return target_labels, target_scores, target_loc
if __name__ == "__main__":
from anchors import ssd_anchor_all_layers
import matplotlib.pyplot as plt
from read_data import Reader
import cv2
reader = Reader(is_training=True)
while (True):
value = reader.generate()
image = value['image'].astype(np.int)
image_copy = image.copy()
true_labels = value['classes']
true_boxes = value['boxes']
image_shape = image.shape
layers_anchors = ssd_anchor_all_layers(image)
target_labels = []
target_scores = []
target_loc = []
t = 0
class Net(object):
def __init__(self, is_training=True):
self.size = cfg.TARGET_SIZE
self.data_path = cfg.DATA_PATH
self.model_path = cfg.MODEL_PATH
self.epoches = cfg.EPOCHES
self.batches = cfg.BATCHES
self.lr = cfg.LEARNING_RATE
self.batch_size = cfg.BATCH_SIZE
self.cls_num = cfg.N_CLASSES
self.reader = Reader()
self.keep_rate = cfg.KEEP_RATE
self.is_training = is_training
self.model_name = cfg.MODEL_NAME
self.x = tf.placeholder(tf.float32, [None, self.size, self.size, 3])
self.y = tf.placeholder(tf.float32, [None, self.cls_num])
self.y_hat = self.resnet(self.x)
self.loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(labels=self.y,
logits=self.y_hat))
self.saver = tf.train.Saver()
self.acc = self.calculate_acc(self.y, self.y_hat)
def calculate_acc(self, labels, logits):
right_pred = tf.equal(tf.argmax(labels, axis=-1),
tf.argmax(logits, axis=-1))
accuracy = tf.reduce_mean(tf.cast(right_pred, tf.float32))
return accuracy
def resnet(self, inputs):
with tf.variable_scope('RESNET'):
net = slim.conv2d(inputs,
64, [7, 7],
2,
scope='conv7x7',
padding='SAME')
net = slim.max_pool2d(net, [2, 2], scope='pool1', padding='SAME')
res = net
# block1
net = slim.repeat(net,
2,
slim.conv2d,
64, [3, 3],
scope='conv1',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block2
net = slim.repeat(net,
2,
slim.conv2d,
64, [3, 3],
scope='conv2',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block3
net = slim.repeat(net,
2,
slim.conv2d,
64, [3, 3],
scope='conv3',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = slim.conv2d(net,
128, [3, 3],
2,
scope='reshape1',
padding='SAME')
# block4
net = slim.conv2d(net,
128, [3, 3],
2,
scope='conv4_3x3',
padding='SAME')
net = slim.conv2d(net,
128, [3, 3],
1,
scope='conv4_1x1',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block5
net = slim.repeat(net,
2,
slim.conv2d,
128, [3, 3],
scope='conv5',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block6
net = slim.repeat(net,
2,
slim.conv2d,
128, [3, 3],
scope='conv6',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block7
net = slim.repeat(net,
2,
slim.conv2d,
128, [3, 3],
scope='conv7',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = slim.conv2d(net,
256, [3, 3],
2,
scope='reshape2',
padding='SAME')
# block8
net = slim.conv2d(net,
256, [3, 3],
2,
scope='conv8_3x3',
padding='SAME')
net = slim.conv2d(net,
256, [3, 3],
1,
scope='conv8_1x1',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block9
net = slim.repeat(net,
2,
slim.conv2d,
256, [3, 3],
scope='conv9',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block10
net = slim.repeat(net,
2,
slim.conv2d,
256, [3, 3],
scope='conv10',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block11
net = slim.repeat(net,
2,
slim.conv2d,
256, [3, 3],
scope='conv11',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = slim.conv2d(net,
512, [3, 3],
2,
scope='reshape3',
padding='SAME')
# block12
net = slim.conv2d(net,
512, [3, 3],
2,
scope='conv12_3x3',
padding='SAME')
net = slim.conv2d(net,
512, [3, 3],
1,
scope='conv12_1x1',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block13
net = slim.repeat(net,
2,
slim.conv2d,
512, [3, 3],
scope='conv13',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block14
net = slim.repeat(net,
2,
slim.conv2d,
512, [3, 3],
scope='conv14',
padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
avg_pool = slim.avg_pool2d(net, [7, 7], scope='avg_pool')
avg_pool = tf.layers.flatten(avg_pool)
logits = tf.layers.dense(avg_pool, 1000)
if self.is_training:
logits = tf.nn.dropout(logits, keep_prob=self.keep_rate)
logits = tf.layers.dense(logits, self.cls_num)
return tf.nn.softmax(logits, name='softmax')
def train_net(self):
if not os.path.exists(self.model_path):
os.makedirs(self.model_path)
self.optimizer = tf.compat.v1.train.AdamOptimizer(self.lr)
self.train_step = self.optimizer.minimize(self.loss)
with tf.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(self.model_path)
if ckpt and ckpt.model_checkpoint_path:
# 如果保存过模型，则在保存的模型的基础上继续训练
self.saver.restore(sess, ckpt.model_checkpoint_path)
print('Model Reload Successfully!')
for epoch in range(self.epoches):
loss_list = []
for batch in range(self.batches):
data = self.reader.generate(self.batch_size)
feed_dict = {
self.x: data['images'],
self.y: data['labels']
}
_, loss = sess.run([self.train_step, self.loss], feed_dict)
loss_list.append(loss)
mean_loss = np.mean(np.array(loss_list))
acc_list = []
for _ in range(10):
test_data = self.reader.generate_test(batch_size=32)
test_dict = {
self.x: test_data['images'],
self.y: test_data['labels']
}
acc = sess.run(self.acc, test_dict)
acc_list.append(acc)
acc = np.mean(np.array(acc_list))
print('Epoch:{} Loss{} Acc:{}'.format(epoch, mean_loss, acc))
self.saver.save(sess, self.model_name)
class Net(object):
def __init__(self, is_training):
self.reader = Reader(is_training)
self.is_training = is_training
self.learning_rate = cfg.LEARNING_RATE
self.batch_size = cfg.BATCH_SIZE
self.class_num = len(cfg.CLASSES)
self.blocks = cfg.BLOCKS
self.ratios = cfg.RATIOS
self.Sk = cfg.Sk
self.x = [tf.placeholder(tf.float32, [None, None, 3])
] * self.batch_size
self.true_labels = [tf.placeholder(tf.float32, [None])
] * self.batch_size
self.true_boxes = [tf.placeholder(tf.float32, [None, 4])
] * self.batch_size
self.pred_loc, self.pred_cls = self.ssd_net(self.x)
self.saver = tf.train.Saver()
def ssd_net(self, inputs):
pred_loc_result = []
pred_score_result = []
for q in range(self.batch_size):
layers = {}
x = tf.expand_dims(inputs[q], axis=0)
with tf.variable_scope('net', reuse=tf.AUTO_REUSE):
# Block 1
net = slim.repeat(x, 2, slim.conv2d, 64, [3, 3], scope='conv1')
net = slim.max_pool2d(net, [2, 2],
scope='pool1',
padding='SAME')
# Block 2
net = slim.repeat(net,
2,
slim.conv2d,
128, [3, 3],
scope='conv2')
net = slim.max_pool2d(net, [2, 2],
scope='pool2',
padding='SAME')
# Block 3
net = slim.repeat(net,
3,
slim.conv2d,
256, [3, 3],
scope='conv3')
net = slim.max_pool2d(net, [2, 2],
scope='pool3',
padding='SAME')
# Block 4
net = slim.repeat(net,
3,
slim.conv2d,
512, [3, 3],
scope='conv4')
layers['block4'] = net
net = slim.max_pool2d(net, [2, 2],
scope='pool4',
padding='SAME')
# Block 5
net = slim.repeat(net,
3,
slim.conv2d,
512, [3, 3],
scope='conv5')
# Block 6
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
# Block 7
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
layers['block7'] = net
# Block 8
with tf.variable_scope('block8'):
net = slim.conv2d(net, 256, [1, 1], scope='conv8_1_1')
net = slim.conv2d(net,
512, [3, 3],
2,
scope='conv8_3_3',
padding='SAME')
layers['block8'] = net
# Block 9
with tf.variable_scope('block9'):
net = slim.conv2d(net, 128, [1, 1], scope='conv9_1_1')
net = slim.conv2d(net,
256, [3, 3],
2,
scope='conv9_3_3',
padding='SAME')
layers['block9'] = net
# Block 10
with tf.variable_scope('block10'):
net = slim.conv2d(net, 128, [1, 1], scope='conv10_1_1')
net = slim.conv2d(net,
256, [3, 3],
2,
scope='conv10_3_3',
padding='SAME')
layers['block10'] = net
# Block 11
with tf.variable_scope('block11'):
net = slim.conv2d(net, 128, [1, 1], scope='conv11_1_1')
net = slim.conv2d(net,
256, [3, 3],
2,
scope='conv11_3_3',
padding='SAME')
layers['block11'] = net
# Block 12
with tf.variable_scope('block12'):
net = slim.conv2d(net, 128, [1, 1], scope='conv12_1_1')
net = slim.conv2d(net,
256, [3, 3],
2,
scope='conv12_3_3',
padding='SAME')
layers['block12'] = net
self.layers = layers
pred_loc = []
pred_score = []
for i, block in enumerate(self.blocks):
with tf.variable_scope(block + '_box'):
loc, score = self.ssd_multibox_layer(
layers[block], self.class_num, self.ratios[i],
self.Sk[i])
pred_loc.append(loc)
pred_score.append(score)
pred_loc_result.append(pred_loc)
pred_score_result.append(pred_score)
return pred_loc_result, pred_score_result
def ssd_multibox_layer(self, inputs, class_num, ratio, size):
num_anchors = len(size) + len(ratio)
num_loc = num_anchors * 4
num_cls = num_anchors * class_num
# loc
loc_pred = slim.conv2d(inputs,
num_loc, [3, 3],
activation_fn=None,
scope='conv_loc')
# cls
cls_pred = slim.conv2d(inputs,
num_cls, [3, 3],
activation_fn=None,
scope='conv_cls')
loc_pred = tf.reshape(loc_pred, (-1, 4))
cls_pred = tf.reshape(cls_pred, (-1, class_num))
# softmax
cls_pred = slim.softmax(cls_pred, scope='softmax')
return loc_pred, cls_pred
def train_net(self):
self.optimizer = tf.compat.v1.train.MomentumOptimizer(
learning_rate=self.learning_rate, momentum=0.9)
self.loss = []
for q in range(self.batch_size):
anchors = tf.numpy_function(ssd_anchor_all_layers, [self.x],
[tf.float32] * 7)
# self._anchors = ssd_anchor_all_layers(self.x)
target_labels = []
target_scores = []
target_loc = []
for i in range(7):
t_labels, t_scores, t_loc = tf.numpy_function(
ssd_bboxes_encode, [
anchors[i], self.true_boxes[q], self.true_labels[q],
self.class_num
], [tf.float32, tf.float32, tf.float32])
target_labels.append(t_labels)
target_scores.append(t_scores)
target_loc.append(t_loc)
total_cross_pos, total_cross_neg, total_loc = loss_layer(
(self.pred_loc[q], self.pred_cls[q]), target_labels,
target_scores, target_loc)
loss = tf.add(tf.add(total_cross_pos, total_cross_neg), total_loc)
self.loss.append(loss)
self.loss = tf.reduce_mean(self.loss, axis=0)
self.train_step = self.optimizer.minimize(self.loss)
with tf.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(cfg.MODEL_PATH)
if ckpt and ckpt.model_checkpoint_path:
# 如果保存过模型，则在保存的模型的基础上继续训练
self.saver.restore(sess, ckpt.model_checkpoint_path)
print('Model Reload Successfully!')
for i in range(cfg.EPOCHES):
batch_images = []
batch_labels = []
batch_boxes = []
for batch in range(self.batch_size):
value = self.reader.generate()
image = value['image']
true_labels = value['classes']
true_boxes = value['boxes']
batch_images.append(image)
batch_labels.append(true_labels)
batch_boxes.append(true_boxes)
feed_dict = {
self.x: batch_images,
self.true_labels: batch_labels,
self.true_boxes: batch_boxes
}
loss_value, _ = sess.run([self.loss, self.train_step],
feed_dict)
self.saver.save(sess, os.path.join(cfg.MODEL_PATH,
'model.ckpt'))
print('epoch:{}, loss:{}'.format(self.reader.epoch,
loss_value))
class Net(object):
def __init__(self, is_training):
self.is_training = is_training
self.epoches = cfg.EPOCHES
self.learning_rate = cfg.LEARNING_RATE
self.num_classes = len(cfg.CLASSES)
self.model_path = cfg.MODEL_PATH
self.batch_size = cfg.BATCH_SIZE
self.target_size = cfg.TARGET_SIZE
self.keep_rate = cfg.KEEP_RATE
self.reader = Reader(is_training=is_training)
self.x = tf.placeholder(
tf.float32, [None, self.target_size, self.target_size, 3])
self.y0 = tf.placeholder(
tf.int32, [None, 14, 14])
self.y1 = tf.placeholder(
tf.int32, [None, 28, 28])
self.y2 = tf.placeholder(
tf.int32, [None, 56, 56])
self.y3 = tf.placeholder(
tf.int32, [None, self.target_size, self.target_size])
self.y = [self.y0, self.y1, self.y2, self.y3]
self.y_hat = self.network(self.x)
self.loss = self.sample_loss(self.y, self.y_hat)
self.saver = tf.train.Saver()
def sample_loss(self, labels, logits):
losses = []
for i in range(4):
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=labels[i], logits=logits[i]
)
losses.append(tf.reduce_mean(loss))
return tf.reduce_mean(losses)
def network(self, inputs):
num_classes = self.num_classes
train = self.is_training
with tf.variable_scope('RESNET'):
net = slim.conv2d(inputs, 64, [7, 7],
2, scope='conv7x7', padding='SAME')
pool1 = net
net = slim.max_pool2d(net, [2, 2], scope='pool1', padding='SAME')
pool2 = net
# block1
net = slim.repeat(net, 2, slim.conv2d, 64, [3, 3],
scope='conv1', padding='SAME')
net = tf.add(net, pool2)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block2
net = slim.repeat(net, 2, slim.conv2d, 64, [3, 3],
scope='conv2', padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block3
net = slim.repeat(net, 2, slim.conv2d, 64, [3, 3],
scope='conv3', padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = slim.conv2d(net, 128, [3, 3], 2,
scope='reshape1', padding='SAME')
# block4
net = slim.conv2d(net, 128, [3, 3], 2,
scope='conv4_3x3', padding='SAME')
pool3 = net
net = slim.conv2d(net, 128, [3, 3], 1,
scope='conv4_1x1', padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block5
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3],
scope='conv5', padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block6
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3],
scope='conv6', padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block7
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3],
scope='conv7', padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = slim.conv2d(net, 256, [3, 3], 2,
scope='reshape2', padding='SAME')
# block8
net = slim.conv2d(net, 256, [3, 3], 2,
scope='conv8_3x3', padding='SAME')
pool4 = net
net = slim.conv2d(net, 256, [3, 3], 1,
scope='conv8_1x1', padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block9
net = slim.repeat(net, 2, slim.conv2d, 256, [3, 3],
scope='conv9', padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block10
net = slim.repeat(net, 2, slim.conv2d, 256, [3, 3],
scope='conv10', padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block11
net = slim.repeat(net, 2, slim.conv2d, 256, [3, 3],
scope='conv11', padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = slim.conv2d(net, 512, [3, 3], 2,
scope='reshape3', padding='SAME')
# block12
net = slim.conv2d(net, 512, [3, 3], 2,
scope='conv12_3x3', padding='SAME')
pool5 = net
net = slim.conv2d(net, 512, [3, 3], 1,
scope='conv12_1x1', padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block13
net = slim.repeat(net, 2, slim.conv2d, 512, [3, 3],
scope='conv13', padding='SAME')
net = tf.add(net, res)
net = tf.layers.batch_normalization(net, training=self.is_training)
res = net
# block14
net = slim.repeat(net, 2, slim.conv2d, 512, [3, 3],
scope='conv14', padding='SAME')
# fc1
net = slim.conv2d(net, 4096, [1, 1],
1, scope='fc1', padding='SAME')
# fc2
net = slim.conv2d(net, 4096, [1, 1],
1, scope='fc2', padding='SAME')
# fc3
net = slim.conv2d(net, 1000, [1, 1],
1, scope='fc3', padding='SAME')
# up_pool 1
net = self._unpool_layer(net,
shape=tf.shape(pool4),
num_classes=num_classes,
name='up_pool1',
ksize=4, stride=2)
self.test = tf.nn.softmax(net, axis=-1)
score_pool1 = slim.conv2d(
pool4, num_classes, [1, 1], 1, scope='score_pool1')
net = tf.add(net, score_pool1)
up_pool1 = tf.nn.softmax(net, axis=-1)
# up_pool 2
net = self._unpool_layer(net,
shape=tf.shape(pool3),
num_classes=num_classes,
name='up_pool2',
ksize=4, stride=2)
score_pool2 = slim.conv2d(
pool3, num_classes, [1, 1], 1, scope='score_pool2')
net = tf.add(net, score_pool2)
up_pool2 = tf.nn.softmax(net, axis=-1)
# up_pool 3
net = self._unpool_layer(net,
shape=tf.shape(pool2),
num_classes=num_classes,
name='up_pool3',
ksize=4, stride=2)
score_pool3 = slim.conv2d(
pool2, num_classes, [1, 1], 1, scope='score_pool3')
net = tf.add(net, score_pool3)
up_pool3 = tf.nn.softmax(net, axis=-1)
# up_pool 4
logits = self._unpool_layer(net,
shape=tf.shape(inputs),
num_classes=num_classes,
name='up_pool4',
ksize=8, stride=4)
logits = tf.nn.softmax(logits, axis=-1)
return up_pool1, up_pool2, up_pool3, logits
def _unpool_layer(self, bottom, shape,
num_classes, name,
ksize=4, stride=2):
strides = [1, stride, stride, 1]
with tf.variable_scope(name):
in_features = bottom.get_shape()[3].value
new_shape = [shape[0], shape[1], shape[2], num_classes]
output_shape = tf.stack(new_shape)
f_shape = [ksize, ksize, num_classes, in_features]
'''weights = tf.get_variable(
'W', f_shape, tf.float32, xavier_initializer())'''
weights = self.get_deconv_filter(f_shape)
deconv = tf.nn.conv2d_transpose(bottom, weights, output_shape,
strides=strides, padding='SAME')
return deconv
def get_deconv_filter(self, f_shape):
# 双线性插值
width = f_shape[0]
height = f_shape[1]
f = ceil(width/2.0)
c = (2 * f - 1 - f % 2) / (2.0 * f)
bilinear = np.zeros([f_shape[0], f_shape[1]])
for x in range(width):
for y in range(height):
value = (1 - abs(x / f - c)) * (1 - abs(y / f - c))
bilinear[x, y] = value
weights = np.zeros(f_shape)
for i in range(f_shape[2]):
weights[:, :, i, i] = bilinear
init = tf.constant_initializer(value=weights+0.1,
dtype=tf.float32)
return tf.get_variable(name="up_filter", initializer=init,
shape=weights.shape)
def run_test(self):
with tf.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(cfg.MODEL_PATH)
if ckpt and ckpt.model_checkpoint_path:
# 如果保存过模型，则在保存的模型的基础上继续训练
self.saver.restore(sess, ckpt.model_checkpoint_path)
print('Model Reload Successfully!')
value = self.reader.generate(1)
images = value['images']
labels = value['labels']
feed_dict = {self.x: images}
v = sess.run(self.y_hat, feed_dict)
for i in range(4):
a1 = np.squeeze(np.argmax(v[i], axis=-1))
a2 = np.squeeze(labels[i])
tmp = np.hstack((a1, a2))
plt.imshow(tmp)
plt.show()
t = tf.get_default_graph().get_tensor_by_name('RESNET/up_pool1/up_filter:0')
t = sess.run(t)
def train_net(self):
if not os.path.exists(self.model_path):
os.makedirs(self.model_path)
self.optimizer = tf.compat.v1.train.MomentumOptimizer(
learning_rate=self.learning_rate, momentum=0.9)
self.optimizer = tf.compat.v1.train.AdamOptimizer(self.learning_rate)
self.train_step = self.optimizer.minimize(self.loss)
with tf.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(cfg.MODEL_PATH)
if ckpt and ckpt.model_checkpoint_path:
# 如果保存过模型，则在保存的模型的基础上继续训练
self.saver.restore(sess, ckpt.model_checkpoint_path)
print('Model Reload Successfully!')
for i in range(cfg.EPOCHES):
loss_list = []
for batch in range(cfg.BATCHES):
value = self.reader.generate(self.batch_size)
images = value['images']
labels = value['labels']
feed_dict = {self.x: images,
self.y0: labels[0],
self.y1: labels[1],
self.y2: labels[2],
self.y3: labels[3]}
_, loss, pred = sess.run(
[self.train_step, self.loss, self.y_hat], feed_dict)
loss_list.append(loss)
print('batch:{} loss:{}'.format(batch, loss), end='\r')
loss_values = np.array(loss_list) # (64, 3)
loss_values = np.mean(loss_values)
with open('./result.txt', 'a') as f:
f.write(str(loss_values)+'\n')
self.saver.save(sess, os.path.join(
cfg.MODEL_PATH, 'model.ckpt'))
print('epoch:{} loss:{}'.format(
self.reader.epoch, loss_values))
class Net(object):
def __init__(self, is_training):
self.reader = Reader(is_training)
self.is_training = is_training
self.learning_rate = cfg.LEARNING_RATE
self.class_num = len(cfg.CLASSES)
self.weight_decay = cfg.WEIGHT_DECAY
self.blocks = cfg.BLOCKS
self.ratios = cfg.RATIOS
self.keep_rate = cfg.KEEP_RATE
self.model_path = cfg.MODEL_PATH
self.momentum = cfg.MOMENTUM
self.Sk = cfg.Sk
self.x = tf.placeholder(tf.float32, [None, None, 3])
self.true_labels = tf.placeholder(tf.float32, [None])
self.true_boxes = tf.placeholder(tf.float32, [None, 4])
self.output = self.ssd_net(
tf.expand_dims(self.x, axis=0)
)
self.anchors = tf.numpy_function(
ssd_anchor_all_layers, [self.x],
[tf.float32]*7
)
self.saver = tf.train.Saver()
def ssd_net(self, inputs, scope='ssd_512_vgg'):
layers = {}
with tf.variable_scope(scope, 'ssd_512_vgg', [inputs], reuse=None):
with slim.arg_scope([slim.conv2d],
activation_fn=tf.nn.relu,
weights_regularizer=slim.l2_regularizer(self.weight_decay)):
# Block 1
net = slim.repeat(inputs, 2, slim.conv2d,
64, [3, 3], scope='conv1')
net = slim.max_pool2d(
net, [2, 2], scope='pool1', padding='SAME')
# Block 2
net = slim.repeat(net, 2, slim.conv2d,
128, [3, 3], scope='conv2')
net = slim.max_pool2d(
net, [2, 2], scope='pool2', padding='SAME')
# Block 3
net = slim.repeat(net, 3, slim.conv2d,
256, [3, 3], scope='conv3')
net = slim.max_pool2d(
net, [2, 2], scope='pool3', padding='SAME')
# net = tf.layers.batch_normalization(net, training=self.is_training)
# Block 4
net = slim.repeat(net, 3, slim.conv2d,
512, [3, 3], scope='conv4')
layers['block4'] = net
net = slim.max_pool2d(
net, [2, 2], scope='pool4', padding='SAME')
# Block 5
net = slim.repeat(net, 3, slim.conv2d,
512, [3, 3], scope='conv5')
# Block 6
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
# Dropout
if self.is_training:
net = tf.nn.dropout(net, keep_prob=self.keep_rate)
# Block 7
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
layers['block7'] = net
# Dropout
if self.is_training:
net = tf.nn.dropout(net, keep_prob=self.keep_rate)
# Block 8
with tf.variable_scope('block8'):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = slim.conv2d(net, 512, [3, 3], 2,
scope='conv3x3', padding='SAME')
layers['block8'] = net
# Block 9
with tf.variable_scope('block9'):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net, 256, [3, 3], 2,
scope='conv3x3', padding='SAME')
layers['block9'] = net
# Block 10
with tf.variable_scope('block10'):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net, 256, [3, 3], 2,
scope='conv3x3', padding='SAME')
layers['block10'] = net
# Block 11
with tf.variable_scope('block11'):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net, 256, [3, 3], 2,
scope='conv3x3', padding='SAME')
layers['block11'] = net
# Block 12
with tf.variable_scope('block12'):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net, 256, [4, 4], 2,
scope='conv4x4', padding='SAME')
layers['block12'] = net
self.layers = layers
pred_loc = []
pred_score = []
for i, block in enumerate(self.blocks):
with tf.variable_scope(block+'_box'):
loc, score = self.ssd_multibox_layer(
layers[block], self.class_num, self.ratios[i], self.Sk[i]
)
pred_loc.append(loc)
pred_score.append(score)
return pred_loc, pred_score
def ssd_multibox_layer(self, inputs, class_num, ratio, size):
num_anchors = len(size) + len(ratio)
num_loc = num_anchors * 4
num_cls = num_anchors * class_num
# loc
loc_pred = slim.conv2d(
inputs, num_loc, [3, 3], activation_fn=None, scope='conv_loc')
# cls
cls_pred = slim.conv2d(
inputs, num_cls, [3, 3], activation_fn=None, scope='conv_cls')
loc_pred = tf.reshape(loc_pred, (-1, 4))
cls_pred = tf.reshape(cls_pred, (-1, class_num))
# softmax
cls_pred = slim.softmax(cls_pred, scope='softmax')
return loc_pred, cls_pred
def train_net(self):
if not os.path.exists(self.model_path):
os.makedirs(self.model_path)
self.target_labels = []
self.target_scores = []
self.target_loc = []
for i in range(7):
target_labels, target_scores, target_loc = tf.numpy_function(
ssd_bboxes_encode, [self.anchors[i], self.true_boxes,
self.true_labels, self.class_num],
[tf.float32, tf.float32, tf.float32]
)
self.target_labels.append(target_labels)
self.target_scores.append(target_scores)
self.target_loc.append(target_loc)
self.total_cross_pos, self.total_cross_neg, self.total_loc = loss_layer(
self.output, self.target_labels, self.target_scores, self.target_loc
)
self.loss = tf.add(
tf.add(self.total_cross_pos, self.total_cross_neg), self.total_loc
)
# gradients = self.optimizer.compute_gradients(self.loss)
self.optimizer = tf.compat.v1.train.MomentumOptimizer(
learning_rate=self.learning_rate, momentum=self.momentum)
# self.optimizer = tf.compat.v1.train.AdamOptimizer(self.learning_rate)
self.train_step = self.optimizer.minimize(self.loss)
with tf.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(self.model_path)
if ckpt and ckpt.model_checkpoint_path:
# 如果保存过模型，则在保存的模型的基础上继续训练
self.saver.restore(sess, ckpt.model_checkpoint_path)
log.info('Model Reload Successfully!')
for i in range(cfg.EPOCHES):
loss_list = []
for batch in range(cfg.BATCHES):
value = self.reader.generate()
image = value['image'] - cfg.PIXEL_MEANS
true_labels = value['classes']
true_boxes = value['boxes']
feed_dict = {self.x: image,
self.true_labels: true_labels,
self.true_boxes: true_boxes}
test = sess.run(self.target_scores, feed_dict)
total_pos = 0
for v in test:
if np.max(v) > cfg.THRESHOLD:
total_pos += 1
if total_pos == 0:
continue
try:
sess.run(self.train_step, feed_dict)
loss_0, loss_1, loss_2 = sess.run(
[self.total_cross_pos, self.total_cross_neg, self.total_loc], feed_dict)
except KeyboardInterrupt:
self.saver.save(sess, os.path.join(
self.model_path, 'model.ckpt'))
print('Model Saved.')
raise KeyboardInterrupt
loss_list.append(
np.array([loss_0, loss_1, loss_2])
)
# print('batch:{},pos_loss:{},neg_loss:{},loc_loss:{}'.format(
# batch, loss_0, loss_1, loss_2
# ), end='\r')
loss_values = np.array(loss_list) # (64, 3)
loss_values = np.mean(loss_values, axis=0)
with open('./result.txt', 'a') as f:
f.write(str(loss_values)+'\n')
self.saver.save(sess, os.path.join(
self.model_path, 'model.ckpt'))
print('epoch:{},pos_loss:{},neg_loss:{},loc_loss:{}'.format(
self.reader.epoch, loss_values[0], loss_values[1], loss_values[2]
))
class Net(object):
def __init__(self, is_training):
self.is_training = is_training
self.epoches = cfg.EPOCHES
self.learning_rate = cfg.LEARNING_RATE
self.num_classes = len(cfg.CLASSES)
self.model_path = cfg.MODEL_PATH
self.batch_size = cfg.BATCH_SIZE
self.target_size = cfg.TARGET_SIZE
self.keep_rate = cfg.KEEP_RATE
self.reader = Reader(is_training=is_training)
self.x = tf.placeholder(tf.float32,
[None, self.target_size, self.target_size, 3])
self.y = tf.placeholder(tf.int32,
[None, self.target_size, self.target_size])
self.loss_weght = tf.placeholder(
tf.float32, [None, self.target_size, self.target_size])
self.y_hat = self.network(self.x)
self.loss = self.sample_loss(self.y, self.y_hat, self.loss_weght)
self.saver = tf.train.Saver()
def sample_loss(self, labels, logits, loss_weight):
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels,
logits=logits)
loss = tf.losses.compute_weighted_loss(losses=loss,
weights=loss_weight)
return tf.reduce_mean(loss)
def network(self, inputs):
num_classes = self.num_classes
train = self.is_training
with tf.variable_scope('vgg'):
conv1_1 = self._conv_layer(inputs, 64, 'conv1_1')
conv1_2 = self._conv_layer(conv1_1, 64, 'conv1_2')
pool1 = self._max_pool(conv1_2, 'pool1')
conv2_1 = self._conv_layer(pool1, 128, 'conv2_1')
conv2_2 = self._conv_layer(conv2_1, 128, 'conv2_2')
pool2 = self._max_pool(conv2_2, 'pool2')
conv3_1 = self._conv_layer(pool2, 256, 'conv3_1')
conv3_2 = self._conv_layer(conv3_1, 256, 'conv3_2')
conv3_3 = self._conv_layer(conv3_2, 256, 'conv3_3')
pool3 = self._max_pool(conv3_3, 'pool3')
conv4_1 = self._conv_layer(pool3, 512, 'conv4_1')
conv4_2 = self._conv_layer(conv4_1, 512, 'conv4_2')
conv4_3 = self._conv_layer(conv4_2, 512, 'conv4_3')
pool4 = self._max_pool(conv4_3, 'pool4')
conv5_1 = self._conv_layer(pool4, 512, 'conv5_1')
conv5_2 = self._conv_layer(conv5_1, 512, 'conv5_2')
conv5_3 = self._conv_layer(conv5_2, 512, 'conv5_3')
pool5 = self._max_pool(conv5_3, 'pool5')
fc6 = self._conv_layer(pool5, 4096, k_size=1, name='fc6')
if train:
fc6 = tf.nn.dropout(fc6, self.keep_rate)
fc7 = self._conv_layer(fc6, 4096, k_size=1, name='fc7')
if train:
fc7 = tf.nn.dropout(fc7, self.keep_rate)
fc8 = self._conv_layer(fc7, 1000, k_size=1, name='fc8')
upscore1 = self._upscore_layer(fc8,
shape=tf.shape(pool4),
num_classes=num_classes,
name='upscore1',
ksize=4,
stride=2)
score_pool4 = self._conv_layer(pool4,
num_classes,
k_size=1,
name='score_pool4')
fuse_pool4 = tf.add(upscore1, score_pool4)
# self.fuse_pool4 = self.score_pool4
upscore2 = self._upscore_layer(fuse_pool4,
shape=tf.shape(inputs),
num_classes=num_classes,
name='upscore2',
ksize=32,
stride=16)
return tf.nn.softmax(upscore2, axis=-1)
def _max_pool(self, bottom, name):
pool = slim.max_pool2d(bottom, [2, 2], scope=name, padding='SAME')
return pool
def _conv_layer(self, bottom, filters, name, k_size=3):
conv = slim.conv2d(bottom,
filters, [k_size, k_size],
scope=name,
padding='SAME')
return conv
def _upscore_layer(self,
bottom,
shape,
num_classes,
name,
ksize=4,
stride=2):
strides = [1, stride, stride, 1]
with tf.variable_scope(name):
in_features = bottom.get_shape()[3].value
if shape is None:
# Compute shape out of Bottom
in_shape = tf.shape(bottom)
h = ((in_shape[1] - 1) * stride) + 1
w = ((in_shape[2] - 1) * stride) + 1
new_shape = [in_shape[0], h, w, num_classes]
else:
new_shape = [shape[0], shape[1], shape[2], num_classes]
output_shape = tf.stack(new_shape)
f_shape = [ksize, ksize, num_classes, in_features]
weights = self.get_deconv_filter_norm(f_shape)
deconv = tf.nn.conv2d_transpose(bottom,
weights,
output_shape,
strides=strides,
padding='SAME')
return deconv
def get_deconv_filter(self, f_shape):
width = f_shape[0]
height = f_shape[1]
f = ceil(width / 2.0)
c = (2 * f - 1 - f % 2) / (2.0 * f)
bilinear = np.zeros([f_shape[0], f_shape[1]])
for x in range(width):
for y in range(height):
value = (1 - abs(x / f - c)) * (1 - abs(y / f - c))
bilinear[x, y] = value
weights = np.zeros(f_shape)
for i in range(f_shape[2]):
weights[:, :, i, i] = bilinear
init = tf.constant_initializer(value=weights, dtype=tf.float32)
return tf.get_variable(name='up_filter',
initializer=init,
shape=weights.shape)
def get_deconv_filter_norm(self, f_shape):
weights = tf.Variable(tf.truncated_normal(f_shape, mean=0.,
stddev=0.1))
return weights
def train_net(self):
if not os.path.exists(self.model_path):
os.makedirs(self.model_path)
self.optimizer = tf.compat.v1.train.MomentumOptimizer(
learning_rate=self.learning_rate, momentum=0.9)
# self.optimizer = tf.compat.v1.train.AdamOptimizer(self.learning_rate)
self.train_step = self.optimizer.minimize(self.loss)
with tf.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(cfg.MODEL_PATH)
if ckpt and ckpt.model_checkpoint_path:
# 如果保存过模型，则在保存的模型的基础上继续训练
self.saver.restore(sess, ckpt.model_checkpoint_path)
print('Model Reload Successfully!')
for i in range(cfg.EPOCHES):
loss_list = []
for batch in range(cfg.BATCHES):
value = self.reader.generate(self.batch_size)
images = value['images']
labels = value['labels']
weights = value['weights']
feed_dict = {
self.x: images,
self.y: labels,
self.loss_weght: weights
}
_, loss = sess.run([self.train_step, self.loss], feed_dict)
loss_list.append(loss)
print('batch:{} loss:{}'.format(batch, loss), end='\r')
loss_values = np.array(loss_list) # (64, 3)
loss_values = np.mean(loss_values)
with open('./result.txt', 'a') as f:
f.write(str(loss_values) + '\n')
self.saver.save(sess, os.path.join(cfg.MODEL_PATH,
'model.ckpt'))
print('epoch:{} loss:{}'.format(self.reader.epoch,
loss_values))
def test(self, image_path):
if not isinstance(image_path, list):
image_path = [image_path]
self.is_training = False
with tf.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(cfg.MODEL_PATH)
if ckpt and ckpt.model_checkpoint_path:
# 如果保存过模型，则在保存的模型的基础上继续训练
self.saver.restore(sess, ckpt.model_checkpoint_path)
print('Model Reload Successfully!')
for path in image_path:
image = self.reader.read_image(path)
image, _ = self.reader.resize_image(image, with_label=False)
image, _ = random_crop(image, None, with_label=False)
image = np.expand_dims(image, axis=0)
label_image = sess.run(
self.y_hat,
feed_dict={self.x: image - self.reader.pixel_means})
label_image = np.squeeze(image)
label_image = np.argmax(label_image, axis=-1)
label_image = self.reader.decode_label(label_image)
image = np.squeeze(image).astype(np.int)
result = np.vstack([image, label_image])
plt.imshow(result)
plt.show()
class Net(object):
def __init__(self, is_training=True):
self.is_training = is_training
if self.is_training:
self.reader = Reader()
self.batch_size = 16
self.lr = 2e-4
self.wd = 5e-3
self.epoches = 100
self.batches = 64
self.size = 96
self.label_num = 30
self.x = tf.placeholder(tf.float32, [None, self.size, self.size, 1])
self.y = tf.placeholder(tf.float32, [None, self.label_num])
self.y_hat = self.network(self.x)
self.model_path = './model'
self.ckpt_path = os.path.join(self.model_path, 'model.ckpt')
self.saver = tf.train.Saver()
def loss_layer(self, y, y_hat):
loss = tf.reduce_sum(tf.square(y - y_hat))
return loss
def network(self, inputs):
with tf.variable_scope('net'):
with slim.arg_scope([slim.conv2d],
activation_fn=tf.nn.relu,
weights_regularizer=slim.l2_regularizer(self.wd)):
# Block init
net = slim.conv2d(inputs, 1024, [3, 3],
2, scope='conv_init', padding='SAME')
# Block 1
net = slim.repeat(net, 2, slim.conv2d,
64, [3, 3], scope='conv1', padding='SAME')
net = slim.max_pool2d(
net, [2, 2], scope='pool1', padding='SAME')
net = tf.layers.batch_normalization(
net, trainable=self.is_training, name='BN_block1')
# Block 2
net = slim.repeat(net, 2, slim.conv2d,
128, [3, 3], scope='conv2')
net = slim.max_pool2d(
net, [2, 2], scope='pool2', padding='SAME')
net = tf.layers.batch_normalization(
net, trainable=self.is_training, name='BN_block2')
# Block 3
net = slim.repeat(net, 3, slim.conv2d,
256, [3, 3], scope='conv3')
net = slim.max_pool2d(
net, [2, 2], scope='pool3', padding='SAME')
net = tf.layers.batch_normalization(
net, trainable=self.is_training, name='BN_block3')
# Block 4
net = slim.repeat(net, 3, slim.conv2d,
512, [3, 3], scope='conv4')
net = slim.max_pool2d(
net, [2, 2], scope='pool4', padding='SAME')
net = tf.layers.batch_normalization(
net, trainable=self.is_training, name='BN_block4')
# Block 5
net = slim.repeat(net, 3, slim.conv2d,
512, [3, 3], scope='conv5')
net = tf.layers.batch_normalization(
net, trainable=self.is_training, name='BN_block5')
# Block 6
net = slim.conv2d(net, 1024, [3, 3],
2, scope='conv6')
net = tf.layers.batch_normalization(
net, trainable=self.is_training, name='BN_block6')
net = tf.layers.flatten(net)
logits = tf.layers.dense(net, self.label_num)
if self.is_training:
logits = tf.layers.dropout(logits, rate=1-0.2)
# logits = tf.nn.tanh(logits)
return logits
def train_net(self):
if not os.path.exists(self.model_path):
os.makedirs(self.model_path)
self.loss = self.loss_layer(self.y, self.y_hat)
self.optimizer = tf.compat.v1.train.AdamOptimizer(self.lr)
self.train_step = self.optimizer.minimize(self.loss)
with tf.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(self.model_path)
if ckpt and ckpt.model_checkpoint_path:
# 如果保存过模型，则在保存的模型的基础上继续训练
self.saver.restore(sess, ckpt.model_checkpoint_path)
print('Model Reload Successfully!')
for epoch in range(self.epoches):
loss_list = []
for batch in range(self.batch_size):
images, labels = self.reader.generate(self.batch_size)
feed_dict = {
self.x: images,
self.y: labels
}
loss_value = sess.run(self.loss, feed_dict)
_ = sess.run(self.train_step, feed_dict)
loss_list.append(loss_value)
loss = np.mean(np.array(loss_list))
self.saver.save(sess, self.ckpt_path)
print('epoch:{} loss:{}'.format(epoch, loss))
with open('./losses.txt', 'a') as f:
f.write(str(loss)+'\n')
def test_net(self, image, sess):
image = image.reshape((1, self.size, self.size, 1)) - 127.5
points = sess.run(self.y_hat, feed_dict={self.x: image})
points = (points * self.size).astype(np.int)
return np.squeeze(points)