cnn model - tutorial 2 - cv trick

cnn-model.txt

classes = ['dogs','cats']
num_classes = len(classes)

# refer to this link for why need this complicated path declaration 
# https://stackoverflow.com/questions/37619246/python-glob-glob-always-returns-empty-list
train_path = '/home/*datduyn/*Documents/*dataset/*dog-cat/*training_set/'

#validation split
validation_size = 0.2

#batch size
batch_size = 32

#image size 
img_size = 128 

num_channels = 3

#TODO: download the class from the link procide above. 
data = read_train_sets(train_path,img_size,classes,validation_size=validation_size)

print(data)
#start the session 
session = tf.Session()

#place holder and input(x) 
x = tf.placeholder(tf.float32,shape=[None,img_size,img_size,num_channels],name='x')

y_true = tf.placeholder(tf.float32, shape=[None, num_classes], name='y_true')
y_true_cls = tf.argmax(y_true,dimension=1)



##Network graph params
filter_size_conv1 = 3 
num_filters_conv1 = 32

filter_size_conv2 = 3
num_filters_conv2 = 32

filter_size_conv3 = 3
num_filters_conv3 = 64
    
fc_layer_size = 128



# create the weight quickly 
def create_weights(shape):
    return tf.Variable(tf.truncated_normal(shape,stddev=0.05))

#create bias quickly 
def create_bias(size):
    return tf.Variable(tf.constant(0.05,shape=[size]))


# create convolutional layer 
def create_convolutional_layer(input, 
                        num_input_channels,
                        conv_filter_size,
                        num_filters):
    
    # define weight that will be train 
    weights = create_weights(shape=[conv_filter_size,conv_filter_size,num_input_channels,num_filters])
    
    #create trained bias 
    biases = create_bias(num_filters)
    
    # creating convolutional layer 
    layer = tf.nn.conv2d(input=input,
                    filter=weights,
                    strides=[1,1,1,1], 
                    padding='SAME')
    
    #update bias 
    layer+= biases
    
    # max pool to reduce size 
    layer = tf.nn.max_pool(value=layer,
                    ksize=[1,2,2,1],
                           strides=[1,2,2,1],# move 2 pixel x,y dir
                           padding='SAME')
    
    #output of pooling is fed to RELU(activation function)
    layer=tf.nn.relu(layer)
    
    return layer
    
'''
The output of a convolution layer is multi-dimensional tensor. 
we want to convert this into a 1-D tensor. called flattening layer
'''
def create_flatten_layer(layer):
    layer_shape=layer.get_shape()
    num_features=layer_shape[1:4].num_elements()# TODO: figure out what this line mean 
    layer = tf.reshape(layer,[-1,num_features])# 1d array with num_features elements
    
    return layer

'''
create fully connected layer 
'''
def create_fc_layer(input,
            num_inputs,
            num_outputs,
            use_relu=True):
    
    #define trainable weight and biases 
    weights=create_weights(shape=[num_inputs,num_outputs])
    biases=create_bias(num_outputs)
    
    layer = tf.matmul(input,weights) + biases 
    
    if use_relu:
        layer = tf.nn.relu(layer)
        
    return layer

'''
network design

note:
filter_size_conv* : global variable from lib
num_filters_conv*: global variable from lib
'''


# create 3 layer for the network 
layer_conv1 = create_convolutional_layer(input=x,
                num_input_channels=num_channels,
                conv_filter_size=filter_size_conv1,
                num_filters=num_filters_conv1)

layer_conv2 = create_convolutional_layer(input=layer_conv1,
                num_input_channels=num_filters_conv1,
                conv_filter_size=filter_size_conv2,
                num_filters=num_filters_conv2) 

layer_conv3 = create_convolutional_layer(input=layer_conv2,
                num_input_channels=num_filters_conv2,
                conv_filter_size=filter_size_conv3,
                num_filters=num_filters_conv3)

# flatten the layer(convert to 1d array)
layer_flat = create_flatten_layer(layer_conv3)

# create connected layer 
layer_fc1 = create_fc_layer(input=layer_flat,
                    num_inputs=layer_flat.get_shape()[1:4].num_elements(),
                    num_outputs=fc_layer_size,
                    use_relu=True)

layer_fc2 = create_fc_layer(input=layer_fc1,
                    num_inputs=fc_layer_size,
                    num_outputs=num_classes,
                    use_relu=False)
                            

'''
Prediction:
'''
#get the probability of each class to output the fully connected layer(use softmax)
#y_pred contains the predicted probability of each class for each input image. 
y_pred = tf.nn.softmax(layer_fc2,name="y_pred")

#pick the highest probability of the all probability
y_pred_cls = tf.argmax(y_pred,dimension=1)

#init global variable from session 
session.run(tf.global_variables_initializer())

# minimize cost(by weight) using cross entropy
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=layer_fc2,
                                            labels=y_true)

#minimize cost by finding mean of cross entropy function area. 
cost = tf.reduce_mean(cross_entropy)

# Optimization:Adamoptimizer
optimizer = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(cost)

#check for accuracy 
correct_prediction = tf.equal(y_pred_cls, y_true_cls)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

session.run(tf.global_variables_initializer()) 

'''
show the progress
'''
def show_progress(epoch, feed_dict_train, feed_dict_validate, val_loss):
    
    #report  the validation accuracy
    acc = session.run(accuracy, feed_dict=feed_dict_train)
    
    #report the accuracy for the train image 
    val_acc = session.run(accuracy, feed_dict=feed_dict_validate)
    
    #print out to console(report result)
    msg = "Training Epoch {0} --- Training Accuracy: {1:>6.1%}, Validation Accuracy: {2:>6.1%},  Validation Loss: {3:.3f}"
    print(msg.format(epoch + 1, acc, val_acc, val_loss)) 
    
# create a saver to save model
saver = tf.train.Saver()

# global total_iterations
total_iterations = 0

# train the model 
def train(num_iteration):
    global total_iterations
    
    for i in range(total_iterations,
                   total_iterations + num_iteration):
        
        #use function from dataset class(.train.next_batch() return # of next batch)
        x_batch, y_true_batch, _, cls_batch = data.train.next_batch(batch_size)
        x_valid_batch, y_valid_batch, _, valid_cls_batch = data.valid.next_batch(batch_size)

        # train dict
        feed_dict_tr = {x: x_batch,
                    y_true: y_true_batch}
        
        # validation dict 
        feed_dict_val = {x: x_valid_batch,
                    y_true: y_valid_batch}
        
        #report result
        session.run(optimizer, feed_dict=feed_dict_tr)
        
        if i % int(data.train.num_examples/batch_size) == 0:
            val_loss = session.run(cost, feed_dict=feed_dict_val)
            epoch = int(i / int(data.train.num_examples/batch_size))
            
            show_progress(epoch, feed_dict_tr, feed_dict_val, val_loss)
            saver.save(session, './dogs-cats-model')
            
    total_iteration += num_iteration
        
train(num_iteration=3000)