• Getting Started
  • Prerequisites
  • Installing
• Running the model
  • Dataset
  • Training
    • Original Training
    • Customize Training
  • Testing
    • Original Testing
    • Customize Testing

• python 3.6.5
• conda 4.5.4
• tensorflow 1.10.0
• opencv-python 3.4.2
• Windows 7+ / Mac OS / Ubuntu 14.04+

If you haven't installed anaconda on your computer, here is the URL link to Download Anaconda After installed anaconda, tensorflow and opencv can be installed by:

$ pip install tensorflow
$ pip install opencv-python

Make sure they are updated to the latest version. You can check their version by:

$ python -c 'import tensorflow as tf; print(tf.__version__)'
$ python -c 'import cv2; print(cv2.__version__)'

• Creating your own dataset is really convenient. In data directory there are two directories train and test.

  All you have to do is to categorize your picture set in different directories respectively and store them in train.

  Lenet.py will use the name of these directories as labels and train them accordingly. It is all the same for test dataset.

  cat and dog directories are just examples, you can delete them after cloning this repository.

• Pictures must be in .jpg format.

• If you are using dataset collected by yourself, the dataset are preferably renamed and resized already.

If you only want to use the simple Lenet model with 2 convolutional layers, all you have to do is to run LeNet.py by python:

$ python LeNet.py

Then the program will load all the images in train and test first. The reading process will be printed on the terminal:

reading the image:data/train/dog/dog.731.jpg
reading the image:data/train/dog/dog.514.jpg
reading the image:data/train/cat/cat.793.jpg
reading the image:data/train/cat/cat.996.jpg
reading the image:data/train/cat/cat.436.jpg

The training process will start as soon as all the images are successfully loaded.
Depending on different configurations, the result of every epoch is going to show within different duration:

train loss: 1.8417279117135887
train acc:0.65571
test loss: 1.898901087524247
test acc:0.588808

The default epoch number is 50 times. After the training process, a .pb file will be generated in the original directory and know you can use test.py to access the .pb file.

If you want to use tensorboard, the tfevent file will be stored in log directory

I marked all the places that can be customized by #TODO

• #TODO 1:

  w = 32
  h = 32
  c = 3

Those are parameters for reshaping images. Depending on your computer configurations and your expected accuracy, you can change w and h. For colored pictures, c should be 3(RGB)

• #TODO 2:

pb_file_path = "Lenet.pb"

You can change the path for storing .pb file and the name of the file.

• #TODO 3: All the major changes such as the number of convolutional layers, the number of kernels in each layers and the size of each kernel can all be modified here.
  The basic structure for one convolutional layers + one maxpooling layers is as followed:

with tf.variable_scope('layer1-conv1'):
        conv1_weights = tf.get_variable('weight',[5,5,c,6],initializer=tf.truncated_normal_initializer(stddev=0.1))
        conv1_biases = tf.get_variable('bias',[6],initializer=tf.constant_initializer(0.0))
        conv1 = tf.nn.conv2d(input_tensor,conv1_weights,strides=[1,1,1,1],padding='VALID')
        relu1 = tf.nn.relu(tf.nn.bias_add(conv1,conv1_biases))
with tf.name_scope('layer2-pool1'):
        pool1 = tf.nn.max_pool(relu1,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')

You can increase the number of layers by simply copy and paste the whole structure and change the name of those variables.
Other parameters can also be changed according to your need.

• TODO 4: Full connected layers can also be modified.
  The basic structure for one full connected layer is as followed:

with tf.variable_scope('layer5-fc1'):
        fc1_weights = tf.get_variable('weight',[nodes,120],initializer=tf.truncated_normal_initializer(stddev=0.1))
        if regularizer != None:
            tf.add_to_collection('losses',regularizer(fc1_weights))
        fc1_biases = tf.get_variable('bias',[120],initializer=tf.constant_initializer(0.1))
        fc1 = tf.nn.relu(tf.matmul(reshaped,fc1_weights) + fc1_biases)
        if train:
            fc1 = tf.nn.dropout(fc1,0.5)

• TODO 5 and TODO 6:

cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y,labels=y_)

cross_entropy_mean = tf.reduce_mean(cross_entropy)

loss = cross_entropy_mean + tf.add_n(tf.get_collection('losses'))

train_op = tf.train.AdamOptimizer(0.001).minimize(loss)

I am using cross entropy mean as my loss function and AdamOptimizer as my optimizer. You can change them into whatever you like.

• TODO 7:

    train_num = 50
    batch_size = 12

Again, those two variables are decided by your computer configurations.

Testing is much simpler.
First, go to test.py and find #TODO a

labels = {'[0]': 'cat', '[1]': 'dog'}

Change keys and values in labels according to the labels of your dataset
Then find #TODO d

if __name__ = '__main__':
    #recognize("img_path", "sample.pb")

Uncomment the recognize function and write the path of your testing img in img path and rename the .pb file.
Then run test.py in your terminal:

$ python test.py

The predicted label should be shown on the terminal:

label: cat

• TODO c: You can add further application based on the predicted label or import test as a module in other python program.