ( sym , ctx_list )
2016-03-19 23:45:52 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2016-06-09 01:32:07 +09:00
def test_deconvolution_with_type ( ) :
2018-01-02 10:47:41 -08:00
# Test basic deconvolution without exercising stride, pad or dilation.
# 1D deconvolution
sym = mx . sym . Deconvolution ( num_filter = 3 , kernel = ( 3 , ) , name = ' deconv ' )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' deconv_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' deconv_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' deconv_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' deconv_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' deconv_data ' : np . float32 } } ]
# wider tolerance needed for true-fp16 test above
tol = { np . dtype ( np . float16 ) : 0.3 ,
np . dtype ( np . float32 ) : 1e-3 ,
np . dtype ( np . float64 ) : 1e-5 ,
np . dtype ( np . uint8 ) : 0 ,
np . dtype ( np . int32 ) : 0 }
check_consistency ( sym , ctx_list , tol = tol )
check_consistency ( sym , ctx_list , tol = tol , grad_req = " add " )
# 2D deconvolution
2016-06-10 11:31:17 +09:00
sym = mx . sym . Deconvolution ( num_filter = 2 , kernel = ( 3 , 3 ) , name = ' deconv ' )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' deconv_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' deconv_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' deconv_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' deconv_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' deconv_data ' : np . float32 } } ]
2017-04-18 22:00:04 -07:00
# wider tolerance needed for true-fp16 test above
tol = { np . dtype ( np . float16 ) : 0.3 ,
np . dtype ( np . float32 ) : 1e-3 ,
np . dtype ( np . float64 ) : 1e-5 ,
np . dtype ( np . uint8 ) : 0 ,
np . dtype ( np . int32 ) : 0 }
check_consistency ( sym , ctx_list , tol = tol )
check_consistency ( sym , ctx_list , tol = tol , grad_req = " add " )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-04-18 22:00:04 -07:00
def test_deconvolution_options ( ) :
2018-01-02 10:47:41 -08:00
# 1D deconvolution
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' deconv_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' deconv_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' deconv_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' deconv_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 7 ) , ' type_dict ' : { ' deconv_data ' : np . float32 } } ]
# Pad > 0
sym = mx . sym . Deconvolution ( layout = ' NCW ' , num_filter = 3 , kernel = ( 3 , ) , pad = ( 1 , ) , name = ' deconv ' )
sym_no_cudnn = mx . sym . Deconvolution ( num_filter = 3 , kernel = ( 3 , ) , pad = ( 1 , ) , cudnn_off = True , name = ' deconv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
# Stride > 1
sym = mx . sym . Deconvolution ( layout = ' NCW ' , num_filter = 3 , kernel = ( 3 , ) , stride = ( 2 , ) , name = ' deconv ' )
sym_no_cudnn = mx . sym . Deconvolution ( num_filter = 3 , kernel = ( 3 , ) , stride = ( 2 , ) , cudnn_off = True , name = ' deconv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
# Dilate > 1
sym = mx . sym . Deconvolution ( layout = ' NCW ' , num_filter = 3 , kernel = ( 3 , ) , dilate = ( 2 , ) , name = ' deconv ' )
sym_no_cudnn = mx . sym . Deconvolution ( num_filter = 3 , kernel = ( 3 , ) , dilate = ( 2 , ) , cudnn_off = True , name = ' deconv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
2017-04-18 22:00:04 -07:00
# 2D deconvolution
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' deconv_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' deconv_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' deconv_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' deconv_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' deconv_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' deconv_data ' : np . float32 } } ]
# Pad > 0
sym = mx . sym . Deconvolution ( num_filter = 2 , kernel = ( 3 , 3 ) , pad = ( 1 , 1 ) , name = ' deconv ' )
sym_no_cudnn = mx . sym . Deconvolution ( num_filter = 2 , kernel = ( 3 , 3 ) , pad = ( 1 , 1 ) , cudnn_off = True , name = ' deconv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
# Stride > 1
sym = mx . sym . Deconvolution ( num_filter = 2 , kernel = ( 3 , 3 ) , stride = ( 2 , 2 ) , name = ' deconv ' )
sym_no_cudnn = mx . sym . Deconvolution ( num_filter = 2 , kernel = ( 3 , 3 ) , stride = ( 2 , 2 ) , cudnn_off = True , name = ' deconv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
# Dilate > 1
sym = mx . sym . Deconvolution ( num_filter = 2 , kernel = ( 3 , 3 ) , dilate = ( 2 , 2 ) , name = ' deconv ' )
sym_no_cudnn = mx . sym . Deconvolution ( num_filter = 2 , kernel = ( 3 , 3 ) , dilate = ( 2 , 2 ) , cudnn_off = True , name = ' deconv ' )
check_consistency_NxM ( [ sym , sym_no_cudnn ] , ctx_list )
2018-01-02 10:47:41 -08:00
# # 3D deconvolution (not yet enabled)
2017-04-18 22:00:04 -07:00
# ctx_list = [{'ctx': mx.cpu(0), 'conv_data': (2, 2, 5, 7, 7), 'type_dict': {'conv_data': np.float64}},
# {'ctx': mx.cpu(0), 'conv_data': (2, 2, 5, 7, 7), 'type_dict': {'conv_data': np.float64}},
# {'ctx': mx.gpu(0), 'conv_data': (2, 2, 5, 7, 7), 'type_dict': {'conv_data': np.float64}},
# {'ctx': mx.gpu(0), 'conv_data': (2, 2, 5, 7, 7), 'type_dict': {'conv_data': np.float32}}]
# # Pad > 0
# sym = mx.sym.Convolution(num_filter=3, kernel=(2,3,3), pad=(1,1,1), name='conv')
# sym_no_cudnn = mx.sym.Convolution(num_filter=3, kernel=(2,3,3), pad=(1,1,1), cudnn_off=True, name='conv')
# check_consistency_NxM([sym, sym_no_cudnn], ctx_list)
# # Stride > 1
# sym = mx.sym.Convolution(num_filter=3, kernel=(2,3,3), stride=(2,2,2), name='conv')
# sym_no_cudnn = mx.sym.Convolution(num_filter=3, kernel=(2,3,3), stride=(2,2,2), cudnn_off=True, name='conv')
# check_consistency_NxM([sym, sym_no_cudnn], ctx_list)
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( 1234 )
2017-01-28 00:45:17 +08:00
def test_bilinear_sampler_with_type ( ) :
data = mx . sym . Variable ( ' data ' )
grid = mx . sym . Variable ( ' grid ' )
sym = mx . sym . BilinearSampler ( data = data , grid = grid )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' data ' : ( 1 , 5 , 10 , 10 ) , ' grid ' : ( 1 , 2 , 10 , 10 ) ,
' type_dict ' : { ' data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' data ' : ( 1 , 5 , 10 , 10 ) , ' grid ' : ( 1 , 2 , 10 , 10 ) ,
' type_dict ' : { ' data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' data ' : ( 1 , 5 , 10 , 10 ) , ' grid ' : ( 1 , 2 , 10 , 10 ) ,
' type_dict ' : { ' data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' data ' : ( 1 , 5 , 10 , 10 ) , ' grid ' : ( 1 , 2 , 10 , 10 ) ,
' type_dict ' : { ' data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' data ' : ( 1 , 5 , 10 , 10 ) , ' grid ' : ( 1 , 2 , 10 , 10 ) ,
' type_dict ' : { ' data ' : np . float32 } } ]
check_consistency ( sym , ctx_list )
check_consistency ( sym , ctx_list , grad_req = " add " )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-01-28 00:45:17 +08:00
def test_grid_generator_with_type ( ) :
data = mx . sym . Variable ( ' data ' )
sym = mx . sym . GridGenerator ( data = data , transform_type = ' affine ' , target_shape = ( 20 , 20 ) )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' data ' : ( 3 , 6 ) , ' type_dict ' : { ' data ' : np . float32 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' data ' : ( 3 , 6 ) , ' type_dict ' : { ' data ' : np . float32 } } ]
check_consistency ( sym , ctx_list )
check_consistency ( sym , ctx_list , grad_req = " add " )
sym = mx . sym . GridGenerator ( data = data , transform_type = ' warp ' , target_shape = ( 20 , 20 ) )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' data ' : ( 3 , 2 , 20 , 20 ) , ' type_dict ' : { ' data ' : np . float32 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' data ' : ( 3 , 2 , 20 , 20 ) , ' type_dict ' : { ' data ' : np . float32 } } ]
check_consistency ( sym , ctx_list )
check_consistency ( sym , ctx_list , grad_req = " add " )
2018-02-18 03:11:58 -08:00
2017-08-28 20:09:41 -07:00
@unittest.skip ( " test fails intermittently. temporarily disabled till it gets fixed. tracked at https://github.com/apache/incubator-mxnet/issues/7645 " )
2018-02-18 03:11:58 -08:00
@with_seed ( 1234 )
2017-08-12 08:08:36 +08:00
def test_spatial_transformer_with_type ( ) :
data = mx . sym . Variable ( ' data ' )
loc = mx . sym . Flatten ( data )
loc = mx . sym . FullyConnected ( data = loc , num_hidden = 10 )
loc = mx . sym . Activation ( data = loc , act_type = ' relu ' )
loc = mx . sym . FullyConnected ( data = loc , num_hidden = 6 )
sym = mx . sym . SpatialTransformer ( data = data , loc = loc , target_shape = ( 10 , 10 ) ,
transform_type = " affine " , sampler_type = " bilinear " )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' data ' : ( 1 , 5 , 10 , 10 ) , ' type_dict ' : { ' data ' : np . float32 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' data ' : ( 1 , 5 , 10 , 10 ) , ' type_dict ' : { ' data ' : np . float32 } } ]
check_consistency ( sym , ctx_list )
check_consistency ( sym , ctx_list , grad_req = " add " )
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
2017-03-21 23:01:57 -07:00
# Checking max pooling consistency over the data sets of different float types is problematic
# as one max value in a float32 data set may not be the max value in a float16 data set.
# This function will not be called.
2018-02-18 03:11:58 -08:00
@with_seed ( 1234 )
2016-11-17 22:19:38 -08:00
def test_pooling_with_type ( ) :
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : ( 10 , 2 , 10 , 10 ) , ' type_dict ' : { ' pool_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : ( 10 , 2 , 10 , 10 ) , ' type_dict ' : { ' pool_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : ( 10 , 2 , 10 , 10 ) , ' type_dict ' : { ' pool_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : ( 10 , 2 , 10 , 10 ) , ' type_dict ' : { ' pool_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : ( 10 , 2 , 10 , 10 ) , ' type_dict ' : { ' pool_data ' : np . float32 } } ]
sym = mx . sym . Pooling ( name = ' pool ' , kernel = ( 3 , 3 ) , stride = ( 2 , 2 ) , pool_type = ' max ' )
check_consistency ( sym , ctx_list )
sym = mx . sym . Pooling ( name = ' pool ' , kernel = ( 3 , 3 ) , pad = ( 1 , 1 ) , pool_type = ' avg ' )
check_consistency ( sym , ctx_list )
2016-12-23 23:55:49 -08:00
# this is unstable
# sym = mx.sym.Pooling(name='pool', kernel=(5,5), pad=(2,2), pool_type='max')
# check_consistency(sym, ctx_list)
2016-11-17 22:19:38 -08:00
sym = mx . sym . Pooling ( name = ' pool ' , kernel = ( 3 , 3 ) , pad = ( 1 , 1 ) , pool_type = ' sum ' )
check_consistency ( sym , ctx_list )
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-03-21 23:01:57 -07:00
def test_pooling_versions ( ) :
def test_pooling_versions_helper ( pool_op_list , data , kernel , pool_type , pad , stride ,
pooling_convention = ' valid ' , global_pool = False ) :
ctx_list = [ ]
sym_list = [ ]
# PoolingV1 cpu
if ' pool_v1_cpu ' in pool_op_list :
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
if not global_pool :
sym_list . append ( mx . sym . Pooling_v1 ( kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = pooling_convention , name = ' pool ' ) )
else :
sym_list . append ( mx . sym . Pooling_v1 ( kernel = kernel , pool_type = pool_type , global_pool = True , name = ' pool ' ) )
# PoolingV1 gpu
if ' pool_v1_gpu ' in pool_op_list :
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
if not global_pool :
sym_list . append ( mx . sym . Pooling_v1 ( kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = pooling_convention , name = ' pool ' ) )
else :
sym_list . append ( mx . sym . Pooling_v1 ( kernel = kernel , pool_type = pool_type , global_pool = True , name = ' pool ' ) )
# Pooling cpu
if ' pool_cpu ' in pool_op_list :
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
if not global_pool :
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = pooling_convention , name = ' pool ' ) )
else :
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pool_type = pool_type , global_pool = True , name = ' pool ' ) )
# Pooling gpu
if ' pool_gpu ' in pool_op_list :
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
if not global_pool :
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = pooling_convention , cudnn_off = True , name = ' pool ' ) )
else :
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pool_type = pool_type , global_pool = True , cudnn_off = True ,
name = ' pool ' ) )
# CuDNNPooling
if ' pool_cudnn ' in pool_op_list :
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
if not global_pool :
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = pooling_convention , cudnn_off = False , name = ' pool ' ) )
else :
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pool_type = pool_type , global_pool = True , cudnn_off = False ,
name = ' pool ' ) )
check_consistency ( sym_list , ctx_list )
def test_1d_pooling ( pool_type ) :
data = ( 2 , 3 , 20 )
kernel = ( 4 , )
pad = ( 0 , )
stride = ( 1 , )
test_pooling_versions_helper ( pool_op_list = [ ' pool_cpu ' , ' pool_gpu ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = ' valid ' , global_pool = False )
pad = ( 2 , )
stride = ( 2 , )
test_pooling_versions_helper ( pool_op_list = [ ' pool_cpu ' , ' pool_gpu ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = ' valid ' , global_pool = False )
pad = ( 0 , )
stride = ( 1 , )
test_pooling_versions_helper ( pool_op_list = [ ' pool_cpu ' , ' pool_gpu ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = ' full ' , global_pool = False )
pad = ( 2 , )
stride = ( 2 , )
test_pooling_versions_helper ( pool_op_list = [ ' pool_cpu ' , ' pool_gpu ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = ' full ' , global_pool = False )
test_pooling_versions_helper ( pool_op_list = [ ' pool_cpu ' , ' pool_gpu ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
global_pool = True )
def test_2d_pooling ( pool_type ) :
data = ( 2 , 3 , 20 , 20 )
kernel = ( 4 , 5 )
pad = ( 0 , 0 )
stride = ( 1 , 1 )
test_pooling_versions_helper ( pool_op_list = [ ' pool_v1_cpu ' , ' pool_v1_gpu ' , ' pool_cpu ' , ' pool_gpu ' , ' pool_cudnn ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = ' valid ' , global_pool = False )
# pool_v1 has bugs when pad is not 0, do not test PoolingV1 here
pad = ( 2 , 3 )
stride = ( 2 , 3 )
test_pooling_versions_helper ( pool_op_list = [ ' pool_cpu ' , ' pool_gpu ' , ' pool_cudnn ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = ' valid ' , global_pool = False )
pad = ( 0 , 0 )
stride = ( 1 , 1 )
test_pooling_versions_helper ( pool_op_list = [ ' pool_v1_cpu ' , ' pool_v1_gpu ' , ' pool_cpu ' , ' pool_gpu ' , ' pool_cudnn ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = ' full ' , global_pool = False )
# pool_v1 has bugs when pad is not 0, do not test PoolingV1 here
pad = ( 2 , 3 )
stride = ( 2 , 3 )
test_pooling_versions_helper ( pool_op_list = [ ' pool_cpu ' , ' pool_gpu ' , ' pool_cudnn ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = ' full ' , global_pool = False )
test_pooling_versions_helper ( pool_op_list = [ ' pool_v1_cpu ' , ' pool_v1_gpu ' , ' pool_cpu ' , ' pool_gpu ' , ' pool_cudnn ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
global_pool = True )
def test_3d_pooling ( pool_type ) :
data = ( 2 , 3 , 20 , 20 , 20 )
kernel = ( 4 , 5 , 3 )
pad = ( 0 , 0 , 0 )
stride = ( 1 , 1 , 1 )
test_pooling_versions_helper ( pool_op_list = [ ' pool_cpu ' , ' pool_gpu ' , ' pool_cudnn ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = ' valid ' , global_pool = False )
pad = ( 2 , 3 , 3 )
stride = ( 2 , 3 , 1 )
test_pooling_versions_helper ( pool_op_list = [ ' pool_cpu ' , ' pool_gpu ' , ' pool_cudnn ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = ' valid ' , global_pool = False )
pad = ( 0 , 0 , 0 )
stride = ( 1 , 1 , 1 )
test_pooling_versions_helper ( pool_op_list = [ ' pool_cpu ' , ' pool_gpu ' , ' pool_cudnn ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = ' full ' , global_pool = False )
pad = ( 2 , 3 , 3 )
stride = ( 2 , 3 , 1 )
test_pooling_versions_helper ( pool_op_list = [ ' pool_cpu ' , ' pool_gpu ' , ' pool_cudnn ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = ' full ' , global_pool = False )
test_pooling_versions_helper ( pool_op_list = [ ' pool_cpu ' , ' pool_gpu ' , ' pool_cudnn ' ] ,
data = data , kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
global_pool = True )
test_1d_pooling ( ' max ' )
test_1d_pooling ( ' avg ' )
test_1d_pooling ( ' sum ' )
test_2d_pooling ( ' max ' )
test_2d_pooling ( ' avg ' )
test_2d_pooling ( ' sum ' )
test_3d_pooling ( ' max ' )
test_3d_pooling ( ' avg ' )
test_3d_pooling ( ' sum ' )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2018-02-09 09:53:16 +08:00
def test_global_pooling ( ) :
def test_1d_pooling ( pool_type ) :
data = ( 2 , 3 , 20 )
kernel = ( 4 , )
pad = ( 2 , )
stride = ( 2 , )
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
ctx_list = [ ]
sym_list = [ ]
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
pooling_convention = ' valid '
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-04-10 02:49:24 +08:00
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , name = ' pool ' ) )
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , cudnn_off = False , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , cudnn_off = False , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-04-10 02:49:24 +08:00
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , cudnn_off = False , name = ' pool ' ) )
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , cudnn_off = True , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , cudnn_off = True , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-04-10 02:49:24 +08:00
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , cudnn_off = True , name = ' pool ' ) )
2018-02-09 09:53:16 +08:00
check_consistency ( sym_list , ctx_list )
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
def test_2d_pooling ( pool_type ) :
data = ( 2 , 3 , 20 , 20 )
kernel = ( 4 , 4 )
pad = ( 2 , 2 )
stride = ( 2 , 2 )
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
ctx_list = [ ]
sym_list = [ ]
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
pooling_convention = ' valid '
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling_v1 ( kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling_v1 ( kernel = kernel , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-04-10 02:49:24 +08:00
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling_v1 ( pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , name = ' pool ' ) )
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-04-10 02:49:24 +08:00
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , name = ' pool ' ) )
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , cudnn_off = False , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , cudnn_off = False , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-04-10 02:49:24 +08:00
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , cudnn_off = False , name = ' pool ' ) )
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pad = pad , stride = stride , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , cudnn_off = True , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-02-09 09:53:16 +08:00
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( kernel = kernel , pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , cudnn_off = True , name = ' pool ' ) )
2018-03-30 17:57:39 -07:00
2018-04-10 02:49:24 +08:00
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' pool_data ' : data , ' type_dict ' : { ' pool_data ' : np . float32 } } )
sym_list . append ( mx . sym . Pooling ( pool_type = pool_type ,
pooling_convention = pooling_convention , global_pool = True , cudnn_off = True , name = ' pool ' ) )
2018-02-09 09:53:16 +08:00
check_consistency ( sym_list , ctx_list )
test_1d_pooling ( ' max ' )
test_1d_pooling ( ' avg ' )
test_1d_pooling ( ' sum ' )
test_2d_pooling ( ' max ' )
test_2d_pooling ( ' avg ' )
test_2d_pooling ( ' sum ' )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2016-06-10 11:31:17 +09:00
def test_upsampling_with_type ( ) :
2017-02-08 01:01:35 +08:00
sym = mx . sym . UpSampling ( scale = 2 , num_filter = 2 , name = ' up ' , sample_type = ' nearest ' , num_args = 1 )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' up_arg0 ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' up_arg0 ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' up_arg0 ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' up_arg0 ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' up_arg0 ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' up_arg0 ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' up_arg0 ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' up_arg0 ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' up_arg0 ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' up_arg0 ' : np . float32 } } ]
check_consistency ( sym , ctx_list )
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-02-08 01:01:35 +08:00
def test_upsampling_bilinear_with_type ( ) :
sym = mx . sym . UpSampling ( scale = 2 , num_filter = 2 , name = ' up ' , sample_type = ' bilinear ' , num_args = 1 )
2017-02-09 01:24:53 +08:00
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' up_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' up_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' up_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' up_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' up_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' up_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' up_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' up_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' up_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' up_data ' : np . float32 } } ]
2016-06-10 11:31:17 +09:00
check_consistency ( sym , ctx_list )
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2016-06-10 11:31:17 +09:00
def test_concat_with_type ( ) :
sym = mx . sym . Concat ( name = ' concat ' , num_args = 2 )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' concat_arg1 ' : ( 2 , 10 ) , ' concat_arg0 ' : ( 2 , 10 ) ,
' type_dict ' : { ' concat_arg0 ' : np . float64 , ' concat_arg1 ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' concat_arg1 ' : ( 2 , 10 ) , ' concat_arg0 ' : ( 2 , 10 ) ,
' type_dict ' : { ' concat_arg0 ' : np . float32 , ' concat_arg1 ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' concat_arg1 ' : ( 2 , 10 ) , ' concat_arg0 ' : ( 2 , 10 ) ,
' type_dict ' : { ' concat_arg0 ' : np . float16 , ' concat_arg1 ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' concat_arg1 ' : ( 2 , 10 ) , ' concat_arg0 ' : ( 2 , 10 ) ,
' type_dict ' : { ' concat_arg0 ' : np . float64 , ' concat_arg1 ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' concat_arg1 ' : ( 2 , 10 ) , ' concat_arg0 ' : ( 2 , 10 ) ,
' type_dict ' : { ' concat_arg0 ' : np . float32 , ' concat_arg1 ' : np . float32 } } ]
check_consistency ( sym , ctx_list )
2016-06-09 01:32:07 +09:00
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2016-06-12 12:26:12 +09:00
def test_elementwisesum_with_type ( ) :
2017-05-11 21:36:16 -07:00
dev_types = [ [ mx . gpu ( 0 ) , [ np . float64 , np . float32 , np . float16 ] ] ,
[ mx . cpu ( 0 ) , [ np . float64 , np . float32 ] ] ]
for num_args in range ( 1 , 6 ) :
ews_arg_shape = { }
for i in range ( num_args ) :
ews_arg_shape [ ' ews_arg ' + str ( i ) ] = ( 2 , 10 )
sym = mx . sym . ElementWiseSum ( name = ' ews ' , num_args = num_args )
ctx_list = [ ]
for dev , types in dev_types :
for dtype in types :
ews_arg_dtype = { ' type_dict ' : { } }
for i in range ( num_args ) :
ews_arg_dtype [ ' type_dict ' ] [ ' ews_arg ' + str ( i ) ] = dtype
ctx_elem = { ' ctx ' : dev }
ctx_elem . update ( ews_arg_shape )
ctx_elem . update ( ews_arg_dtype )
ctx_list . append ( ctx_elem )
2016-06-12 12:26:12 +09:00
check_consistency ( sym , ctx_list )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2016-06-10 15:58:43 +09:00
def test_reshape_with_type ( ) :
sym = mx . sym . Reshape ( name = ' reshape ' , shape = ( - 1 , 1 , 1 , 0 ) )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' reshape_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' reshape_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' reshape_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' reshape_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' reshape_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' reshape_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' reshape_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' reshape_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' reshape_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' reshape_data ' : np . float32 } } ]
check_consistency ( sym , ctx_list )
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2016-06-10 16:13:47 +09:00
def test_blockgrad_with_type ( ) :
sym = mx . sym . BlockGrad ( name = ' bg ' )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' bg_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' bg_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' bg_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' bg_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' bg_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' bg_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' bg_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' bg_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' bg_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' bg_data ' : np . float32 } } ]
check_consistency ( sym , ctx_list )
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2016-06-11 10:46:25 +09:00
def test_swapaxis_with_type ( ) :
sym = mx . sym . SwapAxis ( name = ' swap ' , dim1 = 1 )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' swap_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' swap_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' swap_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' swap_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' swap_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' swap_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' swap_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' swap_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' swap_data ' : ( 2 , 2 , 2 , 10 ) , ' type_dict ' : { ' swap_data ' : np . float32 } } ]
check_consistency ( sym , ctx_list )
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2016-03-19 23:45:52 -07:00
def test_fullyconnected_with_type ( ) :
sym = mx . sym . FullyConnected ( num_hidden = 3 , name = ' inner ' )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' inner_data ' : ( 2 , 10 ) , ' type_dict ' : { ' inner_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' inner_data ' : ( 2 , 10 ) , ' type_dict ' : { ' inner_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' inner_data ' : ( 2 , 10 ) , ' type_dict ' : { ' inner_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' inner_data ' : ( 2 , 10 ) , ' type_dict ' : { ' inner_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' inner_data ' : ( 2 , 10 ) , ' type_dict ' : { ' inner_data ' : np . float32 } } ]
2016-05-27 23:27:19 -07:00
check_consistency ( sym , ctx_list )
2017-08-17 21:16:51 -07:00
# Sizes are divisible by 8 to test TensorCore on Volta GPU.
sym = mx . sym . FullyConnected ( num_hidden = 8 , name = ' inner ' )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' inner_data ' : ( 16 , 24 ) , ' type_dict ' : { ' inner_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' inner_data ' : ( 16 , 24 ) , ' type_dict ' : { ' inner_data ' : np . float32 } } ]
check_consistency ( sym , ctx_list )
2016-03-19 23:45:52 -07:00
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2016-03-19 23:45:52 -07:00
def test_activation_with_type ( ) :
sym = mx . sym . Activation ( name = ' act ' , act_type = ' sigmoid ' )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' act_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' act_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' act_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' act_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' act_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' act_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' act_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' act_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' act_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' act_data ' : np . float32 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' act_data ' : ( 2 , 2 , 10 , 10 ) , ' type_dict ' : { ' act_data ' : np . float16 } } ]
2016-05-27 23:27:19 -07:00
check_consistency ( sym , ctx_list )
2015-10-24 15:57:42 -07:00
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2018-02-15 14:44:34 -08:00
def test_lrn ( ) :
sym = mx . sym . LRN ( alpha = 0.0001 , beta = 0.75 , knorm = 2 , nsize = 5 , name = ' lrn ' )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' lrn_data ' : ( 2 , 6 , 10 , 10 ) , ' type_dict ' : { ' lrn_data ' : np . float32 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' lrn_data ' : ( 2 , 6 , 10 , 10 ) , ' type_dict ' : { ' lrn_data ' : np . float32 } } ]
check_consistency ( sym , ctx_list )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2016-06-24 11:13:26 +08:00
def test_embedding_with_type ( ) :
2017-04-13 22:54:27 -07:00
def test_embedding_helper ( data_types , weight_types , low_pad , high_pad ) :
NVD = [ [ 20 , 10 , 20 ] , [ 200 , 10 , 300 ] ]
for N , V , D in NVD :
sym = mx . sym . Embedding ( name = ' embedding ' , input_dim = V , output_dim = D )
ctx_list = [ ]
for data_type in data_types :
for weight_type in weight_types :
ctx_list . append ( { ' ctx ' : mx . gpu ( 0 ) , ' embedding_data ' : ( N , ) ,
' type_dict ' : { ' embedding_data ' : data_type , ' embedding_weight ' : weight_type } } )
ctx_list . append ( { ' ctx ' : mx . cpu ( 0 ) , ' embedding_data ' : ( N , ) ,
' type_dict ' : { ' embedding_data ' : data_type , ' embedding_weight ' : weight_type } } )
arg_params = { ' embedding_data ' : np . random . randint ( low = - low_pad , high = V + high_pad , size = ( N , ) ) }
check_consistency ( sym , ctx_list , grad_req = { ' embedding_data ' : ' null ' , ' embedding_weight ' : ' write ' } ,
arg_params = arg_params )
data_types = [ np . float16 , np . float32 , np . float64 , np . int32 ]
weight_types = [ np . float16 , np . float32 , np . float64 ]
test_embedding_helper ( data_types , weight_types , 5 , 5 )
data_types = [ np . uint8 ]
weight_types = [ np . float16 , np . float32 , np . float64 ]
test_embedding_helper ( data_types , weight_types , 0 , 5 )
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
2017-12-13 23:52:32 +01:00
@unittest.skip ( " test fails intermittently. temporarily disabled till it gets fixed. tracked at https://github.com/apache/incubator-mxnet/issues/8288 " )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-03-01 01:58:32 +08:00
def test_svmoutput_with_type ( ) :
sym = mx . sym . SVMOutput ( name = ' svmoutput ' , use_linear = True )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' svmoutput_data ' : ( 20 , 10 ) , ' type_dict ' : { ' svmoutput_data ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' svmoutput_data ' : ( 20 , 10 ) , ' type_dict ' : { ' svmoutput_data ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) , ' svmoutput_data ' : ( 20 , 10 ) , ' type_dict ' : { ' svmoutput_data ' : np . float16 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' svmoutput_data ' : ( 20 , 10 ) , ' type_dict ' : { ' svmoutput_data ' : np . float64 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' svmoutput_data ' : ( 20 , 10 ) , ' type_dict ' : { ' svmoutput_data ' : np . float32 } } ,
{ ' ctx ' : mx . cpu ( 0 ) , ' svmoutput_data ' : ( 20 , 10 ) , ' type_dict ' : { ' svmoutput_data ' : np . float16 } } ]
check_consistency ( sym , ctx_list )
2017-03-17 12:42:11 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-01-19 01:45:49 +08:00
def test_take_with_type ( ) :
sym = mx . sym . take ( name = ' take ' )
for data_ndim in range ( 2 , 5 ) :
for idx_ndim in range ( 1 , 4 ) :
data_shape = ( )
for _ in range ( data_ndim ) :
data_shape + = ( np . random . randint ( low = 3 , high = 6 ) , )
idx_shape = ( )
for _ in range ( idx_ndim ) :
2017-03-12 15:37:21 -07:00
idx_shape + = ( np . random . randint ( low = 3 , high = 5 ) , )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) , ' take_indices ' : idx_shape ,
' take_a ' : data_shape ,
' type_dict ' : { ' take_indices ' : np . float64 ,
2017-01-19 01:45:49 +08:00
' take_a ' : np . float64 } } ,
2017-03-12 15:37:21 -07:00
{ ' ctx ' : mx . gpu ( 0 ) , ' take_indices ' : idx_shape ,
' take_a ' : data_shape ,
' type_dict ' : { ' take_indices ' : np . float32 ,
2017-01-19 01:45:49 +08:00
' take_a ' : np . float32 } } ,
2017-03-12 15:37:21 -07:00
{ ' ctx ' : mx . gpu ( 0 ) , ' take_indices ' : idx_shape ,
' take_a ' : data_shape ,
' type_dict ' : { ' take_indices ' : np . float16 ,
2017-01-19 01:45:49 +08:00
' take_a ' : np . float16 } } ,
2017-03-12 15:37:21 -07:00
{ ' ctx ' : mx . cpu ( 0 ) , ' take_indices ' : idx_shape ,
' take_a ' : data_shape ,
' type_dict ' : { ' take_indices ' : np . float64 ,
2017-01-19 01:45:49 +08:00
' take_a ' : np . float64 } } ,
2017-03-12 15:37:21 -07:00
{ ' ctx ' : mx . cpu ( 0 ) , ' take_indices ' : idx_shape ,
' take_a ' : data_shape ,
' type_dict ' : { ' take_indices ' : np . float32 ,
2017-01-19 01:45:49 +08:00
' take_a ' : np . float32 } } ,
2017-03-12 15:37:21 -07:00
{ ' ctx ' : mx . cpu ( 0 ) , ' take_indices ' : idx_shape ,
' take_a ' : data_shape ,
' type_dict ' : { ' take_indices ' : np . float16 ,
2017-01-19 01:45:49 +08:00
' take_a ' : np . float16 } } ]
2017-03-12 15:37:21 -07:00
arg_params = { ' take_indices ' : np . random . randint ( low = 0 ,
high = data_shape [ 0 ] ,
size = idx_shape ) ,
2017-01-19 01:45:49 +08:00
' take_a ' : np . random . normal ( size = data_shape ) }
2017-03-12 15:37:21 -07:00
check_consistency ( sym , ctx_list ,
2017-01-19 01:45:49 +08:00
grad_req = { ' take_indices ' : ' null ' ,
' take_a ' : ' write ' } ,
arg_params = arg_params )
2017-03-17 12:42:11 -07:00
2017-02-16 11:15:26 -08:00
def check_rnn_consistency ( cell1 , cell2 ) :
dshape = ( 32 , 5 , 200 )
data = mx . sym . Variable ( ' data ' )
sym1 , _ = cell1 . unroll ( 5 , data , merge_outputs = True )
mod1 = mx . mod . Module ( sym1 , label_names = None , context = mx . gpu ( 0 ) )
mod1 . bind ( data_shapes = [ ( ' data ' , dshape ) ] , label_shapes = None )
sym2 , _ = cell2 . unroll ( 5 , data , merge_outputs = True )
mod2 = mx . mod . Module ( sym2 , label_names = None , context = mx . gpu ( 0 ) )
mod2 . bind ( data_shapes = [ ( ' data ' , dshape ) ] , label_shapes = None )
mod1 . init_params ( )
args , auxs = mod1 . get_params ( )
args = cell1 . unpack_weights ( args )
args = cell2 . pack_weights ( args )
mod2 . set_params ( args , auxs )
batch = mx . io . DataBatch ( data = [ mx . random . uniform ( shape = dshape ) ] , label = [ ] )
2017-03-22 00:11:16 -07:00
mod1 . forward ( batch , is_train = False )
mod2 . forward ( batch , is_train = False )
2017-02-16 11:15:26 -08:00
2017-02-18 21:15:39 +00:00
assert_allclose ( mod1 . get_outputs ( ) [ 0 ] . asnumpy ( ) , mod2 . get_outputs ( ) [ 0 ] . asnumpy ( ) , rtol = 1e-2 , atol = 1e-4 )
2017-02-16 11:15:26 -08:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-02-16 11:15:26 -08:00
def test_rnn ( ) :
fused = mx . rnn . FusedRNNCell ( 100 , num_layers = 2 , mode = ' rnn_relu ' , prefix = ' ' )
stack = mx . rnn . SequentialRNNCell ( )
stack . add ( mx . rnn . RNNCell ( 100 , activation = ' relu ' , prefix = ' l0_ ' ) )
stack . add ( mx . rnn . RNNCell ( 100 , activation = ' relu ' , prefix = ' l1_ ' ) )
check_rnn_consistency ( fused , stack )
check_rnn_consistency ( stack , fused )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-02-16 11:15:26 -08:00
def test_lstm ( ) :
fused = mx . rnn . FusedRNNCell ( 100 , num_layers = 2 , mode = ' lstm ' , prefix = ' ' )
stack = mx . rnn . SequentialRNNCell ( )
stack . add ( mx . rnn . LSTMCell ( 100 , prefix = ' l0_ ' ) )
stack . add ( mx . rnn . LSTMCell ( 100 , prefix = ' l1_ ' ) )
check_rnn_consistency ( fused , stack )
check_rnn_consistency ( stack , fused )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-03-22 00:52:36 +01:00
def test_lstm_forget_bias ( ) :
forget_bias = 2.0
fused = mx . rnn . FusedRNNCell ( 10 , forget_bias = forget_bias , num_layers = 2 , mode = ' lstm ' , prefix = ' ' )
dshape = ( 32 , 1 , 20 )
data = mx . sym . Variable ( ' data ' )
sym , _ = fused . unroll ( 1 , data , merge_outputs = True )
mod = mx . mod . Module ( sym , label_names = None , context = mx . gpu ( 0 ) )
mod . bind ( data_shapes = [ ( ' data ' , dshape ) ] , label_shapes = None )
mod . init_params ( )
args , auxs = mod . get_params ( )
args = fused . unpack_weights ( args )
bias_name = next ( x for x in args if x . endswith ( ' f_bias ' ) )
expected_bias = forget_bias * np . ones ( 10 , )
assert_allclose ( args [ bias_name ] . asnumpy ( ) , expected_bias )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-03-09 01:31:05 +01:00
def test_gru ( ) :
fused = mx . rnn . FusedRNNCell ( 100 , num_layers = 2 , mode = ' gru ' , prefix = ' ' )
stack = mx . rnn . SequentialRNNCell ( )
stack . add ( mx . rnn . GRUCell ( 100 , prefix = ' l0_ ' ) )
stack . add ( mx . rnn . GRUCell ( 100 , prefix = ' l1_ ' ) )
check_rnn_consistency ( fused , stack )
check_rnn_consistency ( stack , fused )
2017-03-22 00:52:36 +01:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-03-12 15:37:21 -07:00
def test_bidirectional ( ) :
fused = mx . rnn . FusedRNNCell ( 100 , num_layers = 2 , mode = ' gru ' , prefix = ' ' ,
bidirectional = True )
stack = mx . rnn . SequentialRNNCell ( )
stack . add ( mx . rnn . BidirectionalCell (
mx . rnn . GRUCell ( 100 , prefix = ' l0_ ' ) ,
mx . rnn . GRUCell ( 100 , prefix = ' r0_ ' ) ,
output_prefix = ' bi_gru_0_ ' ) )
stack . add ( mx . rnn . BidirectionalCell (
mx . rnn . GRUCell ( 100 , prefix = ' l1_ ' ) ,
mx . rnn . GRUCell ( 100 , prefix = ' r1_ ' ) ,
output_prefix = ' bi_gru_1_ ' ) )
check_rnn_consistency ( fused , stack )
check_rnn_consistency ( stack , fused )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-03-13 09:34:21 -07:00
def test_unfuse ( ) :
for mode in [ ' rnn_tanh ' , ' rnn_relu ' , ' lstm ' , ' gru ' ] :
2017-03-22 00:11:16 -07:00
fused = mx . rnn . FusedRNNCell (
100 , num_layers = 2 , mode = mode ,
prefix = ' test_ %s ' % mode ,
bidirectional = True ,
dropout = 0.5 )
2017-03-13 09:34:21 -07:00
stack = fused . unfuse ( )
check_rnn_consistency ( fused , stack )
check_rnn_consistency ( stack , fused )
2017-03-09 01:31:05 +01:00
2018-02-18 03:11:58 -08:00
@with_seed ( 1234 )
2017-06-18 01:53:37 +08:00
def test_psroipooling_with_type ( ) :
arg_params = {
' psroipool_rois ' : np . array ( [ [ 0 , 10 , 22 , 161 , 173 ] , [ 0 , 20 , 15 , 154 , 160 ] ] ) }
# plain psroipooling
2017-08-29 10:34:56 -07:00
sym = mx . sym . contrib . PSROIPooling ( spatial_scale = 0.0625 , output_dim = 2 , pooled_size = 3 , name = ' psroipool ' )
2017-06-18 01:53:37 +08:00
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) ,
' psroipool_data ' : ( 1 , 18 , 14 , 14 ) ,
' psroipool_rois ' : ( 2 , 5 ) ,
' type_dict ' : { ' psroipool_data ' : np . float64 , ' psroipool_rois ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) ,
' psroipool_data ' : ( 1 , 18 , 14 , 14 ) ,
' psroipool_rois ' : ( 2 , 5 ) ,
' type_dict ' : { ' psroipool_data ' : np . float32 , ' psroipool_rois ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) ,
' psroipool_data ' : ( 1 , 18 , 14 , 14 ) ,
' psroipool_rois ' : ( 2 , 5 ) ,
' type_dict ' : { ' psroipool_data ' : np . float16 , ' psroipool_rois ' : np . float16 } } ,
]
check_consistency ( sym , ctx_list , grad_req = { ' psroipool_data ' : ' write ' ,
' psroipool_rois ' : ' null ' } , arg_params = arg_params )
2018-02-18 03:11:58 -08:00
@with_seed ( 1234 )
2017-06-18 01:53:37 +08:00
def test_deformable_psroipooling_with_type ( ) :
arg_params = {
' deformable_psroipool_rois ' : np . array ( [ [ 0 , 10 , 22 , 161 , 173 ] , [ 0 , 20 , 15 , 154 , 160 ] ] ) }
# deformable psroipooling
2017-08-29 10:34:56 -07:00
sym = mx . sym . contrib . DeformablePSROIPooling ( spatial_scale = 0.0625 , sample_per_part = 4 , group_size = 3 , pooled_size = 3 ,
2017-06-18 01:53:37 +08:00
output_dim = 2 , trans_std = 0.1 , no_trans = False , name = ' deformable_psroipool ' )
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) ,
' deformable_psroipool_data ' : ( 1 , 18 , 14 , 14 ) ,
' deformable_psroipool_rois ' : ( 2 , 5 ) ,
' deformable_psroipool_trans ' : ( 2 , 4 , 3 , 3 ) ,
' type_dict ' : { ' deformable_psroipool_data ' : np . float64 , ' deformable_psroipool_rois ' : np . float64 ,
' deformable_psroipool_trans ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) ,
' deformable_psroipool_data ' : ( 1 , 18 , 14 , 14 ) ,
' deformable_psroipool_rois ' : ( 2 , 5 ) ,
' deformable_psroipool_trans ' : ( 2 , 4 , 3 , 3 ) ,
' type_dict ' : { ' deformable_psroipool_data ' : np . float32 , ' deformable_psroipool_rois ' : np . float32 ,
' deformable_psroipool_trans ' : np . float32 } } ,
{ ' ctx ' : mx . gpu ( 0 ) ,
' deformable_psroipool_data ' : ( 1 , 18 , 14 , 14 ) ,
' deformable_psroipool_rois ' : ( 2 , 5 ) ,
' deformable_psroipool_trans ' : ( 2 , 4 , 3 , 3 ) ,
' type_dict ' : { ' deformable_psroipool_data ' : np . float16 , ' deformable_psroipool_rois ' : np . float16 ,
' deformable_psroipool_trans ' : np . float16 } } ,
]
check_consistency ( sym , ctx_list , grad_req = { ' deformable_psroipool_data ' : ' write ' ,
' deformable_psroipool_rois ' : ' null ' ,
' deformable_psroipool_trans ' : ' write ' } , arg_params = arg_params )
2018-02-18 03:11:58 -08:00
@with_seed ( 1234 )
2017-06-18 01:53:37 +08:00
def test_deformable_convolution_with_type ( ) :
2017-08-29 10:34:56 -07:00
sym = mx . sym . contrib . DeformableConvolution ( num_filter = 3 , kernel = ( 3 , 3 ) , name = ' deformable_conv ' )
2017-06-18 01:53:37 +08:00
# since atomicAdd does not support fp16 (which deformable conv uses in backward), we do not test fp16 here
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) ,
' deformable_conv_data ' : ( 2 , 2 , 10 , 10 ) ,
' deformable_conv_offset ' : ( 2 , 18 , 8 , 8 ) ,
' type_dict ' : { ' deformable_conv_data ' : np . float64 , ' deformable_conv_offset ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) ,
' deformable_conv_data ' : ( 2 , 2 , 10 , 10 ) ,
' deformable_conv_offset ' : ( 2 , 18 , 8 , 8 ) ,
' type_dict ' : { ' deformable_conv_data ' : np . float32 , ' deformable_conv_offset ' : np . float32 } } ,
# {'ctx': mx.gpu(0),
# 'deformable_conv_data': (2, 2, 10, 10),
# 'deformable_conv_offset': (2, 18, 8, 8),
# 'type_dict': {'deformable_conv_data': np.float16, 'deformable_conv_offset': np.float16}},
]
# wider tolerance needed for true-fp16 NCHW test above
tol = { np . dtype ( np . float16 ) : 0.5 ,
np . dtype ( np . float32 ) : 1e-3 ,
np . dtype ( np . float64 ) : 1e-5 ,
np . dtype ( np . uint8 ) : 0 ,
np . dtype ( np . int32 ) : 0 }
check_consistency ( sym , ctx_list , tol = tol )
# test ability to turn off training on bias
check_consistency ( sym , ctx_list , grad_req = { ' deformable_conv_data ' : ' write ' ,
' deformable_conv_offset ' : ' write ' ,
' deformable_conv_weight ' : ' write ' ,
' deformable_conv_bias ' : ' null ' } , tol = tol )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-06-18 01:53:37 +08:00
def test_deformable_convolution_options ( ) :
# 2D convolution
# Pad > 0
# since atomicAdd does not support fp16 (which deformable conv uses in backward), we do not test fp16 here
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) ,
' deformable_conv_data ' : ( 2 , 2 , 7 , 7 ) ,
' deformable_conv_offset ' : ( 2 , 18 , 7 , 7 ) ,
' type_dict ' : { ' deformable_conv_data ' : np . float64 , ' deformable_conv_offset ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) ,
' deformable_conv_data ' : ( 2 , 2 , 7 , 7 ) ,
' deformable_conv_offset ' : ( 2 , 18 , 7 , 7 ) ,
' type_dict ' : { ' deformable_conv_data ' : np . float32 , ' deformable_conv_offset ' : np . float32 } } ,
# {'ctx': mx.gpu(0),
# 'deformable_conv_data': (2, 2, 7, 7),
# 'deformable_offset': (2, 18, 7, 7),
# 'type_dict': {'deformable_conv_data': np.float16, 'deformable_offset': np.float16}},
]
2017-08-29 10:34:56 -07:00
sym = mx . sym . contrib . DeformableConvolution ( num_filter = 3 , kernel = ( 3 , 3 ) , pad = ( 1 , 1 ) , name = ' deformable_conv ' )
2017-06-18 01:53:37 +08:00
check_consistency ( sym , ctx_list )
# Stride > 1
# since atomicAdd does not support fp16 (which deformable conv uses in backward), we do not test fp16 here
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) ,
' deformable_conv_data ' : ( 2 , 2 , 7 , 7 ) ,
' deformable_conv_offset ' : ( 2 , 18 , 3 , 3 ) ,
' type_dict ' : { ' deformable_conv_data ' : np . float64 , ' deformable_conv_offset ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) ,
' deformable_conv_data ' : ( 2 , 2 , 7 , 7 ) ,
' deformable_conv_offset ' : ( 2 , 18 , 3 , 3 ) ,
' type_dict ' : { ' deformable_conv_data ' : np . float32 , ' deformable_conv_offset ' : np . float32 } } ,
# {'ctx': mx.gpu(0),
# 'deformable_conv_data': (2, 2, 7, 7),
# 'deformable_conv_offset': (2, 18, 3, 3),
# 'type_dict': {'deformable_conv_data': np.float16, 'deformable_offset': np.float16}},
]
2017-08-29 10:34:56 -07:00
sym = mx . sym . contrib . DeformableConvolution ( num_filter = 3 , kernel = ( 3 , 3 ) , stride = ( 2 , 2 ) , name = ' deformable_conv ' )
2017-06-18 01:53:37 +08:00
check_consistency ( sym , ctx_list )
# Dilate > 1
# since atomicAdd does not support fp16 (which deformable conv uses in backward), we do not test fp16 here
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) ,
' deformable_conv_data ' : ( 2 , 2 , 7 , 7 ) ,
' deformable_conv_offset ' : ( 2 , 18 , 3 , 3 ) ,
' type_dict ' : { ' deformable_conv_data ' : np . float64 , ' deformable_conv_offset ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) ,
' deformable_conv_data ' : ( 2 , 2 , 7 , 7 ) ,
' deformable_conv_offset ' : ( 2 , 18 , 3 , 3 ) ,
' type_dict ' : { ' deformable_conv_data ' : np . float32 , ' deformable_conv_offset ' : np . float32 } } ,
# {'ctx': mx.gpu(0),
# 'deformable_conv_data': (2, 2, 7, 7),
# 'deformable_conv_offset': (2, 18, 3, 3),
# 'type_dict': {'deformable_conv_data': np.float16, 'deformable_offset': np.float16}},
]
2017-08-29 10:34:56 -07:00
sym = mx . sym . contrib . DeformableConvolution ( num_filter = 3 , kernel = ( 3 , 3 ) , dilate = ( 2 , 2 ) , name = ' deformable_conv ' )
2017-06-18 01:53:37 +08:00
check_consistency ( sym , ctx_list )
# Deformable group > 1
# since atomicAdd does not support fp16 (which deformable conv uses in backward), we do not test fp16 here
ctx_list = [ { ' ctx ' : mx . gpu ( 0 ) ,
' deformable_conv_data ' : ( 2 , 2 , 7 , 7 ) ,
' deformable_conv_offset ' : ( 2 , 36 , 5 , 5 ) ,
' type_dict ' : { ' deformable_conv_data ' : np . float64 , ' deformable_conv_offset ' : np . float64 } } ,
{ ' ctx ' : mx . gpu ( 0 ) ,
' deformable_conv_data ' : ( 2 , 2 , 7 , 7 ) ,
' deformable_conv_offset ' : ( 2 , 36 , 5 , 5 ) ,
' type_dict ' : { ' deformable_conv_data ' : np . float32 , ' deformable_conv_offset ' : np . float32 } } ,
# {'ctx': mx.gpu(0),
# 'deformable_conv_data': (2, 2, 7, 7),
# 'deformable_conv_offset': (2, 36, 5, 5),
# 'type_dict': {'deformable_conv_data': np.float16, 'deformable_offset': np.float16}},
]
2017-08-29 10:34:56 -07:00
sym = mx . sym . contrib . DeformableConvolution ( num_filter = 4 , kernel = ( 3 , 3 ) , num_deformable_group = 2 ,
2017-06-18 01:53:37 +08:00
name = ' deformable_conv ' )
2017-05-31 09:56:32 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-05-26 15:00:17 -07:00
def test_residual_fused ( ) :
cell = mx . rnn . ResidualCell (
mx . rnn . FusedRNNCell ( 50 , num_layers = 3 , mode = ' lstm ' ,
prefix = ' rnn_ ' , dropout = 0.5 ) )
inputs = [ mx . sym . Variable ( ' rnn_t %d _data ' % i ) for i in range ( 2 ) ]
outputs , _ = cell . unroll ( 2 , inputs , merge_outputs = None )
assert sorted ( cell . params . _params . keys ( ) ) == \
[ ' rnn_parameters ' ]
args , outs , auxs = outputs . infer_shape ( rnn_t0_data = ( 10 , 50 ) , rnn_t1_data = ( 10 , 50 ) )
assert outs == [ ( 10 , 2 , 50 ) ]
outputs = outputs . eval ( ctx = mx . gpu ( 0 ) ,
rnn_t0_data = mx . nd . ones ( ( 10 , 50 ) , ctx = mx . gpu ( 0 ) ) + 5 ,
rnn_t1_data = mx . nd . ones ( ( 10 , 50 ) , ctx = mx . gpu ( 0 ) ) + 5 ,
rnn_parameters = mx . nd . zeros ( ( 61200 , ) , ctx = mx . gpu ( 0 ) ) )
expected_outputs = np . ones ( ( 10 , 2 , 50 ) ) + 5
assert np . array_equal ( outputs [ 0 ] . asnumpy ( ) , expected_outputs )
2018-02-18 03:11:58 -08:00
2017-06-26 22:37:11 -07:00
def check_rnn_layer ( layer ) :
2017-07-10 09:20:03 -07:00
layer . collect_params ( ) . initialize ( ctx = [ mx . cpu ( 0 ) , mx . gpu ( 0 ) ] )
2017-06-26 22:37:11 -07:00
with mx . gpu ( 0 ) :
x = mx . nd . ones ( ( 10 , 16 , 30 ) )
states = layer . begin_state ( 16 )
go , gs = layer ( x , states )
2017-06-05 10:07:12 -07:00
2017-06-26 22:37:11 -07:00
with mx . cpu ( 0 ) :
x = mx . nd . ones ( ( 10 , 16 , 30 ) )
states = layer . begin_state ( 16 )
co , cs = layer ( x , states )
2017-06-05 10:07:12 -07:00
2018-02-18 03:11:58 -08:00
# atol of 1e-6 required, as exposed by seed 2124685726
assert_almost_equal ( go . asnumpy ( ) , co . asnumpy ( ) , rtol = 1e-2 , atol = 1e-6 )
2017-06-26 22:37:11 -07:00
for g , c in zip ( gs , cs ) :
2018-02-18 03:11:58 -08:00
assert_almost_equal ( g . asnumpy ( ) , c . asnumpy ( ) , rtol = 1e-2 , atol = 1e-6 )
2017-06-05 10:07:12 -07:00
2018-03-09 21:46:34 -08:00
def check_rnn_layer_w_rand_inputs ( layer ) :
layer . collect_params ( ) . initialize ( ctx = [ mx . cpu ( 0 ) , mx . gpu ( 0 ) ] )
x = mx . nd . uniform ( shape = ( 10 , 16 , 30 ) )
with mx . gpu ( 0 ) :
x = x . copyto ( mx . gpu ( 0 ) )
states = layer . begin_state ( 16 )
go , gs = layer ( x , states )
with mx . cpu ( 0 ) :
x = x . copyto ( mx . cpu ( 0 ) )
states = layer . begin_state ( 16 )
co , cs = layer ( x , states )
assert_almost_equal ( go . asnumpy ( ) , co . asnumpy ( ) , rtol = 1e-2 , atol = 1e-6 )
for g , c in zip ( gs , cs ) :
assert_almost_equal ( g . asnumpy ( ) , c . asnumpy ( ) , rtol = 1e-2 , atol = 1e-6 )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-06-26 22:37:11 -07:00
def test_rnn_layer ( ) :
2017-07-10 10:55:46 -07:00
check_rnn_layer ( gluon . rnn . RNN ( 100 , num_layers = 3 ) )
check_rnn_layer ( gluon . rnn . RNN ( 100 , activation = ' tanh ' , num_layers = 3 ) )
check_rnn_layer ( gluon . rnn . LSTM ( 100 , num_layers = 3 ) )
check_rnn_layer ( gluon . rnn . GRU ( 100 , num_layers = 3 ) )
2017-06-05 10:07:12 -07:00
2017-07-10 10:55:46 -07:00
check_rnn_layer ( gluon . rnn . LSTM ( 100 , num_layers = 3 , bidirectional = True ) )
2018-03-09 21:46:34 -08:00
check_rnn_layer_w_rand_inputs ( gluon . rnn . LSTM ( 100 , num_layers = 3 , bidirectional = True ) )
2017-05-31 09:56:32 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-07-16 17:41:32 -07:00
def test_sequence_reverse ( ) :
check_sequence_reverse ( mx . gpu ( 0 ) )
2018-02-18 03:11:58 -08:00
2017-10-14 19:44:32 -07:00
@unittest.skip ( " Test fails intermittently. Temporarily disabled until fixed. Tracked at https://github.com/apache/incubator-mxnet/issues/8211 " )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-10-14 19:44:32 -07:00
def test_autograd_save_memory ( ) :
x = mx . nd . zeros ( ( 128 , 512 , 512 ) , ctx = mx . gpu ( 0 ) )
x . attach_grad ( )
2017-07-16 17:41:32 -07:00
2017-10-14 19:44:32 -07:00
with mx . autograd . record ( ) :
for i in range ( 200 ) :
x = x + 1
x . wait_to_read ( )
x . backward ( )
2017-08-15 12:24:35 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-08-30 10:28:17 -07:00
def test_gluon_ctc_consistency ( ) :
2017-09-11 10:05:09 -07:00
loss = mx . gluon . loss . CTCLoss ( )
2017-08-30 10:28:17 -07:00
data = mx . nd . arange ( 0 , 4 , repeat = 40 , ctx = mx . gpu ( 0 ) ) . reshape ( ( 2 , 20 , 4 ) ) . flip ( axis = 0 )
2017-09-11 10:05:09 -07:00
cpu_label = mx . nd . array ( [ [ 2 , 1 , - 1 , - 1 ] , [ 3 , 2 , 2 , - 1 ] ] , ctx = mx . cpu ( 0 ) )
gpu_label = mx . nd . array ( [ [ 2 , 1 , - 1 , - 1 ] , [ 3 , 2 , 2 , - 1 ] ] , ctx = mx . gpu ( 0 ) )
2017-08-30 10:28:17 -07:00
cpu_data = data . copy ( ) . as_in_context ( mx . cpu ( 0 ) )
cpu_data . attach_grad ( )
with mx . autograd . record ( ) :
l_cpu = loss ( cpu_data , cpu_label )
l_cpu . backward ( )
gpu_data = data . copyto ( mx . gpu ( 0 ) )
gpu_data . attach_grad ( )
with mx . autograd . record ( ) :
l_gpu = loss ( gpu_data , gpu_label )
l_gpu . backward ( )
assert_almost_equal ( cpu_data . grad . asnumpy ( ) , gpu_data . grad . asnumpy ( ) , atol = 1e-3 , rtol = 1e-3 )
2017-08-15 12:24:35 -07:00
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-09-26 12:10:26 -07:00
def test_cuda_rtc ( ) :
source = r '''
extern " C " __global__ void axpy(const float *x, float *y, float alpha) {
int i = threadIdx.x + blockIdx.x * blockDim.x;
y[i] += alpha * x[i];
}
extern " C " __global__ void saxpy(const float *x, float *y, float alpha) {
extern __shared__ float smem[];
int i = threadIdx.x + blockIdx.x * blockDim.x;
smem[threadIdx.x] = x[i];
y[i] += alpha * smem[threadIdx.x];
}
'''
module = mx . rtc . CudaModule ( source )
axpy = module . get_kernel ( " axpy " , " const float *x, float *y, float alpha " )
x = mx . nd . ones ( ( 10 , ) , ctx = mx . gpu ( 0 ) )
y = mx . nd . zeros ( ( 10 , ) , ctx = mx . gpu ( 0 ) )
axpy . launch ( [ x , y , 3.0 ] , mx . gpu ( 0 ) , ( 1 , 1 , 1 ) , ( 10 , 1 , 1 ) )
assert ( y . asnumpy ( ) == 3 ) . all ( )
saxpy = module . get_kernel ( " saxpy " , " const float *x, float *y, float alpha " )
saxpy . launch ( [ x , y , 4.0 ] , mx . gpu ( 0 ) , ( 1 , 1 , 1 ) , ( 10 , 1 , 1 ) , 10 )
assert ( y . asnumpy ( ) == 7 ) . all ( )
saxpy . launch ( [ x , y , 5.0 ] , mx . gpu ( 0 ) , ( 2 , 1 , 1 ) , ( 5 , 1 , 1 ) , 5 )
assert ( y . asnumpy ( ) == 12 ) . all ( )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-11-25 13:58:18 -08:00
def test_global_norm_clip_multi_device ( ) :
x1 = mx . nd . ones ( ( 3 , 3 ) , ctx = mx . gpu ( 0 ) )
2017-11-26 13:58:45 -08:00
x2 = mx . nd . ones ( ( 4 , 4 ) , ctx = mx . cpu ( 0 ) )
2017-11-25 13:58:18 -08:00
norm = gluon . utils . clip_global_norm ( [ x1 , x2 ] , 1.0 )
assert norm == 5.0
assert_almost_equal ( x1 . asnumpy ( ) , np . ones ( ( 3 , 3 ) ) / 5 )
assert_almost_equal ( x2 . asnumpy ( ) , np . ones ( ( 4 , 4 ) ) / 5 )
2018-02-18 03:11:58 -08:00
@with_seed ( )
2017-10-02 11:48:51 -07:00
def test_cross_device_autograd ( ) :
x = mx . nd . random . uniform ( shape = ( 10 , ) )
x . attach_grad ( )
with mx . autograd . record ( ) :
y = mx . nd . tanh ( x )
y = y . copyto ( mx . gpu ( 0 ) )
y = mx . nd . tanh ( y )
y = y . copyto ( mx . cpu ( 0 ) )
y = mx . nd . tanh ( y )
y = y . copyto ( mx . gpu ( 0 ) )
y = y . copyto ( mx . gpu ( 0 ) )
y . backward ( )
dx = x . grad . asnumpy ( )
x . grad [ : ] = 0
with mx . autograd . record ( ) :
y = x
for i in range ( 3 ) :
y = mx . nd . tanh ( y )
y . backward ( )
assert_almost_equal ( dx , x . grad . asnumpy ( ) )
2018-03-23 01:51:15 +08:00
@unittest.skip ( " JIRA issue: https://issues.apache.org/jira/projects/MXNET/issues/MXNET-130 " )
2018-03-21 06:56:55 +08:00
@with_seed ( )
def test_multi_proposal_op ( ) :
# paramters
feature_stride = 16
scales = ( 8 , 16 , 32 )
ratios = ( 0.5 , 1 , 2 )
rpn_pre_nms_top_n = 12000
rpn_post_nms_top_n = 2000
threshold = 0.7
rpn_min_size = feature_stride
feat_len = ( 1000 + 15 ) / / 16
H , W = feat_len , feat_len
num_anchors = len ( scales ) * len ( ratios )
count_anchors = H * W * num_anchors
def get_new_data ( batch_size , ctx ) :
'''
cls_prob: (batch_size, 2 * num_anchors, H, W)
bbox_pred: (batch_size, 4 * num_anchors, H, W)
im_info: (batch_size, 3)
'''
dtype = np . float32
cls_prob = mx . nd . empty ( ( batch_size , 2 * num_anchors , H , W ) , dtype = dtype , ctx = ctx )
bbox_pred = mx . nd . empty ( ( batch_size , 4 * num_anchors , H , W ) , dtype = dtype , ctx = ctx )
im_info = mx . nd . empty ( ( batch_size , 3 ) , dtype = dtype , ctx = ctx )
cls = [ 1.0 * ( i + 1 ) / cls_prob . size for i in range ( cls_prob . size ) ]
np . random . shuffle ( cls )
cls_prob = mx . nd . reshape ( mx . nd . array ( cls , dtype = dtype , ctx = ctx ) , shape = cls_prob . shape )
bbox_pred = mx . nd . array ( np . random . randint ( - 2 , 3 , size = bbox_pred . shape ) , dtype = dtype , ctx = ctx )
for i in range ( batch_size ) :
im_size = np . random . randint ( 600 , feat_len * feature_stride , size = ( 2 , ) )
im_scale = np . random . randint ( 80 , 100 ) / 100.0
im_info [ i , : ] = [ im_size [ 0 ] , im_size [ 1 ] , im_scale ]
return cls_prob , bbox_pred , im_info
def check_proposal_consistency ( op , batch_size ) :
'''
op is mx.nd.contrib.Proposal or mx.nd.contrib.MultiProposal
'''
cls_prob , bbox_pred , im_info = get_new_data ( batch_size , mx . cpu ( 0 ) )
rois_cpu , score_cpu = op (
cls_score = cls_prob ,
bbox_pred = bbox_pred ,
im_info = im_info ,
feature_stride = feature_stride ,
scales = scales ,
ratios = ratios ,
rpn_pre_nms_top_n = rpn_pre_nms_top_n ,
rpn_post_nms_top_n = rpn_post_nms_top_n ,
threshold = threshold ,
rpn_min_size = rpn_min_size , output_score = True )
gpu_ctx = mx . gpu ( 0 )
# copy data to gpu from cpu
cls_prob_gpu = cls_prob . as_in_context ( gpu_ctx )
bbox_pred_gpu = bbox_pred . as_in_context ( gpu_ctx )
im_info_gpu = im_info . as_in_context ( gpu_ctx )
rois_gpu , score_gpu = op (
cls_score = cls_prob_gpu ,
bbox_pred = bbox_pred_gpu ,
im_info = im_info_gpu ,
feature_stride = feature_stride ,
scales = scales ,
ratios = ratios ,
rpn_pre_nms_top_n = rpn_pre_nms_top_n ,
rpn_post_nms_top_n = rpn_post_nms_top_n ,
threshold = threshold ,
rpn_min_size = rpn_min_size , output_score = True )
rois_cpu_np = rois_cpu . asnumpy ( )
rois_gpu_np = rois_gpu . asnumpy ( )
score_cpu_np = score_cpu . asnumpy ( )
score_gpu_np = score_gpu . asnumpy ( )
assert_almost_equal ( score_cpu_np , score_gpu_np , atol = 1e-3 , rtol = 1e-3 )
assert_almost_equal ( rois_cpu_np , rois_gpu_np , atol = 1e-3 , rtol = 1e-3 )
check_proposal_consistency ( mx . nd . contrib . Proposal , 1 )
check_proposal_consistency ( mx . nd . contrib . MultiProposal , 20 )
2018-01-30 10:45:25 -08:00
# The following 2 functions launch 0-thread kernels, an error that should be caught and signaled.
def kernel_error_check_imperative ( ) :
os . environ [ ' MXNET_ENGINE_TYPE ' ] = ' NaiveEngine '
a = mx . nd . array ( [ 1 , 2 , 3 ] , ctx = mx . gpu ( 0 ) )
b = mx . nd . array ( [ ] , ctx = mx . gpu ( 0 ) )
c = ( a / b ) . asnumpy ( )
def kernel_error_check_symbolic ( ) :
os . environ [ ' MXNET_ENGINE_TYPE ' ] = ' NaiveEngine '
a = mx . sym . Variable ( ' a ' )
b = mx . sym . Variable ( ' b ' )
c = a / b
f = c . bind ( mx . gpu ( 0 ) , { ' a ' : mx . nd . array ( [ 1 , 2 , 3 ] , ctx = mx . gpu ( 0 ) ) ,
' b ' : mx . nd . array ( [ ] , ctx = mx . gpu ( 0 ) ) } )
f . forward ( )
g = f . outputs [ 0 ] . asnumpy ( )
def test_kernel_error_checking ( ) :
# Running tests that may throw exceptions out of worker threads will stop CI testing
# if not run in a separate process (with its own address space for CUDA compatibility).
try :
mpctx = mp . get_context ( ' spawn ' )
except :
print ( ' SKIP: python %s . %s lacks the required process fork-exec support ... ' %
sys . version_info [ 0 : 2 ] , file = sys . stderr , end = ' ' )
else :
with discard_stderr ( ) :
for f in [ kernel_error_check_imperative , kernel_error_check_symbolic ] :
p = mpctx . Process ( target = f )
p . start ( )
p . join ( )
assert p . exitcode != 0 , \
" Expected a synchronous kernel error from %s (), none seen. " % f . __name__
2018-04-03 10:33:56 -07:00
def test_incorrect_gpu ( ) :
# Try setting dev_id to a really big number
assert_raises ( MXNetError , mx . nd . ones , ( 2 , 2 ) , ctx = mx . gpu ( 100001 ) )
2018-01-30 10:45:25 -08:00
2018-04-09 14:43:53 -07:00
@with_seed ( )
def test_batchnorm_backwards_notrain ( ) :
for ctx in [ mx . cpu ( 0 ) , mx . gpu ( 0 ) ] :
for cudnn_o in [ False , True ] :
B , C , H , W = 4 , 3 , 2 , 2
x = mx . nd . random . poisson ( 1 , shape = ( B , C , H , W ) ) . as_in_context ( ctx )
gamma = mx . nd . random . normal ( shape = ( C ) ) . as_in_context ( ctx )
beta = mx . nd . random . normal ( shape = ( C ) ) . as_in_context ( ctx )
mean = mx . nd . random . normal ( shape = ( C ) ) . as_in_context ( ctx )
std = mx . nd . random . normal ( shape = ( C ) ) . as_in_context ( ctx )
x . attach_grad ( )
with autograd . record ( False ) :
y = mx . ndarray . BatchNorm ( x , gamma , beta , mean , std . square ( ) ,
fix_gamma = False , cudnn_off = cudnn_o )
loss = y . square ( ) . sum ( )
loss . backward ( train_mode = False )
2017-05-31 09:56:32 -07:00
if __name__ == ' __main__ ' :
import nose
nose . runmodule ( )
