-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
208 lines (174 loc) · 9.01 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
"""
Basic idea of FOTS:
0. Input: [1, 480, 640, 3]
1. Extract features from ResNet50/Mobilenet (sharedconv) --> [1, 120, 160, 32]
2. Input sharedconv to DetectionModel to detect text regions --> [1, 120, 160, 1] and [1, 120, 160, 5]
3. RoIRotate to crop and rotate sharedconv for text recognition --> [5, 8, 64, 32] [text_regions, height, width, chan]
4. Input RoIRotate to CRNN and recognize chars in every text region --> [5, 64, 78] [text_regions, width, chars]
logits 64
logit_length 64
blank_index 62
TODO
4. detection kernel size 3? originally it was 1 Keep 1
5. ctc loss mask?: https://github.com/yu20103983/FOTS/blob/master/FOTS/recognizer.py
7. recognition max pool 2d pool_size https://github.com/yu20103983/FOTS/blob/master/FOTS/recognizer.py keep as in paper
10. rotate expand_w = 60?
11. CRNN more same convs..? if i do this, the rnn outputs same fore each batch, how does that make sence?
12. Expand y_true boundries in classification - shrink_poly R param
13. There is no gradient link between backbone and recognition..? !!!!!
Sparse to dense etc non differentiable
https://github.com/Mainak431/List-of-Differentiable--OPs-and-Non-differentiable-OPs--in-Tensorflow
It looks like will have to do that on linux
https://www.tensorflow.org/graphics/api_docs/python/tfg/geometry/transformation/rotation_matrix_3d/from_axis_angle?authuser=2&hl=lt
https://databricks.com/tensorflow/tensorflow-in-3d
"""
import tensorflow as tf
import cv2
import numpy as np
# from icdar import generator
from synthtext import generator
from config import CHAR_VECTOR
from model_backbone import Backbone
from model_detection import Detection
from model_roirotate import RoIRotate, tfa_enabled
from model_recognition import Recognition
from utils import decode_to_text
# init
cpkt_dir = 'checkpoints/'
load_models = True
loss_hist = []
input_shape = [640, 640, 3] # 160 / 480 / 640 / 800
model_sharedconv = Backbone(backbone='mobilenet', input_shape=input_shape)
model_detection = Detection()
model_RoIrotate = RoIRotate(tfa_enabled=tfa_enabled)
model_recognition = Recognition(num_classes=len(CHAR_VECTOR)+1, training=True, drop_prob=0.001)
optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001)
save_iter = 100
iter = 0
if load_models:
model_sharedconv.load_weights(cpkt_dir + 'sharedconv')
model_detection.load_weights(cpkt_dir + 'detection')
model_recognition.load_weights(cpkt_dir + 'recognition')
# -------- #
data_gen = generator(input_size=input_shape[0],
batch_size=1,
min_text_size=8,
random_scale=np.array([1.0]))
for x_batch in data_gen:
with tf.GradientTape() as tape:
# forward-prop
sharedconv = model_sharedconv(x_batch['images'].copy())
f_score_, geo_score_ = model_detection(sharedconv)
features, ws = model_RoIrotate(sharedconv, x_batch['rboxes'], expand_px=1)
logits = model_recognition(features)
# loss
loss_cls = model_detection.loss_classification2(x_batch['score_maps'], f_score_, x_batch['training_masks'])
loss_iou, loss_angle = model_detection.loss_regression(x_batch['geo_maps'], geo_score_, x_batch['score_maps'], x_batch['training_masks'])
loss_recongition = model_recognition.loss_recognition(y=x_batch['text_labels_sparse'], logits=logits, ws=ws)
loss_cls *= 100 # 2
loss_iou *= 100 # 2
loss_angle *= 100 # 60
loss_recongition /= 1 # 10
model_loss = (loss_cls + loss_iou + loss_angle + loss_recongition)
# backward-prop
grads = tape.gradient(model_loss,
model_sharedconv.trainable_variables +
model_detection.trainable_variables +
model_recognition.trainable_variables)
optimizer.apply_gradients(zip(grads,
model_sharedconv.trainable_variables +
model_detection.trainable_variables +
model_recognition.trainable_variables))
print(iter, loss_cls.numpy(), loss_iou.numpy(), loss_angle.numpy(),
loss_recongition.numpy(), sum([len(i[0]) for i in x_batch['rboxes']]))
iter += 1
loss_hist.append([loss_cls.numpy(), loss_iou.numpy(), loss_angle.numpy(), loss_recongition.numpy()])
# recognition results
y_true = tf.sparse.to_dense(tf.SparseTensor(*x_batch['text_labels_sparse'])).numpy()
decoded, _ = tf.nn.ctc_greedy_decoder(logits.numpy().transpose((1, 0, 2)), sequence_length=[logits.shape[1]] * logits.shape[0])
decoded = tf.sparse.to_dense(decoded[0]).numpy()
print([decode_to_text(CHAR_VECTOR, [j for j in i if j != 0]) for i in decoded[:4, :]],
[decode_to_text(CHAR_VECTOR, [j for j in i if j != 0]) for i in y_true[:4, :]])
# save
if iter % save_iter == 0:
model_sharedconv.save_weights(cpkt_dir + 'sharedconv')
model_detection.save_weights(cpkt_dir + 'detection')
model_recognition.save_weights(cpkt_dir + 'recognition')
with open('loss.txt', ['a' if load_models else 'w'][0]) as file:
for line in loss_hist:
[file.write(str(i) + ' ') for i in line]
file.write('\n')
loss_hist = []
# -------- #
# -------- #
# -------- #
data_gen = generator(input_size=input_shape[0], batch_size=1, min_text_size=10, random_scale=np.array([1.0]))
x_batch = data_gen.__next__()
sharedconv = model_sharedconv(x_batch['images'].copy())
f_score_, geo_score_ = model_detection(sharedconv)
features, ws = model_RoIrotate(sharedconv, x_batch['rboxes'], expand_px=2)
logits = model_recognition(features)
cv2.imshow('img', cv2.resize(x_batch['images'][0, ::], (640, 640)).astype(np.uint8))
cv2.imshow('scr', cv2.resize(x_batch['score_maps'][0, ::]*255, (640, 640)).astype(np.uint8))
cv2.imshow('pre', cv2.resize(f_score_[0, ::].numpy()*255, (640, 640)).astype(np.uint8))
cv2.imshow('msk', cv2.resize(x_batch['training_masks'][0, ::]*255, (640, 640)).astype(np.uint8))
[cv2.imshow(str(i), cv2.resize(geo_score_[0, :, :, i].numpy()*255, (512, 512)).astype(np.uint8)) for i in [0, 1, 2, 3, 4]]
cv2.waitKey(0)
cv2.destroyAllWindows()
# ------- #
cv2.imshow('i', cv2.resize(x_batch['images'][0, ::], (512, 512)).astype(np.uint8))
cv2.waitKey(1)
for i in range(32):
cv2.imshow(str(i), cv2.resize(sharedconv[-1, :, :, i].numpy()*255, (512, 512)).astype(np.uint8))
cv2.imshow(str(i)+'_crop', (features.numpy() * 255).astype(np.uint8)[0, :, :, i])
cv2.waitKey(0)
cv2.destroyAllWindows()
# ------- #
decoded, log_prob = tf.nn.ctc_greedy_decoder(logits.numpy().transpose((1, 0, 2)), sequence_length=[64]*logits.shape[0])
decoded = tf.sparse.to_dense(decoded[0]).numpy()
print([decode_to_text(CHAR_VECTOR, [j for j in i if j != 0]) for i in decoded])
y_true = tf.sparse.to_dense(tf.SparseTensor(*x_batch['text_labels_sparse'])).numpy()
print([decode_to_text(CHAR_VECTOR, [j for j in i if j != 0]) for i in y_true])
# ------------------------- #
# --------- #
# DEBUGGING #
# --------- #
data_gen = generator(input_size=input_shape[0], batch_size=1, min_text_size=5)
x_batch = next(data_gen)
stride = 1
shape = tuple([int(i / stride) for i in x_batch['images'].shape[1:3]])
dummy_input = cv2.resize(x_batch['images'][0, :, :, :].copy(), shape)[np.newaxis, :, :, :]
RoIrotate_test = RoIRotate(tfa_enabled=tfa_enabled, features_stride=stride)
features, ws = RoIrotate_test(dummy_input, x_batch['rboxes'], plot=True, expand_px=1)
# --- #
data_gen = generator(input_size=input_shape[0], batch_size=1, min_text_size=4)
x_batch = data_gen.__next__()
y_true = tf.sparse.to_dense(tf.SparseTensor(*x_batch['text_labels_sparse'])).numpy()
print([decode_to_text(CHAR_VECTOR, [j for j in i if j != 0]) for i in y_true[:4, :]])
cv2.imshow('img', cv2.resize(x_batch['images'][0, ::], (512, 512)).astype(np.uint8))
cv2.waitKey(0)
cv2.destroyAllWindows()
# ------------------------- #
# check overlapping
orig_img = x_batch['images'][0].copy()
shrd_con = cv2.resize(sharedconv.numpy().copy()[0, :, :, :], orig_img.shape[:2])
for i in range(shrd_con.shape[-1]):
overlay = np.tile(shrd_con[:, :, i].copy()[:, :, np.newaxis], 3) * 255*4
combined = cv2.addWeighted(orig_img, 0.4, overlay, 0.3, 1)
cv2.imshow('{}'.format(i), combined.astype(np.uint8))
cv2.waitKey(0)
cv2.destroyAllWindows()
# ----------------------- #
# check mobilenet layers
model = tf.keras.applications.MobileNetV2(include_top=False, input_shape=input_shape)
layer_ids = [154, 119, 57, 30]
backbone_layers = tf.keras.models.Model(
inputs=model.input,
outputs=[model.get_layer(index=i).output for i in layer_ids])
input = tf.keras.applications.mobilenet.preprocess_input(x_batch['images'][0].copy()[np.newaxis, :, :, :])
output = backbone_layers(input)
for i in range(36):
conv = cv2.resize(output[3].numpy().copy()[0, :, :, i], (512, 512)) * 255
cv2.imshow('{}'.format(i), conv.astype(np.uint8))
cv2.waitKey(0)
cv2.destroyAllWindows()