foxik/predibajk

Classification from image data

PrediBajk

The repository contains the following scripts:

• download_images.sh: given an input tsv file with URL in the first column, download the images, resize them to height 480px, save them as jpgs, and print the output tsv with the image path prepended as the first column.

• model_train.py: train a model using an EfficientNetV2 backbone. The trained is saved as Keras model in SavedModel format. The script has the following options:

  • --augment (default flip,rotate,scale,colors): augmentation operations to apply to the training data
  • --batch_size (default 8): batch size to use; 8 works on a 16GB GPU with a small variant of EfficientNetV2, 4 works on a 24GB GPU with a medium EfficientNetV2
  • --data_classes (default all): which classes from the input data file to predict, or all to predict all of them
  • --data_folds (default 0): number of folds to use during cross-validation; value of 0 indicates not to use cross-validation
  • --data_fold (default 0): zero-based index of a fold to use during cross-validation
  • --data_path (default bikes.tsv): input data file in tsv format
  • --dropout (default 0.0): dropout rate to use
  • --efficientnetv2_size (default s): EfficientNetV2 size (s or m)
  • --output (default model): path where the trained model should be saved
  • --schedule (default frozen:30:1e-3,finetune:30:1e-4): training schedule to use
  • --seed (default 42): random seed to use
  • --threads (default 4): CPU threads to use (for augmentation and other CPU computation); 0 indicates all physical cores

  The script has been developed using tensorflow==2.8.0, tensorflow-addons==0.16.1, and tensorflow-hub==0.12.0, but newer (consistent) versions are expected to work too.

• predict.py: predict the classes for the given images. The first argument is the directory with the model, all remaining arguments are paths to the filed to predict. Furthermore, the following options are supported:

  • --batch_size (default 8): the batch size to use during prediction
  • --threads (default 4): CPU threads to use; 0 indicates all physical cores

  The output is a tsv file, the first column image_path is the path to the file being predicted, and other columns are the predicted classes.

  Only the tensorflow package is required for prediction.

• predict_jackknife.py: when a model (given as the first argument) was trained via cross-validation, perform the prediction on the development fold (the one not used during training). Output a tsv file with both gold and predicted classes for each file in the development fold.

• visualize.py: an internal script used to visualize the cross-validation results. The input (given as the first argument) is a tsv file where columns from the third are expected to be pairs of gold and predicted values.