2020-03-23 22:01:54 +08:00
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed 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.
2020-04-26 19:55:57 +08:00
2024-06-19 22:00:57 +08:00
from __future__ import annotations
2024-08-10 01:59:21 +08:00
from typing import TYPE_CHECKING
2024-06-19 22:00:57 +08:00
2025-06-24 14:16:09 +08:00
import paddle
2023-11-02 14:22:52 +08:00
from paddle import _C_ops , _legacy_C_ops , get_flags , in_dynamic_mode , pir
2023-10-18 14:55:07 +08:00
from paddle . base . framework import _global_flags , in_dynamic_or_pir_mode
2022-11-29 18:50:04 +08:00
from paddle . device import (
2023-04-21 11:56:51 +08:00
get_all_custom_device_type ,
2022-11-29 18:50:04 +08:00
is_compiled_with_cuda ,
is_compiled_with_rocm ,
)
2023-03-20 19:07:02 +08:00
from paddle . tensor . manipulation import reshape
2022-12-12 13:09:14 +08:00
from paddle . tensor . math import _add_with_axis
2026-01-30 20:20:08 +08:00
from paddle . utils . decorator_utils import param_one_alias
2022-11-29 18:50:04 +08:00
2023-09-07 17:26:19 +08:00
from . . . base . data_feeder import check_dtype , check_variable_and_dtype
from . . . base . layer_helper import LayerHelper
remove fluid.initializer.UniformInitializer, ConstantInitializer, NormalInitializer, TruncatedNormalInitializer, XavierInitializer, BilinearInitializer, MSRAInitializer, NumpyArrayInitializer and calculate_gain.. (#49498)
* move UniformInitializer and ConstantInitializer
* more modify
* circular import resolved
* another circular import resolved?
* more circular import 2
* circular import 3
* change import paddle in metric.py
* BuildStrategy import from fluid
* modify the framework import path in common.py
* change rnn.py import, from static to original framework
* change import static in the nn folder
* default_main_program should import from common_ops_import
* add import paddle in param_attr.py
* use core not paddle module for using VarDesc
* another old uniform
* mistake that use Uniform instead of UniformInitializer
* modify UniformInitializer doc
* move fluid.NormalInitializer to nn.initializer.NormalInitializer
* remove import of Normal in fluid.layers.nn.py
* remove more import of old Normal
* remove more import of old Normal
* sample code modify and tests modify import
* is_listen_failed passing arg should be log file
* problem solved
* a mistake solved
* comments resoleved and remove paddle.fluid.initializer.TruncatedNormalInitializer
* remove paddle.fluid.initializer.XavierInitializer and paddle.fluid.initializer.MSRAInitializer
* remove paddle.fluid.initializer.BilinearInitializer NumpyArrayInitializer and set_global_initializer
* change fluid to static
* change static to fluid to avoid circular import in distributed_strategy.py
* fix example code and test_initializer
* ValueType
* sample code fix
* change set_global_initializer back to fluid
* put paddle.static.BuildStrategy.ReduceStrategy into the fuction to avoid circular import
* remove calculate_gain, delete BilinearInitializer and revert set_global_initializer
* change the time of using UniformInitializer, ConstantInitializer, NormalInitializer, TruncatedNormalInitializer, XavierInitializer, MSRAInitializer, NumpyArrayInitializer as few as possible
* fix argument incampatible
* fix more arg incompatible
* fix test_prelu_op_xpu.py Constant
* fix inaccurate doc
* more doc fix: default value
2023-02-01 21:38:27 +08:00
from . . . common_ops_import import Variable
2020-08-22 19:57:15 +08:00
from . . . device import get_cudnn_version
2023-04-21 11:56:51 +08:00
from . . . framework import no_grad
2023-03-09 20:30:52 +08:00
from . . . tensor . manipulation import squeeze , unsqueeze
from . . . utils import (
2022-10-23 20:01:27 +08:00
_contain_var ,
_convert_to_tensor_list ,
2022-11-29 18:50:04 +08:00
_is_symmetric_padding ,
convert_to_list ,
2022-10-23 20:01:27 +08:00
)
2020-04-04 03:04:07 +08:00
2024-06-29 11:36:39 +08:00
if TYPE_CHECKING :
2024-08-10 01:59:21 +08:00
from collections . abc import Sequence
2024-06-29 11:36:39 +08:00
from paddle import Tensor
from paddle . _typing import (
DataLayout1D ,
DataLayout2D ,
DataLayout3D ,
DataLayoutND ,
Size1 ,
Size2 ,
Size3 ,
Size4 ,
Size6 ,
)
from . common import _PaddingSizeMode
2021-04-29 19:31:40 +08:00
__all__ = [ ]
2020-04-04 03:04:07 +08:00
def _is_list_or_tuple ( input ) :
return isinstance ( input , ( list , tuple ) )
def _zero_padding_in_batch_and_channel ( padding , channel_last ) :
if channel_last :
return list ( padding [ 0 ] ) == [ 0 , 0 ] and list ( padding [ - 1 ] ) == [ 0 , 0 ]
else :
return list ( padding [ 0 ] ) == [ 0 , 0 ] and list ( padding [ 1 ] ) == [ 0 , 0 ]
def _exclude_padding_in_batch_and_channel ( padding , channel_last ) :
padding_ = padding [ 1 : - 1 ] if channel_last else padding [ 2 : ]
padding_ = [ elem for pad_a_dim in padding_ for elem in pad_a_dim ]
return padding_
def _update_padding_nd ( padding , channel_last , num_dims ) :
if isinstance ( padding , str ) :
padding = padding . upper ( )
if padding not in [ " SAME " , " VALID " ] :
raise ValueError (
2023-09-22 10:14:38 +08:00
f " Unknown padding: ' { padding } ' . It can only be ' SAME ' or ' VALID ' . "
2022-10-23 20:01:27 +08:00
)
2020-04-04 03:04:07 +08:00
if padding == " VALID " :
padding_algorithm = " VALID "
padding = [ 0 ] * num_dims
else :
padding_algorithm = " SAME "
padding = [ 0 ] * num_dims
elif _is_list_or_tuple ( padding ) :
# for padding like
# [(pad_before, pad_after), (pad_before, pad_after), ...]
# padding for batch_dim and channel_dim included
if len ( padding ) == 2 + num_dims and _is_list_or_tuple ( padding [ 0 ] ) :
if not _zero_padding_in_batch_and_channel ( padding , channel_last ) :
raise ValueError (
2023-09-22 10:14:38 +08:00
f " Non-zero padding( { padding } ) in the batch or channel dimensions "
" is not supported. "
2022-10-23 20:01:27 +08:00
)
2020-04-04 03:04:07 +08:00
padding_algorithm = " EXPLICIT "
2022-06-05 10:58:58 +08:00
padding = _exclude_padding_in_batch_and_channel (
2022-10-23 20:01:27 +08:00
padding , channel_last
)
2021-07-26 15:10:15 +08:00
if _is_symmetric_padding ( padding , num_dims ) :
2020-04-04 03:04:07 +08:00
padding = padding [ 0 : : 2 ]
# for padding like [pad_before, pad_after, pad_before, pad_after, ...]
elif len ( padding ) == 2 * num_dims and isinstance ( padding [ 0 ] , int ) :
padding_algorithm = " EXPLICIT "
2021-07-26 15:10:15 +08:00
padding = convert_to_list ( padding , 2 * num_dims , ' padding ' )
if _is_symmetric_padding ( padding , num_dims ) :
2020-04-04 03:04:07 +08:00
padding = padding [ 0 : : 2 ]
# for padding like [pad_d1, pad_d2, ...]
elif len ( padding ) == num_dims and isinstance ( padding [ 0 ] , int ) :
padding_algorithm = " EXPLICIT "
2021-07-26 15:10:15 +08:00
padding = convert_to_list ( padding , num_dims , ' padding ' )
2020-04-04 03:04:07 +08:00
else :
2023-03-31 10:11:56 +08:00
raise ValueError ( f " In valid padding: { padding } " )
2020-04-04 03:04:07 +08:00
# for integer padding
else :
padding_algorithm = " EXPLICIT "
2021-07-26 15:10:15 +08:00
padding = convert_to_list ( padding , num_dims , ' padding ' )
2023-06-09 10:14:23 +08:00
if not all ( p > = 0 for p in padding ) :
2021-05-31 20:12:04 +08:00
raise ValueError (
2023-11-02 18:28:29 +08:00
f " Invalid padding, all value should be larger than or equal to 0, but received: { padding } "
2022-10-23 20:01:27 +08:00
)
2020-04-04 03:04:07 +08:00
return padding , padding_algorithm
2026-03-03 14:52:13 +08:00
_MEMORY_FORMAT_CONTIGUOUS = 0
_MEMORY_FORMAT_CHANNELS_LAST = 1
_MEMORY_FORMAT_CHANNELS_LAST_3D = 2
def _cudnn_conv_suggest_memory_format (
input : paddle . Tensor , weight : paddle . Tensor , data_format : str = " NCHW "
) - > int :
# Disable NHWC for float64 input/weight
if input . dtype == paddle . float64 or weight . dtype == paddle . float64 :
return _MEMORY_FORMAT_CONTIGUOUS
cudnn_version = get_cudnn_version ( )
weight_ndim = weight . ndim
input_memory_format_is_cl = data_format == " NHWC "
weight_memory_format_is_cl = False
can_use_cudnn_channels_last_2d = (
( cudnn_version > = 7603 )
and ( weight_ndim == 4 )
and ( input_memory_format_is_cl or weight_memory_format_is_cl )
)
if can_use_cudnn_channels_last_2d :
return _MEMORY_FORMAT_CHANNELS_LAST
can_use_cudnn_channels_last_3d = (
( cudnn_version > = 8005 )
and ( weight_ndim == 5 )
and ( input_memory_format_is_cl or weight_memory_format_is_cl )
)
if can_use_cudnn_channels_last_3d :
return _MEMORY_FORMAT_CHANNELS_LAST_3D
return _MEMORY_FORMAT_CONTIGUOUS
def _is_cudnn_supported (
x : paddle . Tensor ,
weight : paddle . Tensor ,
data_format : str ,
start_use_cudnn : bool ,
) - > bool :
if not start_use_cudnn :
return False
if not ( paddle . is_compiled_with_cuda ( ) and x . place . is_gpu_place ( ) ) :
return False
cudnn_version = get_cudnn_version ( )
is_low_precision = x . dtype in [ paddle . bfloat16 , paddle . float16 ]
# cuDNN Version Specific Bugs (9.8 - 9.14) for 3D Conv
if (
90800 < = cudnn_version < 91500
and _cudnn_conv_suggest_memory_format ( x , weight , data_format )
== _MEMORY_FORMAT_CONTIGUOUS
and is_low_precision
and weight . ndim == 5
) :
kernel_is_trivial = True
for k in weight . shape [ 2 : ] :
if k != 1 :
kernel_is_trivial = False
break
if not kernel_is_trivial :
return False
return True
2022-10-23 20:01:27 +08:00
def _conv_nd (
2024-06-29 03:38:56 +08:00
x : Tensor ,
weight : Tensor ,
bias : Tensor | None = None ,
stride : int | Sequence [ int ] = 1 ,
padding : _PaddingSizeMode | int | Sequence [ int ] | Sequence [ Size2 ] = 0 ,
2022-10-23 20:01:27 +08:00
padding_algorithm = None ,
2024-06-29 03:38:56 +08:00
dilation : int | Sequence [ int ] = 1 ,
groups : int = 1 ,
data_format : DataLayoutND = " NCHW " ,
channel_dim : int = 1 ,
op_type : str = " conv2d " ,
use_cudnn : bool = True ,
name : str | None = None ,
) - > Tensor :
2025-12-25 17:17:52 +08:00
use_accuracy_compatible = paddle . get_flags (
[ " FLAGS_use_accuracy_compatible_kernel " ]
) . get (
" FLAGS_use_accuracy_compatible_kernel " , False
) # Due to the poor performance of NHWC, we transpose the input to NCHW.
2023-10-18 14:55:07 +08:00
if in_dynamic_or_pir_mode ( ) and op_type == " conv2d " :
2026-03-03 14:52:13 +08:00
# TODO: alignment with PyTorch 2.9.1 use_cudnn logic, will remove in future
if (
in_dynamic_mode ( )
and use_accuracy_compatible
and not _is_cudnn_supported ( x , weight , data_format , use_cudnn )
) :
# x = x._use_gpudnn(False)
return _C_ops . slow_conv2d_dilated (
x ,
weight ,
bias ,
stride ,
padding ,
padding_algorithm ,
dilation ,
groups ,
data_format ,
)
2022-10-23 20:01:27 +08:00
pre_bias = _C_ops . conv2d (
x ,
weight ,
stride ,
padding ,
padding_algorithm ,
dilation ,
2022-10-31 21:47:59 -05:00
groups ,
2022-10-23 20:01:27 +08:00
data_format ,
)
2022-04-06 08:58:59 +08:00
if bias is not None :
2022-11-30 18:34:20 +08:00
new_shape = [ 1 ] * len ( x . shape )
new_shape [ channel_dim ] = - 1
bias = bias . reshape ( new_shape )
2024-01-18 11:28:47 +08:00
# TODO(qili93): temporary for ascend npu performance to be removed along with npu_identity op
2023-04-21 11:56:51 +08:00
if (
_global_flags ( ) [ ' FLAGS_npu_storage_format ' ]
and ' npu ' in get_all_custom_device_type ( )
) :
with no_grad ( ) :
bias_storage = _C_ops . npu_identity (
bias , 3
) # ACL_FORMAT_NC1HWC0 = 3
bias_storage . _share_underline_tensor_to ( bias )
2022-11-28 13:26:54 +08:00
return _C_ops . add ( pre_bias , bias )
2022-04-06 08:58:59 +08:00
else :
return pre_bias
2022-06-01 10:03:49 +08:00
2023-10-18 14:55:07 +08:00
if in_dynamic_or_pir_mode ( ) and op_type == " depthwise_conv2d " :
2025-12-25 17:17:52 +08:00
if use_accuracy_compatible and is_compiled_with_cuda ( ) :
return _C_ops . depthwise_conv2d_bias (
x ,
weight ,
bias ,
stride ,
padding ,
padding_algorithm ,
groups ,
dilation ,
data_format ,
)
else :
pre_bias = _C_ops . depthwise_conv2d (
x ,
weight ,
stride ,
padding ,
padding_algorithm ,
groups ,
dilation ,
data_format ,
)
if bias is not None :
new_shape = [ 1 ] * len ( x . shape )
new_shape [ channel_dim ] = - 1
bias = bias . reshape ( new_shape )
return _C_ops . add ( pre_bias , bias )
else :
return pre_bias
if in_dynamic_or_pir_mode ( ) and op_type == " conv3d " :
2026-03-03 14:52:13 +08:00
# TODO: alignment with PyTorch 2.9.1 use_cudnn logic, will remove in future
if (
in_dynamic_mode ( )
and use_accuracy_compatible
and not _is_cudnn_supported ( x , weight , data_format , use_cudnn )
) :
# x = x._use_gpudnn(False)
return _C_ops . slow_conv3d_dilated (
x ,
weight ,
bias ,
stride ,
padding ,
padding_algorithm ,
groups ,
dilation ,
data_format ,
)
2025-12-25 17:17:52 +08:00
pre_bias = _C_ops . conv3d (
2022-10-23 20:01:27 +08:00
x ,
weight ,
stride ,
padding ,
padding_algorithm ,
groups ,
dilation ,
data_format ,
)
2022-06-01 10:03:49 +08:00
if bias is not None :
2022-11-30 18:34:20 +08:00
new_shape = [ 1 ] * len ( x . shape )
new_shape [ channel_dim ] = - 1
bias = bias . reshape ( new_shape )
return _C_ops . add ( pre_bias , bias )
2022-06-01 10:03:49 +08:00
else :
return pre_bias
2025-12-25 17:17:52 +08:00
if in_dynamic_or_pir_mode ( ) and op_type == " depthwise_conv3d " :
return _C_ops . depthwise_conv3d_bias (
2022-10-23 20:01:27 +08:00
x ,
weight ,
2025-12-25 17:17:52 +08:00
bias ,
2022-10-23 20:01:27 +08:00
stride ,
padding ,
padding_algorithm ,
groups ,
dilation ,
data_format ,
)
2022-06-01 10:03:49 +08:00
2022-02-22 15:07:11 +08:00
if in_dynamic_mode ( ) :
2022-10-23 20:01:27 +08:00
attrs = (
' strides ' ,
stride ,
' paddings ' ,
padding ,
' dilations ' ,
dilation ,
' groups ' ,
groups ,
' use_cudnn ' ,
use_cudnn ,
' fuse_relu_before_depthwise_conv ' ,
False ,
" padding_algorithm " ,
padding_algorithm ,
" data_format " ,
data_format ,
)
2022-08-26 16:11:47 +08:00
pre_bias = getattr ( _legacy_C_ops , op_type ) ( x , weight , * attrs )
2020-11-23 11:50:49 +08:00
if bias is not None :
2022-12-12 13:09:14 +08:00
out = _add_with_axis ( pre_bias , bias , axis = channel_dim )
2020-11-23 11:50:49 +08:00
else :
out = pre_bias
else :
inputs = { ' Input ' : [ x ] , ' Filter ' : [ weight ] }
attrs = {
' strides ' : stride ,
' paddings ' : padding ,
' dilations ' : dilation ,
' groups ' : groups ,
' use_cudnn ' : use_cudnn ,
' fuse_relu_before_depthwise_conv ' : False ,
" padding_algorithm " : padding_algorithm ,
2022-10-23 20:01:27 +08:00
" data_format " : data_format ,
2020-11-23 11:50:49 +08:00
}
2022-10-23 20:01:27 +08:00
check_variable_and_dtype (
2023-04-21 15:16:59 +08:00
x , ' x ' , [ ' float16 ' , ' uint16 ' , ' float32 ' , ' float64 ' ] , op_type
2022-10-23 20:01:27 +08:00
)
2020-11-23 11:50:49 +08:00
helper = LayerHelper ( op_type , * * locals ( ) )
dtype = helper . input_dtype ( input_param_name = ' x ' )
pre_bias = helper . create_variable_for_type_inference ( dtype )
outputs = { " Output " : [ pre_bias ] }
2022-10-23 20:01:27 +08:00
helper . append_op (
type = op_type , inputs = inputs , outputs = outputs , attrs = attrs
)
2020-11-23 11:50:49 +08:00
if bias is not None :
out = helper . create_variable_for_type_inference ( dtype )
2023-03-20 19:07:02 +08:00
x_shape = list ( pre_bias . shape )
y_shape = list ( bias . shape )
if channel_dim == - 1 or len ( x_shape ) == len ( y_shape ) :
helper . append_op (
type = ' elementwise_add ' ,
inputs = { ' X ' : [ pre_bias ] , ' Y ' : [ bias ] } ,
outputs = { ' Out ' : [ out ] } ,
2023-12-11 19:38:07 +08:00
attrs = { ' axis ' : - 1 } ,
2023-03-20 19:07:02 +08:00
)
else :
2025-08-21 01:51:31 +08:00
assert len ( x_shape ) > len ( y_shape ) , (
' The length of pre_bias must greater than the length of bias '
)
2023-03-20 19:07:02 +08:00
padding = len ( x_shape ) - len ( y_shape ) - channel_dim
bias = reshape (
bias , [ 1 ] * channel_dim + y_shape + [ 1 ] * padding
)
helper . append_op (
type = ' elementwise_add ' ,
inputs = { ' X ' : [ pre_bias ] , ' Y ' : [ bias ] } ,
outputs = { ' Out ' : [ out ] } ,
2023-12-11 19:38:07 +08:00
attrs = { ' axis ' : - 1 } ,
2023-03-20 19:07:02 +08:00
)
2020-11-23 11:50:49 +08:00
else :
out = pre_bias
return out
2026-01-30 20:20:08 +08:00
@param_one_alias ( [ " x " , " input " ] )
2022-10-23 20:01:27 +08:00
def conv1d (
2024-06-19 22:00:57 +08:00
x : Tensor ,
weight : Tensor ,
bias : Tensor | None = None ,
stride : Size1 = 1 ,
2024-06-27 20:19:16 +08:00
padding : _PaddingSizeMode | Size1 | Size2 | Sequence [ Size2 ] = 0 ,
2024-06-19 22:00:57 +08:00
dilation : Size1 = 1 ,
groups : int = 1 ,
data_format : DataLayout1D = ' NCL ' ,
name : str | None = None ,
) - > Tensor :
2020-11-24 14:53:51 +08:00
r """
2020-08-23 18:41:20 +08:00
The convolution1D layer calculates the output based on the input, filter
and strides, paddings, dilations, groups parameters. Input and
Output are in NCL format, where N is batch size, C is the number of
channels, L is the length of the feature.
Filter is in MCK format, where M is the number of output image channels,
C is the number of input image channels, K is the size of the kernel.
If the groups is greater than 1, C will equal the number of input image
channels divided by the groups. If bias attribution and activation type
are provided, bias is added to the output of the convolution, and the
corresponding activation function is applied to the final result.
For each input :math:`X`, the equation is:
.. math::
2020-11-27 10:27:17 +08:00
Out = \ sigma (W \ ast X + b)
2020-08-23 18:41:20 +08:00
Where:
* :math:`X`: Input value, a tensor with NCL format.
* :math:`W`: Kernel value, a tensor with MCK format.
* :math:` \ \ ast`: Convolution operation.
* :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
* :math:` \ \ sigma`: Activation function.
* :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
Example:
- Input:
Input shape: :math:`(N, C_ {in} , L_ {in} )`
Filter shape: :math:`(C_ {out} , C_ {in} , L_f)`
- Output:
Output shape: :math:`(N, C_ {out} , L_ {out} )`
Where
.. math::
2020-11-27 10:27:17 +08:00
L_ {out} = \ frac { (L_ {in} + 2 * padding - (dilation * (L_f - 1) + 1))} {stride} + 1
2020-08-23 18:41:20 +08:00
2025-09-18 17:06:08 +08:00
.. note::
Alias Support: The parameter name ``input`` can be used as an alias for ``x``.
2020-08-23 18:41:20 +08:00
Args:
2022-09-14 21:56:19 +08:00
x (Tensor): The input is 3-D Tensor with shape [N, C, L], the data type
2020-08-23 18:41:20 +08:00
of input is float16 or float32 or float64.
2025-09-18 17:06:08 +08:00
Alias: ``input``.
2020-08-23 18:41:20 +08:00
weight (Tensor): The convolution kernel with shape [M, C/g, K], where M is
2022-09-14 21:56:19 +08:00
the number of output channels, g is the number of groups, K is the kernel ' s size.
2020-08-23 18:41:20 +08:00
bias (Tensor, optional): The bias with shape [M,]. Default: None.
2021-04-27 13:41:05 +08:00
stride (int|list|tuple, optional): The stride size. If stride is a list/tuple, it must
2020-08-23 18:41:20 +08:00
contain one integers, (stride_size). Default: 1.
2025-10-13 02:29:05 +08:00
padding (int|str|tuple|list, optional): The padding size.
Padding could be in one of the following forms.
2020-08-23 18:41:20 +08:00
1. a string in [ ' valid ' , ' same ' ].
2. an int, which means the feature map is zero paded by size of `padding` on both sides.
3. a list[int] or tuple[int] whose length is 1, which means the feature map is zero paded by size of `padding[0]` on both sides.
4. a list[int] or tuple[int] whose length is 2. It has the form [pad_before, pad_after].
5. a list or tuple of pairs of ints. It has the form [[pad_before, pad_after], [pad_before, pad_after], ...]. Note that, the batch dimension and channel dimension are also included. Each pair of integers correspond to the amount of padding for a dimension of the input. Padding in batch dimension and channel dimension should be [0, 0] or (0, 0).
2025-10-13 02:29:05 +08:00
2020-08-23 18:41:20 +08:00
The default value is 0.
2021-04-27 13:41:05 +08:00
dilation (int|list|tuple, optional): The dilation size. If dilation is a list/tuple, it must
2020-08-23 18:41:20 +08:00
contain one integer, (dilation_size). Default: 1.
groups (int, optional): The groups number of the conv1d function. According to grouped
convolution in Alex Krizhevsky ' s Deep CNN paper: when group=2,
the first half of the filters is only connected to the first half
of the input channels, while the second half of the filters is only
connected to the second half of the input channels. Default: 1.
2022-09-14 21:56:19 +08:00
data_format (str, optional): Specify the data format of the input, and the data format of the output
2020-08-23 18:41:20 +08:00
will be consistent with that of the input. An optional string from: ` " NCL " `, ` " NLC " `.
The default is ` " NCL " `. When it is ` " NCL " `, the data is stored in the order of:
`[batch_size, input_channels, feature_length]`.
2022-09-14 21:56:19 +08:00
name(str, optional): For detailed information, please refer
to :ref:`api_guide_Name`. Usually name is no need to set and
2020-08-23 18:41:20 +08:00
None by default.
Returns:
2022-09-14 21:56:19 +08:00
A tensor representing the conv1d, whose data type is the
2020-08-23 18:41:20 +08:00
same with input.
Examples:
2026-02-05 05:25:41 +08:00
.. code-block:: pycon
2020-08-23 18:41:20 +08:00
2023-08-31 11:09:42 +08:00
>>> import paddle
>>> import paddle.nn.functional as F
2026-02-05 05:25:41 +08:00
>>> x = paddle.to_tensor(
... [
... [
... [4, 8, 1, 9],
... [7, 2, 0, 9],
... [6, 9, 2, 6],
... ],
... ],
... dtype= " float32 " ,
... )
>>> w = paddle.to_tensor(
... [
... [
... [9, 3, 4],
... [0, 0, 7],
... [2, 5, 6],
... ],
... [
... [0, 3, 4],
... [2, 9, 7],
... [5, 6, 8],
... ],
... ],
... dtype= " float32 " ,
... )
2023-08-31 11:09:42 +08:00
>>> y = F.conv1d(x, w)
>>> print(y)
Tensor(shape=[1, 2, 2], dtype=float32, place=Place(cpu), stop_gradient=True,
[[[133., 238.],
[160., 211.]]])
2020-08-23 18:41:20 +08:00
"""
cudnn_version = get_cudnn_version ( )
if cudnn_version is not None :
use_cudnn = True
else :
use_cudnn = False
if data_format not in [ " NCL " , " NLC " ] :
2022-10-23 20:01:27 +08:00
raise ValueError (
" Attr(data_format) should be ' NCL ' or ' NLC ' . "
2023-09-22 10:14:38 +08:00
f " Received Attr(data_format): { data_format } . "
2022-10-23 20:01:27 +08:00
)
2020-08-23 18:41:20 +08:00
2022-10-23 20:01:27 +08:00
channel_last = data_format == " NLC "
2020-08-23 18:41:20 +08:00
channel_dim = - 1 if channel_last else 1
conv2d_data_format = " NHWC " if channel_last else " NCHW "
2021-09-07 19:28:09 +08:00
if len ( x . shape ) != 3 :
raise ValueError (
2023-09-22 10:14:38 +08:00
f " Input x should be 3D tensor, but received x with the shape of { x . shape } "
2022-10-23 20:01:27 +08:00
)
2020-08-23 18:41:20 +08:00
num_channels = x . shape [ channel_dim ]
num_filters = weight . shape [ 0 ]
2024-08-10 12:07:14 +08:00
if groups == 0 :
raise ValueError ( " The groups of conv1d should not be zero " )
if num_channels % groups != 0 and (
in_dynamic_mode ( ) or ( num_channels != - 1 and groups != - 1 )
) :
2020-08-23 18:41:20 +08:00
raise ValueError (
2024-08-10 12:07:14 +08:00
" The channel of input must be divisible by groups, "
2024-04-01 10:20:33 +08:00
f " received: the channel of input is { num_channels } , the shape of input is { x . shape } "
f " , the groups is { groups } "
2022-10-23 20:01:27 +08:00
)
2024-08-10 12:07:14 +08:00
if num_filters % groups != 0 and (
in_dynamic_mode ( ) or ( num_filters != - 1 and groups != - 1 )
) :
2020-08-23 18:41:20 +08:00
raise ValueError (
2024-08-10 12:07:14 +08:00
" The number of filters must be divisible by groups, "
2024-04-01 10:20:33 +08:00
f " received: the number of filters is { num_filters } , the shape of weight is { weight . shape } "
f " , the groups is { groups } "
2022-10-23 20:01:27 +08:00
)
2020-08-23 18:41:20 +08:00
# update attrs
padding , padding_algorithm = _update_padding_nd ( padding , channel_last , 1 )
2022-02-10 13:53:01 +08:00
2020-08-23 18:41:20 +08:00
if len ( padding ) == 2 :
2024-08-12 18:54:52 +08:00
padding = [ 0 , 0 , * padding ]
2020-08-23 18:41:20 +08:00
elif len ( padding ) == 1 :
2024-08-12 18:54:52 +08:00
padding = [ 0 , * padding ]
2020-08-23 18:41:20 +08:00
else :
raise ValueError (
2023-11-02 18:28:29 +08:00
f " The size of padding ' s dimension should be 1 or 2. But got padding= { padding } "
2022-10-23 20:01:27 +08:00
)
2024-08-12 18:54:52 +08:00
stride = [ 1 , * convert_to_list ( stride , 1 , " stride " ) ]
dilation = [ 1 , * convert_to_list ( dilation , 1 , " dilation " ) ]
2023-07-31 15:05:15 +08:00
from . . . tensor . creation import assign as paddle_assign
2025-06-24 14:16:09 +08:00
# cpu not support float16, need to convert dtype.
float16_convert = False
if paddle . device . get_device ( ) == " cpu " :
if weight . dtype == paddle . float16 :
float16_convert = True
weight = weight . astype ( x . dtype )
if bias is not None and bias . dtype == paddle . float16 :
float16_convert = True
bias = bias . astype ( x . dtype )
2023-07-31 15:05:15 +08:00
weight = paddle_assign ( weight )
2022-02-10 13:53:01 +08:00
weight = unsqueeze ( weight , axis = [ - 2 ] )
2020-08-23 18:41:20 +08:00
l_type = " conv2d "
2022-02-10 13:53:01 +08:00
# When "groups==num_channels and num_filters% num_channels == 0" using depthwise_conv2d has better performance
2022-10-23 20:01:27 +08:00
if (
is_compiled_with_cuda ( )
and num_channels == groups
and num_channels != 1
and num_filters % num_channels == 0
) :
2020-08-23 18:41:20 +08:00
l_type = ' depthwise_conv2d '
use_cudnn = False
2024-01-18 11:28:47 +08:00
squeeze_axis = - 3 if channel_last else - 2
x = unsqueeze ( x , axis = [ squeeze_axis ] )
2023-10-26 10:37:23 +08:00
if in_dynamic_or_pir_mode ( ) :
2022-10-31 21:47:59 -05:00
if l_type == ' conv2d ' :
out = _C_ops . conv2d (
x ,
weight ,
stride ,
padding ,
padding_algorithm ,
dilation ,
groups ,
conv2d_data_format ,
)
else :
2022-11-16 14:36:17 +08:00
out = _C_ops . depthwise_conv2d (
2022-10-31 21:47:59 -05:00
x ,
weight ,
stride ,
padding ,
padding_algorithm ,
groups ,
dilation ,
conv2d_data_format ,
False ,
- 1 ,
False ,
False ,
)
2022-08-22 18:37:13 +08:00
if bias is not None :
2022-12-12 13:09:14 +08:00
out = _add_with_axis ( out , bias , axis = channel_dim )
2020-08-23 18:41:20 +08:00
else :
inputs = { ' Input ' : [ x ] , ' Filter ' : [ weight ] }
attrs = {
' strides ' : stride ,
' paddings ' : padding ,
' dilations ' : dilation ,
' groups ' : groups ,
' use_cudnn ' : use_cudnn ,
' fuse_relu_before_depthwise_conv ' : False ,
" padding_algorithm " : padding_algorithm ,
2022-10-23 20:01:27 +08:00
" data_format " : conv2d_data_format ,
2020-08-23 18:41:20 +08:00
}
2022-10-23 20:01:27 +08:00
check_variable_and_dtype (
x , ' input ' , [ ' float16 ' , ' float32 ' , ' float64 ' ] , ' conv2d '
)
2020-08-23 18:41:20 +08:00
helper = LayerHelper ( l_type , * * locals ( ) )
2020-11-20 14:07:08 +08:00
dtype = helper . input_dtype ( input_param_name = ' x ' )
2020-08-23 18:41:20 +08:00
out = helper . create_variable_for_type_inference ( dtype )
outputs = { " Output " : [ out ] }
2022-10-23 20:01:27 +08:00
helper . append_op (
type = l_type , inputs = inputs , outputs = outputs , attrs = attrs
)
2020-08-23 18:41:20 +08:00
if bias is not None :
2022-12-12 13:09:14 +08:00
out = _add_with_axis ( out , bias , axis = channel_dim )
2024-01-18 11:28:47 +08:00
out = squeeze ( out , axis = [ squeeze_axis ] )
2025-06-24 14:16:09 +08:00
if float16_convert :
# out is float16
out = out . astype ( paddle . float16 )
2020-08-23 18:41:20 +08:00
return out
2026-01-30 20:20:08 +08:00
@param_one_alias ( [ " x " , " input " ] )
2022-10-23 20:01:27 +08:00
def conv2d (
2024-06-19 22:00:57 +08:00
x : Tensor ,
weight : Tensor ,
bias : Tensor | None = None ,
stride : Size2 = 1 ,
2024-06-27 20:19:16 +08:00
padding : _PaddingSizeMode | Size2 | Size4 | Sequence [ Size2 ] = 0 ,
2024-06-19 22:00:57 +08:00
dilation : Size2 = 1 ,
groups : int = 1 ,
data_format : DataLayout2D = " NCHW " ,
name : str | None = None ,
) - > Tensor :
2020-11-24 14:53:51 +08:00
r """
2020-05-14 11:27:10 +08:00
2020-04-04 03:04:07 +08:00
The convolution2D layer calculates the output based on the input, filter
and strides, paddings, dilations, groups parameters. Input and
Output are in NCHW or NHWC format, where N is batch size, C is the number of
channels, H is the height of the feature, and W is the width of the feature.
Filter is in MCHW format, where M is the number of output image channels,
C is the number of input image channels, H is the height of the filter,
and W is the width of the filter. If the groups is greater than 1,
C will equal the number of input image channels divided by the groups.
Please refer to UFLDL ' s `convolution
<http://ufldl.stanford.edu/tutorial/supervised/FeatureExtractionUsingConvolution/>`_
for more details.
If bias attribution and activation type are provided, bias is added to the
output of the convolution, and the corresponding activation function is
applied to the final result.
For each input :math:`X`, the equation is:
2020-08-29 16:04:19 +08:00
.. math::
2020-04-04 03:04:07 +08:00
2021-04-22 13:57:47 +08:00
Out = \ sigma (W \ ast X + b)
2020-04-04 03:04:07 +08:00
Where:
* :math:`X`: Input value, a tensor with NCHW or NHWC format.
* :math:`W`: Filter value, a tensor with MCHW format.
* :math:` \ \ ast`: Convolution operation.
* :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
* :math:` \ \ sigma`: Activation function.
* :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
Example:
- Input:
Input shape: :math:`(N, C_ {in} , H_ {in} , W_ {in} )`
Filter shape: :math:`(C_ {out} , C_ {in} , H_f, W_f)`
- Output:
Output shape: :math:`(N, C_ {out} , H_ {out} , W_ {out} )`
Where
2020-08-29 16:04:19 +08:00
.. math::
2020-04-04 03:04:07 +08:00
2021-04-22 13:57:47 +08:00
H_ {out} &= \ frac { (H_ {in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))} {strides[0]} + 1 \ \ \ \
W_ {out} &= \ frac { (W_ {in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))} {strides[1]} + 1
2020-04-04 03:04:07 +08:00
2025-09-18 17:06:08 +08:00
.. note::
Alias Support: The parameter name ``input`` can be used as an alias for ``x``.
2020-04-04 03:04:07 +08:00
Args:
2022-09-14 21:56:19 +08:00
x (Tensor): The input is 4-D Tensor with shape [N, C, H, W], the data type
2020-04-04 03:04:07 +08:00
of input is float16 or float32 or float64.
2025-09-18 17:06:08 +08:00
Alias: ``input``.
2020-08-23 14:21:05 +08:00
weight (Tensor): The convolution kernel with shape [M, C/g, kH, kW], where M is
2020-04-04 03:04:07 +08:00
the number of output channels, g is the number of groups, kH is the filter ' s
2022-09-14 21:56:19 +08:00
height, kW is the filter ' s width.
2020-08-23 14:21:05 +08:00
bias (Tensor, optional): The bias with shape [M,].
2022-12-07 20:02:53 +08:00
stride (int|list|tuple, optional): The stride size. It means the stride in convolution.
2022-09-14 21:56:19 +08:00
If stride is a list/tuple, it must contain two integers, (stride_height, stride_width).
2020-08-23 14:21:05 +08:00
Otherwise, stride_height = stride_width = stride. Default: stride = 1.
2022-12-07 20:02:53 +08:00
padding (string|int|list|tuple, optional): The padding size. It means the number of zero-paddings
2020-04-04 03:04:07 +08:00
on both sides for each dimension.If `padding` is a string, either ' VALID ' or
' SAME ' which is the padding algorithm. If padding size is a tuple or list,
it could be in three forms: `[pad_height, pad_width]` or
2022-09-14 21:56:19 +08:00
`[pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`, and when
`data_format` is ` " NCHW " `, `padding` can be in the form `[[0,0], [0,0],
2020-04-04 03:04:07 +08:00
[pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
2020-09-03 13:27:41 +08:00
when `data_format` is ` " NHWC " `, `padding` can be in the form
2020-04-04 03:04:07 +08:00
`[[0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
Default: padding = 0.
2022-12-07 20:02:53 +08:00
dilation (int|list|tuple, optional): The dilation size. It means the spacing between the kernel
2022-09-14 21:56:19 +08:00
points. If dilation is a list/tuple, it must contain two integers, (dilation_height,
dilation_width). Otherwise, dilation_height = dilation_width = dilation.
2020-04-04 03:04:07 +08:00
Default: dilation = 1.
2022-12-07 20:02:53 +08:00
groups (int, optional): The groups number of the Conv2D Layer. According to grouped
2020-04-04 03:04:07 +08:00
convolution in Alex Krizhevsky ' s Deep CNN paper: when group=2,
the first half of the filters is only connected to the first half
of the input channels, while the second half of the filters is only
connected to the second half of the input channels. Default: groups=1.
2022-09-14 21:56:19 +08:00
data_format (str, optional): Specify the data format of the input, and the data format of the output
2020-04-04 03:04:07 +08:00
will be consistent with that of the input. An optional string from: ` " NCHW " `, ` " NHWC " `.
The default is ` " NCHW " `. When it is ` " NCHW " `, the data is stored in the order of:
`[batch_size, input_channels, input_height, input_width]`.
2022-09-14 21:56:19 +08:00
name(str, optional): For detailed information, please refer
to :ref:`api_guide_Name`. Usually name is no need to set and
2020-04-04 03:04:07 +08:00
None by default.
Returns:
2022-09-14 21:56:19 +08:00
A Tensor representing the conv2d result, whose data type is the same with input.
2020-04-04 03:04:07 +08:00
Examples:
[CodeStyle][Xdoctest][8,12,15,16,18-20,22-26,29-32,34-36,38-40,42,43,45-48,50,51,60-65,75,80,82,83,85-87,89-94,99-141,143,145,147-167,169-187,207-220,257,258,260-275,277-313,315-325][API Compatibility] Update shape output format in documentation examples (#76574)
---------
Co-authored-by: SigureMo <sigure.qaq@gmail.com>
2025-11-26 10:31:46 +08:00
.. code-block:: pycon
2020-04-04 03:04:07 +08:00
2023-08-31 11:09:42 +08:00
>>> import paddle
>>> import paddle.nn.functional as F
2020-04-04 03:04:07 +08:00
2023-08-31 11:09:42 +08:00
>>> x_var = paddle.randn((2, 3, 8, 8), dtype= ' float32 ' )
>>> w_var = paddle.randn((6, 3, 3, 3), dtype= ' float32 ' )
2020-08-23 14:21:05 +08:00
2023-08-31 11:09:42 +08:00
>>> y_var = F.conv2d(x_var, w_var)
2020-08-23 14:21:05 +08:00
2023-08-31 11:09:42 +08:00
>>> print(y_var.shape)
[CodeStyle][Xdoctest][8,12,15,16,18-20,22-26,29-32,34-36,38-40,42,43,45-48,50,51,60-65,75,80,82,83,85-87,89-94,99-141,143,145,147-167,169-187,207-220,257,258,260-275,277-313,315-325][API Compatibility] Update shape output format in documentation examples (#76574)
---------
Co-authored-by: SigureMo <sigure.qaq@gmail.com>
2025-11-26 10:31:46 +08:00
paddle.Size([2, 6, 6, 6])
2020-04-04 03:04:07 +08:00
"""
# entry checks
if data_format not in [ " NCHW " , " NHWC " ] :
2022-10-23 20:01:27 +08:00
raise ValueError (
" Attr(data_format) should be ' NCHW ' or ' NHWC ' . "
2023-09-22 10:14:38 +08:00
f " Received Attr(data_format): { data_format } . "
2022-10-23 20:01:27 +08:00
)
2020-04-04 03:04:07 +08:00
2022-10-23 20:01:27 +08:00
channel_last = data_format == " NHWC "
2020-04-04 03:04:07 +08:00
channel_dim = - 1 if channel_last else 1
2021-09-07 19:28:09 +08:00
if len ( x . shape ) != 4 :
raise ValueError (
2023-09-22 10:14:38 +08:00
f " Input x should be 4D tensor, but received x with the shape of { x . shape } "
2022-10-23 20:01:27 +08:00
)
2020-08-23 14:21:05 +08:00
num_channels = x . shape [ channel_dim ]
2020-04-04 03:04:07 +08:00
num_filters = weight . shape [ 0 ]
2024-08-10 12:07:14 +08:00
if groups == 0 :
raise ValueError ( " The groups of conv2d should be not be zero. " )
if num_channels % groups != 0 and (
in_dynamic_mode ( ) or ( num_channels != - 1 and groups != - 1 )
) :
2020-04-04 03:04:07 +08:00
raise ValueError (
2024-08-10 12:07:14 +08:00
" The channel of input must be divisible by groups, "
2024-04-01 10:20:33 +08:00
f " received: the channel of input is { num_channels } , the shape of input is { x . shape } "
f " , the groups is { groups } "
2022-10-23 20:01:27 +08:00
)
2024-08-10 12:07:14 +08:00
if num_filters % groups != 0 and (
in_dynamic_mode ( ) or ( num_filters != - 1 and groups != - 1 )
) :
2020-04-04 03:04:07 +08:00
raise ValueError (
2024-08-10 12:07:14 +08:00
" The number of filters must be divisible by groups, "
2024-04-01 10:20:33 +08:00
f " received: the number of filters is { num_filters } , the shape of weight is { weight . shape } "
f " , the groups is { groups } "
2022-10-23 20:01:27 +08:00
)
2020-04-04 03:04:07 +08:00
2020-08-23 14:21:05 +08:00
cudnn_version = get_cudnn_version ( )
2022-10-23 20:01:27 +08:00
use_cudnn = (
2026-03-03 14:52:13 +08:00
is_compiled_with_cuda ( )
and cudnn_version is not None
and not get_flags ( " FLAGS_conv2d_disable_cudnn " ) [
" FLAGS_conv2d_disable_cudnn "
]
2022-10-23 20:01:27 +08:00
)
2026-03-03 14:52:13 +08:00
use_accuracy_compatible = paddle . get_flags (
[ " FLAGS_use_accuracy_compatible_kernel " ]
) . get ( " FLAGS_use_accuracy_compatible_kernel " , False )
2020-08-23 14:21:05 +08:00
2020-04-04 03:04:07 +08:00
# update attrs
padding , padding_algorithm = _update_padding_nd ( padding , channel_last , 2 )
2021-07-26 15:10:15 +08:00
stride = convert_to_list ( stride , 2 , ' stride ' )
dilation = convert_to_list ( dilation , 2 , ' dilation ' )
2020-04-04 03:04:07 +08:00
l_type = " conv2d "
2022-10-23 20:01:27 +08:00
if (
num_channels == groups
and num_channels != 1
and num_filters % num_channels == 0
) :
2020-04-04 03:04:07 +08:00
l_type = ' depthwise_conv2d '
2022-02-22 15:07:11 +08:00
if is_compiled_with_rocm ( ) :
2021-04-13 12:09:41 +08:00
use_cudnn = True
else :
use_cudnn = False
2022-04-06 08:58:59 +08:00
else :
2026-03-03 14:52:13 +08:00
if in_dynamic_mode ( ) and not use_accuracy_compatible :
2022-10-23 20:01:27 +08:00
pre_bias = _C_ops . conv2d (
x ,
weight ,
stride ,
padding ,
padding_algorithm ,
dilation ,
2022-10-31 21:47:59 -05:00
groups ,
2022-10-23 20:01:27 +08:00
data_format ,
)
2022-04-06 08:58:59 +08:00
if bias is not None :
2022-11-28 13:26:54 +08:00
channel_dim = (
channel_dim + len ( x . shape )
if channel_dim < 0
else channel_dim
)
if len ( bias . shape ) < len ( x . shape ) :
bias = _C_ops . reshape (
bias ,
[ 1 for i in range ( channel_dim ) ]
+ bias . shape
+ [ 1 for i in range ( len ( x . shape ) - channel_dim - 1 ) ] ,
)
2024-01-18 11:28:47 +08:00
# TODO(qili93): temporary for ascend npu performance to be removed along with npu_identity op
2023-04-21 11:56:51 +08:00
if (
_global_flags ( ) [ ' FLAGS_npu_storage_format ' ]
and ' npu ' in get_all_custom_device_type ( )
) :
with no_grad ( ) :
bias_storage = _C_ops . npu_identity (
bias , 3
) # ACL_FORMAT_NC1HWC0 = 3
bias_storage . _share_underline_tensor_to ( bias )
2022-11-28 13:26:54 +08:00
return _C_ops . add ( pre_bias , bias )
2022-04-06 08:58:59 +08:00
else :
return pre_bias
2022-10-23 20:01:27 +08:00
return _conv_nd (
x ,
weight ,
bias ,
stride ,
padding ,
padding_algorithm ,
dilation ,
groups ,
data_format ,
channel_dim ,
l_type ,
use_cudnn ,
name ,
)
2026-03-31 05:10:34 -07:00
@param_one_alias ( [ " x " , " input " ] )
2022-10-23 20:01:27 +08:00
def conv1d_transpose (
2024-06-19 22:00:57 +08:00
x : Tensor ,
weight : Tensor ,
bias : Tensor | None = None ,
stride : Size1 = 1 ,
2024-06-29 03:38:56 +08:00
padding : _PaddingSizeMode | Size1 | Size2 | Sequence [ Size2 ] = 0 ,
2024-06-19 22:00:57 +08:00
output_padding : Size1 = 0 ,
groups : int = 1 ,
dilation : Size1 = 1 ,
output_size : Size1 | None = None ,
data_format : DataLayout1D = " NCL " ,
name : str | None = None ,
) - > Tensor :
2020-11-24 14:53:51 +08:00
r """
2020-08-23 22:10:12 +08:00
The 1-D convolution transpose layer calculates the output based on the input,
filter, and dilation, stride, padding. Input(Input) and output(Output)
are in ' NCL ' format or ' NLC ' where N is batch size, C is the number of channels,
L is the length of the feature. The details of convolution transpose
layer, please refer to the following explanation and references
`therein <https://arxiv.org/pdf/1603.07285.pdf>`_.
If bias attribution and activation type are provided, bias is added to
the output of the convolution, and the corresponding activation function
is applied to the final result.
For each input :math:`X`, the equation is:
.. math::
2020-11-27 10:27:17 +08:00
Out = \ sigma (W \ ast X + b)
2020-08-23 22:10:12 +08:00
Where:
* :math:`X`: Input value, a 3-D Tensor with ' NCL ' format or ' NLC ' format.
* :math:`W`: Filter value, a 3-D Tensor with ' MCK ' format.
* :math:` \ \ ast`: Convolution operation.
* :math:`b`: Bias value, a 2-D Tensor with shape [M, 1].
* :math:` \ \ sigma`: Activation function.
* :math:`Out`: Output value, a 3-D Tensor with data format ' NCL ' or ' NLC ' , the shape of :math:`Out` and :math:`X` may be different.
Example:
- Input:
Input shape: :math:`(N, C_ {in} , L_ {in} )`
Filter shape: :math:`(C_ {in} , C_ {out} , L_f)`
- Output:
Output shape: :math:`(N, C_ {out} , L_ {out} )`
Where
.. math::
2022-12-23 19:01:32 +08:00
L^ \ prime_ {out} &= (L_ {in} - 1) * stride - 2 * padding + dilation * (L_f - 1) + 1 \ \
2020-08-23 22:10:12 +08:00
L_ {out} & \ in [ L^ \ prime_ {out} , L^ \ prime_ {out} + stride ]
Note:
The conv1d_transpose can be seen as the backward of the conv1d. For conv1d,
when stride > 1, conv1d maps multiple input shape to the same output shape,
so for conv1d_transpose, when stride > 1, input shape maps multiple output shape.
If output_size is None, :math:`L_ {out} = L^ \ prime_ {out} `;
else, the :math:`L_ {out} ` of the output size must between :math:`L^ \ prime_ {out} `
2021-04-22 13:57:47 +08:00
and :math:`L^ \ prime_ {out} + stride`.
2020-08-23 22:10:12 +08:00
Args:
x(Tensor): 3-D tensor with [N, C, L] or [N, L, C] format,
its data type is float32 or float64.
weight(Tensor): The convolution kernel, a Tensor with shape [C, M/g, K],
where M is the number of output channels(filters), g is the number of groups,
K is the size of the kernel.
bias(Tensor, optional): The bias, a Tensor with shape [M, ].
stride(int|tuple|list, optional): The stride size. It means the stride in transposed convolution.
2021-04-27 13:41:05 +08:00
If stride is a list/tuple, it must contain one integer, `(stride_size)`.
2020-08-23 22:10:12 +08:00
Default: stride = 1.
padding(int|list|str|tuple, optional): The padding size. The padding argument effectively adds
`dilation * (kernel - 1)` amount of zero-padding on both sides of input. If `padding` is a
string, either ' VALID ' or ' SAME ' supported, which is the padding algorithm.
If `padding` is a tuple or list, it could be in two forms:
`[pad]` or `[pad_left, pad_right]`. Default: padding = 0.
output_padding(int|list|tuple, optional): The count of zeros to be added to tail of each dimension.
2021-04-27 13:41:05 +08:00
If it is a list/tuple, it must contain one integer. Default: 0.
2020-08-23 22:10:12 +08:00
groups(int, optional): The groups number of the conv1d transpose function. Inspired by
grouped convolution in Alex Krizhevsky ' s Deep CNN paper, in which
when group=2, the first half of the filters is only connected to the
first half of the input channels, while the second half of the
filters is only connected to the second half of the input channels.
Default: groups = 1.
dilation(int|tuple|list, optional): The dilation size. It means the spacing between the kernel points.
2021-04-27 13:41:05 +08:00
If dilation is a list/tuple, it must contain one integer, `(dilation_size)`.
2020-08-23 22:10:12 +08:00
Default: dilation = 1.
output_size(int|tuple|list, optional): The output image size. If output size is a
2021-04-27 13:41:05 +08:00
tuple/list, it must contain one integer, `(feature_length)`. None if use
2021-04-22 13:57:47 +08:00
filter_size(shape of weight), padding, and stride to calculate output_size.
2022-09-14 21:56:19 +08:00
data_format (str, optional): Specify the data format of the input, and the data format of the output
2020-08-23 22:10:12 +08:00
will be consistent with that of the input. An optional string from: ` " NCL " `, ` " NLC " `.
The default is ` " NCL " `. When it is ` " NCL " `, the data is stored in the order of:
`[batch_size, input_channels, input_length]`.
2022-09-14 21:56:19 +08:00
name(str, optional): For detailed information, please refer
to :ref:`api_guide_Name`. Usually name is no need to set and
2020-08-23 22:10:12 +08:00
None by default.
Returns:
2025-06-30 17:37:39 +08:00
A tensor representing the result of 1-D transpose convolution, whose
2020-08-23 22:10:12 +08:00
data type is the same with input. And its shape is (num_batches, channels, length)
when data_format is ` " NCL " ` and (num_batches, length, channels) when data_format is
` " NLC " `.
Examples:
2026-02-05 05:25:41 +08:00
.. code-block:: pycon
2020-08-23 22:10:12 +08:00
2023-08-31 11:09:42 +08:00
>>> import paddle
>>> import paddle.nn.functional as F
2022-09-14 21:56:19 +08:00
2023-08-31 11:09:42 +08:00
>>> # shape: (1, 2, 4)
>>> x = paddle.to_tensor([[[4, 0, 9, 7],
>>> [8, 0, 9, 2,]]], dtype= " float32 " )
>>> # shape: (2, 1, 2)
>>> w = paddle.to_tensor([[[7, 0]],
>>> [[4, 2]]], dtype= " float32 " )
2022-11-03 14:00:26 +08:00
2023-08-31 11:09:42 +08:00
>>> y = F.conv1d_transpose(x, w)
>>> print(y)
Tensor(shape=[1, 1, 5], dtype=float32, place=Place(cpu), stop_gradient=True,
[[[60., 16., 99., 75., 4. ]]])
2020-08-23 22:10:12 +08:00
"""
cudnn_version = get_cudnn_version ( )
if cudnn_version is not None :
use_cudnn = True
else :
use_cudnn = False
if data_format not in [ ' NCL ' , ' NLC ' ] :
raise ValueError (
" Attr(data_format) of conv2d_transpose got wrong value: "
2023-09-22 10:14:38 +08:00
f " received { data_format } , but only ' NCL ' or ' NLC ' are supported. "
2022-10-23 20:01:27 +08:00
)
channel_last = data_format == " NLC "
2020-08-23 22:10:12 +08:00
channel_dim = - 1 if channel_last else 1
2021-09-07 19:28:09 +08:00
if len ( x . shape ) != 3 :
raise ValueError (
2023-09-22 10:14:38 +08:00
f " Input x should be 3D tensor, but received x with the shape of { x . shape } "
2022-10-23 20:01:27 +08:00
)
2020-08-23 22:10:12 +08:00
num_channels = x . shape [ channel_dim ]
2024-08-10 12:07:14 +08:00
if groups == 0 :
raise ValueError ( " The groups of conv1d_transpose should not be zero. " )
if num_channels % groups != 0 and (
in_dynamic_mode ( ) or ( num_channels != - 1 and groups != - 1 )
) :
2020-08-23 22:10:12 +08:00
raise ValueError (
" the channel of input must be divisible by groups, "
2024-04-01 10:20:33 +08:00
f " received: the channel of input is { num_channels } , the shape of input is { x . shape } "
f " , the groups is { groups } "
2022-10-23 20:01:27 +08:00
)
2020-08-23 22:10:12 +08:00
# update attrs
padding , padding_algorithm = _update_padding_nd ( padding , channel_last , 1 )
if len ( padding ) == 2 :
2024-08-12 18:54:52 +08:00
padding = [ * padding , 0 , 0 ]
2020-08-23 22:10:12 +08:00
elif len ( padding ) == 1 :
2024-08-12 18:54:52 +08:00
padding = [ * padding , 0 ]
2020-08-23 22:10:12 +08:00
else :
raise ValueError (
2023-09-22 10:14:38 +08:00
f " The size of padding ' s dimension should 1 or 2. But got padding= { padding } "
2022-10-23 20:01:27 +08:00
)
2020-08-23 22:10:12 +08:00
2024-08-12 18:54:52 +08:00
stride = [ * convert_to_list ( stride , 1 , " stride " ) , 1 ]
dilation = [ * convert_to_list ( dilation , 1 , " dilation " ) , 1 ]
2020-08-23 22:10:12 +08:00
if output_size is None :
output_size = [ ]
else :
2020-09-03 13:27:41 +08:00
if output_padding != 0 :
2022-10-23 20:01:27 +08:00
raise ValueError (
2025-08-18 00:05:11 +08:00
' output_padding option is mutually exclusive with output_size '
2022-10-23 20:01:27 +08:00
)
2020-09-03 13:27:41 +08:00
if isinstance ( output_size , ( list , tuple , int ) ) :
2024-08-12 18:54:52 +08:00
output_size = [ * convert_to_list ( output_size , 1 , ' output_size ' ) , 1 ]
2020-09-03 13:27:41 +08:00
else :
raise ValueError (
2022-10-23 20:01:27 +08:00
" output_size should be int, or list, tuple of ints "
)
2020-09-03 13:27:41 +08:00
if output_padding == 0 :
output_padding = [ ]
else :
2024-08-12 18:54:52 +08:00
output_padding = [
* convert_to_list ( output_padding , 1 , ' output_padding ' ) ,
0 ,
]
2020-09-03 13:27:41 +08:00
if len ( output_padding ) > 0 and output_padding [ 0 ] > stride [ 0 ] :
raise ValueError (
" The size of output_padding should not be greater than stride. "
2023-09-22 10:14:38 +08:00
f " But got output_padding= { output_padding [ 0 ] } and stride= { stride [ 0 ] } "
2022-10-23 20:01:27 +08:00
)
2020-08-23 22:10:12 +08:00
2022-12-30 05:05:19 -06:00
if len ( weight . shape ) != 3 :
raise ValueError (
2024-04-01 10:20:33 +08:00
f ' Input weight should be 3D tensor, but received weight with the shape of { weight . shape } '
2022-12-30 05:05:19 -06:00
)
2020-08-23 22:10:12 +08:00
op_type = ' conv2d_transpose '
num_filters = weight . shape [ 1 ]
2022-10-23 20:01:27 +08:00
if (
num_channels == groups
and num_channels != 1
and num_filters == 1
and not use_cudnn
) :
2020-08-23 22:10:12 +08:00
op_type = ' depthwise_conv2d_transpose '
use_cudnn = False
squeeze_axis = - 2 if channel_last else - 1
conv2d_data_format = " NHWC " if channel_last else " NCHW "
2021-07-26 15:10:15 +08:00
x = unsqueeze ( x , axis = [ squeeze_axis ] )
weight = unsqueeze ( weight , axis = [ - 1 ] )
2020-08-23 22:10:12 +08:00
2024-07-12 10:26:25 +08:00
if in_dynamic_or_pir_mode ( ) :
2022-10-23 20:01:27 +08:00
out = getattr ( _C_ops , op_type ) (
x ,
weight ,
stride ,
padding ,
output_padding ,
output_size ,
padding_algorithm ,
groups ,
dilation ,
conv2d_data_format ,
)
2022-08-22 18:37:13 +08:00
if bias is not None :
2022-12-12 13:09:14 +08:00
out = _add_with_axis ( out , bias , axis = channel_dim )
2020-08-23 22:10:12 +08:00
else :
inputs = { ' Input ' : [ x ] , ' Filter ' : [ weight ] }
attrs = {
2020-09-03 13:27:41 +08:00
' output_padding ' : output_padding ,
2020-08-23 22:10:12 +08:00
' output_size ' : output_size ,
' strides ' : stride ,
' paddings ' : padding ,
' padding_algorithm ' : padding_algorithm ,
' dilations ' : dilation ,
' groups ' : groups ,
' use_cudnn ' : use_cudnn ,
2022-10-23 20:01:27 +08:00
' data_format ' : conv2d_data_format ,
2020-08-23 22:10:12 +08:00
}
2022-10-23 20:01:27 +08:00
check_variable_and_dtype (
x , ' input ' , [ ' float16 ' , ' float32 ' , ' float64 ' ] , ' conv2d_transpose '
)
2020-08-23 22:10:12 +08:00
helper = LayerHelper ( op_type , * * locals ( ) )
2020-11-20 14:07:08 +08:00
dtype = helper . input_dtype ( input_param_name = ' x ' )
2020-08-23 22:10:12 +08:00
out = helper . create_variable_for_type_inference ( dtype )
outputs = { " Output " : [ out ] }
2022-10-23 20:01:27 +08:00
helper . append_op (
type = op_type , inputs = inputs , outputs = outputs , attrs = attrs
)
2020-08-23 22:10:12 +08:00
if bias is not None :
2022-12-12 13:09:14 +08:00
out = _add_with_axis ( out , bias , axis = channel_dim )
2020-08-23 22:10:12 +08:00
2021-07-26 15:10:15 +08:00
out = squeeze ( out , axis = [ squeeze_axis ] )
2020-08-23 22:10:12 +08:00
return out
2026-03-31 05:10:34 -07:00
@param_one_alias ( [ " x " , " input " ] )
2022-10-23 20:01:27 +08:00
def conv2d_transpose (
2024-06-19 22:00:57 +08:00
x : Tensor ,
weight : Tensor ,
bias : Tensor | None = None ,
stride : Size2 = 1 ,
2024-06-27 20:19:16 +08:00
padding : _PaddingSizeMode | Size2 | Size4 | Sequence [ Size2 ] = 0 ,
2024-06-19 22:00:57 +08:00
output_padding : Size2 = 0 ,
dilation : Size2 = 1 ,
groups : int = 1 ,
output_size : Size2 | None = None ,
data_format : DataLayout2D = ' NCHW ' ,
name : str | None = None ,
) - > Tensor :
2020-11-24 14:53:51 +08:00
r """
2020-05-14 11:27:10 +08:00
2020-04-04 03:04:07 +08:00
The convolution2D transpose layer calculates the output based on the input,
filter, and dilations, strides, paddings. Input(Input) and output(Output)
are in NCHW or NHWC format. Where N is batch size, C is the number of channels,
H is the height of the feature, and W is the width of the feature.
Parameters(dilations, strides, paddings) are two elements. These two elements
represent height and width, respectively. The details of convolution transpose
layer, please refer to the following explanation and references
`therein <https://arxiv.org/pdf/1603.07285.pdf>`_.
If bias attribution and activation type are provided, bias is added to
the output of the convolution, and the corresponding activation function
is applied to the final result.
2023-09-28 11:30:42 +08:00
See more detail in :ref:`api_paddle_nn_Conv2DTranspose` .
2020-04-04 03:04:07 +08:00
For each input :math:`X`, the equation is:
2020-08-29 16:04:19 +08:00
.. math::
2020-04-04 03:04:07 +08:00
2021-04-22 13:57:47 +08:00
Out = \ sigma (W \ ast X + b)
2020-04-04 03:04:07 +08:00
Where:
* :math:`X`: Input value, a 4-D Tensor with NCHW or NHWC format.
* :math:`W`: Filter value, a 4-D Tensor with MCHW format.
* :math:` \ \ ast`: Convolution operation.
* :math:`b`: Bias value, a 2-D Tensor with shape [M, 1].
* :math:` \ \ sigma`: Activation function.
* :math:`Out`: Output value, a 4-D Tensor with data format ' NCHW ' or ' NHWC ' , the shape of :math:`Out` and :math:`X` may be different.
Example:
- Input:
Input shape: :math:`(N, C_ {in} , H_ {in} , W_ {in} )`
Filter shape: :math:`(C_ {in} , C_ {out} , H_f, W_f)`
- Output:
Output shape: :math:`(N, C_ {out} , H_ {out} , W_ {out} )`
Where
2020-08-29 16:04:19 +08:00
.. math::
2020-04-04 03:04:07 +08:00
2022-12-23 19:01:32 +08:00
H^ \ prime_ {out} &= (H_ {in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (H_f - 1) + 1 \ \
W^ \ prime_ {out} &= (W_ {in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (W_f - 1) + 1 \ \
H_ {out} & \ in [ H^ \ prime_ {out} , H^ \ prime_ {out} + strides[0] ] \ \
2020-04-04 03:04:07 +08:00
W_ {out} & \ in [ W^ \ prime_ {out} , W^ \ prime_ {out} + strides[1] ]
Note:
2022-09-14 21:56:19 +08:00
The conv2d_transpose can be seen as the backward of the conv2d. For conv2d,
when stride > 1, conv2d maps multiple input shape to the same output shape,
2020-04-04 03:04:07 +08:00
so for conv2d_transpose, when stride > 1, input shape maps multiple output shape.
2022-09-14 21:56:19 +08:00
If output_size is None, :math:`H_ {out} = H^ \ prime_ {out} , W_ {out} = W^ \ prime_ {out} `;
else, the :math:`H_ {out} ` of the output size must between :math:`H^ \ prime_ {out} `
and :math:`H^ \ prime_ {out} + strides[0]`, and the :math:`W_ {out} ` of the output size must
2021-04-22 13:57:47 +08:00
between :math:`W^ \ prime_ {out} ` and :math:`W^ \ prime_ {out} + strides[1]`.
2020-04-04 03:04:07 +08:00
Args:
2020-08-22 19:57:15 +08:00
x(Tensor): 4-D Tensor with [N, C, H, W] or [N, H, W, C] format,
2020-04-04 03:04:07 +08:00
whose data type is float32 or float64.
2020-08-22 19:57:15 +08:00
weight(Tensor): The convolution kernel, a Tensor with shape [C, M/g, kH, kW],
2020-04-04 03:04:07 +08:00
where M is the number of output channels(filters), g is the number of groups,
kH is the height of the kernel, and kW is the width of the kernel.
2020-08-22 19:57:15 +08:00
bias(Tensor, optional): The bias, a Tensor with shape [M, ].
2022-09-14 21:56:19 +08:00
stride(int|list|tuple, optional): The stride size. It means the stride in transposed convolution.
If stride is a list/tuple, it must contain two integers, (stride_height, stride_width).
2020-08-22 19:57:15 +08:00
Otherwise, stride_height = stride_width = stride. Default: stride = 1.
2022-09-14 21:56:19 +08:00
padding(str|int|list|tuple, optional): The padding size. It means the number of zero-paddings
on both sides for each dimension. If `padding` is a string, either ' VALID ' or
2020-08-29 16:04:19 +08:00
' SAME ' which is the padding algorithm. If padding size is a tuple or list,
2022-09-14 21:56:19 +08:00
it could be in three forms: `[pad_height, pad_width]` or
2020-08-29 16:04:19 +08:00
`[pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`,
2022-09-14 21:56:19 +08:00
and when `data_format` is ` " NCHW " `, `padding` can be in the form
2020-08-29 16:04:19 +08:00
`[[0,0], [0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
2022-09-14 21:56:19 +08:00
when `data_format` is ` " NHWC " `, `padding` can be in the form
2020-04-04 03:04:07 +08:00
`[[0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
Default: padding = 0.
2020-08-22 19:57:15 +08:00
output_padding(int|list|tuple, optional): Additional size added to one side
of each dimension in the output shape. Default: 0.
2020-10-21 17:32:38 +08:00
groups(int, optional): The groups number of the Conv2D transpose layer. Inspired by
2020-04-04 03:04:07 +08:00
grouped convolution in Alex Krizhevsky ' s Deep CNN paper, in which
when group=2, the first half of the filters is only connected to the
first half of the input channels, while the second half of the
filters is only connected to the second half of the input channels.
Default: groups = 1.
2022-09-14 21:56:19 +08:00
dilation(int|list|tuple, optional): The dilation size. It means the spacing between the kernel points.
If dilation is a list/tuple, it must contain two integers, (dilation_height, dilation_width).
2020-08-29 16:04:19 +08:00
Otherwise, dilation_height = dilation_width = dilation. Default: dilation = 1.
2020-08-22 19:57:15 +08:00
output_size(int|tuple|list, optional): The output image size. If output size is a
2021-04-27 13:41:05 +08:00
tuple/list, it must contain two integers, (image_height, image_width). None if use
2021-04-22 13:57:47 +08:00
filter_size(shape of weight), padding, and stride to calculate output_size.
2022-09-14 21:56:19 +08:00
data_format (str, optional): Specify the data format of the input, and the data format of the output
2020-04-04 03:04:07 +08:00
will be consistent with that of the input. An optional string from: ` " NCHW " `, ` " NHWC " `.
The default is ` " NCHW " `. When it is ` " NCHW " `, the data is stored in the order of:
`[batch_size, input_channels, input_height, input_width]`.
2022-09-14 21:56:19 +08:00
name(str, optional): For detailed information, please refer
to :ref:`api_guide_Name`. Usually name is no need to set and
2020-04-04 03:04:07 +08:00
None by default.
Returns:
2020-10-26 11:53:18 +08:00
A Tensor representing the conv2d_transpose, whose
2022-09-14 21:56:19 +08:00
data type is the same with input and shape is (num_batches, channels, out_h,
out_w) or (num_batches, out_h, out_w, channels). The tensor variable storing
2020-08-22 19:57:15 +08:00
transposed convolution result.
2020-04-04 03:04:07 +08:00
Examples:
[CodeStyle][Xdoctest][8,12,15,16,18-20,22-26,29-32,34-36,38-40,42,43,45-48,50,51,60-65,75,80,82,83,85-87,89-94,99-141,143,145,147-167,169-187,207-220,257,258,260-275,277-313,315-325][API Compatibility] Update shape output format in documentation examples (#76574)
---------
Co-authored-by: SigureMo <sigure.qaq@gmail.com>
2025-11-26 10:31:46 +08:00
.. code-block:: pycon
2020-04-04 03:04:07 +08:00
2023-08-31 11:09:42 +08:00
>>> import paddle
>>> import paddle.nn.functional as F
2020-04-04 03:04:07 +08:00
2023-08-31 11:09:42 +08:00
>>> x_var = paddle.randn((2, 3, 8, 8), dtype= ' float32 ' )
>>> w_var = paddle.randn((3, 6, 3, 3), dtype= ' float32 ' )
2020-04-04 03:04:07 +08:00
2023-08-31 11:09:42 +08:00
>>> y_var = F.conv2d_transpose(x_var, w_var)
2020-04-04 03:04:07 +08:00
2023-08-31 11:09:42 +08:00
>>> print(y_var.shape)
[CodeStyle][Xdoctest][8,12,15,16,18-20,22-26,29-32,34-36,38-40,42,43,45-48,50,51,60-65,75,80,82,83,85-87,89-94,99-141,143,145,147-167,169-187,207-220,257,258,260-275,277-313,315-325][API Compatibility] Update shape output format in documentation examples (#76574)
---------
Co-authored-by: SigureMo <sigure.qaq@gmail.com>
2025-11-26 10:31:46 +08:00
paddle.Size([2, 6, 10, 10])
2020-04-04 03:04:07 +08:00
"""
if data_format not in [ ' NCHW ' , ' NHWC ' ] :
raise ValueError (
" Attr(data_format) of conv2d_transpose got wrong value: "
2023-09-22 10:14:38 +08:00
f " received { data_format } , but only ' NCHW ' or ' NHWC ' are supported. "
2022-10-23 20:01:27 +08:00
)
channel_last = data_format == " NHWC "
2020-04-04 03:04:07 +08:00
channel_dim = - 1 if channel_last else 1
2021-09-07 19:28:09 +08:00
if len ( x . shape ) != 4 :
raise ValueError (
2023-09-22 10:14:38 +08:00
f " Input x should be 4D tensor, but received x with the shape of { x . shape } "
2022-10-23 20:01:27 +08:00
)
2023-02-02 16:52:00 +08:00
if len ( weight . shape ) != 4 :
raise ValueError (
2024-04-01 10:20:33 +08:00
f " Input weight should be 4D tensor, but received weight with the shape of { weight . shape } "
2023-02-02 16:52:00 +08:00
)
2020-08-22 19:57:15 +08:00
num_channels = x . shape [ channel_dim ]
2024-08-10 12:07:14 +08:00
if groups == 0 :
raise ValueError ( " The groups of conv2d_transpose should not be zero. " )
if num_channels % groups != 0 and (
in_dynamic_mode ( ) or ( num_channels != - 1 and groups != - 1 )
) :
2020-04-04 03:04:07 +08:00
raise ValueError (
" the channel of input must be divisible by groups, "
2024-04-01 10:20:33 +08:00
f " received: the channel of input is { num_channels } , the shape of input is { x . shape } "
f " , the groups is { groups } "
2022-10-23 20:01:27 +08:00
)
2020-08-22 19:57:15 +08:00
cudnn_version = get_cudnn_version ( )
2022-10-23 20:01:27 +08:00
use_cudnn = (
True
if ( is_compiled_with_cuda ( ) and cudnn_version is not None )
else False
)
2020-04-04 03:04:07 +08:00
# update attrs
padding , padding_algorithm = _update_padding_nd ( padding , channel_last , 2 )
2021-07-26 15:10:15 +08:00
stride = convert_to_list ( stride , 2 , ' stride ' )
dilation = convert_to_list ( dilation , 2 , ' dilation ' )
2020-08-22 19:57:15 +08:00
2020-04-04 03:04:07 +08:00
if output_size is None :
output_size = [ ]
else :
2020-08-22 19:57:15 +08:00
if output_padding != 0 :
2022-10-23 20:01:27 +08:00
raise ValueError (
2025-08-18 00:05:11 +08:00
' output_padding option is mutually exclusive with output_size '
2022-10-23 20:01:27 +08:00
)
2022-09-07 15:48:37 +08:00
if isinstance ( output_size , ( list , tuple ) ) :
if _contain_var ( output_size ) :
output_size = _convert_to_tensor_list ( output_size )
else :
output_size = convert_to_list ( output_size , 2 , ' output_size ' )
elif isinstance ( output_size , int ) :
2021-07-26 15:10:15 +08:00
output_size = convert_to_list ( output_size , 2 , ' output_size ' )
2023-12-11 14:10:01 +08:00
elif isinstance ( output_size , ( Variable , pir . Value ) ) :
2022-10-23 20:01:27 +08:00
check_dtype (
output_size . dtype ,
' output_size ' ,
[ ' int32 ' , ' int64 ' ] ,
' conv2d_transpose ' ,
)
if len ( output_size . shape ) == 1 and (
output_size . shape [ 0 ] == 1 or output_size . shape [ 0 ] == 2
) :
2022-09-07 15:48:37 +08:00
if output_size . shape [ 0 ] == 1 :
output_size = [ output_size , output_size ]
else :
raise ValueError (
2022-10-23 20:01:27 +08:00
" output_size must contain one or two integers. "
)
2020-08-22 19:57:15 +08:00
else :
raise ValueError (
2022-09-07 15:48:37 +08:00
" output_size should be int or Tensor or list, tuple of ints or Tensor "
)
2020-08-22 19:57:15 +08:00
if output_padding == 0 :
output_padding = [ ]
else :
2021-07-26 15:10:15 +08:00
output_padding = convert_to_list ( output_padding , 2 , ' output_padding ' )
2020-04-04 03:04:07 +08:00
op_type = ' conv2d_transpose '
num_filters = weight . shape [ 1 ]
2025-12-25 17:17:52 +08:00
use_accuracy_compatible = paddle . get_flags (
[ " FLAGS_use_accuracy_compatible_kernel " ]
) . get ( " FLAGS_use_accuracy_compatible_kernel " , False )
if (
not use_accuracy_compatible
and num_channels == groups
and num_channels != 1
and num_filters == 1
) :
2020-04-04 03:04:07 +08:00
op_type = ' depthwise_conv2d_transpose '
2020-08-22 19:57:15 +08:00
use_cudnn = False
2020-04-04 03:04:07 +08:00
2023-11-02 14:22:52 +08:00
if in_dynamic_or_pir_mode ( ) :
2022-10-23 20:01:27 +08:00
op = (
_C_ops . conv2d_transpose
if op_type == ' conv2d_transpose '
else _C_ops . depthwise_conv2d_transpose
)
pre_bias = op (
x ,
weight ,
stride ,
padding ,
output_padding ,
output_size ,
padding_algorithm ,
groups ,
dilation ,
data_format ,
)
2022-04-03 11:31:39 +08:00
if bias is not None :
2022-12-12 13:09:14 +08:00
return _add_with_axis ( pre_bias , bias , axis = channel_dim )
2022-04-03 11:31:39 +08:00
else :
return pre_bias
2020-04-04 03:04:07 +08:00
else :
2020-08-22 19:57:15 +08:00
inputs = { ' Input ' : [ x ] , ' Filter ' : [ weight ] }
2020-04-04 03:04:07 +08:00
attrs = {
2020-08-22 19:57:15 +08:00
' output_padding ' : output_padding ,
2020-04-04 03:04:07 +08:00
' output_size ' : output_size ,
' strides ' : stride ,
' paddings ' : padding ,
' padding_algorithm ' : padding_algorithm ,
' dilations ' : dilation ,
' groups ' : groups ,
' use_cudnn ' : use_cudnn ,
2022-10-23 20:01:27 +08:00
' data_format ' : data_format ,
2020-04-04 03:04:07 +08:00
}
2022-10-23 20:01:27 +08:00
check_variable_and_dtype (
2023-04-21 15:16:59 +08:00
x ,
' x ' ,
[ ' float16 ' , ' uint16 ' , ' float32 ' , ' float64 ' ] ,
' conv2d_transpose ' ,
2022-10-23 20:01:27 +08:00
)
2020-04-04 03:04:07 +08:00
helper = LayerHelper ( op_type , * * locals ( ) )
2020-08-22 19:57:15 +08:00
pre_bias = helper . create_variable_for_type_inference ( x . dtype )
2020-04-04 03:04:07 +08:00
outputs = { " Output " : [ pre_bias ] }
2022-10-23 20:01:27 +08:00
helper . append_op (
type = op_type , inputs = inputs , outputs = outputs , attrs = attrs
)
2020-08-22 19:57:15 +08:00
2020-04-04 03:04:07 +08:00
if bias is not None :
2023-03-20 19:07:02 +08:00
out = helper . create_variable_for_type_inference ( x . dtype )
x_shape = list ( pre_bias . shape )
y_shape = list ( bias . shape )
if channel_dim == - 1 or len ( x_shape ) == len ( y_shape ) :
helper . append_op (
type = ' elementwise_add ' ,
inputs = { ' X ' : [ pre_bias ] , ' Y ' : [ bias ] } ,
outputs = { ' Out ' : [ out ] } ,
2023-12-11 19:38:07 +08:00
attrs = { ' axis ' : - 1 } ,
2023-03-20 19:07:02 +08:00
)
else :
2025-08-21 01:51:31 +08:00
assert len ( x_shape ) > len ( y_shape ) , (
' The length of pre_bias must greater than the length of bias '
)
2023-03-20 19:07:02 +08:00
padding = len ( x_shape ) - len ( y_shape ) - channel_dim
bias = reshape (
bias , [ 1 ] * channel_dim + y_shape + [ 1 ] * padding
)
helper . append_op (
type = ' elementwise_add ' ,
inputs = { ' X ' : [ pre_bias ] , ' Y ' : [ bias ] } ,
outputs = { ' Out ' : [ out ] } ,
2023-12-11 19:38:07 +08:00
attrs = { ' axis ' : - 1 } ,
2023-03-20 19:07:02 +08:00
)
2020-04-04 03:04:07 +08:00
else :
2020-08-22 19:57:15 +08:00
out = pre_bias
2020-04-04 03:04:07 +08:00
return out
2026-01-30 20:20:08 +08:00
@param_one_alias ( [ " x " , " input " ] )
2022-10-23 20:01:27 +08:00
def conv3d (
2024-06-19 22:00:57 +08:00
x : Tensor ,
weight : Tensor ,
bias : Tensor | None = None ,
stride : Size3 = 1 ,
2024-06-29 03:38:56 +08:00
padding : _PaddingSizeMode | Size3 | Size6 | Sequence [ Size2 ] = 0 ,
2024-06-19 22:00:57 +08:00
dilation : Size3 = 1 ,
groups : int = 1 ,
data_format : DataLayout3D = " NCDHW " ,
name : str | None = None ,
) - > Tensor :
2020-11-24 14:53:51 +08:00
r """
2020-05-14 11:27:10 +08:00
2020-04-04 03:04:07 +08:00
The convolution3D layer calculates the output based on the input, filter
and strides, paddings, dilations, groups parameters. Input(Input) and
Output(Output) are in NCDHW or NDHWC format. Where N is batch size C is the number of
channels, D is the depth of the feature, H is the height of the feature,
2024-01-18 11:28:47 +08:00
and W is the width of the feature. Convolution3D is similar with Convolution2D
2020-04-04 03:04:07 +08:00
but adds one dimension(depth). If bias attribution and activation type are
provided, bias is added to the output of the convolution, and the
corresponding activation function is applied to the final result.
For each input :math:`X`, the equation is:
2020-08-29 16:04:19 +08:00
.. math::
2020-04-04 03:04:07 +08:00
2021-04-22 13:57:47 +08:00
Out = \ sigma (W \ ast X + b)
2020-04-04 03:04:07 +08:00
In the above equation:
* :math:`X`: Input value, a tensor with NCDHW or NDHWC format.
* :math:`W`: Filter value, a tensor with MCDHW format.
* :math:` \ \ ast`: Convolution operation.
* :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
* :math:` \ \ sigma`: Activation function.
* :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
Example:
- Input:
Input shape: :math:`(N, C_ {in} , D_ {in} , H_ {in} , W_ {in} )`
Filter shape: :math:`(C_ {out} , C_ {in} , D_f, H_f, W_f)`
- Output:
Output shape: :math:`(N, C_ {out} , D_ {out} , H_ {out} , W_ {out} )`
Where
2020-08-29 16:04:19 +08:00
.. math::
2020-04-04 03:04:07 +08:00
2022-12-23 19:01:32 +08:00
D_ {out} &= \ frac { (D_ {in} + 2 * paddings[0] - (dilations[0] * (D_f - 1) + 1))} {strides[0]} + 1 \ \
H_ {out} &= \ frac { (H_ {in} + 2 * paddings[1] - (dilations[1] * (H_f - 1) + 1))} {strides[1]} + 1 \ \
W_ {out} &= \ frac { (W_ {in} + 2 * paddings[2] - (dilations[2] * (W_f - 1) + 1))} {strides[2]} + 1
2020-04-04 03:04:07 +08:00
2025-09-18 17:06:08 +08:00
.. note::
Alias Support: The parameter name ``input`` can be used as an alias for ``x``.
2020-04-04 03:04:07 +08:00
Args:
2022-09-14 21:56:19 +08:00
x (Tensor): The input is 5-D Tensor with shape [N, C, D, H, W], the data
2020-04-04 03:04:07 +08:00
type of input is float16 or float32 or float64.
2025-09-18 17:06:08 +08:00
Alias: ``input``.
2020-12-04 23:05:04 +08:00
weight (Tensor): The convolution kernel, a Tensor with shape [M, C/g, kD, kH, kW],
2020-04-04 03:04:07 +08:00
where M is the number of filters(output channels), g is the number of groups,
kD, kH, kW are the filter ' s depth, height and width respectively.
2020-08-23 14:21:05 +08:00
bias (Tensor, optional): The bias, a Tensor of shape [M, ].
2022-09-14 21:56:19 +08:00
stride (int|list|tuple, optional): The stride size. It means the stride in convolution. If stride is a
list/tuple, it must contain three integers, (stride_depth, stride_height, stride_width).
2020-08-23 14:21:05 +08:00
Otherwise, stride_depth = stride_height = stride_width = stride. Default: stride = 1.
2022-09-14 21:56:19 +08:00
padding (string|int|list|tuple, optional): The padding size. It means the number of zero-paddings
2020-04-04 03:04:07 +08:00
on both sides for each dimension. If `padding` is a string, either ' VALID ' or
' SAME ' which is the padding algorithm. If padding size is a tuple or list,
it could be in three forms: `[pad_depth, pad_height, pad_width]` or
`[pad_depth_front, pad_depth_back, pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`,
2020-09-03 13:27:41 +08:00
and when `data_format` is ` " NCDHW " `, `padding` can be in the form
2020-04-04 03:04:07 +08:00
`[[0,0], [0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
2020-09-03 13:27:41 +08:00
when `data_format` is ` " NDHWC " `, `padding` can be in the form
2020-04-04 03:04:07 +08:00
`[[0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
Default: padding = 0.
2022-09-14 21:56:19 +08:00
dilation (int|list|tuple, optional): The dilation size. It means the spacing between the kernel points.
2021-04-27 13:41:05 +08:00
If dilation is a list/tuple, it must contain three integers, (dilation_depth, dilation_height,
2022-09-14 21:56:19 +08:00
dilation_width). Otherwise, dilation_depth = dilation_height = dilation_width = dilation.
2020-04-04 03:04:07 +08:00
Default: dilation = 1.
2022-06-07 11:23:01 +08:00
groups (int, optional): The groups number of the Conv3D Layer. According to grouped
2020-04-04 03:04:07 +08:00
convolution in Alex Krizhevsky ' s Deep CNN paper: when group=2,
the first half of the filters is only connected to the first half
of the input channels, while the second half of the filters is only
connected to the second half of the input channels. Default: groups=1
2022-09-14 21:56:19 +08:00
data_format (str, optional): Specify the data format of the input, and the data format of the output
2022-06-07 11:23:01 +08:00
will be consistent with that of the input. An optional string from: ` " NCDHW " `, ` " NDHWC " `.
The default is ` " NCDHW " `. When it is ` " NCDHW " `, the data is stored in the order of:
`[batch_size, input_channels, input_depth, input_height, input_width]`.
2022-09-14 21:56:19 +08:00
name(str|None, optional): For detailed information, please refer
to :ref:`api_guide_Name`. Usually name is no need to set and
2020-04-04 03:04:07 +08:00
None by default.
Returns:
2022-09-14 21:56:19 +08:00
A Tensor representing the conv3d, whose data type is
the same with input. If act is None, the tensor storing the
convolution result, and if act is not None, the tensor storing
2020-04-04 03:04:07 +08:00
convolution and non-linearity activation result.
Examples:
[CodeStyle][Xdoctest][8,12,15,16,18-20,22-26,29-32,34-36,38-40,42,43,45-48,50,51,60-65,75,80,82,83,85-87,89-94,99-141,143,145,147-167,169-187,207-220,257,258,260-275,277-313,315-325][API Compatibility] Update shape output format in documentation examples (#76574)
---------
Co-authored-by: SigureMo <sigure.qaq@gmail.com>
2025-11-26 10:31:46 +08:00
.. code-block:: pycon
2020-04-04 03:04:07 +08:00
2023-08-31 11:09:42 +08:00
>>> import paddle
>>> import paddle.nn.functional as F
2020-04-04 03:04:07 +08:00
2023-08-31 11:09:42 +08:00
>>> x_var = paddle.randn((2, 3, 8, 8, 8), dtype= ' float32 ' )
>>> w_var = paddle.randn((6, 3, 3, 3, 3), dtype= ' float32 ' )
2020-04-04 03:04:07 +08:00
2023-08-31 11:09:42 +08:00
>>> y_var = F.conv3d(x_var, w_var)
2020-04-04 03:04:07 +08:00
2023-08-31 11:09:42 +08:00
>>> print(y_var.shape)
[CodeStyle][Xdoctest][8,12,15,16,18-20,22-26,29-32,34-36,38-40,42,43,45-48,50,51,60-65,75,80,82,83,85-87,89-94,99-141,143,145,147-167,169-187,207-220,257,258,260-275,277-313,315-325][API Compatibility] Update shape output format in documentation examples (#76574)
---------
Co-authored-by: SigureMo <sigure.qaq@gmail.com>
2025-11-26 10:31:46 +08:00
paddle.Size([2, 6, 6, 6, 6])
2020-04-04 03:04:07 +08:00
"""
# entry check
if data_format not in [ " NCDHW " , " NDHWC " ] :
raise ValueError (
" Attr(data_format) should be ' NCDHW ' or ' NDHWC ' . Received "
2023-09-22 10:14:38 +08:00
f " Attr(data_format): { data_format } . "
2022-10-23 20:01:27 +08:00
)
2020-04-04 03:04:07 +08:00
2022-10-23 20:01:27 +08:00
channel_last = data_format == " NDHWC "
2020-04-04 03:04:07 +08:00
channel_dim = - 1 if channel_last else 1
2021-09-07 19:28:09 +08:00
if len ( x . shape ) != 5 :
raise ValueError (
2023-09-22 10:14:38 +08:00
f " Input x should be 5D tensor, but received x with the shape of { x . shape } "
2022-10-23 20:01:27 +08:00
)
2020-08-23 14:21:05 +08:00
num_channels = x . shape [ channel_dim ]
2020-04-04 03:04:07 +08:00
num_filters = weight . shape [ 0 ]
2024-08-10 12:07:14 +08:00
if groups == 0 :
raise ValueError ( " The groups of conv3d should not be 0. " )
if num_channels % groups != 0 and (
in_dynamic_mode ( ) or ( num_channels != - 1 and groups != - 1 )
) :
2020-04-04 03:04:07 +08:00
raise ValueError (
" The number of input channels must be divisible by Attr(groups). "
2023-09-22 10:14:38 +08:00
f " Received: number of channels( { num_channels } ), groups( { groups } ). "
2022-10-23 20:01:27 +08:00
)
2024-08-10 12:07:14 +08:00
if num_filters % groups != 0 and (
in_dynamic_mode ( ) or ( num_filters != - 1 and groups != - 1 )
) :
2020-04-04 03:04:07 +08:00
raise ValueError (
" The number of filters must be divisible by Attr(groups). "
2023-09-22 10:14:38 +08:00
f " Received: number of filters( { num_filters } ), groups( { groups } ). "
2022-10-23 20:01:27 +08:00
)
2020-04-04 03:04:07 +08:00
2020-08-23 14:21:05 +08:00
cudnn_version = get_cudnn_version ( )
2022-10-23 20:01:27 +08:00
use_cudnn = (
2026-03-03 14:52:13 +08:00
is_compiled_with_cuda ( )
and cudnn_version is not None
and not get_flags ( " FLAGS_conv3d_disable_cudnn " ) [
" FLAGS_conv3d_disable_cudnn "
]
2022-10-23 20:01:27 +08:00
)
2020-08-23 14:21:05 +08:00
2020-04-04 03:04:07 +08:00
padding , padding_algorithm = _update_padding_nd ( padding , channel_last , 3 )
2021-07-26 15:10:15 +08:00
stride = convert_to_list ( stride , 3 , ' stride ' )
dilation = convert_to_list ( dilation , 3 , ' dilation ' )
2020-04-04 03:04:07 +08:00
op_type = " conv3d "
2025-12-25 17:17:52 +08:00
use_accuracy_compatible = paddle . get_flags (
[ " FLAGS_use_accuracy_compatible_kernel " ]
) . get ( " FLAGS_use_accuracy_compatible_kernel " , False )
if (
use_accuracy_compatible
and is_compiled_with_cuda ( )
and num_channels == groups
and num_channels != 1
and num_filters % num_channels == 0
) :
op_type = ' depthwise_conv3d '
2020-04-04 03:04:07 +08:00
2022-10-23 20:01:27 +08:00
return _conv_nd (
x ,
weight ,
bias ,
stride ,
padding ,
padding_algorithm ,
dilation ,
groups ,
data_format ,
channel_dim ,
op_type ,
use_cudnn ,
name ,
)
2026-03-31 05:10:34 -07:00
@param_one_alias ( [ " x " , " input " ] )
2022-10-23 20:01:27 +08:00
def conv3d_transpose (
2024-06-19 22:00:57 +08:00
x : Tensor ,
weight : Tensor ,
bias : Tensor | None = None ,
stride : Size3 = 1 ,
2024-06-27 20:19:16 +08:00
padding : _PaddingSizeMode | Size3 | Size6 | Sequence [ Size2 ] = 0 ,
2024-06-19 22:00:57 +08:00
output_padding : Size3 = 0 ,
groups : int = 1 ,
dilation : Size3 = 1 ,
output_size : Size3 | None = None ,
data_format : DataLayout3D = ' NCDHW ' ,
name : str | None = None ,
) - > Tensor :
2020-11-24 14:53:51 +08:00
r """
2020-08-22 19:57:15 +08:00
The convolution3d transpose layer calculates the output based on the input,
2020-04-04 03:04:07 +08:00
filter, and dilations, strides, paddings. Input(Input) and output(Output)
are in NCDHW or NDHWC format. Where N is batch size, C is the number of channels,
D is the depth of the feature, H is the height of the feature, and W
is the width of the feature. Parameters(dilations, strides, paddings) are
two elements. These two elements represent height and width, respectively.
The details of convolution transpose layer, please refer to the following
explanation and references `therein <https://arxiv.org/pdf/1603.07285.pdf>`_.
If bias attribution and activation type are provided, bias is added to
the output of the convolution, and the corresponding activation function
is applied to the final result.
2023-09-28 11:30:42 +08:00
See more detail in :ref:`api_paddle_nn_Conv3DTranspose` .
2020-04-04 03:04:07 +08:00
For each input :math:`X`, the equation is:
2020-08-29 16:04:19 +08:00
.. math::
2020-04-04 03:04:07 +08:00
2021-04-22 13:57:47 +08:00
Out = \ sigma (W \ ast X + b)
2020-04-04 03:04:07 +08:00
In the above equation:
* :math:`X`: Input value, a Tensor with NCDHW or NDHWC format.
2022-12-23 19:01:32 +08:00
* :math:`W`: Filter value, a Tensor with NCDHW format.
* :math:` \ ast`: Convolution operation.
2020-04-04 03:04:07 +08:00
* :math:`b`: Bias value, a 2-D Tensor with shape [M, 1].
2022-12-23 19:01:32 +08:00
* :math:` \ sigma`: Activation function.
2020-04-04 03:04:07 +08:00
* :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
Example:
- Input:
Input shape: :math:`(N, C_ {in} , D_ {in} , H_ {in} , W_ {in} )`
Filter shape: :math:`(C_ {in} , C_ {out} , D_f, H_f, W_f)`
- Output:
Output shape: :math:`(N, C_ {out} , D_ {out} , H_ {out} , W_ {out} )`
Where
2020-08-29 16:04:19 +08:00
.. math::
2020-04-04 03:04:07 +08:00
2022-12-23 19:01:32 +08:00
D^ \ prime_ {out} &= (D_ {in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (D_f - 1) + 1 \ \
H^ \ prime_ {out} &= (H_ {in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (H_f - 1) + 1 \ \
W^ \ prime_ {out} &= (W_ {in} - 1) * strides[2] - 2 * paddings[2] + dilations[2] * (W_f - 1) + 1 \ \
D_ {out} & \ in [ D^ \ prime_ {out} , D^ \ prime_ {out} + strides[0] ] \ \
H_ {out} & \ in [ H^ \ prime_ {out} , H^ \ prime_ {out} + strides[1] ] \ \
2020-04-04 03:04:07 +08:00
W_ {out} & \ in [ W^ \ prime_ {out} , W^ \ prime_ {out} + strides[2] ]
Note:
2022-12-23 19:01:32 +08:00
The conv3d_transpose can be seen as the backward of the conv3d. For conv3d,
when stride > 1, conv3d maps multiple input shape to the same output shape,
so for conv3d_transpose, when stride > 1, input shape maps multiple output shape.
If output_size is None, :math:`H_ {out} = H^ \ prime_ {out} , W_ {out} = W^ \ prime_ {out} `;
else, the :math:`D_ {out} ` of the output size must between :math:`D^ \ prime_ {out} ` and
:math:`D^ \ prime_ {out} + strides[0]`, the :math:`H_ {out} ` of the output size must
between :math:`H^ \ prime_ {out} ` and :math:`H^ \ prime_ {out} + strides[1]`, and the
:math:`W_ {out} ` of the output size must between :math:`W^ \ prime_ {out} ` and
:math:`W^ \ prime_ {out} + strides[2]`.
2020-04-04 03:04:07 +08:00
Args:
2022-12-23 19:01:32 +08:00
x (Tensor): The input is 5-D Tensor with shape [N, C, D, H, W] or [N, D, H, W, C], the data type
2020-04-04 03:04:07 +08:00
of input is float32 or float64.
2020-08-22 19:57:15 +08:00
weight (Tensor): The convolution kernel, a Tensor with shape [C, M/g, kD, kH, kW],
2022-12-23 19:01:32 +08:00
where M is the number of filters (output channels), g is the number of groups,
2020-04-04 03:04:07 +08:00
kD, kH, kW are the filter ' s depth, height and width respectively.
2022-12-23 19:01:32 +08:00
bias (Tensor, optional): The bias, a Tensor of shape [M, ]. Default: None.
stride (int|list|tuple, optional): The stride size. It means the stride in transposed convolution.
2022-09-14 21:56:19 +08:00
If stride is a list/tuple, it must contain three integers, (stride_depth, stride_height,
stride_width). Otherwise, stride_depth = stride_height = stride_width = stride.
2022-12-23 19:01:32 +08:00
Default: 1.
padding (str|int|list|tuple, optional): The padding size. It means the number of zero-paddings
2020-08-29 16:04:19 +08:00
on both sides for each dimension. If `padding` is a string, either ' VALID ' or
' SAME ' which is the padding algorithm. If padding size is a tuple or list,
it could be in three forms: `[pad_depth, pad_height, pad_width]` or
2020-04-04 03:04:07 +08:00
`[pad_depth_front, pad_depth_back, pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`,
2020-09-03 13:27:41 +08:00
and when `data_format` is ` " NCDHW " `, `padding` can be in the form
2020-04-04 03:04:07 +08:00
`[[0,0], [0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
2020-09-03 13:27:41 +08:00
when `data_format` is ` " NDHWC " `, `padding` can be in the form
2020-04-04 03:04:07 +08:00
`[[0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
2022-12-23 19:01:32 +08:00
Default: 0.
output_padding (int|list|tuple, optional): Additional size added to one side
2020-08-22 19:57:15 +08:00
of each dimension in the output shape. Default: 0.
2022-12-23 19:01:32 +08:00
groups (int, optional): The groups number of the Conv3D transpose layer. Inspired by
grouped convolution in `Alex Krizhevsky ' s Deep CNN paper <https://papers.nips.cc/paper/2012/file/c399862d3b9d6b76c8436e924a68c45b-Paper.pdf>`_, in which
when groups = 2, the first half of the filters is only connected to the
2020-04-04 03:04:07 +08:00
first half of the input channels, while the second half of the
filters is only connected to the second half of the input channels.
2022-12-23 19:01:32 +08:00
Default: 1.
dilation (int|list|tuple, optional): The dilation size. It means the spacing between the kernel points.
2022-09-14 21:56:19 +08:00
If dilation is a list/tuple, it must contain three integers, (dilation_depth, dilation_height,
dilation_width). Otherwise, dilation_depth = dilation_height = dilation_width = dilation.
2022-12-23 19:01:32 +08:00
Default: 1.
output_size (int|list|tuple, optional): The output image size. If output size is a
2021-04-27 13:41:05 +08:00
list/tuple, it must contain three integers, (image_depth, image_height, image_width).
2021-04-22 13:57:47 +08:00
None if use filter_size(shape of weight), padding, and stride to calculate output_size.
2022-09-14 21:56:19 +08:00
data_format (str, optional): Specify the data format of the input, and the data format of the output
2020-08-29 16:04:19 +08:00
will be consistent with that of the input. An optional string from: ` " NCHW " `, ` " NHWC " `.
2022-12-23 19:01:32 +08:00
When it is ` " NCHW " `, the data is stored in the order of: `[batch_size, input_channels, input_height, input_width]`.
Default: ` " NCHW " `.
name (str, optional): For detailed information, please refer
to :ref:`api_guide_Name`. Usually name is no need to set.
Default: None.
2020-04-04 03:04:07 +08:00
Returns:
2020-10-26 11:53:18 +08:00
A Tensor representing the conv3d_transpose, whose data
2022-09-14 21:56:19 +08:00
type is the same with input and shape is (num_batches, channels, out_d, out_h,
out_w) or (num_batches, out_d, out_h, out_w, channels). If act is None, the tensor
variable storing the transposed convolution result, and if act is not None, the tensor
2020-04-04 03:04:07 +08:00
variable storing transposed convolution and non-linearity activation result.
Examples:
[CodeStyle][Xdoctest][8,12,15,16,18-20,22-26,29-32,34-36,38-40,42,43,45-48,50,51,60-65,75,80,82,83,85-87,89-94,99-141,143,145,147-167,169-187,207-220,257,258,260-275,277-313,315-325][API Compatibility] Update shape output format in documentation examples (#76574)
---------
Co-authored-by: SigureMo <sigure.qaq@gmail.com>
2025-11-26 10:31:46 +08:00
.. code-block:: pycon
2022-09-14 21:56:19 +08:00
2023-08-31 11:09:42 +08:00
>>> import paddle
>>> import paddle.nn.functional as F
2020-04-04 03:04:07 +08:00
2023-08-31 11:09:42 +08:00
>>> x_var = paddle.randn((2, 3, 8, 8, 8), dtype= ' float32 ' )
>>> w_var = paddle.randn((3, 6, 3, 3, 3), dtype= ' float32 ' )
2020-04-04 03:04:07 +08:00
2023-08-31 11:09:42 +08:00
>>> y_var = F.conv3d_transpose(x_var, w_var)
2020-04-04 03:04:07 +08:00
2023-08-31 11:09:42 +08:00
>>> print(y_var.shape)
[CodeStyle][Xdoctest][8,12,15,16,18-20,22-26,29-32,34-36,38-40,42,43,45-48,50,51,60-65,75,80,82,83,85-87,89-94,99-141,143,145,147-167,169-187,207-220,257,258,260-275,277-313,315-325][API Compatibility] Update shape output format in documentation examples (#76574)
---------
Co-authored-by: SigureMo <sigure.qaq@gmail.com>
2025-11-26 10:31:46 +08:00
paddle.Size([2, 6, 10, 10, 10])
2020-04-04 03:04:07 +08:00
"""
# entry checks
if data_format not in [ " NCDHW " , " NDHWC " ] :
raise ValueError (
" Attr(data_format) should be ' NCDHW ' or ' NDHWC ' . Received "
2023-09-22 10:14:38 +08:00
f " Attr(data_format): { data_format } . "
2022-10-23 20:01:27 +08:00
)
2020-04-04 03:04:07 +08:00
2022-10-23 20:01:27 +08:00
channel_last = data_format == " NDHWC "
2020-04-04 03:04:07 +08:00
channel_dim = - 1 if channel_last else 1
2021-09-07 19:28:09 +08:00
if len ( x . shape ) != 5 :
raise ValueError (
2023-09-22 10:14:38 +08:00
f " Input x should be 5D tensor, but received x with the shape of { x . shape } "
2022-10-23 20:01:27 +08:00
)
2023-02-02 16:50:05 +08:00
if len ( weight . shape ) != 5 :
raise ValueError (
2024-04-01 10:20:33 +08:00
f " Input weight should be 5D tensor, but received weight with the shape of { weight . shape } "
2023-02-02 16:50:05 +08:00
)
2020-08-22 19:57:15 +08:00
num_channels = x . shape [ channel_dim ]
2024-08-10 12:07:14 +08:00
if groups == 0 :
raise ValueError ( " The groups of conv3d_transpose should not be zero. " )
if num_channels % groups != 0 and (
in_dynamic_mode ( ) or ( num_channels != - 1 and groups != - 1 )
) :
2020-04-04 03:04:07 +08:00
raise ValueError (
" The number of input channels must be divisible by Attr(groups). "
2023-09-22 10:14:38 +08:00
f " Received: number of channels( { num_channels } ), groups( { groups } ). "
2022-10-23 20:01:27 +08:00
)
2020-04-04 03:04:07 +08:00
padding , padding_algorithm = _update_padding_nd ( padding , channel_last , 3 )
2021-07-26 15:10:15 +08:00
stride = convert_to_list ( stride , 3 , ' stride ' )
dilation = convert_to_list ( dilation , 3 , ' dilation ' )
2020-04-04 03:04:07 +08:00
if output_size is None :
output_size = [ ]
else :
2020-08-22 19:57:15 +08:00
if output_padding != 0 :
2022-10-23 20:01:27 +08:00
raise ValueError (
2025-08-18 00:05:11 +08:00
' output_padding option is mutually exclusive with output_size '
2022-10-23 20:01:27 +08:00
)
2020-08-22 19:57:15 +08:00
if isinstance ( output_size , ( list , tuple , int ) ) :
2021-07-26 15:10:15 +08:00
output_size = convert_to_list ( output_size , 3 , ' output_size ' )
2020-08-22 19:57:15 +08:00
else :
raise ValueError (
2022-10-23 20:01:27 +08:00
" output_size should be int, or list, tuple of ints "
)
2020-08-22 19:57:15 +08:00
if output_padding == 0 :
output_padding = [ ]
else :
2021-07-26 15:10:15 +08:00
output_padding = convert_to_list ( output_padding , 3 , ' output_padding ' )
2020-08-22 19:57:15 +08:00
cudnn_version = get_cudnn_version ( )
2022-10-23 20:01:27 +08:00
# TODO(LielinJiang): whether to use cudnn according to the version of cudnn
use_cudnn = (
True
if ( is_compiled_with_cuda ( ) and cudnn_version is not None )
else False
)
2020-04-04 03:04:07 +08:00
op_type = ' conv3d_transpose '
data_format_ = " NHWC " if channel_last else " NCHW "
2023-12-15 16:33:00 +08:00
if in_dynamic_or_pir_mode ( ) :
2022-10-23 20:01:27 +08:00
pre_bias = _C_ops . conv3d_transpose (
x ,
weight ,
stride ,
padding ,
output_padding ,
output_size ,
padding_algorithm ,
groups ,
dilation ,
data_format_ ,
)
2022-04-03 11:31:39 +08:00
if bias is not None :
2022-12-12 13:09:14 +08:00
return _add_with_axis ( pre_bias , bias , axis = channel_dim )
2022-04-03 11:31:39 +08:00
else :
return pre_bias
2020-04-04 03:04:07 +08:00
else :
2020-08-22 19:57:15 +08:00
inputs = { ' Input ' : [ x ] , ' Filter ' : [ weight ] }
2020-04-04 03:04:07 +08:00
attrs = {
2020-08-22 19:57:15 +08:00
' output_padding ' : output_padding ,
2020-04-04 03:04:07 +08:00
' output_size ' : output_size ,
' paddings ' : padding ,
" padding_algorithm " : padding_algorithm ,
' strides ' : stride ,
' dilations ' : dilation ,
' groups ' : groups ,
' use_cudnn ' : use_cudnn ,
2022-10-23 20:01:27 +08:00
" data_format " : data_format_ ,
2020-04-04 03:04:07 +08:00
}
helper = LayerHelper ( op_type , * * locals ( ) )
2022-10-23 20:01:27 +08:00
check_variable_and_dtype (
x , ' x ' , [ ' float16 ' , ' float32 ' , ' float64 ' ] , ' conv3d '
)
2020-04-04 03:04:07 +08:00
2020-08-22 19:57:15 +08:00
pre_bias = helper . create_variable_for_type_inference ( x . dtype )
2020-04-04 03:04:07 +08:00
outputs = { " Output " : [ pre_bias ] }
2022-10-23 20:01:27 +08:00
helper . append_op (
type = op_type , inputs = inputs , outputs = outputs , attrs = attrs
)
2020-04-04 03:04:07 +08:00
if bias is not None :
2022-12-12 13:09:14 +08:00
out = _add_with_axis ( pre_bias , bias , axis = channel_dim )
2020-04-04 03:04:07 +08:00
else :
2020-08-22 19:57:15 +08:00
out = pre_bias
2020-04-04 03:04:07 +08:00
return out
