-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
275 lines (230 loc) · 9.72 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
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
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
import tensorflow as tf
import tensorflow_datasets as tfds
from tensorflow.keras.models import load_model, Model
from pathlib import Path
from random import randint
import os
import wandb
import sys
import time
from datetime import date, datetime
import numpy as np
import config as c
from lib.my_logging import *
from lib.model import *
from lib.trainer import Trainer
from lib.extFunc import saveArrToFile, loadArrFromFile, save_dictionary, load_dictionary
CUDA_DEVICE_INDEX = '0'
os.environ['CUDA_VISIBLE_DEVICES'] = CUDA_DEVICE_INDEX
physical_devices = tf.config.list_physical_devices('GPU')
try:
tf.config.experimental.set_memory_growth(physical_devices[0], True)
except:
pass
########################################
# INITIALIZATION #
########################################
config = c.config
bool_save_model = True
# bool_save_model = False
bool_load_model = True
bool_load_model = False
# If load_model_for_retraining is True, then we will retrain all models
load_model_for_retraining = True
# load_model_for_retraining = False
inputs = tf.keras.Input(shape=config["input_shape"])
# dd/mm/YY
if bool_load_model is not True:
d1 = datetime.now().strftime("%Y%m%d%H00_"+CUDA_DEVICE_INDEX)
else:
d1 = "202306161500_2"
start_index = 9
CONST_SAVED_MODEL_OUTDIR = f"""./saved_models/{config["dataset_name"]}_tf_e{config["epochs"]}_{d1}"""
CONST_SAVED_MODEL_DIR = CONST_SAVED_MODEL_OUTDIR+"/"+"saved_model_label"
CONST_SAVED_PATTERN_MODEL_DIR = CONST_SAVED_MODEL_OUTDIR+"/"+"saved_pattern_model_label"
CONST_SAVED_FIG_DIR = CONST_SAVED_MODEL_OUTDIR+"/"+"imgs"
Path(f"{CONST_SAVED_MODEL_OUTDIR}").mkdir(parents=True, exist_ok=True)
Path(f"{CONST_SAVED_FIG_DIR}").mkdir(parents=True, exist_ok=True)
config["save_fig_path"] = CONST_SAVED_FIG_DIR
modelSummaryStr = ""
patternModelSummaryStr = ""
########################################
# WANDB #
########################################
def initialize_wandb(config, model_index):
# start a new wandb run to track this script
wandb.init(
# set the wandb project where this run will be logged
project=f"""ban_{config["dataset_name"]}_tf_e{config["epochs"]}_{d1}""",
name=f"M{model_index}",
# track hyperparameters and run metadata
config=config
)
if bool_load_model:
filename = f"{CONST_SAVED_MODEL_OUTDIR}/configuration.json"
config = load_dictionary(filename)
# config["learning_rate"] = 0.002
########################################
# LOGGING #
########################################
logname = f"""{CONST_SAVED_MODEL_OUTDIR}/{config["dataset_name"]}_log_{d1}.log"""
logger = createLogger(logname)
########################################
# LOAD FUNCTION #
########################################
def load_pattern(filename):
pattern = loadArrFromFile(filename)
return pattern
def load_pattern_model(model_index):
models = []
if bool_load_model:
# for i in range(config["num_classes"]):
# if train_one_model is True and i != model_index:
# continue
dir = f"{CONST_SAVED_PATTERN_MODEL_DIR}_{model_index}"
model = load_saved_model(dir)
# models.append(model)
return model
# return models
def load_models(dir):
model = load_saved_model(dir)
model.summary()
output = model.layers[-1].output
tmp_model = tf.keras.models.Model(inputs=model.inputs, outputs=output)
return tmp_model
########################################
# OTHER FUNCTION #
########################################
def generateClassPattern(num_of_nodes, index=0):
patterns = []
if config["use_same_pattern"]:
tf.random.set_seed(5)
for i in range(config["num_of_pattern_per_label"]):
pattern = tf.random.uniform(shape=[num_of_nodes], minval=config["min_node"], maxval=config["max_node"])
# pattern = tf.random.uniform(shape=[num_of_nodes], minval=index*0.001, maxval=index*0.001+0.0001)
# pattern = tf.random.uniform(shape=[num_of_nodes], minval=index*10, maxval=index*10+1)
# pattern = tf.random.uniform(shape=[num_of_nodes], minval=1, maxval=50)
# threshold = (index*0.001) + (((index*0.001+0.0001) - (index*0.001)) / 2)
# pattern = tf.where(rand_tensor > threshold, x=tf.ones_like(rand_tensor), y=tf.zeros_like(rand_tensor))
patterns.append(pattern)
return patterns
def getModelSummaryFromKeras(s):
global modelSummaryStr
modelSummaryStr += f"{s}\n"
def getPatternModelSummaryFromKeras(s):
global patternModelSummaryStr
patternModelSummaryStr += f"{s}\n"
def normalize_img(image, label):
"""Normalizes images: `uint8` -> `float32`."""
return tf.cast(image, tf.float32) / 255., label
def getOptimizer():
decay_steps = int(config["epochs"]*config["num_training_data"]/config["batch_size"])
use_learning_rate = tf.keras.experimental.CosineDecay(config["learning_rate"], decay_steps=decay_steps)
optimizer = tf.keras.optimizers.SGD(use_learning_rate, momentum=0.9)
return optimizer
def get_each_label_dataset(ds_test):
# Filter the data based on each label
test_label_datasets = []
for i in range(config["num_classes"]):
test_ds = ds_test.filter(lambda img, label: label == i)
# test_ds = test_ds.batch(config["batch_size"])
test_ds = test_ds.batch(256)
test_ds = test_ds.prefetch(tf.data.AUTOTUNE)
test_label_datasets.append(test_ds)
return test_label_datasets
def main():
if bool_load_model:
# --- Load the pattern and the pattern model ---
filename = f"{CONST_SAVED_MODEL_OUTDIR}/pattern.txt"
patterns = load_pattern(filename)
else:
patterns = []
if config["use_same_pattern"]:
pattern = generateClassPattern(config["output_nodes"])
for i in range(config["num_classes"]):
patterns.append(pattern)
else:
for i in range(config["num_classes"]):
patterns.append(generateClassPattern(config["output_nodes"], i))
np_patterns = []
for label_patterns in patterns:
tmpPatterns = []
for pattern in label_patterns:
tmp = []
for i in pattern.numpy():
tmp.append(i)
tmpPatterns.append(tmp)
np_patterns.append(tmpPatterns)
# --- Save the pattern into the directory ---
filename = f"{CONST_SAVED_MODEL_OUTDIR}/pattern.txt"
saveArrToFile(filename, np_patterns)
logger.info(f'Pattern: {np_patterns}, {np.shape(np_patterns)}')
# --- Save configuration ---
filename = f"{CONST_SAVED_MODEL_OUTDIR}/configuration.json"
save_dictionary(config, filename)
(ds_train, ds_test), ds_info = tfds.load(
config["dataset_name"],
split=['train', 'test'],
shuffle_files=True,
as_supervised=True,
with_info=True,
)
ds_train = ds_train.map(normalize_img, num_parallel_calls=tf.data.AUTOTUNE)
ds_train = ds_train.cache()
ds_train = ds_train.shuffle(ds_info.splits['train'].num_examples)
# ds_train = ds_train.batch(128)
# ds_train = ds_train.prefetch(tf.data.AUTOTUNE)
# Filter the data based on each label
train_label_datasets = []
for i in range(config["num_classes"]):
train_ds = ds_train.filter(lambda img, label: label == i)
train_ds = train_ds.batch(config["batch_size"])
train_ds = train_ds.prefetch(tf.data.AUTOTUNE)
train_label_datasets.append(train_ds)
ds_test = ds_test.map(normalize_img, num_parallel_calls=tf.data.AUTOTUNE)
# ds_test = ds_test.batch(128)
# ds_test = ds_test.cache()
# ds_test = ds_test.prefetch(tf.data.AUTOTUNE)
test_label_datasets = get_each_label_dataset(ds_test)
print("START TRAINING ...")
logits_per_class = []
for index, ds in enumerate(train_label_datasets):
if bool_load_model:
# This is used to retrain only specific model index
# If load_model_for_retraining is True, then we will retrain all models
if index < start_index and load_model_for_retraining is False:
continue
dir = f"{CONST_SAVED_MODEL_DIR}_{index}"
model = load_saved_model(dir)
pattern_model = load_pattern_model(index)
else:
model = createModel(config)
pattern_model = createPatternModel(config)
pattern = patterns[index]
optimizer = getOptimizer()
pattern_optimizer = tf.keras.optimizers.Adam(learning_rate=0.001)
# Print one of the model
if index == 0:
model.summary(print_fn=getModelSummaryFromKeras)
logger.info(msg=f"1 class model summary:\n {modelSummaryStr}")
pattern_model.summary(print_fn=getPatternModelSummaryFromKeras)
logger.info(msg=f"pattern model summary:\n {patternModelSummaryStr}")
initialize_wandb(config, index)
print(f"Train on class: {index}")
print(f"pattern: {pattern}")
trainer = Trainer(model, optimizer, pattern, pattern_model, pattern_optimizer, model_index=index, config=config)
trainer.start_training(ds, config["epochs"], wandb)
if bool_save_model:
# --- Save normal model ---
save_model(model, index, CONST_SAVED_MODEL_DIR)
# --- Save pattern model ---
save_model(pattern_model, index, CONST_SAVED_PATTERN_MODEL_DIR)
wandb.finish()
if __name__ == "__main__":
logger.info(f'Configurations: {config}')
start_time = time.time()
main()
end_time = time.time()
logger.info(f'Total time = {(end_time - start_time)}')
print(f"logger file: {logname}")
print(f"saved model directory: {CONST_SAVED_MODEL_OUTDIR}")