{ "id": 195408, "title": "effect of \"history_availabilities\"", "url": "/competitions/lyft-motion-prediction-autonomous-vehicles/discussion/195408", "author_name": "", "post_date": "2020-11-05T10:23:12.589161100Z", "votes": 10, "comment_count": 2, "views": 0, "content": "

This has already been mentioned in other post.
\nI did some experiment and presented my experimental results here
\n(using validate.zarr, at every : valid_index = np.arange(0,len(valid_dataset),1000))

\n

\"\"

\n

\"\"

", "messages": [ { "id": "1070083", "postDate": "11/05/2020 10:23:12", "content": "

This has already been mentioned in other post.
\nI did some experiment and presented my experimental results here
\n(using validate.zarr, at every : valid_index = np.arange(0,len(valid_dataset),1000))

\n

\"\"

\n

\"\"

", "rawMarkdown": "This has already been mentioned in other post. \nI did some experiment and presented my experimental results here\n(using validate.zarr, at every : valid_index = np.arange(0,len(valid_dataset),1000))\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F46f08c3d904f023e7474ca7e886ea68f%2FSelection_131.png?generation=1604571766623152&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fa8d7d474ca65602a26120d56ae2e4ccf%2FSelection_132.png?generation=1604571789783019&alt=media)", "votes": null }, { "id": "1070151", "postDate": "11/05/2020 12:40:55", "content": "

\"\"

\n

or train different predictor for different history? or lstm to handle different history seq length?
\ni would want to try writing an augmentation that randomly maskout the history positions (and images) according to the test.zarr probability …

", "rawMarkdown": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fbfd385a6d25cd7daf6e87a7f4d9e48ab%2FSelection_145.png?generation=1604580048370990&alt=media)\n\nor train different predictor for different history? or lstm to handle different history seq length?\ni would want to try writing an augmentation that randomly maskout the history positions (and images) according to the test.zarr probability ...", "votes": null }, { "id": "1084064", "postDate": "11/19/2020 17:28:19", "content": "

Hi, @hengck23. I think the average unavailable history frames in test.zarr or chopped validation set is between 1 and 2 and might be very close to 1, according to my observation. By the way, have you tried training on the masked train.zarr? Is image[:, -3-H:-3] exactly what you do to mask? can we simply mask those frames to 0?

", "rawMarkdown": "Hi, @hengck23. I think the average unavailable history frames in test.zarr or chopped validation set is between 1 and 2 and might be very close to 1, according to my observation. By the way, have you tried training on the masked train.zarr? Is image[:, -3-H:-3] exactly what you do to mask? can we simply mask those frames to 0?", "votes": null } ], "comments": [ { "id": 1070151, "author_name": "hengck23", "author_url": "", "post_date": "11/05/2020 12:40:55", "content": "

\"\"

\n

or train different predictor for different history? or lstm to handle different history seq length?
\ni would want to try writing an augmentation that randomly maskout the history positions (and images) according to the test.zarr probability …

", "votes": null, "replies": [ { "id": 1084064, "author_name": "spicychicken38", "author_url": "", "post_date": "11/19/2020 17:28:19", "content": "

Hi, @hengck23. I think the average unavailable history frames in test.zarr or chopped validation set is between 1 and 2 and might be very close to 1, according to my observation. By the way, have you tried training on the masked train.zarr? Is image[:, -3-H:-3] exactly what you do to mask? can we simply mask those frames to 0?

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1070083": "This has already been mentioned in other post. \nI did some experiment and presented my experimental results here\n(using validate.zarr, at every : valid_index = np.arange(0,len(valid_dataset),1000))\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F46f08c3d904f023e7474ca7e886ea68f%2FSelection_131.png?generation=1604571766623152&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fa8d7d474ca65602a26120d56ae2e4ccf%2FSelection_132.png?generation=1604571789783019&alt=media)", "1070151": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fbfd385a6d25cd7daf6e87a7f4d9e48ab%2FSelection_145.png?generation=1604580048370990&alt=media)\n\nor train different predictor for different history? or lstm to handle different history seq length?\ni would want to try writing an augmentation that randomly maskout the history positions (and images) according to the test.zarr probability ...", "1084064": "Hi, @hengck23. I think the average unavailable history frames in test.zarr or chopped validation set is between 1 and 2 and might be very close to 1, according to my observation. By the way, have you tried training on the masked train.zarr? Is image[:, -3-H:-3] exactly what you do to mask? can we simply mask those frames to 0?" }, "source": "meta" }