# for train_index in range(len(train_out)):
# if train_out[train_index] == train_l[train_index]:
# tr_acc.append(1)
# else:
# tr_acc.append(0)
# # misc.imsave('/Users/wywy/Desktop/train_e1'+'/'+str(chioce_dict.get(train_out[train_index]))+'_'+bytes.decode(tarin_file[train_index]),train_img.reshape([-1,64,64]) [train_index])
#
# train_acc = np.mean(np.array(tr_acc))
#
# print('train iter :{}, train loss:{}, train acc:{}'.format(i,train_loss,train_acc))
# # # #
if i % 100 == 0:
test_img, test_label1, test_name = sess.run(test_data)
test_label = one_hot(test_label1.tolist(), 5)
test_img1 = test_img / 255 - 0.5
test_loss, test_acc, test_out, test_l = sess.run(
[net.loss, net.acc, net.argamx_out, net.argamx_label],
feed_dict={
net.x: test_img1,
net.y_: test_label
})
tes_acc = []
for test_index in range(len(test_out)):
if test_out[test_index] == test_l[test_index]:
tes_acc.append(1)
else:
tes_acc.append(0)
def train(num_batches, batch_size, learning_rate, save_dir=None):
# Build placeholders for the input samples and labels
# 创建输入样本和标签的占位符
width_multiplier = 1
inputs = tf.placeholder(tf.float32, [None, 64, 64, 1], name='input')
labels = tf.placeholder(tf.float32, [None, 5])
# Add placeholder to indicate whether or not we're training the model
# 创建占位符表明当前是否正在训练模型
is_training = tf.placeholder(tf.bool, name='is_training')
# Feed the inputs into a series of 20 convolutional layers
# 把输入数据填充到一系列20个卷积层的神经网络中
layer = inputs
# for layer_i in range(1, 20):
# layer = conv_layer(layer, layer_i, is_training)
#
# # Flatten the output from the convolutional layers
# # 将卷积层输出扁平化处理
# orig_shape = layer.get_shape().as_list()
# layer = tf.reshape(layer, shape=[-1, orig_shape[1]*orig_shape[2]*orig_shape[3]])
#
# # Add one fully connected layer
# # 添加一个具有100个神经元的全连接层
# layer = fully_connected(layer, 100, is_training)
#
# # Create the output layer with 1 node for each
# # 为每一个类别添加一个输出节点
# logits = tf.layers.dense(layer, 5)
# self.is_training = tf.placeholder(tf.bool, [])
net = conv2d(layer,
"conv_1",
round(32 * width_multiplier),
filter_size=3,
strides=2) # ->
net = tf.nn.relu(
tf.layers.batch_normalization(net,
name="conv_1/bn",
training=is_training)) # NB+RELU
net = _depthwise_separable_conv2d(
net, 10, width_multiplier, "ds_conv_2",
istraining=is_training) # ->[N, 112, 112, 64]
net = _depthwise_separable_conv2d(
net,
16,
width_multiplier,
"ds_conv_3",
downsample=True,
istraining=is_training) # ->[N, 56, 56, 128] #32,32
net = _depthwise_separable_conv2d(
net, 16, width_multiplier, "ds_conv_4",
istraining=is_training) # ->[N, 56, 56, 128]
net = _depthwise_separable_conv2d(
net,
32,
width_multiplier,
"ds_conv_5",
downsample=True,
istraining=is_training) # ->[N, 28, 28, 256] #16,16
net = _depthwise_separable_conv2d(
net, 32, width_multiplier, "ds_conv_6",
istraining=is_training) # ->[N, 28, 28, 256]
net = _depthwise_separable_conv2d(
net,
64,
width_multiplier,
"ds_conv_7",
downsample=True,
istraining=is_training) # ->[N, 14, 14, 512] #8,8
net = _depthwise_separable_conv2d(
net, 64, width_multiplier, "ds_conv_8",
istraining=is_training) # ->[N, 14, 14, 64]
net = _depthwise_separable_conv2d(
net, 64, width_multiplier, "ds_conv_9",
istraining=is_training) # ->[N, 14, 14, 64】
net = _depthwise_separable_conv2d(net,
64,
width_multiplier,
"ds_conv_10",
istraining=is_training)
net = _depthwise_separable_conv2d(net,
64,
width_multiplier,
"ds_conv_11",
istraining=is_training)
net = _depthwise_separable_conv2d(
net, 64, width_multiplier, "ds_conv_12",
istraining=is_training) # ->[N, 14, 14, 64]
net = _depthwise_separable_conv2d(
net,
128,
width_multiplier,
"ds_conv_13",
downsample=True,
istraining=is_training) # ->[N, 7, 7, 128] #4,4
net = _depthwise_separable_conv2d(
net, 128, width_multiplier, "ds_conv_14",
istraining=is_training) # ->[N, 7, 7, 1024]
net = avg_pool(net, 2, "avg_pool_15") # ->[N, 1, 1, 128]
net = tf.squeeze(
net, [1, 2],
name="SpatialSqueeze") # 去掉维度为1的维[N, 1, 1, 128] => [N,128]
logits = fc(net, 5, "fc_16") # -> [N, 5]
predictions = tf.nn.softmax(logits)
# Define loss and training operations
# 定义loss 函数和训练操作
model_loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(logits=predictions,
labels=labels))
# Tell TensorFlow to update the population statistics while training
# 通知Tensorflow在训练时要更新均值和方差的分布
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
train_opt = tf.train.AdamOptimizer(learning_rate).minimize(model_loss)
# Create operations to test accuracy
# 创建计算准确度的操作
arg_out = tf.reshape(tf.argmax(predictions, 1), [-1, 1], name='output')
correct_prediction = tf.equal(tf.argmax(predictions, 1),
tf.argmax(labels, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# Train and test the network
# 训练并测试网络模型
train_data = train_shuffle_batch(train_filename, [64, 64, 1], 128)
test_data = test_shuffle_batch(test_filename, [64, 64, 1], 30)
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
saver.restore(sess, './save/test_model.dpk')
for batch_i in range(num_batches):
# batch_xs, batch_ys = mnist.train.next_batch(batch_size)
batch_xs, batch_ys, tarin_file = sess.run(train_data)
batch_xs = batch_xs / 255 - 0.5
# batch_xs=[cv2.resize(pic, (28, 28), interpolation=cv2.INTER_CUBIC).reshape([28,28,1]) for pic in batch_xs]
batch_ys = one_hot(batch_ys.tolist(), 5)
# train this batch
# 训练样本批次
sess.run(train_opt, {
inputs: np.array(batch_xs),
labels: batch_ys,
is_training: True
})
# Periodically check the validation or training loss and accuracy
# 定期检查训练或验证集上的loss和精确度
if batch_i % 100 == 0:
test_img, test_label1, test_name = sess.run(test_data)
test_label = one_hot(test_label1.tolist(), 5)
test_img1 = test_img / 255 - 0.5
for tt, test_image in enumerate(test_img1):
# test_image1 = cv2.resize(test_image, (28, 28), interpolation=cv2.INTER_CUBIC).reshape([-1,28,28,1])
test_image1 = test_image.reshape([-1, 64, 64, 1])
test_label11 = test_label[tt].reshape([-1, 5])
loss, acc, test_out = sess.run(
[model_loss, accuracy, arg_out], {
inputs: test_image1,
labels: test_label11,
is_training: False
})
print(
'Batch: {:>2}: Validation loss: {:>3.5f}, Validation accuracy: {:>3.5f}'
.format(batch_i, loss, acc))
print('label:{},out:{}'.format(test_label1[tt], test_out))
elif batch_i % 25 == 0:
loss, acc = sess.run([model_loss, accuracy], {
inputs: batch_xs,
labels: batch_ys,
is_training: False
})
print(
'Batch: {:>2}: Training loss: {:>3.5f}, Training accuracy: {:>3.5f}'
.format(batch_i, loss, acc))
if save_dir and batch_i % 1000 == 0:
saver.save(sess, save_dir)
print('modle save at ./save/test_model.dpk' 'save succuccy!!')
graph_def = tf.get_default_graph().as_graph_def()
output_graph_def = tf.graph_util.convert_variables_to_constants(
sess, graph_def, ['output'])
with tf.gfile.GFile("./mobile222.pb", 'wb') as f:
f.write(output_graph_def.SerializeToString())
test_data = test_shuffle_batch(test_filename, [64, 64, 1], 1)
init = tf.global_variables_initializer()
saver = tf.train.Saver()
x = []
y = []
chioce_dict = dict(zip([0, 1, 2, 3, 4], list('ABCDX')))
with tf.Session() as sess:
sess.run(init)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
saver.restore(sess, './save/test_model.dpk')
for i in range(100000):
train_img, train_label, tarin_file = sess.run(train_data)
train_img1 = train_img / 255 - 0.5
train_label = one_hot(train_label.tolist(), 5)
_, train_loss, train_acc, train_out, train_l = sess.run(
[net.opt, net.loss, net.acc, net.argamx_out, net.argamx_label],
feed_dict={
net.x: train_img1,
net.y_: train_label
})
tr_acc = []
for train_index in range(len(train_out)):
if train_out[train_index] == train_l[train_index]:
tr_acc.append(1)
else:
tr_acc.append(0)
# misc.imsave('/Users/wywy/Desktop/train_e1'+'/'+str(chioce_dict.get(train_out[train_index]))+'_'+bytes.decode(tarin_file[train_index]),train_img.reshape([-1,64,64]) [train_index])