-
Notifications
You must be signed in to change notification settings - Fork 24
/
train_r3dcnn.py
144 lines (131 loc) · 7.35 KB
/
train_r3dcnn.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
#!/usr/bin/python
# -*- coding: utf-8 -*-
from __future__ import absolute_import;
from __future__ import division;
from __future__ import print_function;
import os;
import numpy as np;
import tensorflow as tf;
#os.environ['CUDA_VISIBLE_DEVICES'] = ''
batch_size = 4;
class_num = 25;
def main(unused_argv):
gesture_classifier = tf.estimator.Estimator(model_fn = action_model_fn, model_dir = "gesture_classifier_model");
tf.logging.set_verbosity(tf.logging.DEBUG);
logging_hook = tf.train.LoggingTensorHook(tensors = {"loss":"loss"}, every_n_iter = 1);
gesture_classifier.train(input_fn = train_input_fn,steps = 200000,hooks = [logging_hook]);
eval_results = gesture_classifier.evaluate(input_fn = eval_input_fn,steps = 1);
print(eval_results);
def parse_function(serialized_example):
feature = tf.parse_single_example(
serialized_example,
features = {
'clips': tf.FixedLenFeature((),dtype = tf.string, default_value = ''),
'label': tf.FixedLenFeature((),dtype = tf.int64, default_value = 0)
}
);
clips = tf.decode_raw(feature['clips'],out_type = tf.uint8);
clips = tf.reshape(clips,[10,8,112,112,3]);
clips = tf.cast(clips, dtype = tf.float32);
sequence_length = tf.constant(clips.get_shape().as_list()[0],dtype = tf.int32);
label = tf.cast(feature['label'], dtype = tf.int32);
label = tf.reshape(label,[1]);
idx = tf.where(tf.not_equal(label,-1));
label = tf.SparseTensor(indices = idx, values = tf.gather_nd(label,idx), dense_shape = tf.cast(label.get_shape(),dtype = tf.int64));
return dict(zip(['data','sequence_lengths'],[clips,sequence_length])),label;
def train_input_fn():
dataset = tf.data.TFRecordDataset(['trainset.tfrecord']);
dataset = dataset.map(parse_function);
dataset = dataset.shuffle(buffer_size = 512);
dataset = dataset.batch(batch_size);
dataset = dataset.repeat(None);
iterator = dataset.make_one_shot_iterator();
features, labels = iterator.get_next();
return features, labels;
def eval_input_fn():
dataset = tf.data.TFRecordDataset(['testset.tfrecord']);
dataset = dataset.map(parse_function);
dataset = dataset.shuffle(buffer_size = 512);
dataset = dataset.batch(batch_size);
dataset = dataset.repeat(None);
iterator = dataset.make_one_shot_iterator();
features, labels = iterator.get_next();
return features, labels;
def action_model_fn(features, labels, mode):
#data.shape = [batch_size = ?, time_steps = 10, depth = 8, height = 112, width = 112, channel = 3]
#labels.shape = [batch_size = ?,label_length = 1]
data = features["data"];
sequence_lengths = features["sequence_lengths"];
timesteps = data.get_shape().as_list()[1];
#create lstm operator object
stacked_lstm = tf.nn.rnn_cell.MultiRNNCell([tf.nn.rnn_cell.BasicLSTMCell(512) for _ in range(1)]);
conv_input = tf.reshape(data,[-1,8,112,112,3]);
#layer 1
c1 = tf.layers.conv3d(conv_input,filters = 64, kernel_size = [3,3,3], padding = "same");
b1 = tf.contrib.layers.layer_norm(c1,activation_fn = tf.nn.relu, trainable = mode == tf.estimator.ModeKeys.TRAIN);
p1 = tf.layers.max_pooling3d(b1,pool_size = [1,2,2], strides = [1,2,2], padding = "same");
#layer 2
c2 = tf.layers.conv3d(p1,filters = 128, kernel_size = [3,3,3], padding = "same");
b2 = tf.contrib.layers.layer_norm(c2,activation_fn = tf.nn.relu, trainable = mode == tf.estimator.ModeKeys.TRAIN);
p2 = tf.layers.max_pooling3d(b2,pool_size = [2,2,2], strides = [2,2,2], padding = "same");
#layer 3
c3a = tf.layers.conv3d(p2,filters = 256, kernel_size = [3,3,3], padding = "same");
b3a = tf.contrib.layers.layer_norm(c3a,activation_fn = tf.nn.relu, trainable = mode == tf.estimator.ModeKeys.TRAIN);
c3b = tf.layers.conv3d(b3a,filters = 256, kernel_size = [3,3,3], padding = "same");
b3b = tf.contrib.layers.layer_norm(c3b,activation_fn = tf.nn.relu, trainable = mode == tf.estimator.ModeKeys.TRAIN);
p3 = tf.layers.max_pooling3d(b3b,pool_size = [2,2,2], strides = [2,2,2], padding = "same");
#layer 4
c4a = tf.layers.conv3d(p3,filters = 512, kernel_size = [3,3,3], padding = "same");
b4a = tf.contrib.layers.layer_norm(c4a,activation_fn = tf.nn.relu, trainable = mode == tf.estimator.ModeKeys.TRAIN);
c4b = tf.layers.conv3d(b4a,filters = 512, kernel_size = [3,3,3], padding = "same");
b4b = tf.contrib.layers.layer_norm(c4b,activation_fn = tf.nn.relu, trainable = mode == tf.estimator.ModeKeys.TRAIN);
p4 = tf.layers.max_pooling3d(b4b,pool_size = [2,2,2], strides = [2,2,2], padding = "same");
#layer 5
c5a = tf.layers.conv3d(p4,filters = 512, kernel_size = [3,3,3], padding = "same");
b5a = tf.contrib.layers.layer_norm(c5a,activation_fn = tf.nn.relu, trainable = mode == tf.estimator.ModeKeys.TRAIN);
c5b = tf.layers.conv3d(b5a,filters = 512, kernel_size = [3,3,3], padding = "same");
b5b = tf.contrib.layers.layer_norm(c5b,activation_fn = tf.nn.relu, trainable = mode == tf.estimator.ModeKeys.TRAIN);
#load c3d pretrained parameters
tf.train.init_from_checkpoint("action_classifier_model",{v.name.split(':')[0]: v for v in tf.contrib.framework.get_variables_to_restore()});
#the last pooling layer of C3D is removed because the input clips length become 8 frames
#r5b.shape = [batch_size * time_steps = ?*10, depth = 1, height = 7, width = 7, channel = 512]
f = tf.layers.flatten(b5b);
d1 = tf.layers.dense(f,units = 512, activation = tf.nn.relu);
dp1 = tf.layers.dropout(d1,training = mode == tf.estimator.ModeKeys.TRAIN);
d2 = tf.layers.dense(dp1,units = 512, activation = tf.nn.relu);
dp2 = tf.layers.dropout(d2,training = mode == tf.estimator.ModeKeys.TRAIN);
#dp2.shape = [batch_size = ?, time_steps = 10, feature_dim = 512]
lstm_input = tf.reshape(dp2,[-1,timesteps,512])
output, _ = tf.nn.dynamic_rnn(stacked_lstm, lstm_input, sequence_length = sequence_lengths, time_major = False, dtype = tf.float32);
output = tf.reshape(output,[-1,512]);
#logits.shape = [batch_size * time_steps = ?*10, num_classes]
logits = tf.layers.dense(output,units = class_num);
#logits.shape = [batch_size = ?, times_steps = 10, num_classes]
logits = tf.reshape(logits,[-1,timesteps,class_num]);
#predict mode
if mode == tf.estimator.ModeKeys.PREDICT:
#logits.shape = [times_steps = 10, batch_size = ?, num_classes]
logits = tf.transpose(logits,[1,0,2]);
decoded,_ = tf.nn.ctc_beam_search_decoder(logits,sequence_lengths);
tf.Print(decoded[0].dense_shape,[decoded[0].dense_shape],message = "shape = ");
predictions = tf.sparse_tensor_to_dense(decoded[0]);
return tf.estimator.EstimatorSpec(mode = mode,predictions = predictions);
#train mode
if mode == tf.estimator.ModeKeys.TRAIN:
loss = tf.nn.ctc_loss(labels,logits,sequence_lengths,time_major = False);
loss = tf.reduce_mean(loss,name = "loss");
optimizer = tf.train.AdamOptimizer(1e-4);
train_op = optimizer.minimize(loss = loss, global_step = tf.train.get_global_step());
return tf.estimator.EstimatorSpec(mode = mode, loss = loss, train_op = train_op);
#eval mode
if mode == tf.estimator.ModeKeys.EVAL:
loss = tf.nn.ctc_loss(labels,logits,sequence_lengths,time_major = False);
loss = tf.reduce_mean(loss,name = "loss");
#logits.shape = [times_steps = 10, batch_size = ?, num_classes]
logits = tf.transpose(logits,[1,0,2]);
decoded,_ = tf.nn.ctc_beam_search_decoder(logits,sequence_lengths);
eval_metric_ops = {"mean_edit_distance": tf.metrics.mean(tf.edit_distance(tf.cast(decoded[0],tf.int32),labels))};
return tf.estimator.EstimatorSpec(mode = mode, loss = loss, eval_metric_ops = eval_metric_ops);
raise Exception('Unknown mode of estimator!');
if __name__ == "__main__":
tf.app.run();