Skip to content

GitLab

Commit b3d6785d authored by myhloli's avatar myhloli
Browse files

refactor(ocr): remove unused code and simplify model architecture 

- Remove unused imports and code
- Simplify model architecture by removing unnecessary components
- Update initialization and forward pass logic
- Rename variables for consistency
parent 3cb156f5
Hide whitespace changes
Inline Side-by-side
Showing with 318 additions and 2627 deletions
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/backbones/rec_mv1_enhance.py 100755 → 100644
View file @ b3d6785d
......@@ -2,7 +2,8 @@ import os, sys
import torch
import torch.nn as nn
import torch.nn.functional as F
from pytorchocr.modeling.common import Activation
from ..common import Activation
class ConvBNLayer(nn.Module):
......
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/backbones/rec_nrtr_mtb.py deleted 100755 → 0
View file @ 3cb156f5
import torch
from torch import nn
class MTB(nn.Module):
def __init__(self, cnn_num, in_channels):
super(MTB, self).__init__()
self.block = nn.Sequential()
self.out_channels = in_channels
self.cnn_num = cnn_num
if self.cnn_num == 2:
for i in range(self.cnn_num):
self.block.add_module(
'conv_{}'.format(i),
nn.Conv2d(
in_channels=in_channels
if i == 0 else 32 * (2**(i - 1)),
out_channels=32 * (2**i),
kernel_size=3,
stride=2,
padding=1))
self.block.add_module('relu_{}'.format(i), nn.ReLU())
self.block.add_module('bn_{}'.format(i),
nn.BatchNorm2d(32 * (2**i)))
def forward(self, images):
x = self.block(images)
if self.cnn_num == 2:
# (b, w, h, c)
x = x.permute(0, 3, 2, 1)
x_shape = x.shape
x = torch.reshape(
x, (x_shape[0], x_shape[1], x_shape[2] * x_shape[3]))
return x
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/backbones/rec_resnet_31.py deleted 100755 → 0
View file @ 3cb156f5
"""
This code is refer from:
https://github.com/open-mmlab/mmocr/blob/main/mmocr/models/textrecog/layers/conv_layer.py
https://github.com/open-mmlab/mmocr/blob/main/mmocr/models/textrecog/backbones/resnet31_ocr.py
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import torch
import torch.nn as nn
import torch.nn.functional as F
# import paddle
# from paddle import ParamAttr
# import paddle.nn as nn
# import paddle.nn.functional as F
__all__ = ["ResNet31"]
def conv3x3(in_channel, out_channel, stride=1):
return nn.Conv2d(
in_channel,
out_channel,
kernel_size=3,
stride=stride,
padding=1,
bias=False)
class BasicBlock(nn.Module):
expansion = 1
def __init__(self, in_channels, channels, stride=1, downsample=False):
super().__init__()
self.conv1 = conv3x3(in_channels, channels, stride)
self.bn1 = nn.BatchNorm2d(channels)
self.relu = nn.ReLU()
self.conv2 = conv3x3(channels, channels)
self.bn2 = nn.BatchNorm2d(channels)
self.downsample = downsample
if downsample:
self.downsample = nn.Sequential(
nn.Conv2d(
in_channels,
channels * self.expansion,
1,
stride,
bias=False),
nn.BatchNorm2d(channels * self.expansion), )
else:
self.downsample = nn.Sequential()
self.stride = stride
def forward(self, x):
residual = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
if self.downsample:
residual = self.downsample(x)
out += residual
out = self.relu(out)
return out
class ResNet31(nn.Module):
'''
Args:
in_channels (int): Number of channels of input image tensor.
layers (list[int]): List of BasicBlock number for each stage.
channels (list[int]): List of out_channels of Conv2d layer.
out_indices (None | Sequence[int]): Indices of output stages.
last_stage_pool (bool): If True, add `MaxPool2d` layer to last stage.
'''
def __init__(self,
in_channels=3,
layers=[1, 2, 5, 3],
channels=[64, 128, 256, 256, 512, 512, 512],
out_indices=None,
last_stage_pool=False):
super(ResNet31, self).__init__()
assert isinstance(in_channels, int)
assert isinstance(last_stage_pool, bool)
self.out_indices = out_indices
self.last_stage_pool = last_stage_pool
# conv 1 (Conv Conv)
self.conv1_1 = nn.Conv2d(
in_channels, channels[0], kernel_size=3, stride=1, padding=1)
self.bn1_1 = nn.BatchNorm2d(channels[0])
self.relu1_1 = nn.ReLU(inplace=True)
self.conv1_2 = nn.Conv2d(
channels[0], channels[1], kernel_size=3, stride=1, padding=1)
self.bn1_2 = nn.BatchNorm2d(channels[1])
self.relu1_2 = nn.ReLU(inplace=True)
# conv 2 (Max-pooling, Residual block, Conv)
self.pool2 = nn.MaxPool2d(
kernel_size=2, stride=2, padding=0, ceil_mode=True)
self.block2 = self._make_layer(channels[1], channels[2], layers[0])
self.conv2 = nn.Conv2d(
channels[2], channels[2], kernel_size=3, stride=1, padding=1)
self.bn2 = nn.BatchNorm2d(channels[2])
self.relu2 = nn.ReLU(inplace=True)
# conv 3 (Max-pooling, Residual block, Conv)
self.pool3 = nn.MaxPool2d(
kernel_size=2, stride=2, padding=0, ceil_mode=True)
self.block3 = self._make_layer(channels[2], channels[3], layers[1])
self.conv3 = nn.Conv2d(
channels[3], channels[3], kernel_size=3, stride=1, padding=1)
self.bn3 = nn.BatchNorm2d(channels[3])
self.relu3 = nn.ReLU(inplace=True)
# conv 4 (Max-pooling, Residual block, Conv)
self.pool4 = nn.MaxPool2d(
kernel_size=(2, 1), stride=(2, 1), padding=0, ceil_mode=True)
self.block4 = self._make_layer(channels[3], channels[4], layers[2])
self.conv4 = nn.Conv2d(
channels[4], channels[4], kernel_size=3, stride=1, padding=1)
self.bn4 = nn.BatchNorm2d(channels[4])
self.relu4 = nn.ReLU(inplace=True)
# conv 5 ((Max-pooling), Residual block, Conv)
self.pool5 = None
if self.last_stage_pool:
self.pool5 = nn.MaxPool2d(
kernel_size=2, stride=2, padding=0, ceil_mode=True)
self.block5 = self._make_layer(channels[4], channels[5], layers[3])
self.conv5 = nn.Conv2d(
channels[5], channels[5], kernel_size=3, stride=1, padding=1)
self.bn5 = nn.BatchNorm2d(channels[5])
self.relu5 = nn.ReLU(inplace=True)
self.out_channels = channels[-1]
def _make_layer(self, input_channels, output_channels, blocks):
layers = []
for _ in range(blocks):
downsample = None
if input_channels != output_channels:
downsample = nn.Sequential(
nn.Conv2d(
input_channels,
output_channels,
kernel_size=1,
stride=1,
bias=False),
nn.BatchNorm2d(output_channels), )
layers.append(
BasicBlock(
input_channels, output_channels, downsample=downsample))
input_channels = output_channels
return nn.Sequential(*layers)
def forward(self, x):
x = self.conv1_1(x)
x = self.bn1_1(x)
x = self.relu1_1(x)
x = self.conv1_2(x)
x = self.bn1_2(x)
x = self.relu1_2(x)
outs = []
for i in range(4):
layer_index = i + 2
pool_layer = getattr(self, 'pool{}'.format(layer_index))
block_layer = getattr(self, 'block{}'.format(layer_index))
conv_layer = getattr(self, 'conv{}'.format(layer_index))
bn_layer = getattr(self, 'bn{}'.format(layer_index))
relu_layer = getattr(self, 'relu{}'.format(layer_index))
if pool_layer is not None:
x = pool_layer(x)
x = block_layer(x)
x = conv_layer(x)
x = bn_layer(x)
x = relu_layer(x)
outs.append(x)
if self.out_indices is not None:
return tuple([outs[i] for i in self.out_indices])
return x
\ No newline at end of file
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/backbones/rec_resnet_fpn.py deleted 100755 → 0
View file @ 3cb156f5
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os, sys
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from pytorchocr.modeling.common import Activation
__all__ = ["ResNetFPN"]
class ResNetFPN(nn.Module):
def __init__(self, in_channels=1, layers=50, **kwargs):
super(ResNetFPN, self).__init__()
supported_layers = {
18: {
'depth': [2, 2, 2, 2],
'block_class': BasicBlock
},
34: {
'depth': [3, 4, 6, 3],
'block_class': BasicBlock
},
50: {
'depth': [3, 4, 6, 3],
'block_class': BottleneckBlock
},
101: {
'depth': [3, 4, 23, 3],
'block_class': BottleneckBlock
},
152: {
'depth': [3, 8, 36, 3],
'block_class': BottleneckBlock
}
}
stride_list = [(2, 2), (2, 2), (1, 1), (1, 1)]
num_filters = [64, 128, 256, 512]
self.depth = supported_layers[layers]['depth']
self.conv = ConvBNLayer(
in_channels=in_channels,
out_channels=64,
kernel_size=7,
stride=2,
act="relu",
name="conv1")
self.block_list = nn.ModuleList()
in_ch = 64
if layers >= 50:
for block in range(len(self.depth)):
for i in range(self.depth[block]):
if layers in [101, 152] and block == 2:
if i == 0:
conv_name = "res" + str(block + 2) + "a"
else:
conv_name = "res" + str(block + 2) + "b" + str(i)
else:
conv_name = "res" + str(block + 2) + chr(97 + i)
bottlenectBlock = BottleneckBlock(
in_channels=in_ch,
out_channels=num_filters[block],
stride=stride_list[block] if i == 0 else 1,
name=conv_name)
in_ch = num_filters[block] * 4
self.block_list.add_module("bottleneckBlock_{}_{}".format(block, i), bottlenectBlock)
else:
for block in range(len(self.depth)):
for i in range(self.depth[block]):
conv_name = "res" + str(block + 2) + chr(97 + i)
if i == 0 and block != 0:
stride = (2, 1)
else:
stride = (1, 1)
basicBlock = BasicBlock(
in_channels=in_ch,
out_channels=num_filters[block],
stride=stride_list[block] if i == 0 else 1,
is_first=block == i == 0,
name=conv_name)
in_ch = basicBlock.out_channels
self.block_list.add_module(conv_name, basicBlock)
out_ch_list = [in_ch // 4, in_ch // 2, in_ch]
self.base_block = nn.ModuleList()
self.conv_trans = []
self.bn_block = []
for i in [-2, -3]:
in_channels = out_ch_list[i + 1] + out_ch_list[i]
bb_0 = nn.Conv2d(
in_channels=in_channels,
out_channels=out_ch_list[i],
kernel_size=1,
bias=True)
self.base_block.add_module("F_{}_base_block_0".format(i), bb_0)
bb_1 = nn.Conv2d(
in_channels=out_ch_list[i],
out_channels=out_ch_list[i],
kernel_size=3,
padding=1,
bias=True)
self.base_block.add_module("F_{}_base_block_1".format(i), bb_1)
bb_2 = nn.Sequential(
nn.BatchNorm2d(out_ch_list[i]),
Activation("relu")
)
self.base_block.add_module("F_{}_base_block_2".format(i), bb_2)
bb_3 = nn.Conv2d(
in_channels=out_ch_list[i],
out_channels=512,
kernel_size=1,
bias=True)
self.base_block.add_module("F_{}_base_block_3".format(i), bb_3)
self.out_channels = 512
def __call__(self, x):
x = self.conv(x)
fpn_list = []
F = []
for i in range(len(self.depth)):
fpn_list.append(np.sum(self.depth[:i + 1]))
for i, block in enumerate(self.block_list):
x = block(x)
for number in fpn_list:
if i + 1 == number:
F.append(x)
base = F[-1]
j = 0
for i, block in enumerate(self.base_block):
if i % 3 == 0 and i < 6:
j = j + 1
b, c, w, h = F[-j - 1].shape
if [w, h] == list(base.shape[2:]):
base = base
else:
base = self.conv_trans[j - 1](base)
base = self.bn_block[j - 1](base)
base = torch.cat([base, F[-j - 1]], dim=1)
base = block(base)
return base
class ConvBNLayer(nn.Module):
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride=1,
groups=1,
act=None,
name=None):
super(ConvBNLayer, self).__init__()
self.conv = nn.Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=2 if stride == (1, 1) else kernel_size,
dilation=2 if stride == (1, 1) else 1,
stride=stride,
padding=(kernel_size - 1) // 2,
groups=groups,
bias=False, )
if name == "conv1":
bn_name = "bn_" + name
else:
bn_name = "bn" + name[3:]
self.bn = nn.BatchNorm2d(out_channels)
self.act = act
if self.act is not None:
self._act = Activation(act_type=self.act, inplace=True)
def __call__(self, x):
x = self.conv(x)
x = self.bn(x)
if self.act is not None:
x = self._act(x)
return x
class ShortCut(nn.Module):
def __init__(self, in_channels, out_channels, stride, name, is_first=False):
super(ShortCut, self).__init__()
self.use_conv = True
if in_channels != out_channels or stride != 1 or is_first == True:
if stride == (1, 1):
self.conv = ConvBNLayer(
in_channels, out_channels, 1, 1, name=name)
else: # stride==(2,2)
self.conv = ConvBNLayer(
in_channels, out_channels, 1, stride, name=name)
else:
self.use_conv = False
def forward(self, x):
if self.use_conv:
x = self.conv(x)
return x
class BottleneckBlock(nn.Module):
def __init__(self, in_channels, out_channels, stride, name):
super(BottleneckBlock, self).__init__()
self.conv0 = ConvBNLayer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
act='relu',
name=name + "_branch2a")
self.conv1 = ConvBNLayer(
in_channels=out_channels,
out_channels=out_channels,
kernel_size=3,
stride=stride,
act='relu',
name=name + "_branch2b")
self.conv2 = ConvBNLayer(
in_channels=out_channels,
out_channels=out_channels * 4,
kernel_size=1,
act=None,
name=name + "_branch2c")
self.short = ShortCut(
in_channels=in_channels,
out_channels=out_channels * 4,
stride=stride,
is_first=False,
name=name + "_branch1")
self.out_channels = out_channels * 4
def forward(self, x):
y = self.conv0(x)
y = self.conv1(y)
y = self.conv2(y)
y = y + self.short(x)
y = F.relu(y)
return y
class BasicBlock(nn.Module):
def __init__(self, in_channels, out_channels, stride, name, is_first):
super(BasicBlock, self).__init__()
self.conv0 = ConvBNLayer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=3,
act='relu',
stride=stride,
name=name + "_branch2a")
self.conv1 = ConvBNLayer(
in_channels=out_channels,
out_channels=out_channels,
kernel_size=3,
act=None,
name=name + "_branch2b")
self.short = ShortCut(
in_channels=in_channels,
out_channels=out_channels,
stride=stride,
is_first=is_first,
name=name + "_branch1")
self.out_channels = out_channels
def forward(self, x):
y = self.conv0(x)
y = self.conv1(y)
y = y + self.short(x)
return F.relu(y)
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/backbones/rec_resnet_vd.py deleted 100755 → 0
View file @ 3cb156f5
import os, sys
import torch
import torch.nn as nn
import torch.nn.functional as F
from pytorchocr.modeling.common import Activation
class ConvBNLayer(nn.Module):
def __init__(
self,
in_channels,
out_channels,
kernel_size,
stride=1,
groups=1,
is_vd_mode=False,
act=None,
name=None, ):
super(ConvBNLayer, self).__init__()
self.act = act
self.is_vd_mode = is_vd_mode
self._pool2d_avg = nn.AvgPool2d(
kernel_size=stride, stride=stride, padding=0, ceil_mode=True)
self._conv = nn.Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=1 if is_vd_mode else stride,
padding=(kernel_size - 1) // 2,
groups=groups,
bias=False)
self._batch_norm = nn.BatchNorm2d(
out_channels,)
if self.act is not None:
self._act = Activation(act_type=act, inplace=True)
def forward(self, inputs):
if self.is_vd_mode:
inputs = self._pool2d_avg(inputs)
y = self._conv(inputs)
y = self._batch_norm(y)
if self.act is not None:
y = self._act(y)
return y
class BottleneckBlock(nn.Module):
def __init__(self,
in_channels,
out_channels,
stride,
shortcut=True,
if_first=False,
name=None):
super(BottleneckBlock, self).__init__()
self.conv0 = ConvBNLayer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
act='relu',
name=name + "_branch2a")
self.conv1 = ConvBNLayer(
in_channels=out_channels,
out_channels=out_channels,
kernel_size=3,
stride=stride,
act='relu',
name=name + "_branch2b")
self.conv2 = ConvBNLayer(
in_channels=out_channels,
out_channels=out_channels * 4,
kernel_size=1,
act=None,
name=name + "_branch2c")
if not shortcut:
self.short = ConvBNLayer(
in_channels=in_channels,
out_channels=out_channels * 4,
kernel_size=1,
stride=stride,
is_vd_mode=not if_first and stride[0] != 1,
name=name + "_branch1")
self.shortcut = shortcut
def forward(self, inputs):
y = self.conv0(inputs)
conv1 = self.conv1(y)
conv2 = self.conv2(conv1)
if self.shortcut:
short = inputs
else:
short = self.short(inputs)
y = short + conv2
y = F.relu(y)
return y
class BasicBlock(nn.Module):
def __init__(self,
in_channels,
out_channels,
stride,
shortcut=True,
if_first=False,
name=None):
super(BasicBlock, self).__init__()
self.stride = stride
self.conv0 = ConvBNLayer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=3,
stride=stride,
act='relu',
name=name + "_branch2a")
self.conv1 = ConvBNLayer(
in_channels=out_channels,
out_channels=out_channels,
kernel_size=3,
act=None,
name=name + "_branch2b")
if not shortcut:
self.short = ConvBNLayer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
stride=stride,
is_vd_mode=not if_first and stride[0] != 1,
name=name + "_branch1")
self.shortcut = shortcut
def forward(self, inputs):
y = self.conv0(inputs)
conv1 = self.conv1(y)
if self.shortcut:
short = inputs
else:
short = self.short(inputs)
y = short + conv1
y = F.relu(y)
return y
class ResNet(nn.Module):
def __init__(self, in_channels=3, layers=50, **kwargs):
super(ResNet, self).__init__()
self.layers = layers
supported_layers = [18, 34, 50, 101, 152, 200]
assert layers in supported_layers, \
"supported layers are {} but input layer is {}".format(
supported_layers, layers)
if layers == 18:
depth = [2, 2, 2, 2]
elif layers == 34 or layers == 50:
depth = [3, 4, 6, 3]
elif layers == 101:
depth = [3, 4, 23, 3]
elif layers == 152:
depth = [3, 8, 36, 3]
elif layers == 200:
depth = [3, 12, 48, 3]
num_channels = [64, 256, 512,
1024] if layers >= 50 else [64, 64, 128, 256]
num_filters = [64, 128, 256, 512]
self.conv1_1 = ConvBNLayer(
in_channels=in_channels,
out_channels=32,
kernel_size=3,
stride=1,
act='relu',
name="conv1_1")
self.conv1_2 = ConvBNLayer(
in_channels=32,
out_channels=32,
kernel_size=3,
stride=1,
act='relu',
name="conv1_2")
self.conv1_3 = ConvBNLayer(
in_channels=32,
out_channels=64,
kernel_size=3,
stride=1,
act='relu',
name="conv1_3")
self.pool2d_max = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
# self.block_list = list()
self.block_list = nn.Sequential()
if layers >= 50:
for block in range(len(depth)):
shortcut = False
for i in range(depth[block]):
if layers in [101, 152, 200] and block == 2:
if i == 0:
conv_name = "res" + str(block + 2) + "a"
else:
conv_name = "res" + str(block + 2) + "b" + str(i)
else:
conv_name = "res" + str(block + 2) + chr(97 + i)
if i == 0 and block != 0:
stride = (2, 1)
else:
stride = (1, 1)
bottleneck_block = BottleneckBlock(in_channels=num_channels[block] if i == 0 else num_filters[block] * 4,
out_channels=num_filters[block],
stride=stride,
shortcut=shortcut,
if_first=block == i == 0,
name=conv_name)
shortcut = True
# self.block_list.append(bottleneck_block)
self.block_list.add_module('bb_%d_%d' % (block, i), bottleneck_block)
self.out_channels = num_filters[block]
else:
for block in range(len(depth)):
shortcut = False
for i in range(depth[block]):
conv_name = "res" + str(block + 2) + chr(97 + i)
if i == 0 and block != 0:
stride = (2, 1)
else:
stride = (1, 1)
basic_block = BasicBlock(in_channels=num_channels[block] if i == 0 else num_filters[block],
out_channels=num_filters[block],
stride=stride,
shortcut=shortcut,
if_first=block == i == 0,
name=conv_name)
shortcut = True
# self.block_list.append(basic_block)
self.block_list.add_module('bb_%d_%d' % (block, i), basic_block)
self.out_channels = num_filters[block]
self.out_pool = nn.MaxPool2d(kernel_size=2, stride=2, padding=0)
def forward(self, inputs):
y = self.conv1_1(inputs)
y = self.conv1_2(y)
y = self.conv1_3(y)
y = self.pool2d_max(y)
for block in self.block_list:
y = block(y)
y = self.out_pool(y)
return y
\ No newline at end of file
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/backbones/rec_svtrnet.py 100755 → 100644
View file @ b3d6785d
import torch
import torch.nn as nn
from pytorchocr.modeling.common import Activation
import numpy as np
import torch
from torch import nn
from ..common import Activation
def drop_path(x, drop_prob=0., training=False):
def drop_path(x, drop_prob=0.0, training=False):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ...
"""
if drop_prob == 0. or not training:
if drop_prob == 0.0 or not training:
return x
keep_prob = torch.as_tensor(1 - drop_prob)
shape = (x.shape[0], ) + (1, ) * (x.ndim - 1)
shape = (x.shape[0],) + (1,) * (x.ndim - 1)
random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype)
random_tensor = torch.floor(random_tensor) # binarize
output = x.divide(keep_prob) * random_tensor
......@@ -19,15 +21,17 @@ def drop_path(x, drop_prob=0., training=False):
class ConvBNLayer(nn.Module):
def __init__(self,
in_channels,
out_channels,
kernel_size=3,
stride=1,
padding=0,
bias_attr=False,
groups=1,
act='gelu'):
def __init__(
self,
in_channels,
out_channels,
kernel_size=3,
stride=1,
padding=0,
bias_attr=False,
groups=1,
act="gelu",
):
super().__init__()
self.conv = nn.Conv2d(
in_channels=in_channels,
......@@ -36,7 +40,8 @@ class ConvBNLayer(nn.Module):
stride=stride,
padding=padding,
groups=groups,
bias=bias_attr)
bias=bias_attr,
)
self.norm = nn.BatchNorm2d(out_channels)
self.act = Activation(act_type=act, inplace=True)
......@@ -48,8 +53,7 @@ class ConvBNLayer(nn.Module):
class DropPath(nn.Module):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
"""
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""
def __init__(self, drop_prob=None):
super(DropPath, self).__init__()
......@@ -68,12 +72,14 @@ class Identity(nn.Module):
class Mlp(nn.Module):
def __init__(self,
in_features,
hidden_features=None,
out_features=None,
act_layer='gelu',
drop=0.):
def __init__(
self,
in_features,
hidden_features=None,
out_features=None,
act_layer="gelu",
drop=0.0,
):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
......@@ -93,11 +99,12 @@ class Mlp(nn.Module):
class ConvMixer(nn.Module):
def __init__(
self,
dim,
num_heads=8,
HW=[8, 25],
local_k=[3, 3], ):
self,
dim,
num_heads=8,
HW=[8, 25],
local_k=[3, 3],
):
super().__init__()
self.HW = HW
self.dim = dim
......@@ -105,9 +112,10 @@ class ConvMixer(nn.Module):
dim,
dim,
local_k,
1, [local_k[0] // 2, local_k[1] // 2],
1,
[local_k[0] // 2, local_k[1] // 2],
groups=num_heads,
)
)
def forward(self, x):
h = self.HW[0]
......@@ -119,16 +127,18 @@ class ConvMixer(nn.Module):
class Attention(nn.Module):
def __init__(self,
dim,
num_heads=8,
mixer='Global',
HW=[8, 25],
local_k=[7, 11],
qkv_bias=False,
qk_scale=None,
attn_drop=0.,
proj_drop=0.):
def __init__(
self,
dim,
num_heads=8,
mixer="Global",
HW=[8, 25],
local_k=[7, 11],
qkv_bias=False,
qk_scale=None,
attn_drop=0.0,
proj_drop=0.0,
):
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
......@@ -143,16 +153,19 @@ class Attention(nn.Module):
W = HW[1]
self.N = H * W
self.C = dim
if mixer == 'Local' and HW is not None:
if mixer == "Local" and HW is not None:
hk = local_k[0]
wk = local_k[1]
mask = torch.ones(H * W, H + hk - 1, W + wk - 1, dtype=torch.float32)
for h in range(0, H):
for w in range(0, W):
mask[h * W + w, h:h + hk, w:w + wk] = 0.
mask_paddle = mask[:, hk // 2:H + hk // 2, wk // 2:W + wk //
2].flatten(1)
mask_inf = torch.full([H * W, H * W], fill_value=float("-Inf"), dtype=torch.float32)
mask[h * W + w, h : h + hk, w : w + wk] = 0.0
mask_paddle = mask[:, hk // 2 : H + hk // 2, wk // 2 : W + wk // 2].flatten(
1
)
mask_inf = torch.full(
[H * W, H * W], fill_value=float("-Inf"), dtype=torch.float32
)
mask = torch.where(mask_paddle < 1, mask_paddle, mask_inf)
self.mask = mask.unsqueeze(0).unsqueeze(1)
# self.mask = mask[None, None, :]
......@@ -165,11 +178,13 @@ class Attention(nn.Module):
else:
_, N, C = x.shape
qkv = self.qkv(x)
qkv = qkv.reshape((-1, N, 3, self.num_heads, C // self.num_heads)).permute(2, 0, 3, 1, 4)
qkv = qkv.reshape((-1, N, 3, self.num_heads, C // self.num_heads)).permute(
2, 0, 3, 1, 4
)
q, k, v = qkv[0] * self.scale, qkv[1], qkv[2]
attn = (q.matmul(k.permute(0, 1, 3, 2)))
if self.mixer == 'Local':
attn = q.matmul(k.permute(0, 1, 3, 2))
if self.mixer == "Local":
attn += self.mask
attn = nn.functional.softmax(attn, dim=-1)
attn = self.attn_drop(attn)
......@@ -181,28 +196,30 @@ class Attention(nn.Module):
class Block(nn.Module):
def __init__(self,
dim,
num_heads,
mixer='Global',
local_mixer=[7, 11],
HW=None,
mlp_ratio=4.,
qkv_bias=False,
qk_scale=None,
drop=0.,
attn_drop=0.,
drop_path=0.,
act_layer='gelu',
norm_layer='nn.LayerNorm',
epsilon=1e-6,
prenorm=True):
def __init__(
self,
dim,
num_heads,
mixer="Global",
local_mixer=[7, 11],
HW=None,
mlp_ratio=4.0,
qkv_bias=False,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
act_layer="gelu",
norm_layer="nn.LayerNorm",
epsilon=1e-6,
prenorm=True,
):
super().__init__()
if isinstance(norm_layer, str):
self.norm1 = eval(norm_layer)(dim, eps=epsilon)
else:
self.norm1 = norm_layer(dim)
if mixer == 'Global' or mixer == 'Local':
if mixer == "Global" or mixer == "Local":
self.mixer = Attention(
dim,
num_heads=num_heads,
......@@ -212,24 +229,26 @@ class Block(nn.Module):
qkv_bias=qkv_bias,
qk_scale=qk_scale,
attn_drop=attn_drop,
proj_drop=drop)
elif mixer == 'Conv':
self.mixer = ConvMixer(
dim, num_heads=num_heads, HW=HW, local_k=local_mixer)
proj_drop=drop,
)
elif mixer == "Conv":
self.mixer = ConvMixer(dim, num_heads=num_heads, HW=HW, local_k=local_mixer)
else:
raise TypeError("The mixer must be one of [Global, Local, Conv]")
self.drop_path = DropPath(drop_path) if drop_path > 0. else Identity()
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else Identity()
if isinstance(norm_layer, str):
self.norm2 = eval(norm_layer)(dim, eps=epsilon)
else:
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp_ratio = mlp_ratio
self.mlp = Mlp(in_features=dim,
hidden_features=mlp_hidden_dim,
act_layer=act_layer,
drop=drop)
self.mlp = Mlp(
in_features=dim,
hidden_features=mlp_hidden_dim,
act_layer=act_layer,
drop=drop,
)
self.prenorm = prenorm
def forward(self, x):
......@@ -243,25 +262,24 @@ class Block(nn.Module):
class PatchEmbed(nn.Module):
""" Image to Patch Embedding
"""
"""Image to Patch Embedding"""
def __init__(self,
img_size=[32, 100],
in_channels=3,
embed_dim=768,
sub_num=2,
patch_size=[4, 4],
mode='pope',
):
def __init__(
self,
img_size=[32, 100],
in_channels=3,
embed_dim=768,
sub_num=2,
patch_size=[4, 4],
mode="pope",
):
super().__init__()
num_patches = (img_size[1] // (2 ** sub_num)) * \
(img_size[0] // (2 ** sub_num))
num_patches = (img_size[1] // (2**sub_num)) * (img_size[0] // (2**sub_num))
self.img_size = img_size
self.num_patches = num_patches
self.embed_dim = embed_dim
self.norm = None
if mode == 'pope':
if mode == "pope":
if sub_num == 2:
self.proj = nn.Sequential(
ConvBNLayer(
......@@ -270,16 +288,19 @@ class PatchEmbed(nn.Module):
kernel_size=3,
stride=2,
padding=1,
act='gelu',
bias_attr=True),
act="gelu",
bias_attr=True,
),
ConvBNLayer(
in_channels=embed_dim // 2,
out_channels=embed_dim,
kernel_size=3,
stride=2,
padding=1,
act='gelu',
bias_attr=True))
act="gelu",
bias_attr=True,
),
)
if sub_num == 3:
self.proj = nn.Sequential(
ConvBNLayer(
......@@ -288,55 +309,66 @@ class PatchEmbed(nn.Module):
kernel_size=3,
stride=2,
padding=1,
act='gelu',
bias_attr=True),
act="gelu",
bias_attr=True,
),
ConvBNLayer(
in_channels=embed_dim // 4,
out_channels=embed_dim // 2,
kernel_size=3,
stride=2,
padding=1,
act='gelu',
bias_attr=True),
act="gelu",
bias_attr=True,
),
ConvBNLayer(
in_channels=embed_dim // 2,
out_channels=embed_dim,
kernel_size=3,
stride=2,
padding=1,
act='gelu',
bias_attr=True))
elif mode == 'linear':
act="gelu",
bias_attr=True,
),
)
elif mode == "linear":
self.proj = nn.Conv2d(
1, embed_dim, kernel_size=patch_size, stride=patch_size)
self.num_patches = img_size[0] // patch_size[0] * img_size[
1] // patch_size[1]
1, embed_dim, kernel_size=patch_size, stride=patch_size
)
self.num_patches = (
img_size[0] // patch_size[0] * img_size[1] // patch_size[1]
)
def forward(self, x):
B, C, H, W = x.shape
assert H == self.img_size[0] and W == self.img_size[1], \
"Input image size ({}*{}) doesn't match model ({}*{}).".format(
H,W,self.img_size[0],self.img_size[1]
)
assert (
H == self.img_size[0] and W == self.img_size[1]
), "Input image size ({}*{}) doesn't match model ({}*{}).".format(
H, W, self.img_size[0], self.img_size[1]
)
x = self.proj(x).flatten(2).permute(0, 2, 1)
return x
class SubSample(nn.Module):
def __init__(self,
in_channels,
out_channels,
types='Pool',
stride=[2, 1],
sub_norm='nn.LayerNorm',
act=None):
def __init__(
self,
in_channels,
out_channels,
types="Pool",
stride=[2, 1],
sub_norm="nn.LayerNorm",
act=None,
):
super().__init__()
self.types = types
if types == 'Pool':
if types == "Pool":
self.avgpool = nn.AvgPool2d(
kernel_size=[3, 5], stride=stride, padding=[1, 2])
kernel_size=[3, 5], stride=stride, padding=[1, 2]
)
self.maxpool = nn.MaxPool2d(
kernel_size=[3, 5], stride=stride, padding=[1, 2])
kernel_size=[3, 5], stride=stride, padding=[1, 2]
)
self.proj = nn.Linear(in_channels, out_channels)
else:
self.conv = nn.Conv2d(
......@@ -345,7 +377,7 @@ class SubSample(nn.Module):
kernel_size=3,
stride=stride,
padding=1,
)
)
self.norm = eval(sub_norm)(out_channels)
if act is not None:
self.act = act()
......@@ -353,8 +385,7 @@ class SubSample(nn.Module):
self.act = None
def forward(self, x):
if self.types == 'Pool':
if self.types == "Pool":
x1 = self.avgpool(x)
x2 = self.maxpool(x)
x = (x1 + x2) * 0.5
......@@ -371,46 +402,51 @@ class SubSample(nn.Module):
class SVTRNet(nn.Module):
def __init__(
self,
img_size=[32, 100],
in_channels=3,
embed_dim=[64, 128, 256],
depth=[3, 6, 3],
num_heads=[2, 4, 8],
mixer=['Local'] * 6 + ['Global'] *
6, # Local atten, Global atten, Conv
local_mixer=[[7, 11], [7, 11], [7, 11]],
patch_merging='Conv', # Conv, Pool, None
mlp_ratio=4,
qkv_bias=True,
qk_scale=None,
drop_rate=0.,
last_drop=0.0,
attn_drop_rate=0.,
drop_path_rate=0.1,
norm_layer='nn.LayerNorm',
sub_norm='nn.LayerNorm',
epsilon=1e-6,
out_channels=192,
out_char_num=25,
block_unit='Block',
act='gelu',
last_stage=True,
sub_num=2,
prenorm=True,
use_lenhead=False,
**kwargs):
self,
img_size=[32, 100],
in_channels=3,
embed_dim=[64, 128, 256],
depth=[3, 6, 3],
num_heads=[2, 4, 8],
mixer=["Local"] * 6 + ["Global"] * 6, # Local atten, Global atten, Conv
local_mixer=[[7, 11], [7, 11], [7, 11]],
patch_merging="Conv", # Conv, Pool, None
mlp_ratio=4,
qkv_bias=True,
qk_scale=None,
drop_rate=0.0,
last_drop=0.0,
attn_drop_rate=0.0,
drop_path_rate=0.1,
norm_layer="nn.LayerNorm",
sub_norm="nn.LayerNorm",
epsilon=1e-6,
out_channels=192,
out_char_num=25,
block_unit="Block",
act="gelu",
last_stage=True,
sub_num=2,
prenorm=True,
use_lenhead=False,
**kwargs
):
super().__init__()
self.img_size = img_size
self.embed_dim = embed_dim
self.out_channels = out_channels
self.prenorm = prenorm
patch_merging = None if patch_merging != 'Conv' and patch_merging != 'Pool' else patch_merging
patch_merging = (
None
if patch_merging != "Conv" and patch_merging != "Pool"
else patch_merging
)
self.patch_embed = PatchEmbed(
img_size=img_size,
in_channels=in_channels,
embed_dim=embed_dim[0],
sub_num=sub_num)
sub_num=sub_num,
)
num_patches = self.patch_embed.num_patches
self.HW = [img_size[0] // (2**sub_num), img_size[1] // (2**sub_num)]
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim[0]))
......@@ -418,81 +454,95 @@ class SVTRNet(nn.Module):
Block_unit = eval(block_unit)
dpr = np.linspace(0, drop_path_rate, sum(depth))
self.blocks1 = nn.ModuleList([
Block_unit(
dim=embed_dim[0],
num_heads=num_heads[0],
mixer=mixer[0:depth[0]][i],
HW=self.HW,
local_mixer=local_mixer[0],
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
act_layer=act,
attn_drop=attn_drop_rate,
drop_path=dpr[0:depth[0]][i],
norm_layer=norm_layer,
epsilon=epsilon,
prenorm=prenorm) for i in range(depth[0])
])
self.blocks1 = nn.ModuleList(
[
Block_unit(
dim=embed_dim[0],
num_heads=num_heads[0],
mixer=mixer[0 : depth[0]][i],
HW=self.HW,
local_mixer=local_mixer[0],
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
act_layer=act,
attn_drop=attn_drop_rate,
drop_path=dpr[0 : depth[0]][i],
norm_layer=norm_layer,
epsilon=epsilon,
prenorm=prenorm,
)
for i in range(depth[0])
]
)
if patch_merging is not None:
self.sub_sample1 = SubSample(
embed_dim[0],
embed_dim[1],
sub_norm=sub_norm,
stride=[2, 1],
types=patch_merging)
types=patch_merging,
)
HW = [self.HW[0] // 2, self.HW[1]]
else:
HW = self.HW
self.patch_merging = patch_merging
self.blocks2 = nn.ModuleList([
Block_unit(
dim=embed_dim[1],
num_heads=num_heads[1],
mixer=mixer[depth[0]:depth[0] + depth[1]][i],
HW=HW,
local_mixer=local_mixer[1],
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
act_layer=act,
attn_drop=attn_drop_rate,
drop_path=dpr[depth[0]:depth[0] + depth[1]][i],
norm_layer=norm_layer,
epsilon=epsilon,
prenorm=prenorm) for i in range(depth[1])
])
self.blocks2 = nn.ModuleList(
[
Block_unit(
dim=embed_dim[1],
num_heads=num_heads[1],
mixer=mixer[depth[0] : depth[0] + depth[1]][i],
HW=HW,
local_mixer=local_mixer[1],
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
act_layer=act,
attn_drop=attn_drop_rate,
drop_path=dpr[depth[0] : depth[0] + depth[1]][i],
norm_layer=norm_layer,
epsilon=epsilon,
prenorm=prenorm,
)
for i in range(depth[1])
]
)
if patch_merging is not None:
self.sub_sample2 = SubSample(
embed_dim[1],
embed_dim[2],
sub_norm=sub_norm,
stride=[2, 1],
types=patch_merging)
types=patch_merging,
)
HW = [self.HW[0] // 4, self.HW[1]]
else:
HW = self.HW
self.blocks3 = nn.ModuleList([
Block_unit(
dim=embed_dim[2],
num_heads=num_heads[2],
mixer=mixer[depth[0] + depth[1]:][i],
HW=HW,
local_mixer=local_mixer[2],
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
act_layer=act,
attn_drop=attn_drop_rate,
drop_path=dpr[depth[0] + depth[1]:][i],
norm_layer=norm_layer,
epsilon=epsilon,
prenorm=prenorm) for i in range(depth[2])
])
self.blocks3 = nn.ModuleList(
[
Block_unit(
dim=embed_dim[2],
num_heads=num_heads[2],
mixer=mixer[depth[0] + depth[1] :][i],
HW=HW,
local_mixer=local_mixer[2],
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
act_layer=act,
attn_drop=attn_drop_rate,
drop_path=dpr[depth[0] + depth[1] :][i],
norm_layer=norm_layer,
epsilon=epsilon,
prenorm=prenorm,
)
for i in range(depth[2])
]
)
self.last_stage = last_stage
if last_stage:
self.avg_pool = nn.AdaptiveAvgPool2d([1, out_char_num])
......@@ -502,8 +552,9 @@ class SVTRNet(nn.Module):
kernel_size=1,
stride=1,
padding=0,
bias=False)
self.hardswish = Activation('hard_swish', inplace=True) #nn.Hardswish()
bias=False,
)
self.hardswish = Activation("hard_swish", inplace=True) # nn.Hardswish()
# self.dropout = nn.Dropout(p=last_drop, mode="downscale_in_infer")
self.dropout = nn.Dropout(p=last_drop)
if not prenorm:
......@@ -511,9 +562,10 @@ class SVTRNet(nn.Module):
self.use_lenhead = use_lenhead
if use_lenhead:
self.len_conv = nn.Linear(embed_dim[2], self.out_channels)
self.hardswish_len = Activation('hard_swish', inplace=True)# nn.Hardswish()
self.dropout_len = nn.Dropout(
p=last_drop)
self.hardswish_len = Activation(
"hard_swish", inplace=True
) # nn.Hardswish()
self.dropout_len = nn.Dropout(p=last_drop)
torch.nn.init.xavier_normal_(self.pos_embed)
self.apply(self._init_weights)
......@@ -521,7 +573,7 @@ class SVTRNet(nn.Module):
def _init_weights(self, m):
# weight initialization
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out')
nn.init.kaiming_normal_(m.weight, mode="fan_out")
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.BatchNorm2d):
......@@ -532,7 +584,7 @@ class SVTRNet(nn.Module):
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.ConvTranspose2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out')
nn.init.kaiming_normal_(m.weight, mode="fan_out")
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.LayerNorm):
......@@ -548,13 +600,17 @@ class SVTRNet(nn.Module):
if self.patch_merging is not None:
x = self.sub_sample1(
x.permute(0, 2, 1).reshape(
[-1, self.embed_dim[0], self.HW[0], self.HW[1]]))
[-1, self.embed_dim[0], self.HW[0], self.HW[1]]
)
)
for blk in self.blocks2:
x = blk(x)
if self.patch_merging is not None:
x = self.sub_sample2(
x.permute(0, 2, 1).reshape(
[-1, self.embed_dim[1], self.HW[0] // 2, self.HW[1]]))
[-1, self.embed_dim[1], self.HW[0] // 2, self.HW[1]]
)
)
for blk in self.blocks3:
x = blk(x)
if not self.prenorm:
......@@ -572,11 +628,11 @@ class SVTRNet(nn.Module):
else:
h = self.HW[0]
x = self.avg_pool(
x.permute(0, 2, 1).reshape(
[-1, self.embed_dim[2], h, self.HW[1]]))
x.permute(0, 2, 1).reshape([-1, self.embed_dim[2], h, self.HW[1]])
)
x = self.last_conv(x)
x = self.hardswish(x)
x = self.dropout(x)
if self.use_lenhead:
return x, len_x
return x
\ No newline at end of file
return x
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/backbones/rec_vitstr.py deleted 100755 → 0
View file @ 3cb156f5
"""
This code is refer from:
https://github.com/roatienza/deep-text-recognition-benchmark/blob/master/modules/vitstr.py
"""
import numpy as np
import torch
import torch.nn as nn
from pytorchocr.modeling.backbones.rec_svtrnet import Block, PatchEmbed
# import paddle
# import paddle.nn as nn
# from ppocr.modeling.backbones.rec_svtrnet import Block, PatchEmbed, zeros_, trunc_normal_, ones_
scale_dim_heads = {'tiny': [192, 3], 'small': [384, 6], 'base': [768, 12]}
class ViTSTR(nn.Module):
def __init__(self,
img_size=[224, 224],
in_channels=1,
scale='tiny',
seqlen=27,
patch_size=[16, 16],
embed_dim=None,
depth=12,
num_heads=None,
mlp_ratio=4,
qkv_bias=True,
qk_scale=None,
drop_path_rate=0.,
drop_rate=0.,
attn_drop_rate=0.,
norm_layer='nn.LayerNorm',
act_layer='gelu',
epsilon=1e-6,
out_channels=None,
**kwargs):
super().__init__()
self.seqlen = seqlen
embed_dim = embed_dim if embed_dim is not None else scale_dim_heads[
scale][0]
num_heads = num_heads if num_heads is not None else scale_dim_heads[
scale][1]
out_channels = out_channels if out_channels is not None else embed_dim
self.patch_embed = PatchEmbed(
img_size=img_size,
in_channels=in_channels,
embed_dim=embed_dim,
patch_size=patch_size,
mode='linear')
num_patches = self.patch_embed.num_patches
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
self.pos_drop = nn.Dropout(p=drop_rate)
dpr = np.linspace(0, drop_path_rate, depth)
self.blocks = nn.ModuleList([
Block(
dim=embed_dim,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_path=dpr[i],
norm_layer=norm_layer,
act_layer=act_layer,
epsilon=epsilon,
prenorm=False) for i in range(depth)
])
self.norm = eval(norm_layer)(embed_dim, eps=epsilon)
self.out_channels = out_channels
torch.nn.init.xavier_normal_(self.pos_embed)
torch.nn.init.xavier_normal_(self.cls_token)
self.apply(self._init_weights)
def _init_weights(self, m):
# weight initialization
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.BatchNorm2d):
nn.init.ones_(m.weight)
nn.init.zeros_(m.bias)
elif isinstance(m, nn.Linear):
nn.init.normal_(m.weight, 0, 0.01)
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.ConvTranspose2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.LayerNorm):
nn.init.ones_(m.weight)
nn.init.zeros_(m.bias)
def forward_features(self, x):
B = x.shape[0]
x = self.patch_embed(x)
# cls_tokens = paddle.tile(self.cls_token, repeat_times=[B, 1, 1])
cls_tokens = self.cls_token.repeat(B, 1, 1)
x = torch.cat((cls_tokens, x), dim=1)
x = x + self.pos_embed
x = self.pos_drop(x)
for blk in self.blocks:
x = blk(x)
x = self.norm(x)
return x
def forward(self, x):
x = self.forward_features(x)
x = x[:, :self.seqlen]
return x.permute(0, 2, 1).unsqueeze(2)
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/backbones/table_mobilenet_v3.py deleted 100755 → 0
View file @ 3cb156f5
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import torch
import torch.nn as nn
import torch.nn.functional as F
__all__ = ['MobileNetV3']
def make_divisible(v, divisor=8, min_value=None):
if min_value is None:
min_value = divisor
new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
if new_v < 0.9 * v:
new_v += divisor
return new_v
def hard_sigmoid(x, slope=0.1666667, offset=0.5,):
return torch.clamp(slope * x + offset, 0., 1.)
def hard_swish(x, inplace=True):
return x * F.relu6(x + 3., inplace=inplace) / 6.
class MobileNetV3(nn.Module):
def __init__(self,
in_channels=3,
model_name='large',
scale=0.5,
disable_se=False,
**kwargs):
"""
the MobilenetV3 backbone network for detection module.
Args:
params(dict): the super parameters for build network
"""
super(MobileNetV3, self).__init__()
self.disable_se = disable_se
if model_name == "large":
cfg = [
# k, exp, c, se, nl, s,
[3, 16, 16, False, 'relu', 1],
[3, 64, 24, False, 'relu', 2],
[3, 72, 24, False, 'relu', 1],
[5, 72, 40, True, 'relu', 2],
[5, 120, 40, True, 'relu', 1],
[5, 120, 40, True, 'relu', 1],
[3, 240, 80, False, 'hardswish', 2],
[3, 200, 80, False, 'hardswish', 1],
[3, 184, 80, False, 'hardswish', 1],
[3, 184, 80, False, 'hardswish', 1],
[3, 480, 112, True, 'hardswish', 1],
[3, 672, 112, True, 'hardswish', 1],
[5, 672, 160, True, 'hardswish', 2],
[5, 960, 160, True, 'hardswish', 1],
[5, 960, 160, True, 'hardswish', 1],
]
cls_ch_squeeze = 960
elif model_name == "small":
cfg = [
# k, exp, c, se, nl, s,
[3, 16, 16, True, 'relu', 2],
[3, 72, 24, False, 'relu', 2],
[3, 88, 24, False, 'relu', 1],
[5, 96, 40, True, 'hardswish', 2],
[5, 240, 40, True, 'hardswish', 1],
[5, 240, 40, True, 'hardswish', 1],
[5, 120, 48, True, 'hardswish', 1],
[5, 144, 48, True, 'hardswish', 1],
[5, 288, 96, True, 'hardswish', 2],
[5, 576, 96, True, 'hardswish', 1],
[5, 576, 96, True, 'hardswish', 1],
]
cls_ch_squeeze = 576
else:
raise NotImplementedError("mode[" + model_name +
"_model] is not implemented!")
supported_scale = [0.35, 0.5, 0.75, 1.0, 1.25]
assert scale in supported_scale, \
"supported scale are {} but input scale is {}".format(supported_scale, scale)
inplanes = 16
# conv1
self.conv = ConvBNLayer(
in_channels=in_channels,
out_channels=make_divisible(inplanes * scale),
kernel_size=3,
stride=2,
padding=1,
groups=1,
if_act=True,
act='hardswish',
name='conv1')
self.stages = nn.ModuleList()
self.out_channels = []
block_list = []
i = 0
inplanes = make_divisible(inplanes * scale)
for (k, exp, c, se, nl, s) in cfg:
se = se and not self.disable_se
start_idx = 2 if model_name == 'large' else 0
if s == 2 and i > start_idx:
self.out_channels.append(inplanes)
self.stages.append(nn.Sequential(*block_list))
block_list = []
block_list.append(
ResidualUnit(
in_channels=inplanes,
mid_channels=make_divisible(scale * exp),
out_channels=make_divisible(scale * c),
kernel_size=k,
stride=s,
use_se=se,
act=nl,
name="conv" + str(i + 2)))
inplanes = make_divisible(scale * c)
i += 1
block_list.append(
ConvBNLayer(
in_channels=inplanes,
out_channels=make_divisible(scale * cls_ch_squeeze),
kernel_size=1,
stride=1,
padding=0,
groups=1,
if_act=True,
act='hardswish',
name='conv_last'))
self.stages.append(nn.Sequential(*block_list))
self.out_channels.append(make_divisible(scale * cls_ch_squeeze))
# for i, stage in enumerate(self.stages):
# self.add_module(module=stage, name="stage{}".format(i))
def forward(self, x):
x = self.conv(x)
out_list = []
for stage in self.stages:
x = stage(x)
out_list.append(x)
return out_list
class ConvBNLayer(nn.Module):
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride,
padding,
groups=1,
if_act=True,
act=None,
name=None):
super(ConvBNLayer, self).__init__()
self.if_act = if_act
self.act = act
self.conv = nn.Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
groups=groups,
bias=False)
self.bn = nn.BatchNorm2d(
out_channels,
)
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
if self.if_act:
if self.act == "relu":
x = F.relu(x)
elif self.act == "hardswish":
x = hard_swish(x)
else:
print("The activation function({}) is selected incorrectly.".
format(self.act))
exit()
return x
class ResidualUnit(nn.Module):
def __init__(self,
in_channels,
mid_channels,
out_channels,
kernel_size,
stride,
use_se,
act=None,
name=''):
super(ResidualUnit, self).__init__()
self.if_shortcut = stride == 1 and in_channels == out_channels
self.if_se = use_se
self.expand_conv = ConvBNLayer(
in_channels=in_channels,
out_channels=mid_channels,
kernel_size=1,
stride=1,
padding=0,
if_act=True,
act=act,
name=name + "_expand")
self.bottleneck_conv = ConvBNLayer(
in_channels=mid_channels,
out_channels=mid_channels,
kernel_size=kernel_size,
stride=stride,
padding=int((kernel_size - 1) // 2),
groups=mid_channels,
if_act=True,
act=act,
name=name + "_depthwise")
if self.if_se:
self.mid_se = SEModule(mid_channels, name=name + "_se")
self.linear_conv = ConvBNLayer(
in_channels=mid_channels,
out_channels=out_channels,
kernel_size=1,
stride=1,
padding=0,
if_act=False,
act=None,
name=name + "_linear")
def forward(self, inputs):
x = self.expand_conv(inputs)
x = self.bottleneck_conv(x)
if self.if_se:
x = self.mid_se(x)
x = self.linear_conv(x)
if self.if_shortcut:
x = torch.add(inputs, x)
return x
class SEModule(nn.Module):
def __init__(self, in_channels, reduction=4, name=""):
super(SEModule, self).__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.conv1 = nn.Conv2d(
in_channels=in_channels,
out_channels=in_channels // reduction,
kernel_size=1,
stride=1,
padding=0,
bias=True)
self.conv2 = nn.Conv2d(
in_channels=in_channels // reduction,
out_channels=in_channels,
kernel_size=1,
stride=1,
padding=0,
bias=True)
def forward(self, inputs):
outputs = self.avg_pool(inputs)
outputs = self.conv1(outputs)
outputs = F.relu(outputs)
outputs = self.conv2(outputs)
outputs = hard_sigmoid(outputs, slope=0.2, offset=0.5)
return inputs * outputs
\ No newline at end of file
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/backbones/table_resnet_vd.py deleted 100755 → 0
View file @ 3cb156f5
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import torch
import torch.nn as nn
import torch.nn.functional as F
from pytorchocr.modeling.common import Activation
__all__ = ["ResNet"]
class ConvBNLayer(nn.Module):
def __init__(
self,
in_channels,
out_channels,
kernel_size,
stride=1,
groups=1,
is_vd_mode=False,
act=None,
name=None, ):
super(ConvBNLayer, self).__init__()
self.is_vd_mode = is_vd_mode
self._pool2d_avg = nn.AvgPool2d(
kernel_size=2, stride=2, padding=0, ceil_mode=True)
self._conv = nn.Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=(kernel_size - 1) // 2,
groups=groups,
bias=False)
if name == "conv1":
bn_name = "bn_" + name
else:
bn_name = "bn" + name[3:]
self._batch_norm = nn.BatchNorm2d(
out_channels,
)
self.act = act
if self.act is not None:
self._act = Activation(act, inplace=True)
def forward(self, inputs):
if self.is_vd_mode:
inputs = self._pool2d_avg(inputs)
y = self._conv(inputs)
y = self._batch_norm(y)
if self.act is not None:
y = self._act(y)
return y
class BottleneckBlock(nn.Module):
def __init__(self,
in_channels,
out_channels,
stride,
shortcut=True,
if_first=False,
name=None):
super(BottleneckBlock, self).__init__()
self.conv0 = ConvBNLayer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
act='relu',
name=name + "_branch2a")
self.conv1 = ConvBNLayer(
in_channels=out_channels,
out_channels=out_channels,
kernel_size=3,
stride=stride,
act='relu',
name=name + "_branch2b")
self.conv2 = ConvBNLayer(
in_channels=out_channels,
out_channels=out_channels * 4,
kernel_size=1,
act=None,
name=name + "_branch2c")
if not shortcut:
self.short = ConvBNLayer(
in_channels=in_channels,
out_channels=out_channels * 4,
kernel_size=1,
stride=1,
is_vd_mode=False if if_first else True,
name=name + "_branch1")
self.shortcut = shortcut
def forward(self, inputs):
y = self.conv0(inputs)
conv1 = self.conv1(y)
conv2 = self.conv2(conv1)
if self.shortcut:
short = inputs
else:
short = self.short(inputs)
y = torch.add(short, conv2)
y = F.relu(y)
return y
class BasicBlock(nn.Module):
def __init__(self,
in_channels,
out_channels,
stride,
shortcut=True,
if_first=False,
name=None):
super(BasicBlock, self).__init__()
self.stride = stride
self.conv0 = ConvBNLayer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=3,
stride=stride,
act='relu',
name=name + "_branch2a")
self.conv1 = ConvBNLayer(
in_channels=out_channels,
out_channels=out_channels,
kernel_size=3,
act=None,
name=name + "_branch2b")
if not shortcut:
self.short = ConvBNLayer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
stride=1,
is_vd_mode=False if if_first else True,
name=name + "_branch1")
self.shortcut = shortcut
def forward(self, inputs):
y = self.conv0(inputs)
conv1 = self.conv1(y)
if self.shortcut:
short = inputs
else:
short = self.short(inputs)
y = torch.add(short, conv1)
y = F.relu(y)
return y
class ResNet(nn.Module):
def __init__(self, in_channels=3, layers=50, **kwargs):
super(ResNet, self).__init__()
self.layers = layers
supported_layers = [18, 34, 50, 101, 152, 200]
assert layers in supported_layers, \
"supported layers are {} but input layer is {}".format(
supported_layers, layers)
if layers == 18:
depth = [2, 2, 2, 2]
elif layers == 34 or layers == 50:
depth = [3, 4, 6, 3]
elif layers == 101:
depth = [3, 4, 23, 3]
elif layers == 152:
depth = [3, 8, 36, 3]
elif layers == 200:
depth = [3, 12, 48, 3]
num_channels = [64, 256, 512,
1024] if layers >= 50 else [64, 64, 128, 256]
num_filters = [64, 128, 256, 512]
self.conv1_1 = ConvBNLayer(
in_channels=in_channels,
out_channels=32,
kernel_size=3,
stride=2,
act='relu',
name="conv1_1")
self.conv1_2 = ConvBNLayer(
in_channels=32,
out_channels=32,
kernel_size=3,
stride=1,
act='relu',
name="conv1_2")
self.conv1_3 = ConvBNLayer(
in_channels=32,
out_channels=64,
kernel_size=3,
stride=1,
act='relu',
name="conv1_3")
self.pool2d_max = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.stages = nn.ModuleList()
self.out_channels = []
if layers >= 50:
for block in range(len(depth)):
block_list = nn.Sequential()
shortcut = False
for i in range(depth[block]):
if layers in [101, 152] and block == 2:
if i == 0:
conv_name = "res" + str(block + 2) + "a"
else:
conv_name = "res" + str(block + 2) + "b" + str(i)
else:
conv_name = "res" + str(block + 2) + chr(97 + i)
bottleneck_block = BottleneckBlock(
in_channels=num_channels[block]
if i == 0 else num_filters[block] * 4,
out_channels=num_filters[block],
stride=2 if i == 0 and block != 0 else 1,
shortcut=shortcut,
if_first=block == i == 0,
name=conv_name
)
shortcut = True
# block_list.append(bottleneck_block)
block_list.add_module('bb_%d_%d' % (block, i), bottleneck_block)
self.out_channels.append(num_filters[block] * 4)
# self.stages.append(nn.Sequential(*block_list))
self.stages.append(block_list)
else:
for block in range(len(depth)):
block_list = nn.Sequential()
shortcut = False
for i in range(depth[block]):
conv_name = "res" + str(block + 2) + chr(97 + i)
basic_block = BasicBlock(
in_channels=num_channels[block]
if i == 0 else num_filters[block],
out_channels=num_filters[block],
stride=2 if i == 0 and block != 0 else 1,
shortcut=shortcut,
if_first=block == i == 0,
name=conv_name
)
shortcut = True
# block_list.append(basic_block)
block_list.add_module('bb_%d_%d' % (block, i), basic_block)
self.out_channels.append(num_filters[block])
# self.stages.append(nn.Sequential(*block_list))
self.stages.append(block_list)
def forward(self, inputs):
y = self.conv1_1(inputs)
y = self.conv1_2(y)
y = self.conv1_3(y)
y = self.pool2d_max(y)
out = []
for block in self.stages:
y = block(y)
out.append(y)
return out
\ No newline at end of file
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/common.py 100755 → 100644
View file @ b3d6785d
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import nn
class Hswish(nn.Module):
def __init__(self, inplace=True):
......@@ -10,7 +9,8 @@ class Hswish(nn.Module):
self.inplace = inplace
def forward(self, x):
return x * F.relu6(x + 3., inplace=self.inplace) / 6.
return x * F.relu6(x + 3.0, inplace=self.inplace) / 6.0
# out = max(0, min(1, slop*x+offset))
# paddle.fluid.layers.hard_sigmoid(x, slope=0.2, offset=0.5, name=None)
......@@ -22,7 +22,8 @@ class Hsigmoid(nn.Module):
def forward(self, x):
# torch: F.relu6(x + 3., inplace=self.inplace) / 6.
# paddle: F.relu6(1.2 * x + 3., inplace=self.inplace) / 6.
return F.relu6(1.2 * x + 3., inplace=self.inplace) / 6.
return F.relu6(1.2 * x + 3.0, inplace=self.inplace) / 6.0
class GELU(nn.Module):
def __init__(self, inplace=True):
......@@ -43,31 +44,33 @@ class Swish(nn.Module):
x.mul_(torch.sigmoid(x))
return x
else:
return x*torch.sigmoid(x)
return x * torch.sigmoid(x)
class Activation(nn.Module):
def __init__(self, act_type, inplace=True):
super(Activation, self).__init__()
act_type = act_type.lower()
if act_type == 'relu':
if act_type == "relu":
self.act = nn.ReLU(inplace=inplace)
elif act_type == 'relu6':
elif act_type == "relu6":
self.act = nn.ReLU6(inplace=inplace)
elif act_type == 'sigmoid':
elif act_type == "sigmoid":
raise NotImplementedError
elif act_type == 'hard_sigmoid':
self.act = Hsigmoid(inplace)#nn.Hardsigmoid(inplace=inplace)#Hsigmoid(inplace)#
elif act_type == 'hard_swish' or act_type == 'hswish':
elif act_type == "hard_sigmoid":
self.act = Hsigmoid(
inplace
) # nn.Hardsigmoid(inplace=inplace)#Hsigmoid(inplace)#
elif act_type == "hard_swish" or act_type == "hswish":
self.act = Hswish(inplace=inplace)
elif act_type == 'leakyrelu':
elif act_type == "leakyrelu":
self.act = nn.LeakyReLU(inplace=inplace)
elif act_type == 'gelu':
elif act_type == "gelu":
self.act = GELU(inplace=inplace)
elif act_type == 'swish':
elif act_type == "swish":
self.act = Swish(inplace=inplace)
else:
raise NotImplementedError
def forward(self, inputs):
return self.act(inputs)
\ No newline at end of file
return self.act(inputs)
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/heads/__init__.py 100755 → 100644
View file @ b3d6785d
......@@ -12,40 +12,32 @@
# See the License for the specific language governing permissions and
# limitations under the License.
__all__ = ['build_head']
__all__ = ["build_head"]
def build_head(config, **kwargs):
# det head
from .det_db_head import DBHead, PFHeadLocal
from .det_east_head import EASTHead
from .det_sast_head import SASTHead
from .det_pse_head import PSEHead
from .det_fce_head import FCEHead
from .e2e_pg_head import PGHead
# rec head
from .rec_ctc_head import CTCHead
from .rec_att_head import AttentionHead
from .rec_srn_head import SRNHead
from .rec_nrtr_head import Transformer
from .rec_sar_head import SARHead
from .rec_can_head import CANHead
from .rec_multi_head import MultiHead
# cls head
from .cls_head import ClsHead
support_dict = [
'DBHead', 'PSEHead', 'EASTHead', 'SASTHead', 'CTCHead', 'ClsHead', 'AttentionHead',
'SRNHead', 'PGHead', 'Transformer', 'TableAttentionHead','SARHead', 'FCEHead',
'CANHead', 'MultiHead', 'PFHeadLocal',
support_dict = [
"DBHead",
"CTCHead",
"ClsHead",
"MultiHead",
"PFHeadLocal",
]
from .table_att_head import TableAttentionHead
module_name = config.pop('name')
assert module_name in support_dict, Exception('head only support {}'.format(
support_dict))
module_name = config.pop("name")
char_num = config.pop("char_num", 6625)
assert module_name in support_dict, Exception(
"head only support {}".format(support_dict)
)
module_class = eval(module_name)(**config, **kwargs)
return module_class
\ No newline at end of file
return module_class
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/heads/cls_head.py 100755 → 100644
View file @ b3d6785d
import os, sys
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import nn
class ClsHead(nn.Module):
"""
......@@ -12,17 +12,12 @@ class ClsHead(nn.Module):
def __init__(self, in_channels, class_dim, **kwargs):
super(ClsHead, self).__init__()
self.training = False
self.pool = nn.AdaptiveAvgPool2d(1)
self.fc = nn.Linear(
in_channels,
class_dim,
bias=True)
self.fc = nn.Linear(in_channels, class_dim, bias=True)
def forward(self, x):
x = self.pool(x)
x = torch.reshape(x, shape=[x.shape[0], x.shape[1]])
x = self.fc(x)
if not self.training:
x = F.softmax(x, dim=1)
return x
\ No newline at end of file
x = F.softmax(x, dim=1)
return x
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/heads/det_db_head.py 100755 → 100644
View file @ b3d6785d
import os, sys
import torch
import torch.nn as nn
import torch.nn.functional as F
from pytorchocr.modeling.common import Activation
from pytorchocr.modeling.backbones.det_mobilenet_v3 import ConvBNLayer
from ..common import Activation
from ..backbones.det_mobilenet_v3 import ConvBNLayer
class Head(nn.Module):
def __init__(self, in_channels, **kwargs):
......@@ -76,13 +75,8 @@ class DBHead(nn.Module):
def forward(self, x):
shrink_maps = self.binarize(x)
if not self.training:
return {'maps': shrink_maps}
return {'maps': shrink_maps}
threshold_maps = self.thresh(x)
binary_maps = self.step_function(shrink_maps, threshold_maps)
y = torch.cat([shrink_maps, threshold_maps, binary_maps], dim=1)
return {'maps': y}
class LocalModule(nn.Module):
def __init__(self, in_c, mid_c, use_distance=True):
......@@ -101,7 +95,7 @@ class PFHeadLocal(DBHead):
super(PFHeadLocal, self).__init__(in_channels, k, **kwargs)
self.mode = mode
self.up_conv = nn.interpolate(scale_factor=2, mode="nearest")
self.up_conv = nn.Upsample(scale_factor=2, mode="nearest")
if self.mode == 'large':
self.cbn_layer = LocalModule(in_channels // 4, in_channels // 4)
elif self.mode == 'small':
......@@ -112,10 +106,4 @@ class PFHeadLocal(DBHead):
base_maps = shrink_maps
cbn_maps = self.cbn_layer(self.up_conv(f), shrink_maps, None)
cbn_maps = F.sigmoid(cbn_maps)
if not self.training:
return {'maps': 0.5 * (base_maps + cbn_maps), 'cbn_maps': cbn_maps}
threshold_maps = self.thresh(x)
binary_maps = self.step_function(shrink_maps, threshold_maps)
y = torch.cat([cbn_maps, threshold_maps, binary_maps], dim=1)
return {'maps': y, 'distance_maps': cbn_maps, 'cbn_maps': binary_maps}
\ No newline at end of file
return {'maps': 0.5 * (base_maps + cbn_maps), 'cbn_maps': cbn_maps}
\ No newline at end of file
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/heads/det_east_head.py deleted 100755 → 0
View file @ 3cb156f5
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import math
import os, sys
import torch
import torch.nn as nn
import torch.nn.functional as F
from pytorchocr.modeling.common import Activation
# import paddle
# from paddle import nn
# import paddle.nn.functional as F
# from paddle import ParamAttr
class ConvBNLayer(nn.Module):
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride,
padding,
groups=1,
if_act=True,
act=None,
name=None):
super(ConvBNLayer, self).__init__()
self.if_act = if_act
self.act = act
self.conv = nn.Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
groups=groups,
bias=False)
self.bn = nn.BatchNorm2d(
out_channels,)
self.act = act
if act is not None:
self._act = Activation(act)
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
if self.act is not None:
x = self._act(x)
return x
class EASTHead(nn.Module):
"""
"""
def __init__(self, in_channels, model_name, **kwargs):
super(EASTHead, self).__init__()
self.model_name = model_name
if self.model_name == "large":
num_outputs = [128, 64, 1, 8]
else:
num_outputs = [64, 32, 1, 8]
self.det_conv1 = ConvBNLayer(
in_channels=in_channels,
out_channels=num_outputs[0],
kernel_size=3,
stride=1,
padding=1,
if_act=True,
act='relu',
name="det_head1")
self.det_conv2 = ConvBNLayer(
in_channels=num_outputs[0],
out_channels=num_outputs[1],
kernel_size=3,
stride=1,
padding=1,
if_act=True,
act='relu',
name="det_head2")
self.score_conv = ConvBNLayer(
in_channels=num_outputs[1],
out_channels=num_outputs[2],
kernel_size=1,
stride=1,
padding=0,
if_act=False,
act=None,
name="f_score")
self.geo_conv = ConvBNLayer(
in_channels=num_outputs[1],
out_channels=num_outputs[3],
kernel_size=1,
stride=1,
padding=0,
if_act=False,
act=None,
name="f_geo")
def forward(self, x):
f_det = self.det_conv1(x)
f_det = self.det_conv2(f_det)
f_score = self.score_conv(f_det)
f_score = torch.sigmoid(f_score)
f_geo = self.geo_conv(f_det)
f_geo = (torch.sigmoid(f_geo) - 0.5) * 2 * 800
pred = {'f_score': f_score, 'f_geo': f_geo}
return pred
\ No newline at end of file
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/heads/det_fce_head.py deleted 100755 → 0
View file @ 3cb156f5
"""
This code is refer from:
https://github.com/open-mmlab/mmocr/blob/main/mmocr/models/textdet/dense_heads/fce_head.py
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
# from paddle import nn
# from paddle import ParamAttr
# import paddle.nn.functional as F
# from paddle.nn.initializer import Normal
# import paddle
from functools import partial
def multi_apply(func, *args, **kwargs):
pfunc = partial(func, **kwargs) if kwargs else func
map_results = map(pfunc, *args)
return tuple(map(list, zip(*map_results)))
class FCEHead(nn.Module):
"""The class for implementing FCENet head.
FCENet(CVPR2021): Fourier Contour Embedding for Arbitrary-shaped Text
Detection.
[https://arxiv.org/abs/2104.10442]
Args:
in_channels (int): The number of input channels.
scales (list[int]) : The scale of each layer.
fourier_degree (int) : The maximum Fourier transform degree k.
"""
def __init__(self, in_channels, fourier_degree=5):
super().__init__()
assert isinstance(in_channels, int)
self.downsample_ratio = 1.0
self.in_channels = in_channels
self.fourier_degree = fourier_degree
self.out_channels_cls = 4
self.out_channels_reg = (2 * self.fourier_degree + 1) * 2
self.out_conv_cls = nn.Conv2d(
in_channels=self.in_channels,
out_channels=self.out_channels_cls,
kernel_size=3,
stride=1,
padding=1,
groups=1,
bias=True)
self.out_conv_reg = nn.Conv2d(
in_channels=self.in_channels,
out_channels=self.out_channels_reg,
kernel_size=3,
stride=1,
padding=1,
groups=1,
bias=True)
def forward(self, feats, targets=None):
cls_res, reg_res = multi_apply(self.forward_single, feats)
level_num = len(cls_res)
outs = {}
if not self.training:
for i in range(level_num):
tr_pred = F.softmax(cls_res[i][:, 0:2, :, :], dim=1)
tcl_pred = F.softmax(cls_res[i][:, 2:, :, :], dim=1)
outs['level_{}'.format(i)] = torch.cat(
[tr_pred, tcl_pred, reg_res[i]], dim=1)
else:
preds = [[cls_res[i], reg_res[i]] for i in range(level_num)]
outs['levels'] = preds
return outs
def forward_single(self, x):
cls_predict = self.out_conv_cls(x)
reg_predict = self.out_conv_reg(x)
return cls_predict, reg_predict
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/heads/det_pse_head.py deleted 100755 → 0
View file @ 3cb156f5
"""
This code is refer from:
https://github.com/whai362/PSENet/blob/python3/models/head/psenet_head.py
"""
# from paddle import nn
from torch import nn
class PSEHead(nn.Module):
def __init__(self, in_channels, hidden_dim=256, out_channels=7, **kwargs):
super(PSEHead, self).__init__()
self.conv1 = nn.Conv2d(
in_channels, hidden_dim, kernel_size=3, stride=1, padding=1)
self.bn1 = nn.BatchNorm2d(hidden_dim)
self.relu1 = nn.ReLU()
self.conv2 = nn.Conv2d(
hidden_dim, out_channels, kernel_size=1, stride=1, padding=0)
def forward(self, x, **kwargs):
out = self.conv1(x)
out = self.relu1(self.bn1(out))
out = self.conv2(out)
return {'maps': out}
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/heads/det_sast_head.py deleted 100755 → 0
View file @ 3cb156f5
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import math
import os, sys
import torch
import torch.nn as nn
import torch.nn.functional as F
from pytorchocr.modeling.common import Activation
# import paddle
# from paddle import nn
# import paddle.nn.functional as F
# from paddle import ParamAttr
class ConvBNLayer(nn.Module):
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride,
groups=1,
if_act=True,
act=None,
name=None):
super(ConvBNLayer, self).__init__()
self.if_act = if_act
self.conv = nn.Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=(kernel_size - 1) // 2,
groups=groups,
bias=False)
self.bn = nn.BatchNorm2d(
out_channels,)
self.act = act
if act is not None:
self._act = Activation(act)
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
if self.act is not None:
x = self._act(x)
return x
class SAST_Header1(nn.Module):
def __init__(self, in_channels, **kwargs):
super(SAST_Header1, self).__init__()
out_channels = [64, 64, 128]
self.score_conv = nn.Sequential(
ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_score1'),
ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_score2'),
ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_score3'),
ConvBNLayer(out_channels[2], 1, 3, 1, act=None, name='f_score4')
)
self.border_conv = nn.Sequential(
ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_border1'),
ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_border2'),
ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_border3'),
ConvBNLayer(out_channels[2], 4, 3, 1, act=None, name='f_border4')
)
def forward(self, x):
f_score = self.score_conv(x)
f_score = torch.sigmoid(f_score)
f_border = self.border_conv(x)
return f_score, f_border
class SAST_Header2(nn.Module):
def __init__(self, in_channels, **kwargs):
super(SAST_Header2, self).__init__()
out_channels = [64, 64, 128]
self.tvo_conv = nn.Sequential(
ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_tvo1'),
ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_tvo2'),
ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_tvo3'),
ConvBNLayer(out_channels[2], 8, 3, 1, act=None, name='f_tvo4')
)
self.tco_conv = nn.Sequential(
ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_tco1'),
ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_tco2'),
ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_tco3'),
ConvBNLayer(out_channels[2], 2, 3, 1, act=None, name='f_tco4')
)
def forward(self, x):
f_tvo = self.tvo_conv(x)
f_tco = self.tco_conv(x)
return f_tvo, f_tco
class SASTHead(nn.Module):
"""
"""
def __init__(self, in_channels, **kwargs):
super(SASTHead, self).__init__()
self.head1 = SAST_Header1(in_channels)
self.head2 = SAST_Header2(in_channels)
def forward(self, x):
f_score, f_border = self.head1(x)
f_tvo, f_tco = self.head2(x)
predicts = {}
predicts['f_score'] = f_score
predicts['f_border'] = f_border
predicts['f_tvo'] = f_tvo
predicts['f_tco'] = f_tco
return predicts
\ No newline at end of file
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/heads/e2e_pg_head.py deleted 100755 → 0
View file @ 3cb156f5
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from pytorchocr.modeling.common import Activation
class ConvBNLayer(nn.Module):
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride,
padding,
groups=1,
if_act=True,
act=None,
name=None):
super(ConvBNLayer, self).__init__()
self.if_act = if_act
self.act = act
self.conv = nn.Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
groups=groups,
bias=False)
self.bn = nn.BatchNorm2d(out_channels)
self.act = act
if self.act is not None:
self._act = Activation(act_type=self.act, inplace=True)
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
if self.act is not None:
x = self._act(x)
return x
class PGHead(nn.Module):
"""
"""
def __init__(self, in_channels, **kwargs):
super(PGHead, self).__init__()
self.conv_f_score1 = ConvBNLayer(
in_channels=in_channels,
out_channels=64,
kernel_size=1,
stride=1,
padding=0,
act='relu',
name="conv_f_score{}".format(1))
self.conv_f_score2 = ConvBNLayer(
in_channels=64,
out_channels=64,
kernel_size=3,
stride=1,
padding=1,
act='relu',
name="conv_f_score{}".format(2))
self.conv_f_score3 = ConvBNLayer(
in_channels=64,
out_channels=128,
kernel_size=1,
stride=1,
padding=0,
act='relu',
name="conv_f_score{}".format(3))
self.conv1 = nn.Conv2d(
in_channels=128,
out_channels=1,
kernel_size=3,
stride=1,
padding=1,
groups=1,
bias=False)
self.conv_f_boder1 = ConvBNLayer(
in_channels=in_channels,
out_channels=64,
kernel_size=1,
stride=1,
padding=0,
act='relu',
name="conv_f_boder{}".format(1))
self.conv_f_boder2 = ConvBNLayer(
in_channels=64,
out_channels=64,
kernel_size=3,
stride=1,
padding=1,
act='relu',
name="conv_f_boder{}".format(2))
self.conv_f_boder3 = ConvBNLayer(
in_channels=64,
out_channels=128,
kernel_size=1,
stride=1,
padding=0,
act='relu',
name="conv_f_boder{}".format(3))
self.conv2 = nn.Conv2d(
in_channels=128,
out_channels=4,
kernel_size=3,
stride=1,
padding=1,
groups=1,
bias=False)
self.conv_f_char1 = ConvBNLayer(
in_channels=in_channels,
out_channels=128,
kernel_size=1,
stride=1,
padding=0,
act='relu',
name="conv_f_char{}".format(1))
self.conv_f_char2 = ConvBNLayer(
in_channels=128,
out_channels=128,
kernel_size=3,
stride=1,
padding=1,
act='relu',
name="conv_f_char{}".format(2))
self.conv_f_char3 = ConvBNLayer(
in_channels=128,
out_channels=256,
kernel_size=1,
stride=1,
padding=0,
act='relu',
name="conv_f_char{}".format(3))
self.conv_f_char4 = ConvBNLayer(
in_channels=256,
out_channels=256,
kernel_size=3,
stride=1,
padding=1,
act='relu',
name="conv_f_char{}".format(4))
self.conv_f_char5 = ConvBNLayer(
in_channels=256,
out_channels=256,
kernel_size=1,
stride=1,
padding=0,
act='relu',
name="conv_f_char{}".format(5))
self.conv3 = nn.Conv2d(
in_channels=256,
out_channels=37,
kernel_size=3,
stride=1,
padding=1,
groups=1,
bias=False)
self.conv_f_direc1 = ConvBNLayer(
in_channels=in_channels,
out_channels=64,
kernel_size=1,
stride=1,
padding=0,
act='relu',
name="conv_f_direc{}".format(1))
self.conv_f_direc2 = ConvBNLayer(
in_channels=64,
out_channels=64,
kernel_size=3,
stride=1,
padding=1,
act='relu',
name="conv_f_direc{}".format(2))
self.conv_f_direc3 = ConvBNLayer(
in_channels=64,
out_channels=128,
kernel_size=1,
stride=1,
padding=0,
act='relu',
name="conv_f_direc{}".format(3))
self.conv4 = nn.Conv2d(
in_channels=128,
out_channels=2,
kernel_size=3,
stride=1,
padding=1,
groups=1,
bias=False)
def forward(self, x):
f_score = self.conv_f_score1(x)
f_score = self.conv_f_score2(f_score)
f_score = self.conv_f_score3(f_score)
f_score = self.conv1(f_score)
f_score = torch.sigmoid(f_score)
# f_border
f_border = self.conv_f_boder1(x)
f_border = self.conv_f_boder2(f_border)
f_border = self.conv_f_boder3(f_border)
f_border = self.conv2(f_border)
f_char = self.conv_f_char1(x)
f_char = self.conv_f_char2(f_char)
f_char = self.conv_f_char3(f_char)
f_char = self.conv_f_char4(f_char)
f_char = self.conv_f_char5(f_char)
f_char = self.conv3(f_char)
f_direction = self.conv_f_direc1(x)
f_direction = self.conv_f_direc2(f_direction)
f_direction = self.conv_f_direc3(f_direction)
f_direction = self.conv4(f_direction)
predicts = {}
predicts['f_score'] = f_score
predicts['f_border'] = f_border
predicts['f_char'] = f_char
predicts['f_direction'] = f_direction
return predicts
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/heads/multiheadAttention.py deleted 100755 → 0
View file @ 3cb156f5
import torch
from torch import nn
import torch.nn.functional as F
from torch.nn import Linear
from torch.nn.init import xavier_uniform_
class MultiheadAttention(nn.Module):
"""Allows the model to jointly attend to information
from different representation subspaces.
See reference: Attention Is All You Need
.. math::
\text{MultiHead}(Q, K, V) = \text{Concat}(head_1,\dots,head_h)W^O
\text{where} head_i = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)
Args:
embed_dim: total dimension of the model
num_heads: parallel attention layers, or heads
"""
def __init__(self,
embed_dim,
num_heads,
dropout=0.,
bias=True,
add_bias_kv=False,
add_zero_attn=False):
super(MultiheadAttention, self).__init__()
self.embed_dim = embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
self.scaling = self.head_dim**-0.5
self.out_proj = Linear(embed_dim, embed_dim, bias=bias)
self._reset_parameters()
self.conv1 = torch.nn.Conv2d(
in_channels=embed_dim, out_channels=embed_dim, kernel_size=(1, 1))
self.conv2 = torch.nn.Conv2d(
in_channels=embed_dim, out_channels=embed_dim, kernel_size=(1, 1))
self.conv3 = torch.nn.Conv2d(
in_channels=embed_dim, out_channels=embed_dim, kernel_size=(1, 1))
def _reset_parameters(self):
xavier_uniform_(self.out_proj.weight)
def forward(self,
query,
key,
value,
key_padding_mask=None,
incremental_state=None,
attn_mask=None):
"""
Inputs of forward function
query: [target length, batch size, embed dim]
key: [sequence length, batch size, embed dim]
value: [sequence length, batch size, embed dim]
key_padding_mask: if True, mask padding based on batch size
incremental_state: if provided, previous time steps are cashed
need_weights: output attn_output_weights
static_kv: key and value are static
Outputs of forward function
attn_output: [target length, batch size, embed dim]
attn_output_weights: [batch size, target length, sequence length]
"""
q_shape = query.shape
src_shape = key.shape
q = self._in_proj_q(query)
k = self._in_proj_k(key)
v = self._in_proj_v(value)
q *= self.scaling
# q = paddle.transpose(
# paddle.reshape(
# q, [q_shape[0], q_shape[1], self.num_heads, self.head_dim]),
# [1, 2, 0, 3])
q = torch.reshape(q, (q_shape[0], q_shape[1], self.num_heads, self.head_dim))
q = q.permute(1, 2, 0, 3)
# k = paddle.transpose(
# paddle.reshape(
# k, [src_shape[0], q_shape[1], self.num_heads, self.head_dim]),
# [1, 2, 0, 3])
k = torch.reshape(k, (src_shape[0], q_shape[1], self.num_heads, self.head_dim))
k = k.permute(1, 2, 0, 3)
# v = paddle.transpose(
# paddle.reshape(
# v, [src_shape[0], q_shape[1], self.num_heads, self.head_dim]),
# [1, 2, 0, 3])
v = torch.reshape(v, (src_shape[0], q_shape[1], self.num_heads, self.head_dim))
v = v.permute(1, 2, 0, 3)
if key_padding_mask is not None:
assert key_padding_mask.shape[0] == q_shape[1]
assert key_padding_mask.shape[1] == src_shape[0]
attn_output_weights = torch.matmul(q,
k.permute(0, 1, 3, 2))
if attn_mask is not None:
attn_mask = torch.unsqueeze(torch.unsqueeze(attn_mask, 0), 0)
attn_output_weights += attn_mask
if key_padding_mask is not None:
attn_output_weights = torch.reshape(
attn_output_weights,
[q_shape[1], self.num_heads, q_shape[0], src_shape[0]])
key = torch.unsqueeze(torch.unsqueeze(key_padding_mask, 1), 2)
key = key.type(torch.float32)
y = torch.full(
size=key.shape, fill_value=float("-Inf"), dtype=torch.float32)
y = torch.where(key == 0., key, y)
attn_output_weights += y
attn_output_weights = F.softmax(
attn_output_weights.type(torch.float32),
dim=-1,
dtype=torch.float32 if attn_output_weights.dtype == torch.float16
else attn_output_weights.dtype)
attn_output_weights = F.dropout(
attn_output_weights, p=self.dropout, training=self.training)
attn_output = torch.matmul(attn_output_weights, v)
attn_output = torch.reshape(
attn_output.permute(2, 0, 1, 3),
[q_shape[0], q_shape[1], self.embed_dim])
attn_output = self.out_proj(attn_output)
return attn_output
def _in_proj_q(self, query):
query = query.permute(1, 2, 0)
query = torch.unsqueeze(query, dim=2)
res = self.conv1(query)
res = torch.squeeze(res, dim=2)
res = res.permute(2, 0, 1)
return res
def _in_proj_k(self, key):
key = key.permute(1, 2, 0)
key = torch.unsqueeze(key, dim=2)
res = self.conv2(key)
res = torch.squeeze(res, dim=2)
res = res.permute(2, 0, 1)
return res
def _in_proj_v(self, value):
value = value.permute(1, 2, 0) #(1, 2, 0)
value = torch.unsqueeze(value, dim=2)
res = self.conv3(value)
res = torch.squeeze(res, dim=2)
res = res.permute(2, 0, 1)
return res
magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/heads/rec_att_head.py deleted 100755 → 0
View file @ 3cb156f5
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from pytorchocr.modeling.common import Activation
class AttentionHead(nn.Module):
def __init__(self, in_channels, out_channels, hidden_size, **kwargs):
super(AttentionHead, self).__init__()
self.input_size = in_channels
self.hidden_size = hidden_size
self.num_classes = out_channels
self.attention_cell = AttentionGRUCell(
in_channels, hidden_size, out_channels, use_gru=False)
self.generator = nn.Linear(hidden_size, out_channels)
def _char_to_onehot(self, input_char, onehot_dim):
input_ont_hot = F.one_hot(input_char.type(torch.int64), onehot_dim)
return input_ont_hot
def forward(self, inputs, targets=None, batch_max_length=25):
batch_size = inputs.size()[0]
num_steps = batch_max_length
hidden = torch.zeros((batch_size, self.hidden_size))
output_hiddens = []
if targets is not None:
for i in range(num_steps):
char_onehots = self._char_to_onehot(
targets[:, i], onehot_dim=self.num_classes)
(outputs, hidden), alpha = self.attention_cell(hidden, inputs,
char_onehots)
output_hiddens.append(torch.unsqueeze(outputs, dim=1))
output = torch.cat(output_hiddens, dim=1)
probs = self.generator(output)
else:
targets = torch.zeros([batch_size], dtype=torch.int32)
probs = None
char_onehots = None
outputs = None
alpha = None
for i in range(num_steps):
char_onehots = self._char_to_onehot(
targets, onehot_dim=self.num_classes)
(outputs, hidden), alpha = self.attention_cell(hidden, inputs,
char_onehots)
probs_step = self.generator(outputs)
if probs is None:
probs = torch.unsqueeze(probs_step, dim=1)
else:
probs = torch.cat(
[probs, torch.unsqueeze(
probs_step, dim=1)], dim=1)
next_input = probs_step.argmax(dim=1)
targets = next_input
return probs
class AttentionGRUCell(nn.Module):
def __init__(self, input_size, hidden_size, num_embeddings, use_gru=False):
super(AttentionGRUCell, self).__init__()
self.i2h = nn.Linear(input_size, hidden_size, bias=False)
self.h2h = nn.Linear(hidden_size, hidden_size)
self.score = nn.Linear(hidden_size, 1, bias=False)
self.rnn = nn.GRUCell(
input_size=input_size + num_embeddings, hidden_size=hidden_size, bias=True)
self.hidden_size = hidden_size
def forward(self, prev_hidden, batch_H, char_onehots):
batch_H_proj = self.i2h(batch_H)
prev_hidden_proj = torch.unsqueeze(self.h2h(prev_hidden), dim=1)
res = torch.add(batch_H_proj, prev_hidden_proj)
res = torch.tanh(res)
e = self.score(res)
alpha = F.softmax(e, dim=1)
alpha = alpha.permute(0, 2, 1)
context = torch.squeeze(torch.matmul(alpha, batch_H), dim=1)
concat_context = torch.cat([context, char_onehots.float()], 1)
cur_hidden = self.rnn(concat_context, prev_hidden)
return (cur_hidden, cur_hidden), alpha
class AttentionLSTM(nn.Module):
def __init__(self, in_channels, out_channels, hidden_size, **kwargs):
super(AttentionLSTM, self).__init__()
self.input_size = in_channels
self.hidden_size = hidden_size
self.num_classes = out_channels
self.attention_cell = AttentionLSTMCell(
in_channels, hidden_size, out_channels, use_gru=False)
self.generator = nn.Linear(hidden_size, out_channels)
def _char_to_onehot(self, input_char, onehot_dim):
input_ont_hot = F.one_hot(input_char.type(torch.int64), onehot_dim)
return input_ont_hot
def forward(self, inputs, targets=None, batch_max_length=25):
batch_size = inputs.shape[0]
num_steps = batch_max_length
hidden = (torch.zeros((batch_size, self.hidden_size)), torch.zeros(
(batch_size, self.hidden_size)))
output_hiddens = []
if targets is not None:
for i in range(num_steps):
# one-hot vectors for a i-th char
char_onehots = self._char_to_onehot(
targets[:, i], onehot_dim=self.num_classes)
hidden, alpha = self.attention_cell(hidden, inputs,
char_onehots)
hidden = (hidden[1][0], hidden[1][1])
output_hiddens.append(torch.unsqueeze(hidden[0], dim=1))
output = torch.cat(output_hiddens, dim=1)
probs = self.generator(output)
else:
targets = torch.zeros([batch_size], dtype=torch.int32)
probs = None
for i in range(num_steps):
char_onehots = self._char_to_onehot(
targets, onehot_dim=self.num_classes)
hidden, alpha = self.attention_cell(hidden, inputs,
char_onehots)
probs_step = self.generator(hidden[0])
hidden = (hidden[1][0], hidden[1][1])
if probs is None:
probs = torch.unsqueeze(probs_step, dim=1)
else:
probs = torch.cat(
[probs, torch.unsqueeze(
probs_step, dim=1)], dim=1)
next_input = probs_step.argmax(dim=1)
targets = next_input
return probs
class AttentionLSTMCell(nn.Module):
def __init__(self, input_size, hidden_size, num_embeddings, use_gru=False):
super(AttentionLSTMCell, self).__init__()
self.i2h = nn.Linear(input_size, hidden_size, bias=False)
self.h2h = nn.Linear(hidden_size, hidden_size)
self.score = nn.Linear(hidden_size, 1, bias=False)
if not use_gru:
self.rnn = nn.LSTMCell(
input_size=input_size + num_embeddings, hidden_size=hidden_size)
else:
self.rnn = nn.GRUCell(
input_size=input_size + num_embeddings, hidden_size=hidden_size)
self.hidden_size = hidden_size
def forward(self, prev_hidden, batch_H, char_onehots):
batch_H_proj = self.i2h(batch_H)
prev_hidden_proj = torch.unsqueeze(self.h2h(prev_hidden[0]), dim=1)
res = torch.add(batch_H_proj, prev_hidden_proj)
res = torch.tanh(res)
e = self.score(res)
alpha = F.softmax(e, dim=1)
alpha = alpha.permute(0, 2, 1)
context = torch.squeeze(torch.matmul(alpha, batch_H), dim=1)
concat_context = torch.cat([context, char_onehots.float()], 1)
cur_hidden = self.rnn(concat_context, prev_hidden)
return cur_hidden, alpha
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment