-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
241 lines (168 loc) · 9.09 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
#%%
import os
import argparse
from datetime import datetime
from tqdm import tqdm
import tensorflow as tf
import tensorflow_probability as tfp
#tf.config.set_visible_devices([], 'GPU')
import sys
sys.path.append('example-cnp/')
from dataloader.load_regression_data_from_arbitrary_gp import RegressionDataGeneratorArbitraryGP
from dataloader.load_regression_data_from_arbitrary_gp_varying_kernel import RegressionDataGeneratorArbitraryGPWithVaryingKernel
from dataloader.load_mnist import load_mnist
from dataloader.load_celeb import load_celeb
from neural_process_model_conditional import NeuralProcessConditional
from neural_process_model_hybrid import NeuralProcessHybrid
from neural_process_model_latent import NeuralProcessLatent
from neural_process_model_hybrid_constrained import NeuralProcessHybridConstrained
from utils.utility import PlotCallback
tfk = tf.keras
tfd = tfp.distributions
# ================================ Parse Training parameters ===============================================
# parser = argparse.ArgumentParser()
# parser.add_argument('-e', '--epochs', type=int, default=120, help='Number of training epochs')
# parser.add_argument('-b', '--batch', type=int, default=1024, help='Batch size for training')
# parser.add_argument('-t', '--task', type=str, default='regression', help='Task to perform : (mnist|regression|celeb|regression_varying)')
# parser.add_argument('-c', '--num_context', type=int, default=100)
# parser.add_argument('-u', '--uniform_sampling', type=bool, default=True)
# parser.add_argument('-m', '--model', type=bool, default='CNP', help='CNP|LNP|HNP|HNPC')
# args = parser.parse_args()
# -------------------------------------------------------------------------------------------------------------------------
# ================================ Training parameters ===============================================
# Regression
args = argparse.Namespace(epochs=160, batch=1024, task='regression', num_context=25, uniform_sampling=True, model='CNP')
# MNIST / Celeb
#args = argparse.Namespace(epochs=30, batch=256, task='mnist', num_context=100, uniform_sampling=True, model='CNP')
LOG_PRIORS = True
# -------------------------------------------------------------------------------------------------------------------------
# =========================== Data Loaders ===========================================================================================
BATCH_SIZE = args.batch
EPOCHS = args.epochs
TRAINING_ITERATIONS = int(100)
TEST_ITERATIONS = int(TRAINING_ITERATIONS/5)
if args.task == 'mnist':
train_ds, test_ds, TRAINING_ITERATIONS, TEST_ITERATIONS = load_mnist(batch_size=BATCH_SIZE, num_context_points=args.num_context, uniform_sampling=args.uniform_sampling)
# Model architecture
z_output_sizes = [500, 500, 500, 1000]
enc_output_sizes = [500, 500, 500, 500]
dec_output_sizes = [500, 500, 500, 2]
elif args.task == 'celeb':
train_ds, test_ds, TRAINING_ITERATIONS, TEST_ITERATIONS = load_celeb(batch_size=BATCH_SIZE, num_context_points=args.num_context, uniform_sampling=args.uniform_sampling)
# Model architecture
z_output_sizes = [500, 500, 500, 1000]
enc_output_sizes = [500, 500, 500, 500]
dec_output_sizes = [500, 500, 500, 6]
elif args.task == 'regression':
data_generator = RegressionDataGeneratorArbitraryGP(
iterations=TRAINING_ITERATIONS,
n_iterations_test=TEST_ITERATIONS,
batch_size=BATCH_SIZE,
min_num_context=3,
max_num_context=40,
min_num_target=2,
max_num_target=40,
min_x_val_uniform=-2,
max_x_val_uniform=2,
kernel_length_scale=0.4
)
train_ds, test_ds = data_generator.load_regression_data()
# Model architecture
z_output_sizes = [500, 500, 500, 1000]
enc_output_sizes = [500, 500, 500, 500]
dec_output_sizes = [500, 500, 500, 2]
elif args.task == 'regression_varying':
data_generator = RegressionDataGeneratorArbitraryGPWithVaryingKernel(
iterations=TRAINING_ITERATIONS,
n_iterations_test=TEST_ITERATIONS,
batch_size=BATCH_SIZE,
min_num_context=3,
max_num_context=40,
min_num_target=2,
max_num_target=40,
min_x_val_uniform=-2,
max_x_val_uniform=2,
min_kernel_length_scale=0.1,
max_kernel_length_scale=1.
)
# Model architecture
z_output_sizes = [500, 500, 500, 1000]
enc_output_sizes = [500, 500, 500, 500]
dec_output_sizes = [500, 500, 500, 2]
train_ds, test_ds = data_generator.train_ds, data_generator.test_ds
# --------------------------------------------------------------------------------------------------------------------------------------------
# ========================================== Define NP Model ===================================================
if args.model == 'CNP':
model = NeuralProcessConditional(enc_output_sizes, dec_output_sizes)
elif args.model == 'HNP':
model = NeuralProcessHybrid(z_output_sizes, enc_output_sizes, dec_output_sizes)
elif args.model == 'LNP':
model = NeuralProcessLatent(z_output_sizes, enc_output_sizes, dec_output_sizes)
elif args.model == 'HNPC':
model = NeuralProcessHybridConstrained(z_output_sizes, enc_output_sizes, dec_output_sizes)
optimizer = tf.keras.optimizers.Adam(1e-3)
# Callbacks
model_path = f'.data/{args.model}_model_{args.task}_context_{args.num_context}_uniform_sampling_{args.uniform_sampling}/' + "cp-{epoch:04d}.ckpt"
time = datetime.now().strftime('%Y%m%d-%H%M%S')
log_dir = os.path.join('.', 'logs', args.model, args.task, time)
writer = tf.summary.create_file_writer(log_dir)
plot_clbk = PlotCallback(logdir=log_dir, ds=test_ds, task=args.task)
cp_callback = tf.keras.callbacks.ModelCheckpoint(filepath=model_path,
save_weights_only=True)
callbacks = [plot_clbk, cp_callback]
for callback in callbacks: callback.model = model
if args.model == 'HNPC':
model.writer = writer
# -----------------------------------------------------------------------------------------------------------
def train_step(model, x, optimizer):
"""Executes one training step and returns the loss.
This function computes the loss and gradients, and uses the latter to
update the model's parameters.
"""
with tf.GradientTape() as tape:
loss = model.compute_loss(x)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
return tf.math.reduce_mean(loss)
#%%
# ============================ Training Loop ===========================================================
epochs = args.epochs
for epoch in range(1, epochs + 1):
with tqdm(total=TRAINING_ITERATIONS, unit='batch') as tepoch:
tepoch.set_description(f"Epoch {epoch}")
# ------------------------------- Training --------------------------------------------
train_loss = tf.keras.metrics.Mean()
for idx, train_x in enumerate(train_ds):
loss = train_step(model, train_x, optimizer)
# VVVVVVVVVVVVVVVVVVV Logging VVVVVVVVVVVVVVVVVVVVVVVVVVVVVVVVVVVV
train_loss(loss)
tepoch.set_postfix({'Batch': idx, 'Train Loss': train_loss.result().numpy()})
tepoch.update(1)
with writer.as_default():
tf.summary.scalar('train_loss', train_loss.result(), step=epoch*TRAINING_ITERATIONS + idx)
# ------------------------------ Testing -----------------------------------------------
test_loss = tf.keras.metrics.Mean()
for idx, test_x in enumerate(test_ds):
loss = model.compute_loss(test_x)
# VVVVVVVVVVVVVVVVVVV Logging VVVVVVVVVVVVVVVVVVVVVVVVVVVVVVVVVVVV
test_loss(loss)
tepoch.set_postfix({'Batch': idx, 'Test Loss': test_loss.result().numpy()})
with writer.as_default():
tf.summary.scalar('test_loss', test_loss.result(), step=epoch*TEST_ITERATIONS + idx)
# ---------------------- Logging Prior & Posterior Distribution -----------------------------------
if LOG_PRIORS and args.model != 'CNP':
(context_x, context_y, target_x), target_y = next(iter(test_ds))
prior_context = tf.concat((context_x, context_y), axis=2)
prior = model.z_encoder_latent(prior_context)
target_context = tf.concat((target_x, target_y), axis=2)
posterior = model.z_encoder_latent(target_context)
with writer.as_default():
tf.summary.histogram('prior/mu', prior.mean(), epoch)
tf.summary.histogram('prior/sigma', prior.stddev(), epoch)
tf.summary.histogram('posterior/mu', posterior.mean(), epoch)
tf.summary.histogram('posterior/sigma', posterior.stddev(), epoch)
# ------------- Some callbacks -----------------------------------------------------------------------
tepoch.set_postfix({'Train loss': train_loss.result().numpy(), 'Test loss': test_loss.result().numpy()})
for callback in callbacks: callback.on_epoch_end(epoch, logs={'loss': train_loss.result()})
writer.flush()
#%%