Welcome to the Deep Learning toolbelt

Here be dragons, and image classifications

Tested on

• Kaggle's Dogs vs. Cats Redux: Kernels Edition
• Kaggle's State Farm Distracted Driver Detection

Main features

1. Use ResNet-50 pretrained on imagenet with a new head layer
2. Cache of heavy computation to disk for fast re-runs
3. Confusion matrix generation
4. Kaggle sample file generation

Prerequisits:

Create directories, and populate them with your data:

• train
  • c0
  • c1 ...
• valid
  • c0
  • c1 ...
• test
  • unknown
• results

Usage

To train, run

python train.py

To predict an image, run

./predict.sh image-url-or-file-path   2> /dev/null

References

• Deep Residual Learning for Image Recognition
• Keras
• Practical Deep Learning for Coders course
• Amazon EC2 P2 Instances

License

• All code in this repository is under the MIT license as specified by the LICENSE file.
• The pretrained ResNet50 imagenet weights are ported from the ones released by Kaiming He under the MIT license.