NazarioR9/GIZ-NLP-Agricultural-Keyword-Spotter

Classify audio utterances in Luganda and English from Uganda. # GIZ NLP Agricultural Keyword Spotter ## 🗒 Description The objective of this challenge is to classify agricultural audio keywords in English and Lugandan from Uganda. For more information about this challenge, have a look on Zindi. ## 😷 The solution 🧠 Our solution consist of a mix of Resnet34 and Resnet50. During the training, we used a single augmentation (SpecAugmentation) strategy from torchlibrosa, which randomly cutout/erase some parts of the spectogram along the frequency and time dimensions. * Resnet34 was the best scorer with a public leaderboard of **0.84** (Logloss). * Resnet50 followed with a public LB of **0.88**. * A simple average of 2 Resnet34 gave us **0.8075** on the public LB. * Our final solution was a weighted average og the Resnet50 and 2 x Resnet34, for a small boost of **-0.06**. NB: We tried other architectures like Densenet, Efficienet, but it seemed that they were overfitting (LB of 1.+) ## 🌴 Repository structure (tree😎) You must follow this structure when runing the notebooks. |----data |       |--- raw |              |--- audio_files.zip |              |--- AdditionalUtterances.zip |              |--- nlp_keywords_29Oct2020.zip |              |--- train.csv |              |--- SampleSubmission.csv |\ |---- Resnet34_base.ipynb\ |---- Resnet34_with_scheduler.ipynb\ |---- Resnet50_base.ipynb\ |---- blend.ipynb\ |---- Readme.md Here is how to run the code: ``` # 1- Run {Resnet34_base, Resnet34_with_scheduler, Resnet50_base}.ipynb # 2- Run blend.ipynb to generate the final submission file ``` ## 🏆 Final Leaderbord 🏅**Position**: 7th 🏅**Score …