-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_cfd.py
159 lines (118 loc) · 5.23 KB
/
train_cfd.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
"""Runs the learner/evaluator."""
import os
import pickle
import argparse
from tqdm import tqdm
import numpy as np
import tensorflow as tf
from tensorflow.keras.optimizers import Adam
import common
import core_model
import cfd_model
from dataset import load_dataset_train
import datetime
gpus = tf.config.experimental.list_physical_devices('GPU')
try:
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
except RuntimeError as e:
print(e)
def frame_to_graph(frame):
"""Builds input graph."""
# construct graph nodes
node_type = tf.one_hot(frame['node_type'][:, 0], common.NodeType.SIZE)
node_features = tf.concat([frame['velocity'], node_type], axis=-1)
# construct graph edges
senders, receivers = common.triangles_to_edges(frame['cells'])
relative_mesh_pos = (tf.gather(frame['mesh_pos'], senders) -
tf.gather(frame['mesh_pos'], receivers))
edge_features = tf.concat([
relative_mesh_pos,
tf.norm(relative_mesh_pos, axis=-1, keepdims=True)], axis=-1)
del frame['cells']
return node_features, edge_features, senders, receivers, frame
def build_model(model, optimizer, dataset, checkpoint=None):
"""Initialize the model"""
node_features, edge_features, senders, receivers, frame = next(iter(dataset))
graph = core_model.MultiGraph(node_features, edge_sets=[core_model.EdgeSet(edge_features, senders, receivers)])
# call the model once to process all input shapes
model.loss(graph, frame)
# get the number of trainable parameters
total = 0
for var in model.trainable_weights:
total += np.prod(var.shape)
print(f'Total trainable parameters: {total}')
if checkpoint:
opt_weights = np.load(f'{checkpoint}_optimizer.npy', allow_pickle=True)
dummy_grads = [tf.zeros_like(w) for w in model.trainable_weights]
optimizer.apply_gradients(zip(dummy_grads, model.trainable_weights))
# only now set the weights of the optimizer and model
optimizer.set_weights(opt_weights)
model.load_weights(checkpoint, by_name=True)
def train(data_path=os.path.join(os.path.dirname(__file__), 'datasets', 'cylinder_flow'),num_steps=1000000, checkpoint = None):
dataset = load_dataset_train(
path=data_path,
split='train',
fields=['velocity'],
add_history=False,
noise_scale=0.02,
noise_gamma=1.0
)
dataset = dataset.map(frame_to_graph, num_parallel_calls=8)
dataset = dataset.prefetch(16)
model = core_model.EncodeProcessDecode(
output_dims=2,
embed_dims=128,
num_layers=3,
num_iterations=15,
num_edge_types=1
)
model = cfd_model.CFDModel(model)
lr = tf.keras.optimizers.schedules.ExponentialDecay(1e-4, decay_steps=num_steps // 2, decay_rate=0.1)
optimizer = Adam(learning_rate=lr)
current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
train_log_dir = 'Visualization/logs/train/' + current_time
train_summary_writer = tf.summary.create_file_writer(train_log_dir)
# build the model
build_model(model, optimizer, dataset, checkpoint = checkpoint)
#@tf.function(jit_compile=True)
@tf.function(experimental_relax_shapes=True)
def warmup(graph, frame):
loss = model.loss(graph, frame)
return loss
#@tf.function(jit_compile=True)
@tf.function(experimental_relax_shapes=True)
def train_step(graph, frame):
with tf.GradientTape() as tape:
loss = model.loss(graph, frame)
grads = tape.gradient(loss, model.trainable_weights)
optimizer.apply_gradients(zip(grads, model.trainable_weights))
return loss
dataset_iter = iter(dataset)
train_loop = tqdm(range(num_steps))
moving_loss = 0
for s in train_loop:
node_features, edge_features, senders, receivers, frame = next(dataset_iter)
graph = core_model.MultiGraph(node_features, edge_sets=[core_model.EdgeSet(edge_features, senders, receivers)])
if s < 1000:
loss = warmup(graph, frame)
else:
loss = train_step(graph, frame)
moving_loss = 0.98 * moving_loss + 0.02 * loss
if s%500 == 0:
with train_summary_writer.as_default():
tf.summary.scalar('loss',loss,step = s) #s for training session
train_loop.set_description(f'Step {s}/{num_steps}, Loss {moving_loss:.5f}')
if s != 0 and s % 10000 == 0:
filename = f'cfd_weights-step{s:07d}-loss{moving_loss:.5f}.hdf5'
model.save_weights(os.path.join(os.path.dirname(__file__), 'datasets', 'cylinder_flow', 'checkpoints', filename))
np.save(os.path.join(os.path.dirname(__file__), 'datasets', 'cylinder_flow', 'checkpoints', f'{filename}_optimizer.npy'), optimizer.get_weights())
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--checkpoint", "-c", help="Path to checkpoint file used to resume training")
parser.add_argument("--data_path", help="Path to dataset")
parser.add_argument("--num_steps", type=int, help="Number of itterations to train (default :1e6)")
args = parser.parse_args()
train(args.data_path, num_steps=args.num_steps, checkpoint=args.checkpoint)
if __name__ == '__main__':
main()