EfficientCornetNet/train.py at master · Tran-Nam/EfficientCornetNet · GitHub

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
import numpy as np
import os

from net.efficentnet import EfficentNet
from helper.generate_gt import gen_gt
from module.loss import offset_loss, focal_loss
# from processing.dataloader import DataLoader
from processing.dataloader_2 import input_fn
import config

_img = np.random.random((1, 512, 512, 3))
_heatmap = np.random.random((1, 128, 128, 4))
_offset = np.random.random((1, 128, 128, 2))
_mask = np.zeros((1, 128, 128))
_mask[:, 20:40, 50:60] = 1

class Train():
    def __init__(self):
        # os.environ['CUDA_VISIBLE_DEVICES'] = '0'
        # self.batchsize = config.BATCH_SIZE
        # self.dataloader = DataLoader(image_dir='../../images')
        # self.dataloader.create_pairs()
        # self.it = iter(self.dataloader.train_ds)

        self.lr = config.LEARNING_RATE
        self.decay_step = config.DECAY_STEP
        self.decay_rate = config.DECAY_RATE
        self.num_steps = 1000
        self.model_dir = config.MODEL_DIR

        self.pretrained = config.PRETRAINED
        self.model_path = config.MODEL_PATH
        # self.save_model_dir = config.MODEL_DIR
        self.net = EfficentNet()
        self.focal_loss = focal_loss
        self.offset_loss = offset_loss

        # self.dataset = input_fn()
        # self.iterator = self.dataset.make_initializable_iterator()
        # self.iterator_init_op = self.iterator.initializer

        # print('Done!')

    def load_ckpt(self, saver, sess, model_path):
        """
        load pretrained weight
        :param saver: saver object
        :param sess: tf session
        :param model_dir: path to checkpoint file
        """
        # ckpt = tf.train.get_checkpoint_state(model_dir)
        # print(model_path)
        if os.path.exists(model_path + '.meta'):
            # ckpt_name = os.path.basename(ckpt.model_checkpoint_path)
            # print(model_path)
            saver.restore(sess, model_path)
            print('Restore model from {}'.format(model_path))
            return True
        else:
            return False


    def train(self):
        # image_batch, heatmap_batch, offset_batch = self.dataloader.next_batch() # load data
        # output = self.net.net(image_batch)
        # heatmap, offset = output
        # steps = tf.Variable(0, name='global_step', trainable=False)
        # lr = tf.train.exponential_decay(self.lr, steps, self.decay_step, self.decay_rate, staircase=True, name='learning_rate')
        # optim = tf.train.AdamOptimizer(learning_rate=1e-3)

        # placeholder
        # image = tf.placeholder(tf.float32, shape=[None, 512, 512, 3], name='image')
        # heatmap = tf.placeholder(tf.float32, shape=[None, 128, 128, 4], name='heatmap')
        # offset = tf.placeholder(tf.float32, shape=[None, 128, 128, 2], name='offset')
        # mask = tf.placeholder(tf.float32, shape=[None, 128, 128], name='mask')
        # gt = (heatmap, offset, mask)

        # batch, init_op = self.iterator.get_next()
        with tf.variable_scope('data_pipeline'):
            batch, iterator_init_op = input_fn()
        image = batch['image']
        heatmap = batch['heatmap']
        offset = batch['offset']
        mask = batch['mask']
        gt = (heatmap, offset, mask)
        # print(image.shape)

        # build graph
        # with tf.variable_scope('', reuse=True):
        output = self.net.net(img=image)
        heatmap_det, offset_det = output
        # loss = self.net.loss(output, gt)
        # loss = self.net.loss(output, gt)
        with tf.variable_scope('loss'):
            f_loss = self.focal_loss(heatmap_det, heatmap)
            f_loss_check = tf.debugging.check_numerics(f_loss, "Focal loss is NaN", name="DEBUG")
            o_loss = self.offset_loss(offset_det, offset, mask)
            o_loss_check = tf.debugging.check_numerics(o_loss, "Focal loss is NaN", name="DEBUG")
            loss = tf.add(f_loss, o_loss)
            loss = tf.cast(loss, dtype=tf.float64)
        # print(loss)
        # print(f_loss)
        # print(o_loss)
        with tf.variable_scope('optimizer'):
            # with tf.control_dependencies([f_loss_check, o_loss_check]):
            global_step = tf.Variable(0, trainable=False, name='global_step')
            # increment_global_step = tf.assign(global_step, global_step + 1)
            lr=tf.train.exponential_decay(self.lr,global_step,self.decay_step,self.decay_rate,staircase=True, name= 'learning_rate')
            optimizer = tf.train.AdamOptimizer(learning_rate=lr)
            train_op = optimizer.minimize(loss, global_step)
        # print(train_op)
        writer = tf.summary.FileWriter('./graphs', tf.get_default_graph())
            # trainable_variables = tf.trainable_variables()
        saver = tf.train.Saver(max_to_keep=3)
        init = tf.global_variables_initializer()

        # f_summary = tf.summary.scalar(name='loss', tensor=loss)
        tf.summary.scalar('focal_loss', f_loss)
        tf.summary.scalar('offset_loss', o_loss)
        tf.summary.scalar('total_loss', loss)
        tf.summary.scalar('lr', lr)
        tf.summary.image('input', image, max_outputs=5)
        tf.summary.image('top_left_gt', heatmap[:, :, :, 0:1], max_outputs=5)
        tf.summary.image('top_left', heatmap_det[:, :, :, 0:1], max_outputs=5)
        merge = tf.summary.merge_all()

        # update = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
        # with tf.control_dependencies(update):


        # saver = tf.train.Saver(max_to_keep=10)
        # config_gpu = tf.ConfigProto(allow_soft_placement=True, log_device_placement=True)
        # config_gpu.gpu_options.allow_growth = True
        # sess = tf.Session(config=config_gpu)
        sess = tf.Session()
        # print('Create session')
        sess.run(init)
        # input()
        sess.run(iterator_init_op)
        # input()
        # sess.run(self.iterator_init_op)
        # sess.run(batch.initialize())

        # print(self.num_steps)
        # epoch = 0
        if self.pretrained:
            if self.load_ckpt(saver, sess, self.model_path):
                print('[*] Load SUCCESS!')
            else:
                print('[*] Load FAIL ...')

        for step in range(self.num_steps):
            # print(step)
            # sess.run(debug)
            loss_, f_loss_, o_loss_, _ = sess.run([loss, f_loss, o_loss, train_op])
            print('step %d, loss %g, focal %g, offset %g'%(step, loss_, f_loss_, o_loss_))

            if step%config.INTERVAL_SAVE==0 and step > 0:
                summary = sess.run(merge)
                writer.add_summary(summary, step)
                saver.save(sess, os.path.join(self.model_dir, str(step) + '.ckpt'))
            # print(loss_)
            # if step%self.interval_save==0 and step>0:
            #     saver.save(sess, self.model_path, epoch)
            #     epoch += 1


if __name__=='__main__':
    t = Train()
    t.train()