I have been using the pre-processing layer in this notebook as part of my training pipeline. It should capture almost all the important highlights made in this competition regarding preprocessing the data. Note that it does not do any feature extraction like some other notebooks but lets you retain all the data for training with very high per-video throughput.
\nThe features provided by the preprocessing layer are:
\nIt has been helpful to me in training my model with simple tweaks to the parameters and should be helpful for someone looking for a quick way to get clean processed data. Give it a read and an upvote if you like it!
", "messages": [ { "id": 2231084, "postDate": "2023-04-23T04:49:10.950Z", "content": "I have been using the pre-processing layer in this notebook as part of my training pipeline. It should capture almost all the important highlights made in this competition regarding preprocessing the data. Note that it does not do any feature extraction like some other notebooks but lets you retain all the data for training with very high per-video throughput.
\nThe features provided by the preprocessing layer are:
\nIt has been helpful to me in training my model with simple tweaks to the parameters and should be helpful for someone looking for a quick way to get clean processed data. Give it a read and an upvote if you like it!
", "rawMarkdown": "I have been using the pre-processing layer in this [notebook](https://www.kaggle.com/code/debman/one-stop-super-fast-tflite-preprocessing-for-gislr?scriptVersionId=126842647) as part of my training pipeline. It should capture almost all the important highlights made in this competition regarding preprocessing the data. Note that it does not do any feature extraction like some other notebooks but lets you retain all the data for training with very high per-video throughput.\n\nThe features provided by the preprocessing layer are:\n\n0. Competition EDA highlights e.g. good sequence sizes to choose for inputs, important landmarks which have a visible effect on accuracy improvement, etc.\n1. Linearly interpolate any missing frames and landmarks sensibly so that nans do not have to be replaced by zeros.\n2. Identify the key postures from large frame sequences which would retain the most important frames in the original data.\n3. Normalize body and hands using conventions from sign language literature about expected signing space in a frame.\n4. Replaces nans using heuristics as far as possible before falling back on replace-by-zeros.\n\nIt has been helpful to me in training my model with simple tweaks to the parameters and should be helpful for someone looking for a quick way to get clean processed data. Give it a read and an upvote if you like it!", "votes": 2 } ], "comments": [], "raw_markdown_by_id": { "2231084": "I have been using the pre-processing layer in this [notebook](https://www.kaggle.com/code/debman/one-stop-super-fast-tflite-preprocessing-for-gislr?scriptVersionId=126842647) as part of my training pipeline. It should capture almost all the important highlights made in this competition regarding preprocessing the data. Note that it does not do any feature extraction like some other notebooks but lets you retain all the data for training with very high per-video throughput.\n\nThe features provided by the preprocessing layer are:\n\n0. Competition EDA highlights e.g. good sequence sizes to choose for inputs, important landmarks which have a visible effect on accuracy improvement, etc.\n1. Linearly interpolate any missing frames and landmarks sensibly so that nans do not have to be replaced by zeros.\n2. Identify the key postures from large frame sequences which would retain the most important frames in the original data.\n3. Normalize body and hands using conventions from sign language literature about expected signing space in a frame.\n4. Replaces nans using heuristics as far as possible before falling back on replace-by-zeros.\n\nIt has been helpful to me in training my model with simple tweaks to the parameters and should be helpful for someone looking for a quick way to get clean processed data. Give it a read and an upvote if you like it!" } }