forked from hqucms/ParticleNet
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtf_keras_model.py
182 lines (151 loc) · 7.41 KB
/
tf_keras_model.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
import tensorflow as tf
from tensorflow import keras
# A shape is (N, P_A, C), B shape is (N, P_B, C)
# D shape is (N, P_A, P_B)
def batch_distance_matrix_general(A, B):
with tf.name_scope('dmat'):
r_A = tf.reduce_sum(A * A, axis=2, keepdims=True)
r_B = tf.reduce_sum(B * B, axis=2, keepdims=True)
m = tf.matmul(A, tf.transpose(B, perm=(0, 2, 1)))
D = r_A - 2 * m + tf.transpose(r_B, perm=(0, 2, 1))
return D
def knn(num_points, k, topk_indices, features):
# topk_indices: (N, P, K)
# features: (N, P, C)
with tf.name_scope('knn'):
queries_shape = tf.shape(features)
batch_size = queries_shape[0]
batch_indices = tf.tile(tf.reshape(tf.range(batch_size), (-1, 1, 1, 1)), (1, num_points, k, 1))
indices = tf.concat([batch_indices, tf.expand_dims(topk_indices, axis=3)], axis=3) # (N, P, K, 2)
return tf.gather_nd(features, indices)
def edge_conv(points, features, num_points, K, channels, with_bn=True, activation='relu', pooling='average', name='edgeconv'):
"""EdgeConv
Args:
K: int, number of neighbors
in_channels: # of input channels
channels: tuple of output channels
pooling: pooling method ('max' or 'average')
Inputs:
points: (N, P, C_p)
features: (N, P, C_0)
Returns:
transformed points: (N, P, C_out), C_out = channels[-1]
"""
with tf.name_scope('edgeconv'):
# distance
D = batch_distance_matrix_general(points, points) # (N, P, P)
_, indices = tf.nn.top_k(-D, k=K + 1) # (N, P, K+1)
indices = indices[:, :, 1:] # (N, P, K)
fts = features
knn_fts = knn(num_points, K, indices, fts) # (N, P, K, C)
knn_fts_center = tf.tile(tf.expand_dims(fts, axis=2), (1, 1, K, 1)) # (N, P, K, C)
knn_fts = tf.concat([knn_fts_center, tf.subtract(knn_fts, knn_fts_center)], axis=-1) # (N, P, K, 2*C)
x = knn_fts
for idx, channel in enumerate(channels):
x = keras.layers.Conv2D(channel, kernel_size=(1, 1), strides=1, data_format='channels_last',
use_bias=False if with_bn else True, kernel_initializer='glorot_normal', name='%s_conv%d' % (name, idx))(x)
if with_bn:
x = keras.layers.BatchNormalization(name='%s_bn%d' % (name, idx))(x)
if activation:
x = keras.layers.Activation(activation, name='%s_act%d' % (name, idx))(x)
if pooling == 'max':
fts = tf.reduce_max(x, axis=2) # (N, P, C')
else:
fts = tf.reduce_mean(x, axis=2) # (N, P, C')
# shortcut
sc = keras.layers.Conv2D(channels[-1], kernel_size=(1, 1), strides=1, data_format='channels_last',
use_bias=False if with_bn else True, kernel_initializer='glorot_normal', name='%s_sc_conv' % name)(tf.expand_dims(features, axis=2))
if with_bn:
sc = keras.layers.BatchNormalization(name='%s_sc_bn' % name)(sc)
sc = tf.squeeze(sc, axis=2)
if activation:
return keras.layers.Activation(activation, name='%s_sc_act' % name)(sc + fts) # (N, P, C')
else:
return sc + fts
def _particle_net_base(points, features=None, mask=None, setting=None, name='particle_net'):
# points : (N, P, C_coord)
# features: (N, P, C_features), optional
# mask: (N, P, 1), optinal
with tf.name_scope(name):
if features is None:
features = points
if mask is not None:
mask = tf.cast(tf.not_equal(mask, 0), dtype='float32') # 1 if valid
coord_shift = tf.multiply(999., tf.cast(tf.equal(mask, 0), dtype='float32')) # make non-valid positions to 99
fts = tf.squeeze(keras.layers.BatchNormalization(name='%s_fts_bn' % name)(tf.expand_dims(features, axis=2)), axis=2)
for layer_idx, layer_param in enumerate(setting.conv_params):
K, channels = layer_param
pts = tf.add(coord_shift, points) if layer_idx == 0 else tf.add(coord_shift, fts)
fts = edge_conv(pts, fts, setting.num_points, K, channels, with_bn=True, activation='relu',
pooling=setting.conv_pooling, name='%s_%s%d' % (name, 'EdgeConv', layer_idx))
if mask is not None:
fts = tf.multiply(fts, mask)
pool = tf.reduce_mean(fts, axis=1) # (N, C)
if setting.fc_params is not None:
x = pool
for layer_idx, layer_param in enumerate(setting.fc_params):
units, drop_rate = layer_param
x = keras.layers.Dense(units, activation='relu')(x)
if drop_rate is not None and drop_rate > 0:
x = keras.layers.Dropout(drop_rate)(x)
out = keras.layers.Dense(setting.num_class, activation='softmax')(x)
return out # (N, num_classes)
else:
return pool
class _DotDict:
pass
def get_particle_net(num_classes, input_shapes):
r"""ParticleNet model from `"ParticleNet: Jet Tagging via Particle Clouds"
<https://arxiv.org/abs/1902.08570>`_ paper.
Parameters
----------
num_classes : int
Number of output classes.
input_shapes : dict
The shapes of each input (`points`, `features`, `mask`).
"""
setting = _DotDict()
setting.num_class = num_classes
# conv_params: list of tuple in the format (K, (C1, C2, C3))
setting.conv_params = [
(16, (64, 64, 64)),
(16, (128, 128, 128)),
(16, (256, 256, 256)),
]
# conv_pooling: 'average' or 'max'
setting.conv_pooling = 'average'
# fc_params: list of tuples in the format (C, drop_rate)
setting.fc_params = [(256, 0.1)]
setting.num_points = input_shapes['points'][0]
points = keras.Input(name='points', shape=input_shapes['points'])
features = keras.Input(name='features', shape=input_shapes['features']) if 'features' in input_shapes else None
mask = keras.Input(name='mask', shape=input_shapes['mask']) if 'mask' in input_shapes else None
outputs = _particle_net_base(points, features, mask, setting, name='ParticleNet')
return keras.Model(inputs=[points, features, mask], outputs=outputs, name='ParticleNet')
def get_particle_net_lite(num_classes, input_shapes):
r"""ParticleNet-Lite model from `"ParticleNet: Jet Tagging via Particle Clouds"
<https://arxiv.org/abs/1902.08570>`_ paper.
Parameters
----------
num_classes : int
Number of output classes.
input_shapes : dict
The shapes of each input (`points`, `features`, `mask`).
"""
setting = _DotDict()
setting.num_class = num_classes
# conv_params: list of tuple in the format (K, (C1, C2, C3))
setting.conv_params = [
(7, (32, 32, 32)),
(7, (64, 64, 64)),
]
# conv_pooling: 'average' or 'max'
setting.conv_pooling = 'average'
# fc_params: list of tuples in the format (C, drop_rate)
setting.fc_params = [(128, 0.1)]
setting.num_points = input_shapes['points'][0]
points = keras.Input(name='points', shape=input_shapes['points'])
features = keras.Input(name='features', shape=input_shapes['features']) if 'features' in input_shapes else None
mask = keras.Input(name='mask', shape=input_shapes['mask']) if 'mask' in input_shapes else None
outputs = _particle_net_base(points, features, mask, setting, name='ParticleNet')
return keras.Model(inputs=[points, features, mask], outputs=outputs, name='ParticleNet')