MNIST Interactive Examples in PyTorch & TensorFlow Eager Mode

Modified from PyTorch MNIST official example. Recreated with TensorFlow under eager execution mode. With detailed comments & interactive interface.

深度學習新手村：PyTorch入門（中文）

Structure

pytorch/
  train.py            # Train the model
  model.py            # The defined model
  app.py              # Interactive predictor
  model               # Pretrained model, will be overriden when you start training
  test_n.png          # Sample images for the use of interactive predictor
  
tensorflow_eager/
  train.py            # Train the model
  model.py            # The defined model
  app.py              # Interactive predictor
  checkpoint, ckpt-*  # Pretrained model, the number after prefix is the final training step
  test_n.png          # Sample images for the use of interactive predictor

Usage

# clone project
$ git clone https://github.com/pyliaorachel/pytorch-mnist-interactive.git
$ cd MNIST-pytorch-tensorflow-eager-interactive

# install dependencies
$ pip3 install -r requirements.txt

# train & test model
$ python3 -m pytorch.train
# ...data will be fetched to ../data/
# ...trained model will be saved to ./pytorch/model
# or
$ python3 -m tensorflow_eager.train
# ...data will be fetched to somewhere
# ...trained model will be saved to ./tensorflow_eager/checkpoint & ./tensorflow_eager/ckpt-*

# test model interactively
$ python3 -m pytorch.app --image=<path-to-image>
# or
$ python3 -m tensorflow_eager.app --image=<path-to-image>

Experiments

Machine Settings

OS	CPU	Memory
MacOS 10.12.4	2 GHz Intel Core i5	16 GB 1867 MHz LPDDR3

Results

	TensorFlow Eager	PyTorch
Time	real 6m4.446s user 13m42.909s sys 1m54.327s	real 3m59.340s user 3m28.285s sys 0m57.395s
Avg. Loss (Test)	0.0610	0.0473
Accuracy	9856/10000 (99%)	9845/10000 (98%)

Avg. Loss and Accuracy are expected to be more or less the same. PyTorch is half the time of TensorFlow's on CPU, while the code complexity is the same.