77summer/mnist_lenet

implement LeNet for MNIST using tensorflow, caffe and PyTorch

MNIST recognition

implement the LeNet for digit recognition using frameworks as follows:

1. Tensorflow
2. Keras
3. Caffe
4. PyTorch

objective

My goal in this project is to get familiar with these frameworks.

Why MNIST

MNIST seems to be a "hello world" in DL. And also a very good choice to get hands on CNN.

Usage

prequiste

you will need many things, but fortunately all is easy to install and cofigue.

• python
• tensorflow
• tensorboard
• tqdm: this was used for logged imfo
• pytorch
• keras

run

run like this

python main.py  --train --train_num 1000 --load --framework tensorflow

• --train : indicate that you want to train the model. If this flag is set on, a --train_num must followed
• --load : use while you want to load a local model
• --framework: choose your framework: tensorflow, pytorch or keras ...

there are other flags, like the --log_dir indicating the directory to store the log info

Experiments

Experiments were made. See meta/doc/report.md for detailed result and analysis.

I made some modifications on the lenet architecture in order to find what matters and what not