-
Notifications
You must be signed in to change notification settings - Fork 2
/
PVT.py
177 lines (154 loc) · 8.82 KB
/
PVT.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
import tensorflow as tf
from model.droppath import droppath
import numpy as np
import tensorflow_addons as tfa
import logging
def get_patch_embed_model(img_size,first_level_patch_size=4,embed_dims=[64, 128, 320, 512]):
num_level = len(embed_dims)
patch_size = [first_level_patch_size] + [2 for _ in range(num_level-1)]
embed_dims = [3] + embed_dims
outputs = []
feat_size_list = []
feat_size = np.array(img_size)
for i in range(1,len(embed_dims)):
input = tf.keras.layers.Input([feat_size[0]*feat_size[1],embed_dims[i-1]])
x = tf.reshape(input,(tf.shape(input)[0],feat_size[0],feat_size[1],embed_dims[i-1]))
x = tf.keras.layers.Conv2D(embed_dims[i],patch_size[i-1],patch_size[i-1])(x)
x = tf.reshape(x,(tf.shape(x)[0],-1,embed_dims[i]))
x = tf.keras.layers.LayerNormalization(epsilon=1e-6)(x)
outputs.append(tf.keras.Model(inputs=input,outputs=x))
feat_size = feat_size // patch_size[i - 1]
feat_size_list.append(feat_size.tolist())
return outputs,feat_size_list
def get_attention_model(feat_size,num_patch,embed_dims,num_heads,sr_ratio,attention_drop_rate,drop_rate,name=None):
input = tf.keras.layers.Input((num_patch, embed_dims))
q_x = tf.keras.layers.Dense(embed_dims, use_bias=True)(input)
q_x = tf.reshape(q_x, (tf.shape(input)[0], -1, num_heads, embed_dims // num_heads))
q = tf.transpose(q_x, perm=(0, 2, 1, 3))
if sr_ratio > 1:
x = tf.reshape(input, (tf.shape(input)[0], feat_size[0], feat_size[1], input.shape[-1]))
x = tf.keras.layers.Conv2D(embed_dims, sr_ratio, sr_ratio)(x)
x = tf.keras.layers.LayerNormalization(epsilon=1e-6)(x)
x = tf.keras.layers.Dense(embed_dims * 2, use_bias=True)(x)
x = tf.reshape(x, (tf.shape(x)[0], -1, 2, num_heads, embed_dims // num_heads))
kv = tf.transpose(x, perm=(2, 0, 3, 1, 4))
k, v = kv[0], kv[1]
else:
k = tf.keras.layers.Dense(embed_dims, use_bias=True)(input)
k = tf.reshape(k, (tf.shape(k)[0], -1, num_heads, embed_dims // num_heads))
k = tf.transpose(k, perm=(0, 2, 1, 3))
v = tf.keras.layers.Dense(embed_dims, use_bias=True)(input)
v = tf.reshape(v, (tf.shape(v)[0], -1, num_heads, embed_dims // num_heads))
v = tf.transpose(v, perm=(0, 2, 1, 3))
head_dim = embed_dims // num_heads
x = tf.matmul(q, tf.transpose(k, perm=(0, 1, 3, 2))) / tf.math.sqrt(tf.cast(head_dim, tf.dtypes.float32))
x = tf.math.softmax(x, axis=-1)
score = tf.keras.layers.Dropout(attention_drop_rate)(x)
x = tf.matmul(score, v)
x = tf.transpose(x, perm=(0, 2, 1, 3))
x = tf.reshape(x, (tf.shape(x)[0], -1, embed_dims))
x = tf.keras.layers.Dense(embed_dims, use_bias=True)(x)
output = tf.keras.layers.Dropout(drop_rate)(x)
return tf.keras.Model(input, output, name=name)
def get_block_attention_model(depth,block_drop_path_rate,mlp_ratio,feat_size,num_patch,embed_dims,num_heads,sr_ratio,attention_drop_rate,drop_rate,name=None):
block_index = 0
block_input = input = tf.keras.layers.Input([num_patch, embed_dims])
for _ in range(depth):
x = tf.keras.layers.LayerNormalization(epsilon=1e-6)(input)
x = get_attention_model(feat_size,num_patch,embed_dims,num_heads,sr_ratio,attention_drop_rate,drop_rate)(x)
attention_output = input + droppath(x, drop_prob=block_drop_path_rate[block_index])
x = tf.keras.layers.LayerNormalization(epsilon=1e-6)(attention_output)
num_mlp_hidden_layers = int(input.shape[-1] * mlp_ratio)
x = tf.keras.layers.Dense(num_mlp_hidden_layers, activation=tfa.activations.gelu)(x)
x = tf.keras.layers.Dropout(drop_rate)(x)
x = tf.keras.layers.Dense(input.shape[-1])(x)
x = tf.keras.layers.Dropout(drop_rate)(x)
mlp_output = attention_output + droppath(x, drop_prob=block_drop_path_rate[block_index])
input = mlp_output
block_index += 1
return tf.keras.Model(block_input, mlp_output, name=name)
class AddClsToken(tf.keras.layers.Layer):
def __init__(self):
super(AddClsToken, self).__init__()
def build(self, input_shape):
self.cls_token = self.add_weight(shape=(1,1,input_shape[-1]),
initializer=tf.keras.initializers.TruncatedNormal(mean=0.0, stddev=0.02, seed=None),trainable=True,dtype=tf.dtypes.float32)
def call(self, x):
cls_token = tf.broadcast_to(self.cls_token, [tf.shape(x)[0], 1, self.cls_token.shape[-1]])
return tf.concat([cls_token,x],axis=1)
class AddPosEmbed(tf.keras.layers.Layer):
def __init__(self,img_len):
super(AddPosEmbed, self).__init__()
self.img_len = img_len
def build(self, input_shape):
self.pos_embed = self.add_weight(shape=[1,self.img_len,input_shape[-1]],
initializer=tf.keras.initializers.TruncatedNormal(mean=0.0, stddev=0.02, seed=None),trainable=True,dtype=tf.dtypes.float32)
def call(self, x):
return x+self.pos_embed
def get_config(self):
config = super().get_config().copy()
config.update({
'img_len': self.img_len,
})
return config
def get_pvt(img_size,num_classes,block_depth,mlp_ratio,drop_path_rate,first_level_patch_size,embed_dims,num_heads,sr_ratio,attention_drop_rate,drop_rate):
block_drop_path_rate = np.linspace(0, drop_path_rate, sum(block_depth))
block_depth_index = 0
input = tf.keras.layers.Input((img_size[0], img_size[1], 3))
patch_embed_model_list,feat_size_list = get_patch_embed_model(img_size,first_level_patch_size,embed_dims)
x = input
x = tf.reshape(x,[tf.shape(input)[0],-1,tf.shape(input)[-1]])
for i in range(len(patch_embed_model_list)):
x = patch_embed_model_list[i](x)
if i == len(patch_embed_model_list)-1:
num_patch = feat_size_list[i][0] * feat_size_list[i][1] + 1
x = AddClsToken()(x)
x = AddPosEmbed(num_patch)(x)
else:
num_patch = feat_size_list[i][0] * feat_size_list[i][1]
x = AddPosEmbed(num_patch)(x)
x = tf.keras.layers.Dropout(drop_rate)(x)
x = get_block_attention_model(block_depth[i],block_drop_path_rate[block_depth_index:block_depth_index+block_depth[i]],mlp_ratio[i],
feat_size_list[i],num_patch,embed_dims[i],num_heads[i],sr_ratio[i],attention_drop_rate,drop_rate)(x)
block_depth_index+=block_depth[i]
x = tf.keras.layers.LayerNormalization(epsilon=1e-6)(x)
output = tf.keras.layers.Dense(num_classes, activation='softmax')(x[:,0])
return tf.keras.Model(input,output,name='PVT')
class PVTNet():
def __init__(self, img_size=224, classes=2,type='tiny',pretrain=None):
self.img_size = img_size
self.classes = classes
self.pretrain = pretrain
self.type = type
def get_model(self):
if self.type=='tiny':
model = get_pvt(img_size=(self.img_size,self.img_size),num_classes=self.classes,
first_level_patch_size=4,
block_depth=[2, 2, 2, 2], mlp_ratio=[8, 8, 4, 4],
embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8],sr_ratio=[8, 4, 2, 1],
drop_path_rate=0.1,attention_drop_rate=0.0,drop_rate=0.02)
elif self.type=='small':
model = get_pvt(img_size=(self.img_size,self.img_size),num_classes=self.classes,
first_level_patch_size=4,
block_depth=[3, 4, 6, 3], mlp_ratio=[8, 8, 4, 4],
embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8],sr_ratio=[8, 4, 2, 1],
drop_path_rate=0.1,attention_drop_rate=0.0,drop_rate=0.0)
elif self.type=='medium':
model = get_pvt(img_size=(self.img_size,self.img_size),num_classes=self.classes,
first_level_patch_size=4,
block_depth=[3, 4, 18, 3], mlp_ratio=[8, 8, 4, 4],
embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8],sr_ratio=[8, 4, 2, 1],
drop_path_rate=0.1,attention_drop_rate=0.0,drop_rate=0.0)
elif self.type=='large':
model = get_pvt(img_size=(self.img_size,self.img_size),num_classes=self.classes,
first_level_patch_size=4,
block_depth=[3, 8, 27, 3], mlp_ratio=[8, 8, 4, 4],
embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8],sr_ratio=[8, 4, 2, 1],
drop_path_rate=0.1,attention_drop_rate=0.0,drop_rate=0.02)
else:
raise ValueError('Unsupported PVT type:{}'.format(self.type))
try:
model.load_weights(self.pretrain)
except:
logging.warning('Failed to load weights file:{}'.format(self.pretrain))
return model