WANN vs ResNet

Initialization

Install the requirements. On Google Colab, only jenkspy and unidip need to be intalled.

pip install -r requirements.txt

Run the download script to obtain the dataset used for the 18 classes we trained on.

sh download.sh

Training

To train the resnet, run the following command.

python train_resnet.py --lr 0.0001 --epochs 50 --batch 1 --name default-experiment --momentum 0.9 --weightdecay 5e-4 --incrlr False --upperlr 0.01

The flags correspond to the following:

• lr = learning rate

• epochs = number of epochs

• batch = batch size (warning: RAM usable quickly increases as batch size increases)

• name = what the model .pth file will be named

• momentum = gradient descent momentum

• weightdecay = gradient descent weight decay

• incrlr = if included, then we increment the learning rate to the upper limit upperlr. If omitted, we do not increment the learning rate.

• upperlr = the upper limit learning rate we iteratively sum to

A similar command exists for training a population of WANN's with the NEAT algorithm.

python train_wann.py --epochs 50 --batch 24 --name default-experiment

Models

Model files are generated for resnet and WANN training. For a name experiment-name, you get model files experiment-name_detector.pth and experiment-name_best_detector.pth for resnet. For WANN, the file experiment-name.json is generated.

Testing

The commands for testing are similar to training. To test the accuracy of the resnet and WANN, call python on the test_resnet.py and test_wann.py files.

References

General sources:

• Gaier - Weight Agnostic NNs
• Lu - Free-hand Sketch Recognition

Specific sources:

• Data
  • Google quickdraw
  • TU-Berlin not used due to the existence of ambiguous categories. The pruned data set due to Schneider was apparently never released.
• numpy NEAT algorithm - ?
• Coding:
  • Sample CNN
  • Adding res connection
  • torch.nn docu