Picnic Classifier
Employ transfer learning to retrain the Resnet101 model on the Picnic dataset. All layers are made trainable, under a limited learning rate.
Requirements
- Google Colaboratory account
Upload data to Google Drive
Create a directory for your project on Google Drive. Move train.ipynb there.
Unpack the picnic dataset into the data/ subdirectory, such that we have
<picnic drive directory>/data/train.tsv
<picnic drive directory>/data/test.tsv
<picnic drive directory>/data/test/
<picnic drive directory>/data/train/
Enable GPU acceleration for notebook
Open train.ipynb on Google Colaboratory.
Go to Edit > Notebook Settings > Hardware Acceleration, and select GPU.
Training and labeling test set
Edit the first cell from train.ipynb to point to the correct dataset directory.
By default, it points to picnic-classifier/data.
Run that cell to mount the Google Drive, you will need to type in an authorization code.
After mounting it, you can click to run all cells. After training is completed, script will
automatically label the test set using the trained model. Results will be written to data/results.tsv.
While training the model, Google Colaboratory will warn you that you are almost reaching the memory limit of the GPU. You can ignore the warning, since the script will indeed use the GPU to the limit.
Troubleshoot
Receiving Input/Output error
If you receive the following message while running the third input cell:
OSError: [Errno 5] Input/output error: '/content/gdrive/My Drive/picnic-classifier/data/train/0.png'
, first double-check the directory provided in the first cell. If it is correct, simply try re-run the notebook. Even after reporting a successful mount, Google Drive may still throw an error when trying to access the data right after authenticating.