-
Notifications
You must be signed in to change notification settings - Fork 0
/
Train_captions.py
196 lines (149 loc) · 6.88 KB
/
Train_captions.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Fri Jun 17 14:52:41 2022
@author: gaurav
"""
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
from pickle import load
import tensorflow as tf
import time
import matplotlib.pyplot as plt
import numpy as np
#from My_XF import Mem_Transformer, create_masks
from Mem_XF import Mem_Transformer, create_masks
data_path = './SCAMET/'
#for Sydney Captions Dataset
train_imgs = load(open(data_path+'/data_preprocess/sc_train_imgs.p',"rb"))
train_capts = load(open(data_path+'/data_preprocess/sc_train_capts.p',"rb"))
val_imgs = load(open(data_path+'/data_preprocess/sc_val_imgs.p',"rb"))
val_capts = load(open(data_path+'/data_preprocess/sc_val_capts.p',"rb"))
tokenizer = load(open(data_path+'/data_preprocess/sc_tokenizer.p',"rb"))
vocab_size = len(tokenizer.word_index)
max_len = 23
batch_size = 128
num_layers = 3
d_model = 512
buffer_size = 512
dff = d_model*4
num_heads = 8
target_vocab_size = vocab_size
input_vocab_size = target_vocab_size
dropout_rate = 0.2
EMB_MAT = None
num_memory = 30 # None, 10, 20, 30, 40, 50
transformer = Mem_Transformer(num_memory, num_layers, d_model, EMB_MAT, num_heads, dff, input_vocab_size, target_vocab_size, pe_input=max_len, pe_target=target_vocab_size, rate=dropout_rate)
img_feature_path = data_path+'/Sydney_captions/Img_features/'
def map_func(img_name, cap):
img_id = img_name.decode('utf-8')
#img_id = img_name.decode('utf-8').split('/')[-1]
img_tensor = np.load(img_feature_path+img_id+'EFB3.npy')
return img_tensor, cap
train_data = tf.data.Dataset.from_tensor_slices((train_imgs, train_capts))
train_data = train_data.map(lambda x1, x2: tf.numpy_function(map_func, [x1, x2], [tf.float32, tf.int32]), num_parallel_calls=tf.data.experimental.AUTOTUNE)
train_data = train_data.shuffle(buffer_size).batch(batch_size).prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
val_data = tf.data.Dataset.from_tensor_slices((val_imgs, val_capts))
val_data = val_data.map(lambda x1, x2: tf.numpy_function(map_func, [x1, x2], [tf.float32, tf.int32]), num_parallel_calls=tf.data.experimental.AUTOTUNE)
val_data = val_data.shuffle(buffer_size).batch(batch_size).prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
del train_imgs, train_capts, val_imgs, val_capts
class CustomSchedule(tf.keras.optimizers.schedules.LearningRateSchedule):
def __init__(self, d_model, warmup_steps=45000):
super(CustomSchedule, self).__init__()
self.d_model = d_model
self.d_model = tf.cast(self.d_model, tf.float32)
self.warmup_steps = warmup_steps
def __call__(self, step):
arg1 = tf.math.rsqrt(step)
arg2 = step * (self.warmup_steps**-1.5)
return tf.math.rsqrt(self.d_model) * tf.math.minimum(arg1, arg2)
learning_rate = CustomSchedule(d_model)
#Defining the Loss function
# using Adam Optimizer for the network
optimizer = tf.keras.optimizers.Adam(learning_rate, beta_1=0.9, beta_2=0.98, epsilon=1e-9)
loss_object = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True, reduction='none')
# custom-loss function
def loss_function(real, pred):
mask = tf.math.logical_not(tf.math.equal(real, 0))
loss_ = loss_object(real, pred)
mask = tf.cast(mask, dtype=loss_.dtype)
loss_ *= mask
#return tf.reduce_mean(loss_)
return tf.reduce_sum(loss_)/tf.reduce_sum(mask)
def accuracy_function(real, pred):
accuracies = tf.equal(real, tf.argmax(pred, axis=2, output_type=real.dtype))
mask = tf.math.logical_not(tf.math.equal(real, 0))
accuracies = tf.math.logical_and(mask, accuracies)
accuracies = tf.cast(accuracies, dtype=tf.float32)
mask = tf.cast(mask, dtype=tf.float32)
return tf.reduce_sum(accuracies)/tf.reduce_sum(mask)
train_loss = tf.keras.metrics.Mean(name='train_loss')
train_accuracy = tf.keras.metrics.Mean(name='train_accuracy')
val_loss = tf.keras.metrics.Mean(name='val_loss')
val_accuracy = tf.keras.metrics.Mean(name='val_accuracy')
chkpt_path = data_path+'/Chk_Path/EN_3DE'
chkpt = tf.train.Checkpoint(transformer=transformer, optimizer=optimizer)
chkpt_manager = tf.train.CheckpointManager(chkpt, chkpt_path, max_to_keep=5)
if chkpt_manager.latest_checkpoint:
print("Found a checkpoint")
#chkpt.restore(chkpt_manager.latest_checkpoint)
def train_step(inp, tar, training):
tar_inp = tar[:, :-1]
tar_real = tar[:, 1:]
_, combined_mask, _ = create_masks(inp, tar_inp)
with tf.GradientTape() as tape:
predictions, _, _ = transformer(inp, tar_inp,
training,
None,
combined_mask,
None)
loss = loss_function(tar_real, predictions)
gradients = tape.gradient(loss, transformer.trainable_variables)
#optimizer.apply_gradients(zip(gradients, transformer.trainable_variables))
optimizer.apply_gradients((grad, var) for (grad, var) in zip(gradients, transformer.trainable_variables) if grad is not None)
train_loss(loss)
train_accuracy(accuracy_function(tar_real, predictions))
def val_step(inp, tar, training):
tar_inp = tar[:, :-1]
tar_real = tar[:, 1:]
_, combined_mask, _ = create_masks(inp, tar_inp)
predictions, _, _ = transformer(inp, tar_inp, training, None, combined_mask, None)
val_loss(loss_function(tar_real, predictions))
val_accuracy(accuracy_function(tar_real, predictions))
#Training of custom Transformer for caption generation
EPOCHS = 60
Val_Loss=[]
Train_Loss=[]
Val_Acc=[]
Train_Acc=[]
Epochs=[]
for epoch in range(EPOCHS):
start = time.time()
train_loss.reset_states()
train_accuracy.reset_states()
val_loss.reset_states()
val_accuracy.reset_states()
for (batch, (t_inp, t_tar)) in enumerate(train_data):
train_step(t_inp, t_tar, training=True)
#t_loss, t_acc = train_step(t_inp, t_tar, training=True)
for (batch, (v_inp, v_tar)) in enumerate(val_data):
val_step(v_inp, v_tar, training=False)
#v_loss, v_acc = train_step(v_inp, v_tar, training=False)
if (epoch + 1) % 5 == 0:
chkpt_save_path = chkpt_manager.save()
print ('Saving checkpoint for epoch {} at {}'.format(epoch+1, chkpt_save_path))
print(f'Epoch:{epoch+1}, Train_loss:{train_loss.result():.4f}, Train_acc:{train_accuracy.result():.4f}, Val_loss:{val_loss.result():.4f}, Val_acc:{val_accuracy.result():.4f} ')
Val_Loss.append(val_loss.result().numpy())
Train_Loss.append(train_loss.result().numpy())
Val_Acc.append(val_accuracy.result().numpy())
Train_Acc.append(train_accuracy.result().numpy())
Epochs.append(epoch+1)
print ('Time taken for epoch: {} secs\n'.format(time.time() - start))
n_par = np.sum([np.prod(v.get_shape().as_list()) for v in transformer.trainable_variables])
print("Number of trainable parameters : ", n_par)
plt.plot(Epochs, Train_Acc, color='green', label=' Train_Acc ')
plt.plot(Epochs, Val_Acc, color='red',label=' Val_Acc ')
plt.plot(Epochs, Train_Loss, color='orange', label=' Train_loss ')
plt.plot(Epochs, Val_Loss, color='blue',label=' Val_loss ')
plt.legend()
plt.show()