-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_face_multistage.py
260 lines (226 loc) · 9.88 KB
/
train_face_multistage.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
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
import tensorflow.keras.backend as K
from tensorflow import keras
import numpy as np
import pandas as pd
import random
import os
import math
#from unet import model
from fcn_multistage53 import model
#from unet_multistage5 import model
from matplotlib import pyplot as plt
from tensorflow.keras.callbacks import ModelCheckpoint, ReduceLROnPlateau, EarlyStopping
from tensorflow.keras.losses import mean_squared_error, categorical_crossentropy
from load_numpy_data_face_multistage import generator
# plots keypoints on face image
def plot_keypoints(img, points):
# display image
plt.imshow(img, cmap='gray')
#plt.imshow(np.float32(img), cmap='gray')
# plot the keypoints
for i in range(0, 42, 2):
#plt.scatter((points[i] + 0.5)*256, (points[i+1]+0.5)*256, color='red')
plt.scatter(points[i], points[i + 1], color='red')
# cv2.circle(img, (int(points[i]), int(points[i + 1])), 3, (0, 255, 0), thickness=-1) # , lineType=-1)#, shift=0)
plt.show()
data_dir = "face"
train_dir = "train"
train_csv = "training.csv"
test_dir = "test"
test_csv = "test.csv"
df_train = pd.read_csv(os.path.join(data_dir, train_csv))
df_test = pd.read_csv(os.path.join(data_dir, test_csv))
n_train = df_train['Image'].size
n_test = df_test['Image'].size
df_kp = df_train.iloc[:,0:30]
idxs = []
img_dict = {}
kp_dict = {}
for i in range(n_train):
if True in df_train.iloc[i, 0:30].isna().values:
continue
else:
idxs.append(i)
img_dict[i] = "train"+str(i)+".png"
# keypoints
kp = df_kp.iloc[i].values.tolist()
kp_dict[i] = kp
random.shuffle(idxs)
# subset = int(0.1*len(idxs))
cutoff_idx = int(0.9*len(idxs))
train_idxs = idxs[0:cutoff_idx]
val_idxs = idxs[cutoff_idx:len(idxs)]
print("\n# of Training Images: {}".format(len(train_idxs)))
print("# of Val Images: {}".format(len(val_idxs)))
transform_dict = {"Flip": False, "Shift": False, "Scale": False, "Rotate": False}
train_generator = generator(os.path.join(data_dir, train_dir),
train_idxs,
img_dict,
kp_dict,
transform_dict=transform_dict,
augment=False,
batch_size=16)
validation_generator = generator(os.path.join(data_dir, train_dir),
val_idxs,
img_dict,
kp_dict,
augment=False,
batch_size=16)
print(len(train_generator), len(validation_generator))
#train_generator = DataGenerator(train_images, train_labels, batch_size = 32)#119 batches(3824/32)
#validation_generator = DataGenerator(test_images, test_labels, batch_size = 32)
for i,j in train_generator:
print(i.shape, j[0].shape)
print(i[0].shape, j[0].shape)
break
id = 1#15
plot_keypoints(i[id], j[0][id])
#input_shape = (368, 368, 3)
#input_shape = (256, 256, 1)
input_shape = (96, 96, 1)
#input_shape = (256, 256, 3)
num_classes = 30
Nkeypoints = 15
print(input_shape)
def get_loss_func():
def mse(x, y):
return mean_squared_error(x, y)
keys = ['output_stage1', 'output_stage2', 'output_stage3', 'output_stage4', 'output_stage5']
losses = dict.fromkeys(keys, mse)
return losses
losses = get_loss_func()
def get_model():
return model(input_shape)
#return model(input_shape=input_shape, num_classes=num_classes)
#return model(input=input_shape, num_classes=num_classes)
model = get_model()
#model.load_weights("vgg16s3.hdf5")
#optimizer = tf.keras.optimizers.Adam(0.1)
#model.compile(optimizer=optimizer, loss="mean_squared_error", metrics=["accuracy"]) # default lr 0.001,1e-3
#model.compile(optimizer="adam", loss="mean_squared_error", metrics=["accuracy"]) # default lr 0.001,1e-3
#model.compile(optimizer=keras.optimizers.Adam(1e-3), loss="mse", metrics=["accuracy"]) #original adam 16% 9%
model.compile(optimizer=keras.optimizers.Adam(1e-3), loss=losses, metrics=["accuracy"]) #original adam 16% 9%
#model.compile(optimizer=keras.optimizers.RMSprop(1e-3), loss="mse", metrics=["accuracy"]) #rmsprop mse 17% 12, 20, 11
#model.compile(optimizer=keras.optimizers.SGD(1e-3), loss="mse", metrics=["accuracy"]) #sgd mse
#model.compile(optimizer=keras.optimizers.RMSprop(1e-3), loss=jaccard, metrics=["accuracy"]) # rmsprop 2%
#model.compile(optimizer=keras.optimizers.Adam(1e-3), loss=jaccard, metrics=["accuracy"]) #regression
#model.compile(optimizer=keras.optimizers.Adam(1e-2), loss="mse", metrics=[tf.keras.metrics.RootMeanSquaredError()]) # default lr 0.001,1e-3# original
#model.compile(optimizer="adam", loss="categorical_crossentropy", metrics=["accuracy"]) # default lr 0.001,1e-3
#lr = 1e-3
callbacks = [ModelCheckpoint("test.hdf5", verbose=1, save_best_only=True),
#ReduceLROnPlateau(monitor="val_loss", patience=5, factor=0.1, verbose=1, min_lr=1e-6, ),# sdnt go below min_lr
EarlyStopping(monitor="val_loss", patience=10, verbose=1)]# try val_root_mean_squared_error
# history = model.fit(x_train, y_train, batch_size=32, verbose=1, epochs= 500, validation_data=(x_test, y_test), shuffle=False) #callbacks=callbacks)#, class_weight=class_weights )
#history = model.fit(train_images, train_labels, batch_size=32, verbose=1, epochs=300, validation_split=0.3,shuffle=False, callbacks=callbacks)#
history = model.fit(train_generator, verbose=1, epochs=500, validation_data=validation_generator, shuffle=True, callbacks=callbacks)#
# history = model.fit(x_train, y_train_cat, batch_size=2, verbose=1, epochs= 10, validation_data=(x_test, y_test_cat), shuffle=False)#, class_weight=class_weights )
# shuffle true sshuffles only the training data for every epoch. but may be we need same for checking imporved models.
#model.save("test.hdf5")
#_, acc = model.evaluate(test_images, test_labels)
#_, acc = model.evaluate(validation_generator)
#print("Accuracy of test set:", (acc * 100.0), "%")
#_, acc = model.evaluate(train_images, train_labels)
#_, acc = model.evaluate(train_generator)
#print("Accuracy of train set:", (acc * 100.0), "%")
# plot train val acc loss
loss = history.history["loss"]
val_loss = history.history["val_loss"]
epochs = range(1, len(loss) + 1)
plt.plot(epochs, loss, "y", label="loss")
plt.plot(epochs, val_loss, "r", label="val loss")
plt.title("model loss")
plt.xlabel("epoch")
plt.ylabel("loss")
plt.grid()
plt.legend()
plt.savefig("loss1.png")
plt.show()
loss = history.history["loss"]
val_loss = history.history["val_loss"]
epochs = range(1, len(loss) + 1)
plt.plot(epochs, loss, label="loss")
plt.plot(epochs, val_loss, label="val loss")
plt.title("model loss")
plt.xlabel("epoch")
plt.ylabel("loss")
plt.grid()
plt.legend()
plt.savefig("loss.png")
plt.show()
plt.close()
stage1_loss = history.history["output_stage1_loss"]
stage2_loss = history.history["output_stage2_loss"]
stage3_loss = history.history["output_stage3_loss"]
stage4_loss = history.history["output_stage4_loss"]
stage5_loss = history.history["output_stage5_loss"]
val_stage1_loss = history.history["val_output_stage1_loss"]
val_stage2_loss = history.history["val_output_stage2_loss"]
val_stage3_loss = history.history["val_output_stage3_loss"]
val_stage4_loss = history.history["val_output_stage4_loss"]
val_stage5_loss = history.history["val_output_stage5_loss"]
epochs = range(1, len(loss) + 1)
plt.plot(epochs, stage1_loss, label="stage1 loss")
plt.plot(epochs, stage2_loss, label="stage2 loss")
plt.plot(epochs, stage3_loss, label="stage3 loss")
plt.plot(epochs, stage4_loss, label="stage4 loss")
plt.plot(epochs, stage5_loss, label="stage5 loss")
plt.plot(epochs, val_stage1_loss, label="val_stage1 loss")
plt.plot(epochs, val_stage2_loss, label="val_stage2 loss")
plt.plot(epochs, val_stage3_loss, label="val_stage3 loss")
plt.plot(epochs, val_stage4_loss, label="val_stage4 loss")
plt.plot(epochs, val_stage5_loss, label="val_stage5 loss")
plt.title("model stage loss")
plt.xlabel("epoch")
plt.ylabel("loss")
plt.grid()
plt.legend()
plt.savefig("stage_loss.png")
plt.show()
plt.close()
stage1_acc = history.history["output_stage1_accuracy"]
stage2_acc = history.history["output_stage2_accuracy"]
stage3_acc = history.history["output_stage3_accuracy"]
stage4_acc = history.history["output_stage4_accuracy"]
stage5_acc = history.history["output_stage5_accuracy"]
val_stage1_acc = history.history["val_output_stage1_accuracy"]
val_stage2_acc = history.history["val_output_stage2_accuracy"]
val_stage3_acc = history.history["val_output_stage3_accuracy"]
val_stage4_acc = history.history["val_output_stage4_accuracy"]
val_stage5_acc = history.history["val_output_stage5_accuracy"]
# plt.plot(epochs, acc, "y", label="Training Accuracy")
# plt.plot(epochs, val_acc, "r", label="Validation Accuracy")
plt.plot(epochs, stage1_acc, label="stage1 acc")
plt.plot(epochs, stage2_acc, label="stage2 acc")
plt.plot(epochs, stage3_acc, label="stage3 acc")
plt.plot(epochs, stage4_acc, label="stage4 acc")
plt.plot(epochs, stage5_acc, label="stage5 acc")
plt.plot(epochs, val_stage1_acc, label="val_stage1 acc")
plt.plot(epochs, val_stage2_acc, label="val_stage2 acc")
plt.plot(epochs, val_stage3_acc, label="val_stage3 acc")
plt.plot(epochs, val_stage4_acc, label="val_stage4 acc")
plt.plot(epochs, val_stage5_acc, label="val_stage5 acc")
plt.title("model stage accuracy")
plt.xlabel("epoch")
plt.ylabel("accuracy")
plt.grid()
plt.legend()
plt.savefig("stage_accuracy.png")
plt.show()
plt.close()
"""
fig = plt.figure(figsize=(15, 15))
# make test images keypoints prediction
points_test = model.predict(test_images)
points_train = model.predict(train_images)
for i in range(16):
ax = fig.add_subplot(4, 4, i + 1, xticks=[], yticks=[])
plot_keypoints(test_images[i], np.squeeze(points_test[i]))
#plot_keypoints(test_images[i], points_test[i])
plt.show()
for i in range(16):
ax = fig.add_subplot(4, 4, i + 1, xticks=[], yticks=[])
plot_keypoints(train_images[i], np.squeeze(points_train[i]))
#plot_keypoints(train_images[i], points_train[i])
plt.show()
"""
a = 1