MissPenguin
5 years ago
committed by
GitHub
commit
2852335e59
@ -0,0 +1,49 @@
|
||||
Global:
|
||||
algorithm: SRN
|
||||
use_gpu: true
|
||||
epoch_num: 72
|
||||
log_smooth_window: 20
|
||||
print_batch_step: 10
|
||||
save_model_dir: output/rec_pvam_withrotate
|
||||
save_epoch_step: 1
|
||||
eval_batch_step: 8000
|
||||
train_batch_size_per_card: 64
|
||||
test_batch_size_per_card: 1
|
||||
image_shape: [1, 64, 256]
|
||||
max_text_length: 25
|
||||
character_type: en
|
||||
loss_type: srn
|
||||
num_heads: 8
|
||||
average_window: 0.15
|
||||
max_average_window: 15625
|
||||
min_average_window: 10000
|
||||
reader_yml: ./configs/rec/rec_benchmark_reader.yml
|
||||
pretrain_weights:
|
||||
checkpoints:
|
||||
save_inference_dir:
|
||||
infer_img:
|
||||
|
||||
Architecture:
|
||||
function: ppocr.modeling.architectures.rec_model,RecModel
|
||||
|
||||
Backbone:
|
||||
function: ppocr.modeling.backbones.rec_resnet50_fpn,ResNet
|
||||
layers: 50
|
||||
|
||||
Head:
|
||||
function: ppocr.modeling.heads.rec_srn_all_head,SRNPredict
|
||||
encoder_type: rnn
|
||||
num_encoder_TUs: 2
|
||||
num_decoder_TUs: 4
|
||||
hidden_dims: 512
|
||||
SeqRNN:
|
||||
hidden_size: 256
|
||||
|
||||
Loss:
|
||||
function: ppocr.modeling.losses.rec_srn_loss,SRNLoss
|
||||
|
||||
Optimizer:
|
||||
function: ppocr.optimizer,AdamDecay
|
||||
base_lr: 0.0001
|
||||
beta1: 0.9
|
||||
beta2: 0.999
|
||||
@ -0,0 +1,172 @@
|
||||
#copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
|
||||
#
|
||||
#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.
|
||||
|
||||
from __future__ import absolute_import
|
||||
from __future__ import division
|
||||
from __future__ import print_function
|
||||
|
||||
import math
|
||||
|
||||
import paddle
|
||||
import paddle.fluid as fluid
|
||||
from paddle.fluid.param_attr import ParamAttr
|
||||
|
||||
|
||||
__all__ = ["ResNet", "ResNet18", "ResNet34", "ResNet50", "ResNet101", "ResNet152"]
|
||||
|
||||
Trainable = True
|
||||
w_nolr = fluid.ParamAttr(
|
||||
trainable = Trainable)
|
||||
train_parameters = {
|
||||
"input_size": [3, 224, 224],
|
||||
"input_mean": [0.485, 0.456, 0.406],
|
||||
"input_std": [0.229, 0.224, 0.225],
|
||||
"learning_strategy": {
|
||||
"name": "piecewise_decay",
|
||||
"batch_size": 256,
|
||||
"epochs": [30, 60, 90],
|
||||
"steps": [0.1, 0.01, 0.001, 0.0001]
|
||||
}
|
||||
}
|
||||
|
||||
class ResNet():
|
||||
def __init__(self, params):
|
||||
self.layers = params['layers']
|
||||
self.params = train_parameters
|
||||
|
||||
|
||||
def __call__(self, input):
|
||||
layers = self.layers
|
||||
supported_layers = [18, 34, 50, 101, 152]
|
||||
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]
|
||||
stride_list = [(2,2),(2,2),(1,1),(1,1)]
|
||||
num_filters = [64, 128, 256, 512]
|
||||
|
||||
conv = self.conv_bn_layer(
|
||||
input=input, num_filters=64, filter_size=7, stride=2, act='relu', name="conv1")
|
||||
F = []
|
||||
if layers >= 50:
|
||||
for block in range(len(depth)):
|
||||
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)
|
||||
conv = self.bottleneck_block(
|
||||
input=conv,
|
||||
num_filters=num_filters[block],
|
||||
stride=stride_list[block] if i == 0 else 1, name=conv_name)
|
||||
F.append(conv)
|
||||
|
||||
base = F[-1]
|
||||
for i in [-2, -3]:
|
||||
b, c, w, h = F[i].shape
|
||||
if (w,h) == base.shape[2:]:
|
||||
base = base
|
||||
else:
|
||||
base = fluid.layers.conv2d_transpose( input=base, num_filters=c,filter_size=4, stride=2,
|
||||
padding=1,act=None,
|
||||
param_attr=w_nolr,
|
||||
bias_attr=w_nolr)
|
||||
base = fluid.layers.batch_norm(base, act = "relu", param_attr=w_nolr, bias_attr=w_nolr)
|
||||
base = fluid.layers.concat([base, F[i]], axis=1)
|
||||
base = fluid.layers.conv2d(base, num_filters=c, filter_size=1, param_attr=w_nolr, bias_attr=w_nolr)
|
||||
base = fluid.layers.conv2d(base, num_filters=c, filter_size=3,padding = 1, param_attr=w_nolr, bias_attr=w_nolr)
|
||||
base = fluid.layers.batch_norm(base, act = "relu", param_attr=w_nolr, bias_attr=w_nolr)
|
||||
|
||||
base = fluid.layers.conv2d(base, num_filters=512, filter_size=1,bias_attr=w_nolr,param_attr=w_nolr)
|
||||
|
||||
return base
|
||||
|
||||
def conv_bn_layer(self,
|
||||
input,
|
||||
num_filters,
|
||||
filter_size,
|
||||
stride=1,
|
||||
groups=1,
|
||||
act=None,
|
||||
name=None):
|
||||
conv = fluid.layers.conv2d(
|
||||
input=input,
|
||||
num_filters=num_filters,
|
||||
filter_size= 2 if stride==(1,1) else filter_size,
|
||||
dilation = 2 if stride==(1,1) else 1,
|
||||
stride=stride,
|
||||
padding=(filter_size - 1) // 2,
|
||||
groups=groups,
|
||||
act=None,
|
||||
param_attr=ParamAttr(name=name + "_weights",trainable = Trainable),
|
||||
bias_attr=False,
|
||||
name=name + '.conv2d.output.1')
|
||||
|
||||
if name == "conv1":
|
||||
bn_name = "bn_" + name
|
||||
else:
|
||||
bn_name = "bn" + name[3:]
|
||||
return fluid.layers.batch_norm(input=conv,
|
||||
act=act,
|
||||
name=bn_name + '.output.1',
|
||||
param_attr=ParamAttr(name=bn_name + '_scale',trainable = Trainable),
|
||||
bias_attr=ParamAttr(bn_name + '_offset',trainable = Trainable),
|
||||
moving_mean_name=bn_name + '_mean',
|
||||
moving_variance_name=bn_name + '_variance', )
|
||||
|
||||
def shortcut(self, input, ch_out, stride, is_first, name):
|
||||
ch_in = input.shape[1]
|
||||
if ch_in != ch_out or stride != 1 or is_first == True:
|
||||
if stride == (1,1):
|
||||
return self.conv_bn_layer(input, ch_out, 1, 1, name=name)
|
||||
else: #stride == (2,2)
|
||||
return self.conv_bn_layer(input, ch_out, 1, stride, name=name)
|
||||
|
||||
else:
|
||||
return input
|
||||
|
||||
def bottleneck_block(self, input, num_filters, stride, name):
|
||||
conv0 = self.conv_bn_layer(
|
||||
input=input, num_filters=num_filters, filter_size=1, act='relu', name=name + "_branch2a")
|
||||
conv1 = self.conv_bn_layer(
|
||||
input=conv0,
|
||||
num_filters=num_filters,
|
||||
filter_size=3,
|
||||
stride=stride,
|
||||
act='relu',
|
||||
name=name + "_branch2b")
|
||||
conv2 = self.conv_bn_layer(
|
||||
input=conv1, num_filters=num_filters * 4, filter_size=1, act=None, name=name + "_branch2c")
|
||||
|
||||
short = self.shortcut(input, num_filters * 4, stride, is_first=False, name=name + "_branch1")
|
||||
|
||||
return fluid.layers.elementwise_add(x=short, y=conv2, act='relu', name=name + ".add.output.5")
|
||||
|
||||
def basic_block(self, input, num_filters, stride, is_first, name):
|
||||
conv0 = self.conv_bn_layer(input=input, num_filters=num_filters, filter_size=3, act='relu', stride=stride,
|
||||
name=name + "_branch2a")
|
||||
conv1 = self.conv_bn_layer(input=conv0, num_filters=num_filters, filter_size=3, act=None,
|
||||
name=name + "_branch2b")
|
||||
short = self.shortcut(input, num_filters, stride, is_first, name=name + "_branch1")
|
||||
return fluid.layers.elementwise_add(x=short, y=conv1, act='relu')
|
||||
@ -0,0 +1,230 @@
|
||||
#copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
|
||||
#
|
||||
#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.
|
||||
|
||||
from __future__ import absolute_import
|
||||
from __future__ import division
|
||||
from __future__ import print_function
|
||||
|
||||
import math
|
||||
|
||||
import paddle
|
||||
import paddle.fluid as fluid
|
||||
from paddle.fluid.param_attr import ParamAttr
|
||||
import numpy as np
|
||||
from .self_attention.model import wrap_encoder
|
||||
from .self_attention.model import wrap_encoder_forFeature
|
||||
gradient_clip = 10
|
||||
|
||||
|
||||
class SRNPredict(object):
|
||||
def __init__(self, params):
|
||||
super(SRNPredict, self).__init__()
|
||||
self.char_num = params['char_num']
|
||||
self.max_length = params['max_text_length']
|
||||
|
||||
self.num_heads = params['num_heads']
|
||||
self.num_encoder_TUs = params['num_encoder_TUs']
|
||||
self.num_decoder_TUs = params['num_decoder_TUs']
|
||||
self.hidden_dims = params['hidden_dims']
|
||||
|
||||
def pvam(self, inputs, others):
|
||||
|
||||
b, c, h, w = inputs.shape
|
||||
conv_features = fluid.layers.reshape(x=inputs, shape=[-1, c, h * w])
|
||||
conv_features = fluid.layers.transpose(x=conv_features, perm=[0, 2, 1])
|
||||
|
||||
#===== Transformer encoder =====
|
||||
b, t, c = conv_features.shape
|
||||
encoder_word_pos = others["encoder_word_pos"]
|
||||
gsrm_word_pos = others["gsrm_word_pos"]
|
||||
|
||||
enc_inputs = [conv_features, encoder_word_pos, None]
|
||||
word_features = wrap_encoder_forFeature(
|
||||
src_vocab_size=-1,
|
||||
max_length=t,
|
||||
n_layer=self.num_encoder_TUs,
|
||||
n_head=self.num_heads,
|
||||
d_key=int(self.hidden_dims / self.num_heads),
|
||||
d_value=int(self.hidden_dims / self.num_heads),
|
||||
d_model=self.hidden_dims,
|
||||
d_inner_hid=self.hidden_dims,
|
||||
prepostprocess_dropout=0.1,
|
||||
attention_dropout=0.1,
|
||||
relu_dropout=0.1,
|
||||
preprocess_cmd="n",
|
||||
postprocess_cmd="da",
|
||||
weight_sharing=True,
|
||||
enc_inputs=enc_inputs, )
|
||||
fluid.clip.set_gradient_clip(
|
||||
fluid.clip.GradientClipByValue(gradient_clip))
|
||||
|
||||
#===== Parallel Visual Attention Module =====
|
||||
b, t, c = word_features.shape
|
||||
|
||||
word_features = fluid.layers.fc(word_features, c, num_flatten_dims=2)
|
||||
word_features_ = fluid.layers.reshape(word_features, [-1, 1, t, c])
|
||||
word_features_ = fluid.layers.expand(word_features_,
|
||||
[1, self.max_length, 1, 1])
|
||||
word_pos_feature = fluid.layers.embedding(gsrm_word_pos,
|
||||
[self.max_length, c])
|
||||
word_pos_ = fluid.layers.reshape(word_pos_feature,
|
||||
[-1, self.max_length, 1, c])
|
||||
word_pos_ = fluid.layers.expand(word_pos_, [1, 1, t, 1])
|
||||
temp = fluid.layers.elementwise_add(
|
||||
word_features_, word_pos_, act='tanh')
|
||||
|
||||
attention_weight = fluid.layers.fc(input=temp,
|
||||
size=1,
|
||||
num_flatten_dims=3,
|
||||
bias_attr=False)
|
||||
attention_weight = fluid.layers.reshape(
|
||||
x=attention_weight, shape=[-1, self.max_length, t])
|
||||
attention_weight = fluid.layers.softmax(input=attention_weight, axis=-1)
|
||||
|
||||
pvam_features = fluid.layers.matmul(attention_weight,
|
||||
word_features) #[b, max_length, c]
|
||||
|
||||
return pvam_features
|
||||
|
||||
def gsrm(self, pvam_features, others):
|
||||
|
||||
#===== GSRM Visual-to-semantic embedding block =====
|
||||
b, t, c = pvam_features.shape
|
||||
word_out = fluid.layers.fc(
|
||||
input=fluid.layers.reshape(pvam_features, [-1, c]),
|
||||
size=self.char_num,
|
||||
act="softmax")
|
||||
#word_out.stop_gradient = True
|
||||
word_ids = fluid.layers.argmax(word_out, axis=1)
|
||||
word_ids.stop_gradient = True
|
||||
word_ids = fluid.layers.reshape(x=word_ids, shape=[-1, t, 1])
|
||||
|
||||
#===== GSRM Semantic reasoning block =====
|
||||
"""
|
||||
This module is achieved through bi-transformers,
|
||||
ngram_feature1 is the froward one, ngram_fetaure2 is the backward one
|
||||
"""
|
||||
pad_idx = self.char_num
|
||||
gsrm_word_pos = others["gsrm_word_pos"]
|
||||
gsrm_slf_attn_bias1 = others["gsrm_slf_attn_bias1"]
|
||||
gsrm_slf_attn_bias2 = others["gsrm_slf_attn_bias2"]
|
||||
|
||||
def prepare_bi(word_ids):
|
||||
"""
|
||||
prepare bi for gsrm
|
||||
word1 for forward; word2 for backward
|
||||
"""
|
||||
word1 = fluid.layers.cast(word_ids, "float32")
|
||||
word1 = fluid.layers.pad(word1, [0, 0, 1, 0, 0, 0],
|
||||
pad_value=1.0 * pad_idx)
|
||||
word1 = fluid.layers.cast(word1, "int64")
|
||||
word1 = word1[:, :-1, :]
|
||||
word2 = word_ids
|
||||
return word1, word2
|
||||
|
||||
word1, word2 = prepare_bi(word_ids)
|
||||
word1.stop_gradient = True
|
||||
word2.stop_gradient = True
|
||||
enc_inputs_1 = [word1, gsrm_word_pos, gsrm_slf_attn_bias1]
|
||||
enc_inputs_2 = [word2, gsrm_word_pos, gsrm_slf_attn_bias2]
|
||||
|
||||
gsrm_feature1 = wrap_encoder(
|
||||
src_vocab_size=self.char_num + 1,
|
||||
max_length=self.max_length,
|
||||
n_layer=self.num_decoder_TUs,
|
||||
n_head=self.num_heads,
|
||||
d_key=int(self.hidden_dims / self.num_heads),
|
||||
d_value=int(self.hidden_dims / self.num_heads),
|
||||
d_model=self.hidden_dims,
|
||||
d_inner_hid=self.hidden_dims,
|
||||
prepostprocess_dropout=0.1,
|
||||
attention_dropout=0.1,
|
||||
relu_dropout=0.1,
|
||||
preprocess_cmd="n",
|
||||
postprocess_cmd="da",
|
||||
weight_sharing=True,
|
||||
enc_inputs=enc_inputs_1, )
|
||||
gsrm_feature2 = wrap_encoder(
|
||||
src_vocab_size=self.char_num + 1,
|
||||
max_length=self.max_length,
|
||||
n_layer=self.num_decoder_TUs,
|
||||
n_head=self.num_heads,
|
||||
d_key=int(self.hidden_dims / self.num_heads),
|
||||
d_value=int(self.hidden_dims / self.num_heads),
|
||||
d_model=self.hidden_dims,
|
||||
d_inner_hid=self.hidden_dims,
|
||||
prepostprocess_dropout=0.1,
|
||||
attention_dropout=0.1,
|
||||
relu_dropout=0.1,
|
||||
preprocess_cmd="n",
|
||||
postprocess_cmd="da",
|
||||
weight_sharing=True,
|
||||
enc_inputs=enc_inputs_2, )
|
||||
gsrm_feature2 = fluid.layers.pad(gsrm_feature2, [0, 0, 0, 1, 0, 0],
|
||||
pad_value=0.)
|
||||
gsrm_feature2 = gsrm_feature2[:, 1:, ]
|
||||
gsrm_features = gsrm_feature1 + gsrm_feature2
|
||||
|
||||
b, t, c = gsrm_features.shape
|
||||
|
||||
gsrm_out = fluid.layers.matmul(
|
||||
x=gsrm_features,
|
||||
y=fluid.default_main_program().global_block().var(
|
||||
"src_word_emb_table"),
|
||||
transpose_y=True)
|
||||
b, t, c = gsrm_out.shape
|
||||
gsrm_out = fluid.layers.softmax(input=fluid.layers.reshape(gsrm_out,
|
||||
[-1, c]))
|
||||
|
||||
return gsrm_features, word_out, gsrm_out
|
||||
|
||||
def vsfd(self, pvam_features, gsrm_features):
|
||||
|
||||
#===== Visual-Semantic Fusion Decoder Module =====
|
||||
b, t, c1 = pvam_features.shape
|
||||
b, t, c2 = gsrm_features.shape
|
||||
combine_features_ = fluid.layers.concat(
|
||||
[pvam_features, gsrm_features], axis=2)
|
||||
img_comb_features_ = fluid.layers.reshape(
|
||||
x=combine_features_, shape=[-1, c1 + c2])
|
||||
img_comb_features_map = fluid.layers.fc(input=img_comb_features_,
|
||||
size=c1,
|
||||
act="sigmoid")
|
||||
img_comb_features_map = fluid.layers.reshape(
|
||||
x=img_comb_features_map, shape=[-1, t, c1])
|
||||
combine_features = img_comb_features_map * pvam_features + (
|
||||
1.0 - img_comb_features_map) * gsrm_features
|
||||
img_comb_features = fluid.layers.reshape(
|
||||
x=combine_features, shape=[-1, c1])
|
||||
|
||||
fc_out = fluid.layers.fc(input=img_comb_features,
|
||||
size=self.char_num,
|
||||
act="softmax")
|
||||
return fc_out
|
||||
|
||||
def __call__(self, inputs, others, mode=None):
|
||||
|
||||
pvam_features = self.pvam(inputs, others)
|
||||
gsrm_features, word_out, gsrm_out = self.gsrm(pvam_features, others)
|
||||
final_out = self.vsfd(pvam_features, gsrm_features)
|
||||
|
||||
_, decoded_out = fluid.layers.topk(input=final_out, k=1)
|
||||
predicts = {
|
||||
'predict': final_out,
|
||||
'decoded_out': decoded_out,
|
||||
'word_out': word_out,
|
||||
'gsrm_out': gsrm_out
|
||||
}
|
||||
|
||||
return predicts
|
||||
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,55 @@
|
||||
#copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
|
||||
#
|
||||
#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.
|
||||
|
||||
from __future__ import absolute_import
|
||||
from __future__ import division
|
||||
from __future__ import print_function
|
||||
|
||||
import math
|
||||
|
||||
import paddle
|
||||
import paddle.fluid as fluid
|
||||
|
||||
|
||||
class SRNLoss(object):
|
||||
def __init__(self, params):
|
||||
super(SRNLoss, self).__init__()
|
||||
self.char_num = params['char_num']
|
||||
|
||||
def __call__(self, predicts, others):
|
||||
predict = predicts['predict']
|
||||
word_predict = predicts['word_out']
|
||||
gsrm_predict = predicts['gsrm_out']
|
||||
label = others['label']
|
||||
lbl_weight = others['lbl_weight']
|
||||
|
||||
casted_label = fluid.layers.cast(x=label, dtype='int64')
|
||||
cost_word = fluid.layers.cross_entropy(
|
||||
input=word_predict, label=casted_label)
|
||||
cost_gsrm = fluid.layers.cross_entropy(
|
||||
input=gsrm_predict, label=casted_label)
|
||||
cost_vsfd = fluid.layers.cross_entropy(
|
||||
input=predict, label=casted_label)
|
||||
|
||||
cost_word = fluid.layers.reshape(
|
||||
x=fluid.layers.reduce_sum(cost_word), shape=[1])
|
||||
cost_gsrm = fluid.layers.reshape(
|
||||
x=fluid.layers.reduce_sum(cost_gsrm), shape=[1])
|
||||
cost_vsfd = fluid.layers.reshape(
|
||||
x=fluid.layers.reduce_sum(cost_vsfd), shape=[1])
|
||||
|
||||
sum_cost = fluid.layers.sum(
|
||||
[cost_word, cost_vsfd * 2.0, cost_gsrm * 0.15])
|
||||
|
||||
return [sum_cost, cost_vsfd, cost_word]
|
||||
Loading…
Reference in new issue