Using Tensorflow's Keras.
ACKNOWLEDGEMENT: Most of this repo is based on alexattia's Kaggle kernel
The model is a Convolutional Neural Network, with 6 Conv2D's and MaxPooling.
The structure can be found here.
This repo uses the Kaggle Simpsons Dataset from alexattia.
One of the main challenges for this dataset was the data preprocessing, as there were several labels missing data, or some labels with two much data (i.e Homer Simpson).
- download the dataset.
- install the requirements
- run the script:
python source/train.py \
--dataset_path path/to/downloaded/dataset \
--output_path path/to/model/output
This script will create a folder output_path/v<VersionNumber>, and will store:
- checkpoints in
output_path/v<Version>/chcks - training history in
output_path/v<Version>/history.csv - tensorboard logs in
output_path/v<Version>/logs - labels in
output_path/v<Version>/labels.txt(the labels resulting from the dataset cleaning)
To test an existing model run the script source/test.py, example:
python source/test.py \
--weights model/v1/chcks/chck_30.h5 \
--model_path model/v1 \
--testset_path path/to/testet
This script will generate the following reports:
- results per image in testset in
output_path/v<Version>/test_results.csv - sklearn class report in
output_path/v<Version>/class_report.csv - confusion matrix in
output_path/v<Version>/confusion_matrix.png - test viz in
output_path/v<Version>/samples.png
The pretrained wheights included, gave the following results when tested:
