forked from lmb-freiburg/hand3d
-
Notifications
You must be signed in to change notification settings - Fork 30
/
NeuralNetwork.py
171 lines (135 loc) · 6.71 KB
/
NeuralNetwork.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
import tensorflow as tf
import numpy as np
import os
import argparse
import shutil
import matplotlib.pyplot as plt
from sklearn.utils import shuffle
from tqdm import tqdm
from ReadData import read_data_files, split_data
N_KEYPOINTS = 63
def parse_args():
parser = argparse.ArgumentParser(description = 'Train neural network')
# You have to give individual csv files based on the order in which position id of poses are made.
parser.add_argument('csv_files', help = 'Comma separated list of paths of training files', type = str)
parser.add_argument('--output-path', dest = 'output_path', type = str, default = None,
help = 'Path of folder where to store the trained model')
parser.add_argument('--train-test-split', dest = 'train_test_split', default = 0.85, type = float,
help = 'Ratio of train to test dataset(0-1)')
parser.add_argument('--learning-rate', dest = 'learning_rate', default = 0.0001, type = float,
help = 'Learning rate of the model')
parser.add_argument('--epochs', dest = 'epochs', type = int, default = 30,
help = 'Number of epochs to train data on')
parser.add_argument('--batch-size', dest = 'batch_size', type = int, default = 64,
help = 'Batch size while training')
parser.add_argument('--max-samples', dest = 'max_samples', type = int, default = 800,
help = 'Maximum number of samples per class allowed')
parser.add_argument('--break-training-at', dest = 'break_training_at', default = 0.96, type = float,
help = 'Break out of training when test accuracy increases than')
args = parser.parse_args()
return args
def fc_layer(X, weight, bias, should_activate = True):
fc = tf.nn.xw_plus_b(X, weight, bias)
if should_activate:
fc = tf.nn.relu(fc)
return fc
def pose_classify_net(X, n_classes):
with tf.variable_scope('Layer1'):
weight = tf.get_variable('weight1', initializer = tf.truncated_normal((N_KEYPOINTS, 22), mean = 0.0, stddev = 0.1))
bias = tf.get_variable('bias1', initializer = tf.ones((22)))
fc1 = fc_layer(X, weight, bias)
with tf.variable_scope('Layer2'):
weight = tf.get_variable('weight2', initializer = tf.truncated_normal((22, n_classes), mean = 0.0, stddev = 0.1))
bias = tf.get_variable('bias2', initializer = tf.ones((n_classes)))
fc2 = fc_layer(fc1, weight, bias, should_activate = False)
return fc2
def evaluate(X_data, y_data, accuracy, X, y, batch_size):
sum_at = 0.0
for offset in range(0, len(y_data), batch_size):
end = offset + batch_size
batch_X, batch_y = X_data[offset: end], y_data[offset: end]
eval_sess = tf.get_default_session()
acc = eval_sess.run(accuracy, feed_dict = {X: batch_X, y: batch_y})
sum_at += acc * len(batch_y)
return sum_at / len(y_data)
def build_and_train_network(train_data, test_data, parameters, output_at):
X_train, y_train = train_data
X_test, y_test = test_data
tf.reset_default_graph()
X = tf.placeholder(tf.float32, shape = (None, N_KEYPOINTS))
y = tf.placeholder(tf.int32, shape = (None))
count = tf.get_variable('count', shape = [], dtype = tf.int32, trainable = False)
one_hot_y = tf.one_hot(y, parameters['n_classes'])
logits = pose_classify_net(X, parameters['n_classes'])
loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels = one_hot_y, logits = logits))
optimizer = tf.train.AdamOptimizer(learning_rate = parameters['learning_rate']).minimize(loss_op, global_step = count)
accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(one_hot_y, 1), tf.argmax(logits, 1)), tf.float32))
saver = tf.train.Saver()
batch_size = parameters['batch_size']
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
train_accuracies = []
test_accuracies = []
for epoch_no in range(parameters['epochs']):
X_train, y_train = shuffle(X_train, y_train)
with tqdm(total = len(y_train)) as pbar:
for offset in range(0, len(y_train), batch_size):
end = offset + batch_size
batch_X, batch_y = X_train[offset: end], y_train[offset: end]
sess.run(optimizer, feed_dict = {X: batch_X, y: batch_y})
pbar.update(len(batch_y))
train_acc = evaluate(X_train, y_train, accuracy, X, y, batch_size)
test_acc = evaluate(X_test, y_test, accuracy, X, y, batch_size)
train_accuracies.append(train_acc)
test_accuracies.append(test_acc)
print('Epoch: {}'.format(epoch_no + 1))
print('Train accuracy: {:.3f}, Test accuracy: {:.3f}'.format(train_acc, test_acc))
if test_acc >= parameters['break_training_at']:
print('Breaking out of training')
break
shutil.rmtree(output_at, ignore_errors = True)
os.mkdir(output_at)
saver.save(sess, '{}/Checkpoint'.format(output_at))
print('Saved checkpoint')
plt.plot(range(len(train_accuracies)), train_accuracies, 'b-', label = 'Train')
plt.plot(range(len(test_accuracies)), test_accuracies, 'r-', label = 'Test')
plt.savefig('{}/accuracy.png'.format(output_at))
def get_value_by_name(name):
tensor = tf.get_default_graph().get_tensor_by_name('{}:0'.format(name))
sess = tf.get_default_session()
val = sess.run(tensor)
reqd_tensor = tf.constant(val, name = name)
return reqd_tensor
def export_created_graph(output_at):
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(output_at))
g = tf.get_default_graph()
g_x = tf.placeholder(tf.float32, shape = [None, N_KEYPOINTS], name = 'input')
weight1 = get_value_by_name('Layer1/weight1')
bias1 = get_value_by_name('Layer1/bias1')
fc1 = fc_layer(g_x, weight1, bias1)
weight2 = get_value_by_name('Layer2/weight2')
bias2 = get_value_by_name('Layer2/bias2')
fc2 = fc_layer(fc1, weight2, bias2, should_activate = False)
output = tf.nn.softmax(fc2, name = 'output')
graph_def = g.as_graph_def()
tf.train.write_graph(graph_def, output_at, 'graph.pb', as_text = False)
print('Saved model')
if __name__ == '__main__':
args = parse_args()
data_files = [os.path.abspath(p.strip()) for p in args.csv_files.split(',')]
X_data, y_data = read_data_files(data_files, args.max_samples)
train_data, test_data = split_data(X_data, y_data, args.train_test_split)
# Keep adding parameters here
parameters = {'batch_size': args.batch_size,
'epochs': args.epochs,
'learning_rate': args.learning_rate,
'n_classes': len(data_files),
'break_training_at': args.break_training_at}
# If output folder is None, will create an output folder
# Also, the output folder will be deleted and recreated again.
output_at = 'Checkpoint' if args.output_path is None else args.output_path
build_and_train_network(train_data, test_data, parameters, output_at)
export_created_graph(output_at)