1. inception v4
import os
import numpy as np
import tensorflow as tf
from data_generator import MNISTData
#reduction表示输出resolution下降
#inception表示相同维度输出！
images_path = "data/MNIST_train_images.gz"
labels_path = "data/MNIST_train_labels.gz"
export_path = "saved_model/"
ckpt_path = "tmp/ckpt"
mnist = MNISTData(images_path, labels_path)
validate_set, train_set = mnist.split(5000)
def main():
model = InceptionModel([None, 28, 28, 1], [None, 10])
#model.train(train_set, epochs=6875)
img, label = validate_set.get_batch(85)
result = model.classify(img)
print(np.argmax(result, axis=1))
print(np.argmax(label, axis=1))
correct_prediction = np.equal(np.argmax(result, 1), np.argmax(label, 1))
accuracy = np.sum(correct_prediction) / correct_prediction.size
print(correct_prediction)
print(accuracy)
return
class InceptionModel():
'''
A base class for an Inception v4 model
'''
def __init__(self, input_shape, output_shape, ckpt_path='train_model/model', model_path='saved_model/model'):
'''
Defines the input and output variables and stores the save locations of the model
Keyword arguments:
input_shape -- The shape of the input tensor, indexed by [sample, row, col, ch]
output_shape -- The shape of the output tensor, indexed by [sample, class]
ckpt_path -- The save path for checkpoints during training
model_path -- The save path for the forward propagation subgraph. The generated model
can be used to classify images without allocating space for the gradients
'''
self.ckpt_path = ckpt_path
self.model_path = model_path
self.num_class = output_shape[1]
self.X = tf.placeholder(tf.float32, input_shape)
self.y = tf.placeholder(tf.float32, output_shape)
self.define_model()
return
def define_model(self):
'''
Defines the Inception v4 model.
'''
self.keep_prob = tf.placeholder(tf.float32)
with tf.variable_scope('input'):
_X = tf.image.resize_images(self.X, [299, 299])
with tf.variable_scope('stem'):
_stem = stem(_X)
_inception_a = {-1: _stem}
for i in range(4):#4次运算，用字典来存放计算结果
with tf.variable_scope('inception_a_'+str(i)):
_inception_a[i] = inception_a(_inception_a[i-1])
with tf.variable_scope('reduction_a'):
_reduction_a = reduction_a(_inception_a[3])
_inception_b = {-1: _reduction_a}
for i in range(7):
with tf.variable_scope('inception_b_'+str(i)):
_inception_b[i] = inception_b(_inception_b[i-1])
with tf.variable_scope('reduction_b'):
_reduction_b = reduction_b(_inception_b[6])
_inception_c = {-1: _reduction_b}
for i in range(3):
with tf.variable_scope('inception_c_'+str(i)):
_inception_c[i] = inception_c(_inception_c[i-1])
pool = tf.nn.avg_pool(_inception_c[2], [1, 8, 8, 1], [1, 1, 1, 1], padding='VALID', name='pool')
pool_f = tf.reshape(pool, [-1, 1536])
pool_drop = tf.nn.dropout(pool_f, self.keep_prob)
self.fc = dense(pool_drop, 'fc', self.num_class)
self.y_hat = tf.nn.softmax(self.fc, name='y_hat')
# Creates a saver object exclusively for the forward propagation subgraph
model_variables = tf.get_collection_ref('tf.GraphKeys.MODEL_VARIABLES')
self.model_saver = tf.train.Saver(model_variables)
return
def train(self, data_generator, batch_size=8, epochs=4000, keep_prob=0.8):
'''
Defines the variables necessary for training then begins training
Keyword arguments:
data_generator -- a data_generator object with an implementation of get_batch, as seen
in the data_generator.py module
batch_size -- the number of samples to be used in each training batch. Keep memory
constraints in mind
epochs -- the number of epochs to be used for training
keep_prob -- the keep probability of the dropout layer to be used for training
'''
self.cross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=self.y, logits=self.fc)
self.train_step = tf.train.AdamOptimizer(1e-4).minimize(self.cross_entropy)
self.correct_prediction = tf.equal(tf.argmax(self.y_hat, 1), tf.argmax(self.y, 1))
self.accuracy = tf.reduce_mean(tf.cast(self.correct_prediction, tf.float32))
# Creates a saver object to generate checkpoints during training. This one also saves
# the gradients and the increment of the Adam Optimizer
self.saver = tf.train.Saver()
with tf.Session() as sess:
if os.path.isdir('train_model'):
self.saver.restore(sess, self.ckpt_path)
else:
sess.run(tf.global_variables_initializer())
for i in range(epochs):
images, labels = data_generator.get_batch(batch_size)
if i % 50 == 0:
train_accuracy = self.accuracy.eval(feed_dict={self.X: images, self.y: labels, self.keep_prob: 1.0})
print('step %d, training accuracy %g' % (i, train_accuracy))
if i % 500 == 0:
self.saver.save(sess, self.ckpt_path)
self.train_step.run(feed_dict={self.X: images, self.y: labels, self.keep_prob: keep_prob})
self.model_saver.save(sess, self.model_path)
return
def classify(self, image):
'''
Classifies the input image based on the trained model in model_path
Keyword arguments:
image -- the image, or image array, indexed by [sample, row, col, ch]
'''
with tf.Session() as sess:
self.model_saver.restore(sess, self.model_path)
y = self.y_hat.eval(feed_dict={self.X: image, self.keep_prob: 1.0})
return y
def stem(tensor):
'''
Generates the graph for the stem subgraph of the Inception v4 model
'''
conv_1 = conv(tensor, 'conv_1', [3, 3, 1, 32], [1, 2, 2, 1], padding='VALID')
conv_2 = conv(conv_1, 'conv_2', [3, 3, 32, 32], [1, 1, 1, 1], padding='VALID')
conv_3 = conv(conv_2, 'conv_3', [3, 3, 32, 64], [1, 1, 1, 1])
conv_4_1 = conv(conv_3, 'conv_4_1', [3, 3, 64, 64], [1, 2, 2, 1], padding='VALID')
conv_4_2 = conv(conv_3, 'conv_4_2', [3, 3, 64, 96], [1, 2, 2, 1], padding='VALID')
concat_1 = tf.concat([conv_4_1, conv_4_2], axis=3, name='concat_1')
conv_5_1_1 = conv(concat_1, 'conv_5_1_1', [1, 1, 160, 64], [1, 1, 1, 1])
conv_5_1_2 = conv(conv_5_1_1, 'conv_5_1_2', [3, 3, 64, 96], [1, 1, 1, 1], padding='VALID')
conv_5_2_1 = conv(concat_1, 'conv_5_2_1', [1, 1, 160, 64], [1, 1, 1, 1])
conv_5_2_2 = conv(conv_5_2_1, 'conv_5_2_2', [7, 1, 64, 64], [1, 1, 1, 1])
conv_5_2_3 = conv(conv_5_2_2, 'conv_5_2_3', [1, 7, 64, 64], [1, 1, 1, 1])
conv_5_2_4 = conv(conv_5_2_3, 'conv_5_2_4', [3, 3, 64, 96], [1, 1, 1, 1], padding='VALID')
concat_2 = tf.concat([conv_5_1_2, conv_5_2_4], axis=3, name='concat_2')
conv_6_1 = conv(concat_2, 'conv_6_1_1', [3, 3, 192, 192], [1, 2, 2, 1], padding='VALID')
pool_6_2 = tf.nn.max_pool(concat_2, [1, 3, 3, 1], [1, 2, 2, 1], padding='VALID', name='pool_6_2')
concat_3 = tf.concat([conv_6_1, pool_6_2], axis=3, name='concat_3')
return concat_3
def inception_a(tensor):
'''
Generates the graph for the Inception A subgraph of the Inception v4 model
'''
pool_1_1 = tf.nn.avg_pool(tensor, [1, 3, 3, 1], [1, 1, 1, 1], padding='SAME', name='pool_1_1')
conv_1_1 = conv(pool_1_1, 'conv_1_1', [1, 1, 384, 96], [1, 1, 1, 1])
conv_2_1 = conv(tensor, 'conv_2_1', [1, 1, 384, 96], [1, 1, 1, 1])
conv_3_1 = conv(tensor, 'conv_3_1', [1, 1, 384, 64], [1, 1, 1, 1])
conv_3_2 = conv(conv_3_1, 'conv_3_2', [3, 3, 64, 96], [1, 1, 1, 1])
conv_4_1 = conv(tensor, 'conv_4_1', [1, 1, 384, 64], [1, 1, 1, 1])
conv_4_2 = conv(conv_4_1, 'conv_4_2', [3, 3, 64, 96], [1, 1, 1, 1])
conv_4_3 = conv(conv_4_2, 'conv_4_3', [3, 3, 96, 96], [1, 1, 1, 1])
concat = tf.concat([conv_1_1, conv_2_1, conv_3_2, conv_4_3], axis=3, name='concat')
return concat
def reduction_a(tensor):
'''
Generates the graph for the Reduction A subgraph of the Inception v4 model
'''
pool_1_1 = tf.nn.max_pool(tensor, [1, 3, 3, 1], [1, 2, 2, 1], padding='VALID', name='pool_1_1')
conv_2_1 = conv(tensor, 'conv_2_1', [3, 3, 384, 384], [1, 2, 2, 1], padding='VALID')
conv_3_1 = conv(tensor, 'conv_3_1', [1, 1, 384, 192], [1, 1, 1, 1])
conv_3_2 = conv(conv_3_1, 'conv_3_2', [3, 3, 192, 224], [1, 1, 1, 1])
conv_3_3 = conv(conv_3_2, 'conv_3_3', [3, 3, 224, 256], [1, 2, 2, 1], padding='VALID')
concat = tf.concat([pool_1_1, conv_2_1, conv_3_3], axis=3, name='concat')
return concat
def inception_b(tensor):
'''
Generates the graph for the Inception B subgraph of the Inception v4 model
'''
pool_1_1 = tf.nn.avg_pool(tensor, [1, 3, 3, 1], [1, 1, 1, 1], padding='SAME', name='pool_1_1')
conv_1_2 = conv(pool_1_1, 'conv_1_2', [1, 1, 1024, 128], [1, 1, 1, 1],)
conv_2_1 = conv(tensor, 'conv_2_1', [1, 1, 1024, 384], [1, 1, 1, 1])
conv_3_1 = conv(tensor, 'conv_3_1', [1, 1, 1024, 192], [1, 1, 1, 1])
conv_3_2 = conv(conv_3_1, 'conv_3_2', [1, 7, 192, 224], [1, 1, 1, 1])
conv_3_3 = conv(conv_3_2, 'conv_3_3', [1, 7, 224, 256], [1, 1, 1, 1])
conv_4_1 = conv(tensor, 'conv_4_1', [1, 1, 1024, 192], [1, 1, 1, 1])
conv_4_2 = conv(conv_4_1, 'conv_4_2', [1, 7, 192, 192], [1, 1, 1, 1])
conv_4_3 = conv(conv_4_2, 'conv_4_3', [7, 1, 192, 224], [1, 1, 1, 1])
conv_4_4 = conv(conv_4_3, 'conv_4_4', [1, 7, 224, 224], [1, 1, 1, 1])
conv_4_5 = conv(conv_4_4, 'conv_4_5', [7, 1, 224, 256], [1, 1, 1, 1])
concat = tf.concat([conv_1_2, conv_2_1, conv_3_3, conv_4_5], axis=3, name='concat')
return concat
def reduction_b(tensor):
'''
Generates the graph for the Reduction B subgraph of the Inception v4 model
'''
pool_1_1 = tf.nn.max_pool(tensor, [1, 3, 3, 1], [1, 2, 2, 1], padding='VALID', name='pool_1_1')
conv_2_1 = conv(tensor, 'conv_2_1', [1, 1, 1024, 192], [1, 1, 1, 1])
conv_2_2 = conv(conv_2_1, 'conv_2_2', [3, 3, 192, 192], [1, 2, 2, 1], padding='VALID')
conv_3_1 = conv(tensor, 'conv_3_1', [1, 1, 1024, 256], [1, 1, 1, 1])
conv_3_2 = conv(conv_3_1, 'conv_3_2', [1, 7, 256, 256], [1, 1, 1, 1])
conv_3_3 = conv(conv_3_2, 'conv_3_3', [7, 1, 256, 320], [1, 1, 1, 1])
conv_3_4 = conv(conv_3_3, 'conv_3_4', [3, 3, 320, 320], [1, 2, 2, 1], padding='VALID')
concat = tf.concat([pool_1_1, conv_2_2, conv_3_4], axis=3, name='concat')
return concat
def inception_c(tensor):
'''
Generates the graph for the Inception C subgraph of the Inception v4 model
'''
pool_1_1 = tf.nn.avg_pool(tensor, [1, 3, 3, 1], [1, 1, 1, 1], padding='SAME', name='pool_1_1')
conv_1_2 = conv(pool_1_1, 'conv_1_2', [1, 1, 1536, 256], [1, 1, 1, 1])
conv_2_1 = conv(tensor, 'conv_2_1', [1, 1, 1536, 256], [1, 1, 1, 1])
conv_3_1 = conv(tensor, 'conv_3_1', [1, 1, 1536, 384], [1, 1, 1, 1])
conv_3_2_1 = conv(conv_3_1, 'conv_3_2_1', [1, 3, 384, 256], [1, 1, 1, 1])
conv_3_2_2 = conv(conv_3_1, 'conv_3_2_2', [3, 1, 384, 256], [1, 1, 1, 1])
conv_4_1 = conv(tensor, 'conv_4_1', [1, 1, 1536, 384], [1, 1, 1, 1])
conv_4_2 = conv(conv_4_1, 'conv_4_2', [1, 3, 384, 448], [1, 1, 1, 1])
conv_4_3 = conv(conv_4_2, 'conv_4_3', [3, 1, 448, 512], [1, 1, 1, 1])
conv_4_3_1 = conv(conv_4_3, 'conv_4_3_1', [1, 3, 512, 256], [1, 1, 1, 1])
conv_4_3_2 = conv(conv_4_3, 'conv_4_3_2', [3, 1, 512, 256], [1, 1, 1, 1])
concat = tf.concat([conv_1_2, conv_2_1, conv_3_2_1, conv_3_2_2, conv_4_3_1, conv_4_3_2], axis=3, name='concat')
return concat
def conv(tensor, name, shape, strides=[1, 1, 1, 1], padding='SAME', activation=tf.nn.relu):
'''
Generates a convolutional layer
Keyword arguments:
tensor -- input tensor. Must be indexed by [sample, row, col, ch]
name -- the name that will be given to the tensorflow Variable in the GraphDef
shape -- the shape of the kernel. Must be indexed by [row, col, num_input_ch, num_output_ch]
strides -- the stride of the convolution. Must be indexed by [sample, row, col, ch]
padding -- if set to 'SAME', the output will have the same height and width as the input. If
set to 'VALID', the output will have its size reduced by the difference between the
tensor size and kernel size
activation -- the activation function to use
'''
W = tf.get_variable(name+"_W", shape)
b = tf.get_variable(name+"_b", shape[-1])
tf.add_to_collection('tf.GraphKeys.MODEL_VARIABLES', W)
tf.add_to_collection('tf.GraphKeys.MODEL_VARIABLES', b)
z = tf.nn.conv2d(tensor, W, strides=strides, padding=padding, name=name+'_z')
h = tf.add(z, b, name=name+'_h')
a = activation(h, name=name+'_a')
return a
def dense(tensor, name, num_out):
'''
Generates a fully connected layer. Does not apply an activation function
Keyword arguments:
tensor -- input tensor. Must be indexed by [sample, ch]
name -- the name that will be given to the tensorflow Variable in the GraphDef
num_out -- the size of the output tensor
'''
W_fc = tf.get_variable('W_fc', [tensor.shape[1], num_out])
b_fc = tf.get_variable('b_fc', [num_out])
tf.add_to_collection('tf.GraphKeys.MODEL_VARIABLES', W_fc)
tf.add_to_collection('tf.GraphKeys.MODEL_VARIABLES', b_fc)
z_fc = tf.matmul(tensor, W_fc, name='z_fc')
h_fc = tf.add(z_fc, b_fc, name='h_fc')
return h_fc
if __name__ == '__main__':
main()
扫一扫
282
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
