cnn_implementation

A simple implementation of CNN using Tensorflow.

Requirements

0. Python 3.x
1. Tensorflow 1.3
2. Numpy 1.11.3
3. PIL 5.1.0

Installation

1. Clone this repository.

2. Install the dependencies. The code should run with TensorFlow 1.0 and newer.

pip install -r requirements.txt  # or make install

Usage

Configuration

First configure your global parameters in global.conf as follows:

• train_data_dir: directory of training data.
• eval_data_dir: directory of evaluation data.
• num_class: number of classes to be classified.
• resize_image_height: resize the input images with the height.
• resize_image_width: resize the input images with the width.
• chnnels: channels of input images.
• batch_size: batch size in each step.
• train_tfrecord_dir: directory of training tfrecord.
• train_data_count: count of training data.
• max_steps: max steps during training.
• eval_log_dir: directory of evaluation log.
• eval_tfrecord_dir: directory of evaluation tfrecord.
• eval_data_count: count of evaluation data.
• model_dir: directory of the trained models.
• moving_average_decay: parameter of moving average decay.
• num_epochs_per_decay: parameter of number of epochs per decay.
• learning_rate_decay_factor: parameter of learning rate decay factor.
• initial_average_decay: parameter of initial average decay.
• tower_name: tower name.
• keep_prob: keep probability in dropout layer.

Customizing network architecture

You can customize your network architecture using network.json with the layer names ("layers"), layers weights ("weights") and biases ("biases") as follows:

{"layers": ["conv1", "conv2", "conv3", "conv4", "conv5", "fc1", "fc2", "fc3"],
 "weights": 
    {"wconv1": [11, 11, 3, 64], 
     "wconv2": [5, 5, 64, 192],
     "wconv3": [3, 3, 192, 384],
     "wconv4": [3, 3, 384, 256], 
     "wconv5": [3, 3, 256, 256], 
     "wfc1": [12544, 4096], 
     "wfc2": [4096, 4096],
     "wfc3": [4096, 10]}, 
 "biases": 
    {"bconv1": [64], 
     "bconv2": [192],  
     "bconv3": [384], 
     "bconv4": [256], 
     "bconv5": [256], 
     "bfc1": [4096], 
     "bfc2": [4096],
     "bfc3": [10]}}

Training model

python train_model.py

You will get:

2018-12-12 20:54:18.709988: step 0, loss = 7.31 (78.9 examples/sec; 0.634 sec/batch)
2018-12-12 20:54:57.119095: step 1, loss = 7.29 (78.9 examples/sec; 0.634 sec/batch)

The model will be evaluated in each 100 steps:

2018-12-12 21:41:39.556357: precision @ 1 = 1.000

Visualization in TensorBoard

To start Tensorflow, run the following command on the console:

#!bash

tensorboard --logdir=./model

Prediction

python predict_inputs.py --input_img ./data/1.png