|
a |
|
b/train_lstm.py |
|
|
1 |
# -*- coding: utf-8 -*- |
|
|
2 |
"""Train_LSTM.ipynb |
|
|
3 |
** |
|
|
4 |
* This file is part of Hybrid CNN-LSTM for COVID-19 Severity Score Prediction paper. |
|
|
5 |
* |
|
|
6 |
* Written by Ankan Ghosh Dastider and Farhan Sadik. |
|
|
7 |
* |
|
|
8 |
* Copyright (c) by the authors under Apache-2.0 License. Some rights reserved, see LICENSE. |
|
|
9 |
*/ |
|
|
10 |
|
|
|
11 |
""" |
|
|
12 |
|
|
|
13 |
''' |
|
|
14 |
Loading frames of the videos sequentially |
|
|
15 |
''' |
|
|
16 |
video_types=['Video 01', 'Video 05', 'Video 06', 'Video 07', 'Video 08', 'Video 09', 'Video 10', 'Video 14', |
|
|
17 |
'Video 15', 'Video 16', 'Video 17', 'Video 20', 'Video 21', 'Video 27', 'Video 29'] |
|
|
18 |
|
|
|
19 |
NUM_VIDEOS = len(video_types) |
|
|
20 |
NUM_FRAMES = 302 |
|
|
21 |
|
|
|
22 |
data_dir_lstm = '' #Link Training Directory videowise |
|
|
23 |
train_dir_lstm = os.path.join(data_dir_lstm) |
|
|
24 |
|
|
|
25 |
train_data_lstm = [] |
|
|
26 |
for defects_id, sp in enumerate(video_types): |
|
|
27 |
temporary = [] |
|
|
28 |
for file in sorted(os.listdir(os.path.join(train_dir_lstm, sp))): |
|
|
29 |
temporary.append(['{}/{}'.format(sp, file), defects_id, sp]) |
|
|
30 |
|
|
|
31 |
total_frames = len(temporary) |
|
|
32 |
index = np.linspace(start = 0, stop = total_frames-1, num = NUM_FRAMES, dtype = int) |
|
|
33 |
|
|
|
34 |
for i in range(NUM_FRAMES): |
|
|
35 |
train_data_lstm.append(temporary[index[i]]) |
|
|
36 |
|
|
|
37 |
train_on_lstm = pd.DataFrame(train_data_lstm, columns=['File', 'FolderID','Video Type']) |
|
|
38 |
train_on_lstm.head(NUM_VIDEOS*NUM_FRAMES) |
|
|
39 |
|
|
|
40 |
video_types=['Video 01', 'Video 05', 'Video 06', 'Video 07', 'Video 08', 'Video 09', 'Video 10', 'Video 14', |
|
|
41 |
'Video 15', 'Video 16', 'Video 17', 'Video 20', 'Video 21', 'Video 27', 'Video 29'] |
|
|
42 |
|
|
|
43 |
data_dir_lstm = '' #Link Training Directory videowise |
|
|
44 |
train_dir_lstm = os.path.join(data_dir_lstm) |
|
|
45 |
|
|
|
46 |
train_data_lstm = [] |
|
|
47 |
for defects_id, sp in enumerate(video_types): |
|
|
48 |
for file in sorted(os.listdir(os.path.join(train_dir_lstm, sp))): |
|
|
49 |
# print(file) |
|
|
50 |
train_data_lstm.append(['{}/{}'.format(sp, file), defects_id, sp]) |
|
|
51 |
|
|
|
52 |
train_on_lstm = pd.DataFrame(train_data_lstm, columns=['File', 'FolderID','Video Type']) |
|
|
53 |
train_on_lstm.head() |
|
|
54 |
|
|
|
55 |
IMAGE_SIZE = 128 |
|
|
56 |
SEED = 42 |
|
|
57 |
|
|
|
58 |
BATCH_SIZE_LSTM = 25 |
|
|
59 |
EPOCHS_LSTM = 120 |
|
|
60 |
|
|
|
61 |
def read_image_lstm(filepath): |
|
|
62 |
return cv2.imread(os.path.join(data_dir_lstm, filepath)) # Loading a color image is the default flag |
|
|
63 |
|
|
|
64 |
#Resize image to target size |
|
|
65 |
def resize_image(newimage, image_size): |
|
|
66 |
return cv2.resize(newimage.copy(), image_size, interpolation=cv2.INTER_AREA) |
|
|
67 |
|
|
|
68 |
from tensorflow.keras.models import load_model |
|
|
69 |
import re |
|
|
70 |
from keras import backend as K |
|
|
71 |
|
|
|
72 |
X_Train_Total = np.zeros((NUM_VIDEOS, NUM_FRAMES, IMAGE_SIZE, IMAGE_SIZE, 3)) |
|
|
73 |
Y_Train_Total = np.zeros((NUM_VIDEOS, NUM_FRAMES, 1)) |
|
|
74 |
k = 0 |
|
|
75 |
j = 0 |
|
|
76 |
for i, file in tqdm(enumerate(train_on_lstm['File'].values)): |
|
|
77 |
if i % NUM_FRAMES == 0 and i != 0 : |
|
|
78 |
k = k + 1 |
|
|
79 |
j = 0 |
|
|
80 |
if k == NUM_VIDEOS: |
|
|
81 |
break |
|
|
82 |
# print(i,file) |
|
|
83 |
newimage = read_image_lstm(file) |
|
|
84 |
if newimage is not None: |
|
|
85 |
# print(k,j) |
|
|
86 |
X_Train_Total[k,j] = resize_image(newimage, (IMAGE_SIZE, IMAGE_SIZE)) |
|
|
87 |
match = re.search('Score(\d)',file) |
|
|
88 |
score = int(match.group(1)) |
|
|
89 |
Y_Train_Total[k,j] = score |
|
|
90 |
#print(file) |
|
|
91 |
#print(score) |
|
|
92 |
#print(Y_test[k,j]) |
|
|
93 |
j = j + 1 |
|
|
94 |
|
|
|
95 |
Y_Train_Total = to_categorical(Y_Train_Total, num_classes=4) |
|
|
96 |
# print(Y_Train_Total) |
|
|
97 |
# Normalize the data |
|
|
98 |
X_Train_Total = X_Train_Total / 255. |
|
|
99 |
print('X_Train_Total Shape: {}'.format(X_Train_Total.shape)) |
|
|
100 |
print('Y_Train_Total Shape: {}'.format(Y_Train_Total.shape)) |
|
|
101 |
|
|
|
102 |
np.random.seed(42) |
|
|
103 |
np.random.shuffle(X_Train_Total) |
|
|
104 |
|
|
|
105 |
np.random.seed(42) |
|
|
106 |
np.random.shuffle(Y_Train_Total) |
|
|
107 |
|
|
|
108 |
print('X_Train_Total Shape: {}'.format(X_Train_Total.shape)) |
|
|
109 |
print('Y_Train_Total Shape: {}'.format(Y_Train_Total.shape)) |
|
|
110 |
|
|
|
111 |
model = load_model('') #Link the CNN weights |
|
|
112 |
model.summary() |
|
|
113 |
|
|
|
114 |
output = np.zeros((NUM_VIDEOS, NUM_FRAMES, 64)) |
|
|
115 |
|
|
|
116 |
for i in range(NUM_VIDEOS): |
|
|
117 |
X_New = X_Train_Total[i] |
|
|
118 |
specific_layer_output = K.function([model.layers[0].input], [model.get_layer('dropout_35').output]) |
|
|
119 |
layer_output = specific_layer_output([X_New])[0] |
|
|
120 |
#print(layer_output.shape) |
|
|
121 |
#print(layer_output) |
|
|
122 |
output[i] = layer_output |
|
|
123 |
|
|
|
124 |
print('Output from CNN Shape: {}'.format(output.shape)) |
|
|
125 |
#custom3 = model.predict(X_Test) |
|
|
126 |
#print(custom3) |
|
|
127 |
|
|
|
128 |
X_Train_Total = output |
|
|
129 |
Y_Train_Total = Y_Train_Total |
|
|
130 |
|
|
|
131 |
print('X_Train_Total Shape: {}'.format(X_Train_Total.shape)) |
|
|
132 |
print('Y_Train_Total Shape: {}'.format(Y_Train_Total.shape)) |
|
|
133 |
|
|
|
134 |
# Split the train and validation sets |
|
|
135 |
X_Train_LSTM, X_Val_LSTM, Y_Train_LSTM, Y_Val_LSTM = train_test_split(X_Train_Total, Y_Train_Total, |
|
|
136 |
test_size=0.2, random_state = SEED) |
|
|
137 |
|
|
|
138 |
from keras.layers import Reshape, LSTM, Lambda, TimeDistributed, Conv1D, MaxPool1D, Dense, Dropout, Flatten, Conv2D, MaxPool2D, BatchNormalization, AveragePooling2D, GlobalAveragePooling2D |
|
|
139 |
|
|
|
140 |
|
|
|
141 |
def build_lstm(): |
|
|
142 |
|
|
|
143 |
input = Input(shape=(NUM_FRAMES, 64)) |
|
|
144 |
|
|
|
145 |
|
|
|
146 |
x = LSTM(1000, return_sequences = True)(input) |
|
|
147 |
x = Dropout(0.5)(x) |
|
|
148 |
|
|
|
149 |
x = LSTM(1000, return_sequences = True)(x) |
|
|
150 |
x = Dropout(0.5)(x) |
|
|
151 |
|
|
|
152 |
x = LSTM(4, return_sequences=True)(x) |
|
|
153 |
# multi output |
|
|
154 |
output = Dense(4,activation = 'softmax', name='root')(x) |
|
|
155 |
|
|
|
156 |
# model |
|
|
157 |
model = Model(input,output) |
|
|
158 |
|
|
|
159 |
optimizer = Adam(lr=0.002, beta_1=0.9, beta_2=0.999, epsilon=0.1, decay=0.0) |
|
|
160 |
model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy']) |
|
|
161 |
model.summary() |
|
|
162 |
|
|
|
163 |
return model |
|
|
164 |
|
|
|
165 |
model_lstm = build_lstm() |
|
|
166 |
annealer = ReduceLROnPlateau(monitor='val_accuracy', factor=0.5, patience=5, verbose=1, min_lr=1e-3) |
|
|
167 |
checkpoint = ModelCheckpoint('model_lstm.h5', verbose=1, save_best_only=True) |
|
|
168 |
# Generates batches of image data with data augmentation |
|
|
169 |
# datagen = ImageDataGenerator(rotation_range=360, # Degree range for random rotations |
|
|
170 |
# width_shift_range=0.2, # Range for random horizontal shifts |
|
|
171 |
# height_shift_range=0.2, # Range for random vertical shifts |
|
|
172 |
# zoom_range=0.2, # Range for random zoom |
|
|
173 |
# horizontal_flip=True, # Randomly flip inputs horizontally |
|
|
174 |
# vertical_flip=True) # Randomly flip inputs vertically |
|
|
175 |
|
|
|
176 |
#datagen.fit(X_train) |
|
|
177 |
# Fits the model on batches with real-time data augmentation |
|
|
178 |
hist = model_lstm.fit(X_Train_LSTM, Y_Train_LSTM, batch_size = BATCH_SIZE_LSTM, |
|
|
179 |
# steps_per_epoch=X_Train_LSTM.shape[0] // BATCH_SIZE, |
|
|
180 |
epochs = EPOCHS_LSTM, |
|
|
181 |
verbose = 2, |
|
|
182 |
callbacks = [annealer, checkpoint], |
|
|
183 |
validation_data = (X_Val_LSTM, Y_Val_LSTM)) |
|
|
184 |
|
|
|
185 |
final_loss_lstm, final_accuracy_lstm = model_lstm.evaluate(X_Val_LSTM, Y_Val_LSTM) |
|
|
186 |
print('Final Loss LSTM: {}, Final Accuracy LSTM: {}'.format(final_loss_lstm, final_accuracy_lstm)) |
|
|
187 |
|
|
|
188 |
score_types = ['Score 0', 'Score 1', 'Score 2', 'Score 3'] |
|
|
189 |
|
|
|
190 |
Y_pred_lstm = model_lstm.predict(X_Val_LSTM) |
|
|
191 |
Y_pred_lstm = np.reshape(Y_pred_lstm, (Y_pred_lstm.shape[0]*Y_pred_lstm.shape[1], Y_pred_lstm.shape[2])) |
|
|
192 |
Y_pred_lstm = np.argmax(Y_pred_lstm, axis=1) |
|
|
193 |
|
|
|
194 |
Y_true_lstm = np.reshape(Y_Val_LSTM, (Y_Val_LSTM.shape[0]*Y_Val_LSTM.shape[1], Y_Val_LSTM.shape[2])) |
|
|
195 |
Y_true_lstm = np.argmax(Y_true_lstm, axis=1) |
|
|
196 |
|
|
|
197 |
#print(Y_pred_lstm.shape) |
|
|
198 |
#print(Y_Val_LSTM.shape) |
|
|
199 |
cm = confusion_matrix(Y_true_lstm, Y_pred_lstm) |
|
|
200 |
plt.figure(figsize=(12, 12)) |
|
|
201 |
ax = sns.heatmap(cm, cmap=plt.cm.Greens, annot=True, square=True, xticklabels=score_types, yticklabels=score_types) |
|
|
202 |
ax.set_ylabel('Actual', fontsize=40) |
|
|
203 |
ax.set_xlabel('Predicted', fontsize=40) |
|
|
204 |
|
|
|
205 |
''' |
|
|
206 |
# accuracy plot |
|
|
207 |
plt.plot(hist.history['accuracy']) |
|
|
208 |
plt.plot(hist.history['val_accuracy']) |
|
|
209 |
plt.title('model accuracy') |
|
|
210 |
plt.ylabel('accuracy') |
|
|
211 |
plt.xlabel('epoch') |
|
|
212 |
plt.legend(['train', 'test'], loc='upper left') |
|
|
213 |
plt.show() |
|
|
214 |
|
|
|
215 |
# loss plot |
|
|
216 |
plt.plot(hist.history['loss']) |
|
|
217 |
plt.plot(hist.history['val_loss']) |
|
|
218 |
plt.title('model loss') |
|
|
219 |
plt.ylabel('loss') |
|
|
220 |
plt.xlabel('epoch') |
|
|
221 |
plt.legend(['train', 'test'], loc='upper left') |
|
|
222 |
plt.show() |
|
|
223 |
''' |