Intro

A project for generic fine-tuning/extending (freeze CNN layers and connect with your own customzied FC layers) pre-trained/existing models, such as AlexNet, VGG-16, ... (More CNN based).

Manual

Step1. Put your datasets somewhere and train.csv under /data/alexnet_finetune

Step2. train.csv should be the format with first column to be the locations of images (I put an example in the folder)

Step3. Run:

# Download Pre-trained model
bash setup.sh

export PYTHONPATH='.'

# You can also take advantages of the Makefile, which actually inspires me dockerize this project if I have time
python apps/finetune_alexnet_train.py

Keys

One dependencies we live on is the pre-trained weights from BVLC. That said you don't wanna mess up with the name scope of those layers you wanna freeze. So be careful. How we load the weights (/services/weights_load_services) should provide you enough information.

You are free to add as many layers as you like and just be aware that conv1 ~ conv5 and fc6~fc8 are those layers (as well as the NAME) you can load a pre-trained weights.

I try very hard to implement what I thought is the best practice for tensorflow -- separate:

• architecture model (models/alexnet.py)

• computation model (models/finetune_graph.py)

• trainer model (models/train.py)

• training app (apps/finetune_alexnet_train.py)

so that you can independently change any part of those without impacting other components.

A better example will be my anther repo -- generic CNN in tensorflow, which may have a better idea of what I am trying to do. Since for this project, you need to sacrifice some graceful implementation due to the constraints of the pre-trained weights organization.

references

1. AlexNet Paper
2. Fine Tuning AlexNet on Tensorflow Example
3. AlexNet Explainations in details