Seed identification Kaggle challenge.
This training and inference script is best described via its help dialog and looking at the source code.
$ git clone git@github.com:kahnvex/seeds.git && cd seeds/
$ pip install -r requirements.txt
$ python3 -m seeds.seeds --help
usage: seeds.py [-h] --name NAME [--epochs EPOCHS] [--batch-size BATCH_SIZE]
[--learning-rate LEARNING_RATE] [--save-path SAVE_PATH]
[--base-model {xception,resnet,densenet121,densenet169}]
[--test] [--load-model] [--img-size IMG_SIZE]
[--outfile OUTFILE]
[--optimizer {adam,rmsprop,adagrad,adadelta}] [--pdist]
[--ensemble ENSEMBLE [ENSEMBLE ...]]
optional arguments:
-h, --help show this help message and exit
--name NAME Name of the model
--epochs EPOCHS, -e EPOCHS
--batch-size BATCH_SIZE, -b BATCH_SIZE
--learning-rate LEARNING_RATE, -l LEARNING_RATE
--save-path SAVE_PATH
Path in which to save the model's h5 file
--base-model {xception,resnet,densenet121,densenet169}
Architecture hyperparameter
--test Classify the test set
--load-model Load the model from disk instead of training
--img-size IMG_SIZE Input image size
--outfile OUTFILE Name of file for results
--optimizer {adam,rmsprop,adagrad,adadelta}
The optimization function to use
--pdist Output the probability distribution as a CSV
--ensemble ENSEMBLE [ENSEMBLE ...]
When testing, ensembe 1+ models using mean
You can see the Kaggle seed identification challenge here.
As of the time of this writing, several models trained by this script have scored in the top 5% of the Kaggle Seedling identification challenge.