tensorflow_mobilenet_v1

This repo mainly use tensorflow 1.8, python 3.6.2, and if it can not work correctly in a higher version, please inform me with a issue, I will update relavent code.

What's new in this repo

1. The model defination of MobileNet-V1 I used is mainly based on the work of timctho. I just modified the varible scope of models.py a bit to make its varible scope consistant with the google's official checkpoint, so that we can use the ImageNet pre-trained model to initialize our model.

2. I add another scripts as follows, so you can use your own dataset to train the MobileNet-v1 model.

   • get_tensor_from_checkpoint.py
   • loaddata.py
   • train_network.py

For more details about MobileNet-V1, please refer to the paper:MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications.

Usage

1. Download the official pre-trained model checkpoint. In this repo I use the pre-trained model with resolution = 224 and width Multiplier = 1.

2. Extract the downloaded file into ./checkpoint, the directory should have this structure: checkpoint/mobilenet_v1_1.0_224_2017_06_14

   After extraction you can use the following script to inspect all the variable name in the checkpoint file.

   python get_tensor_from_checkpoint.py -c "your checkpoint file path"

3. Prepare your train/test dataset,and re-organize the directory as follows:

   Because the resolusion of the model is 224, so you should resize your train/test images to 224 before you use them.

   dataset/
          |/train/
          |      |/class1
          |      |/class2
          |      |/class3
          |      |/class4
          |/trainAnno/
          
    ......
          |/test/
          |     |/class1
          |     |/class2
          |     |/class3
          |     |/class4
          |/testAnno/
          
    ...... 

4. Load the dataset and compute the mean of the train dataset for pre-processing.

   python loaddata.py -d "Your dataset dir" -n "the number of classes you want to classify"

   The last output in your command window will show you the pixel means train_dataset_mean of your train dataset which will use again in the train_network.py

5. Now you can train the MobileNet_v1 :python train_network.py

   Notice: if you don't have sufficient train samples, you can modify the var exclude_vars to freeze more layers to reuse more vars in the original checkpoint. But in my experiments unfreezing more layers may achieve better performence.

Contents updata late

• add a predict.py script
• replace the dataset preparation module using tf.data API. As for why I don't replace it now, I still can't figure out one thing when I parse a tfrecord file. But I think it won't be long, I will update soon.