-
Notifications
You must be signed in to change notification settings - Fork 0
/
DSF.py
237 lines (208 loc) · 8.96 KB
/
DSF.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
import argparse
import numpy as np
import tensorflow as tf
import os
from tensorflow.keras import layers
import tensorflow_addons as tfa
parser = argparse.ArgumentParser(description="DSF")
parser.add_argument('--batch_size', type=int, default=512, help='batch size for training')
parser.add_argument('--learning_rate', type=float, default=0.00001, help='learning rate for training')
parser.add_argument('--weight_decay', type=float, default=0.00001, help='weight decay for training')
parser.add_argument('--epochs', type=int, default=500, help='number of epochs for training')
parser.add_argument('--patience', type=int, default=6, help='early-stopping patience')
parser.add_argument('--num_workers', type=int, default=8, help='number of workers for the train loader')
parser.add_argument('--num_patches', type=int, default=9, help='DSF patch number')
parser.add_argument('--projection_dim', type=int, default=500, help='linear projection dimension')
parser.add_argument('--num_heads', type=int, default=10, help='number of heads')
parser.add_argument('--transformer_layers', type=int, default=2, help='number of transformer layers')
parser.add_argument('--mlp_head_units_0', type=int, default=512, help='mlp head 0 units')
parser.add_argument('--mlp_head_units_1', type=int, default=128, help='mlp head 1 units')
args = parser.parse_args()
# set random seed
np.random.seed(0)
tf.random.set_seed(0)
# load dataset
x_train = # training DDMs
y_train = # training label
print(f"x_train shape: {x_train.shape} - y_train shape: {y_train.shape}")
x_valid = # validation DDMs
y_valid = # validation label
print(f"x_valid shape: {x_valid.shape} - y_valid shape: {y_valid.shape}")
# model save directory
weights_dir = 'logs_DSF/weights/'
if not os.path.exists(weights_dir):
os.makedirs(weights_dir)
board_dir = 'logs_DSF/events/'
if not os.path.exists(board_dir):
os.makedirs(board_dir)
# tr block setup
transformer_units = [args.projection_dim * 4, args.projection_dim]
mlp_head_units = [args.mlp_head_units_0, args.mlp_head_units_1]
initializer = tf.keras.initializers.GlorotUniform()
# mlp sublayer
def mlp(x, hidden_units, dropout_rate):
for units in hidden_units:
x = layers.Dense(units, activation=tf.nn.gelu, kernel_initializer=initializer)(x)
x = layers.Dropout(dropout_rate)(x)
return x
# ddm-wise tokenization
class Patches(layers.Layer):
def __init__(self, patch_size):
super(Patches, self).__init__()
def call(self, images):
batch_size = tf.shape(images)[0]
patches = tf.image.extract_patches(
images=images,
sizes=[1, 17, 11, 1],
strides=[1, 17, 11, 1],
rates=[1, 1, 1, 1],
padding="VALID",
)
patch_dims = patches.shape[-1]
patches = tf.reshape(patches, [batch_size, 9, patch_dims])
return patches
def get_config(self):
config = super().get_config()
return config
# DDA token
class DDAToken(layers.Layer):
def __init__(self):
super(DDAToken, self).__init__()
dda_init = tf.zeros_initializer()
self.hidden_size = 748 # 17*11*4
self.dda = tf.Variable(
name="dda",
initial_value=dda_init(shape=(1, 1, self.hidden_size), dtype="float32"),
trainable=True,
)
def call(self, inputs):
batch_size = tf.shape(inputs)[0]
dda_broadcasted = tf.cast(
tf.broadcast_to(self.dda, [batch_size, 1, self.hidden_size]),
dtype=inputs.dtype,
)
concat = tf.concat([dda_broadcasted, inputs], 1)
return concat
def get_config(self):
config = super().get_config()
return config
class PatchEncoder(layers.Layer):
def __init__(self, num_patches, projection_dim):
super(PatchEncoder, self).__init__()
self.num_patches = num_patches+1
self.projection = layers.Dense(units=projection_dim, kernel_initializer=initializer)
self.position_embedding = layers.Embedding(
input_dim=num_patches, output_dim=projection_dim
)
def call(self, patch):
positions = tf.range(start=0, limit=self.num_patches, delta=1)
encoded = self.projection(patch) + self.position_embedding(positions)
return encoded
def get_config(self):
config = super().get_config()
config.update({
"num_pathces": self.num_patches,
"projection_dim": self.projection,
})
return config
# msa sublayer
@tf.keras.utils.register_keras_serializable()
class MultiHeadSelfAttention(layers.Layer):
def __init__(self, *args, num_heads, **kwargs):
super().__init__(*args, **kwargs)
self.num_heads = num_heads
def build(self, input_shape):
hidden_size = input_shape[-1]
num_heads = self.num_heads
if hidden_size % num_heads != 0:
raise ValueError(
f"embedded dimension not divisible by number of heads"
)
self.hidden_size = hidden_size
self.projection_dim = hidden_size // num_heads
self.query_dense = tf.keras.layers.Dense(hidden_size, name="query")
self.key_dense = tf.keras.layers.Dense(hidden_size, name="key")
self.value_dense = tf.keras.layers.Dense(hidden_size, name="value")
self.combine_heads = tf.keras.layers.Dense(hidden_size, name="out")
def attention(self, query, key, value):
score = tf.matmul(query, key, transpose_b=True)
dim_key = tf.cast(tf.shape(key)[-1], score.dtype)
scaled_score = score / tf.math.sqrt(dim_key)
weights = tf.nn.softmax(scaled_score, axis=-1)
output = tf.matmul(weights, value)
return output, weights
def separate_heads(self, x, batch_size):
x = tf.reshape(x, (batch_size, -1, self.num_heads, self.projection_dim))
return tf.transpose(x, perm=[0, 2, 1, 3])
def call(self, inputs):
batch_size = tf.shape(inputs)[0]
query = self.query_dense(inputs)
key = self.key_dense(inputs)
value = self.value_dense(inputs)
query = self.separate_heads(query, batch_size)
key = self.separate_heads(key, batch_size)
value = self.separate_heads(value, batch_size)
attention, weights = self.attention(query, key, value)
attention = tf.transpose(attention, perm=[0, 2, 1, 3])
concat_attention = tf.reshape(attention, (batch_size, -1, self.hidden_size))
output = self.combine_heads(concat_attention)
return output, weights
def get_config(self):
config = super().get_config()
config.update({"num_heads": self.num_heads})
return config
@classmethod
def from_config(dda, config):
return dda(**config)
def ddm_seq_former():
inputs = layers.Input(shape=(51,33,4))
# create patches
patches = Patches()(inputs)
dda = DDAToken()(patches)
# encode patches
encoded_patches = PatchEncoder(args.num_patches, args.projection_dim)(dda)
# create tr block
for _ in range(args.transformer_layers):
# LN
x1 = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
# msa
attention_output, _ = MultiHeadSelfAttention(num_heads=args.num_heads)(x1)
# skip connection
x2 = layers.Add()([attention_output, encoded_patches])
# LN
x3 = layers.LayerNormalization(epsilon=1e-6)(x2)
# mlp
x3 = mlp(x3, hidden_units=transformer_units, dropout_rate=0.1)
# skip connection
encoded_patches = layers.Add()([x3, x2])
out_repre = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
out_repre = layers.Flatten()(out_repre)
out_repre = layers.Dropout(0.5)(out_repre)
# mlp
out_repre = mlp(out_repre, hidden_units=mlp_head_units, dropout_rate=0.5)
# ws
ws = layers.Dense(9, kernel_initializer=initializer)(out_repre)
# create model
model = tf.keras.Model(inputs=inputs, outputs=ws)
return model
print('---------Training----------')
dsf_model = ddm_seq_former()
optimizer = tfa.optimizers.AdamW(learning_rate=args.learning_rate,
weight_decay=args.weight_decay)
dsf_model.compile(optimizer=optimizer,
loss="mse",
metrics=[tf.keras.metrics.MeanSquaredError(name='MSE')])
callback = tf.keras.callbacks.EarlyStopping(monitor='val_loss',
patience=args.patience)
model_ckt = tf.keras.callbacks.ModelCheckpoint(filepath=os.path.join(weights_dir, 'weights.h5'),
verbose=1,
save_best_only=True)
tfboard = tf.keras.callbacks.TensorBoard(log_dir=board_dir,
write_graph=True,
write_images=True)
dsf_model.fit(x_train, y_train, batch_size=args.batch_size, epochs=args.epochs,
callbacks=[model_ckt, tfboard, callback],
validation_data=(x_valid, y_valid),
shuffle=True, workers=args.num_workers)
dsf_model.save(os.path.join(weights_dir, 'model-DSF.tf'))
print('---------Training Done---------')