This repository was archived by the owner on Mar 15, 2025. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy pathcifar_training_tools.py
159 lines (121 loc) · 5.94 KB
/
cifar_training_tools.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
import os
import logging
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
logging.getLogger('tensorflow').disabled = True
import numpy as np
import tensorflow as tf
import tensorflow_datasets as tfds
from tqdm import tqdm, tqdm_notebook
def cifar_training(model, logdir, run_name, lr_values=[0.01, 0.1, 0.01, 0.001], lr_boundaries=[400, 32000, 48000, 64000],
val_interval=2000, log_interval=200, batch_size=128, nesterov=False):
schedule = tf.keras.optimizers.schedules.PiecewiseConstantDecay(boundaries=lr_boundaries[:-1], values=lr_values)
optimizer = tf.keras.optimizers.SGD(schedule, momentum=0.9, nesterov=nesterov)
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
ds = tfds.load('cifar10', as_supervised=True, in_memory=True)
std = tf.reshape((0.2023, 0.1994, 0.2010), shape=(1, 1, 3))
mean= tf.reshape((0.4914, 0.4822, 0.4465), shape=(1, 1, 3))
def train_prep(x, y):
x = tf.cast(x, tf.float32)/255.
x = tf.image.random_flip_left_right(x)
x = tf.image.pad_to_bounding_box(x, 4, 4, 40, 40)
x = tf.image.random_crop(x, (32, 32, 3))
x = (x - mean) / std
return x, y
def valid_prep(x, y):
x = tf.cast(x, tf.float32)/255.
x = (x - mean) / std
return x, y
ds['train'] = ds['train'].map(train_prep).shuffle(10000).repeat().batch(batch_size).prefetch(-1)
ds['test'] = ds['test'].map(valid_prep).batch(batch_size*4).prefetch(-1)
runid = run_name + '_x' + str(np.random.randint(10000))
writer = tf.summary.create_file_writer(logdir + '/' + runid)
accuracy = tf.metrics.SparseCategoricalAccuracy()
cls_loss = tf.metrics.Mean()
reg_loss = tf.metrics.Mean()
print(f"RUNID: {runid}")
tf.keras.utils.plot_model(model, os.path.join('saved_plots', runid + '.png'))
@tf.function
def step(x, y, training):
with tf.GradientTape() as tape:
r_loss = tf.add_n(model.losses)
outs = model(x, training)
c_loss = loss_fn(y, outs)
loss = c_loss + r_loss
if training:
gradients = tape.gradient(loss, model.trainable_weights)
optimizer.apply_gradients(zip(gradients, model.trainable_weights))
accuracy(y, outs)
cls_loss(c_loss)
reg_loss(r_loss)
training_step = 0
best_validation_acc = 0
epochs = lr_boundaries[-1] // val_interval
for epoch in range(epochs):
for x, y in tqdm(ds['train'].take(val_interval), desc=f'epoch {epoch+1}/{epochs}',
total=val_interval, ncols=100, ascii=True):
training_step += 1
step(x, y, training=True)
if training_step % log_interval == 0:
with writer.as_default():
c_loss, r_loss, err = cls_loss.result(), reg_loss.result(), 1-accuracy.result()
print(f" c_loss: {c_loss:^6.3f} | r_loss: {r_loss:^6.3f} | err: {err:^6.3f}", end='\r')
tf.summary.scalar('train/error_rate', err, training_step)
tf.summary.scalar('train/classification_loss', c_loss, training_step)
tf.summary.scalar('train/regularization_loss', r_loss, training_step)
tf.summary.scalar('train/learnig_rate', optimizer._decayed_lr('float32'), training_step)
cls_loss.reset_states()
reg_loss.reset_states()
accuracy.reset_states()
for x, y in ds['test']:
step(x, y, training=False)
with writer.as_default():
tf.summary.scalar('test/classification_loss', cls_loss.result(), step=training_step)
tf.summary.scalar('test/error_rate', 1-accuracy.result(), step=training_step)
if accuracy.result() > best_validation_acc:
best_validation_acc = accuracy.result()
model.save_weights(os.path.join('saved_models', runid + '.tf'))
cls_loss.reset_states()
accuracy.reset_states()
def cifar_error_test(model, tr_len=20, vd_len=2):
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
optimizer = tf.keras.optimizers.SGD(0.01)
ds = tfds.load('cifar10', as_supervised=True, in_memory=True)
std = tf.reshape((0.2023, 0.1994, 0.2010), shape=(1, 1, 3))
mean= tf.reshape((0.4914, 0.4822, 0.4465), shape=(1, 1, 3))
def train_prep(x, y):
x = tf.cast(x, tf.float32)/255.
x = tf.image.random_flip_left_right(x)
x = tf.image.pad_to_bounding_box(x, 4, 4, 40, 40)
x = tf.image.random_crop(x, (32, 32, 3))
x = (x - mean) / std
return x, y
def valid_prep(x, y):
x = tf.cast(x, tf.float32)/255.
x = (x - mean) / std
return x, y
ds['train'] = ds['train'].map(train_prep).batch(5).take(tr_len).prefetch(-1)
ds['test'] = ds['test'].map(valid_prep).batch(5).take(vd_len).prefetch(-1)
accuracy = tf.metrics.SparseCategoricalAccuracy()
cls_loss = tf.metrics.Mean()
reg_loss = tf.metrics.Mean()
@tf.function
def step(x, y, training):
with tf.GradientTape() as tape:
r_loss = tf.add_n(model.losses)
outs = model(x, training)
c_loss = loss_fn(y, outs)
loss = c_loss + r_loss
if training:
gradients = tape.gradient(loss, model.trainable_weights)
optimizer.apply_gradients(zip(gradients, model.trainable_weights))
accuracy(y, outs)
cls_loss(c_loss)
reg_loss(r_loss)
training_step = 0
for x, y in tqdm(ds['train'], desc=f'test', total=tr_len, ncols=100, ascii=True):
training_step += 1
step(x, y, training=True)
c_loss, r_loss, err = cls_loss.result(), reg_loss.result(), 1-accuracy.result()
print(f" c_loss: {c_loss:^6.3f} | r_loss: {r_loss:^6.3f} | err: {err:^6.3f}", end='\r')
for x, y in ds['test']:
step(x, y, training=False)