Hey all,
\nWe had a question raised on the bonus prize and use of TensorFlow:
\nThe rule stating \"Use TensorFlow 2.X on GPU\" as a qualification for the bonus prize has been clarified to note this means training and inference code.
\nThe TensorFlow team is working with the CSIRO team to install a model based on the winning solution(s) on a TensorFlow-configured device to be used in future research and exploration. Therefore the use of TensorFlow in both training and inference is intended for this bonus prize.
\nI've updated the language for avoidance of doubt.
\nThere are no restrictions to any technologies or libraries used for the main leaderboard prize.
", "messages": [ { "id": 1599844, "postDate": "2021-11-29T22:04:28.033Z", "content": "Hey all,
\nWe had a question raised on the bonus prize and use of TensorFlow:
\nThe rule stating \"Use TensorFlow 2.X on GPU\" as a qualification for the bonus prize has been clarified to note this means training and inference code.
\nThe TensorFlow team is working with the CSIRO team to install a model based on the winning solution(s) on a TensorFlow-configured device to be used in future research and exploration. Therefore the use of TensorFlow in both training and inference is intended for this bonus prize.
\nI've updated the language for avoidance of doubt.
\nThere are no restrictions to any technologies or libraries used for the main leaderboard prize.
", "rawMarkdown": "Hey all,\n\nWe had a question raised on the bonus prize and use of TensorFlow: \n\nThe rule stating \"Use TensorFlow 2.X on GPU\" as a qualification for the bonus prize has been clarified to note this means *training and inference* code.\n\nThe TensorFlow team is working with the CSIRO team to install a model based on the winning solution(s) on a TensorFlow-configured device to be used in future research and exploration. Therefore the use of TensorFlow in both training and inference is intended for this bonus prize. \n\nI've updated the language for avoidance of doubt.\n\nThere are no restrictions to any technologies or libraries used for the main leaderboard prize.", "votes": 40 }, { "id": 1614719, "postDate": "2021-12-11T12:29:28.720Z", "content": "Thanks for clarification!
\nHow about this case?
I wish to know the clarification as well. If the whole training process must be done in tf2, it would disqualify most transfer learning for the bonus prize. Are transfer learning utilizing pre-trained model not trained in tf2 allowed for bonus prize?
", "rawMarkdown": "I wish to know the clarification as well. If the whole training process must be done in tf2, it would disqualify most transfer learning for the bonus prize. Are transfer learning utilizing pre-trained model not trained in tf2 allowed for bonus prize?", "votes": 1 } ] }, { "id": 1687242, "postDate": "2022-02-12T17:30:34.923Z", "content": "Then can we use Deep Learning Keras as well
", "rawMarkdown": "Then can we use Deep Learning Keras as well", "votes": 1 }, { "id": 1653276, "postDate": "2022-01-17T11:59:40.753Z", "content": "@addisonhoward Could you as a summary of this competition record and publish TOP5 (or paid zone) models prediction videos (test dataset) with GT and inference? 😄 In my opinion it would be great summary of competition solution quality.
", "rawMarkdown": "@addisonhoward Could you as a summary of this competition record and publish TOP5 (or paid zone) models prediction videos (test dataset) with GT and inference? 😄 In my opinion it would be great summary of competition solution quality.", "votes": 2 }, { "id": 1629372, "postDate": "2021-12-26T04:19:46.383Z", "content": "Can we use TensorRT with TensorFlow? any updates on this plz
", "rawMarkdown": "Can we use TensorRT with TensorFlow? any updates on this plz", "votes": 2 }, { "id": 1621013, "postDate": "2021-12-17T11:10:08.287Z", "content": "Can we use TensorRT with TensorFlow?
", "rawMarkdown": "Can we use TensorRT with TensorFlow?", "votes": 2 }, { "id": 1671687, "postDate": "2022-02-01T16:50:14.327Z", "content": "\nAre models trained on TPU still eligible for bonus prize? I know it clearly states GPU but in terms of using Kaggle GPU, I am limited to specific models at a low resolution and batch size. The TF2 OD GitHub supports TPUs and would allow me to make stronger models and I'd still be using TensorFlow for training and inference.
", "rawMarkdown": "@addisonhoward \n\nAre models trained on TPU still eligible for bonus prize? I know it clearly states GPU but in terms of using Kaggle GPU, I am limited to specific models at a low resolution and batch size. The TF2 OD GitHub supports TPUs and would allow me to make stronger models and I'd still be using TensorFlow for training and inference. ", "replies": [ { "id": 1671823, "postDate": "2022-02-01T18:43:05.503Z", "content": "Hey there,
\nFrom the hosts: \"The deployment rig contains a Jetson AGX Xavier GPU. Tensorflow engineers plan to tune the winning TF model to run efficiently on the embedded system before deploying and open sourcing. So for this challenge we want you to use a GPU for training and inference.\"
", "rawMarkdown": "Hey there,\n\nFrom the hosts: \"The deployment rig contains a Jetson AGX Xavier GPU. Tensorflow engineers plan to tune the winning TF model to run efficiently on the embedded system before deploying and open sourcing. So for this challenge we want you to use a GPU for training and inference.\"", "votes": 1 }, { "id": 1671868, "postDate": "2022-02-01T18:56:51.780Z", "content": "Thanks for clarifying!
", "rawMarkdown": "Thanks for clarifying!" } ] }, { "id": 1649442, "postDate": "2022-01-14T09:10:30.630Z", "content": "I run two epochs.how to use the file 'fasterrcnn_resnet50_fpn-e1.pth' on the output file?
", "rawMarkdown": "I run two epochs.how to use the file 'fasterrcnn_resnet50_fpn-e1.pth' on the output file?" }, { "id": 1631202, "postDate": "2021-12-28T07:01:43.943Z", "content": "Thanks Addison,Will inference performance only be considered or both training time and inference?
", "rawMarkdown": "Thanks Addison,Will inference performance only be considered or both training time and inference?" }, { "id": 1615686, "postDate": "2021-12-12T15:10:46.960Z", "content": "Thanks Addison,
\nJust a small question. Can you elaborate more on how performance will be calculated?
\nWill inference performance only be considered or both training time and inference?
\nAnd if so do we have to train the model in the submission or we can train it offline and send the model weights with the submission with the training code is not executed on the submission?
The time used for this is the time it takes your submission notebook to execute on the private dataset. You absolutely may train offline and upload your model as a dataset to only do inference in the submission.
", "rawMarkdown": "The time used for this is the time it takes your submission notebook to execute on the private dataset. You absolutely may train offline and upload your model as a dataset to only do inference in the submission.", "votes": 2 }, { "id": 1646909, "postDate": "2022-01-12T06:49:22.023Z", "content": "Hi @wcukierski,
\nI hope, I understood your comment correctly that the speed is measured just for inference on the unseen (private) test dataset which makes sense. However, in the description of the above post \"TensorFlow in both training and inference is intended for this bonus prize\", I find it contradictory as it explicitly mentions training time. Am I missing something here?
\nThanks!
Thanks Addison,
\nCan we use model from tf1_detection_zoo?
Much thanks!
", "rawMarkdown": "Much thanks!" } ], "comments": [ { "id": 1614719, "author_name": "Camaro", "author_url": "", "post_date": "2021-12-11T12:29:28.720000", "content": "Thanks for clarification!
\nHow about this case?
I wish to know the clarification as well. If the whole training process must be done in tf2, it would disqualify most transfer learning for the bonus prize. Are transfer learning utilizing pre-trained model not trained in tf2 allowed for bonus prize?
", "votes": 1, "replies": [] } ] }, { "id": 1687242, "author_name": "Muhammad Ammar Jamshed", "author_url": "", "post_date": "2022-02-12T17:30:34.923000", "content": "Then can we use Deep Learning Keras as well
", "votes": 1, "replies": [] }, { "id": 1653276, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-17T11:59:40.753000", "content": "@addisonhoward Could you as a summary of this competition record and publish TOP5 (or paid zone) models prediction videos (test dataset) with GT and inference? 😄 In my opinion it would be great summary of competition solution quality.
", "votes": 2, "replies": [] }, { "id": 1629372, "author_name": "Anshul Khadse", "author_url": "", "post_date": "2021-12-26T04:19:46.383000", "content": "Can we use TensorRT with TensorFlow? any updates on this plz
", "votes": 2, "replies": [] }, { "id": 1621013, "author_name": "Aleksandr Kruchinin", "author_url": "", "post_date": "2021-12-17T11:10:08.287000", "content": "Can we use TensorRT with TensorFlow?
", "votes": 2, "replies": [] }, { "id": 1671687, "author_name": "outwrest", "author_url": "", "post_date": "2022-02-01T16:50:14.327000", "content": "\nAre models trained on TPU still eligible for bonus prize? I know it clearly states GPU but in terms of using Kaggle GPU, I am limited to specific models at a low resolution and batch size. The TF2 OD GitHub supports TPUs and would allow me to make stronger models and I'd still be using TensorFlow for training and inference.
", "votes": 0, "replies": [ { "id": 1671823, "author_name": "Addison Howard", "author_url": "", "post_date": "2022-02-01T18:43:05.503000", "content": "Hey there,
\nFrom the hosts: \"The deployment rig contains a Jetson AGX Xavier GPU. Tensorflow engineers plan to tune the winning TF model to run efficiently on the embedded system before deploying and open sourcing. So for this challenge we want you to use a GPU for training and inference.\"
", "votes": 1, "replies": [] }, { "id": 1671868, "author_name": "outwrest", "author_url": "", "post_date": "2022-02-01T18:56:51.780000", "content": "Thanks for clarifying!
", "votes": 0, "replies": [] } ] }, { "id": 1649442, "author_name": "jlzhang1620", "author_url": "", "post_date": "2022-01-14T09:10:30.630000", "content": "I run two epochs.how to use the file 'fasterrcnn_resnet50_fpn-e1.pth' on the output file?
", "votes": 0, "replies": [] }, { "id": 1631202, "author_name": "fei", "author_url": "", "post_date": "2021-12-28T07:01:43.943000", "content": "Thanks Addison,Will inference performance only be considered or both training time and inference?
", "votes": 0, "replies": [] }, { "id": 1615686, "author_name": "A.Bashar Eter", "author_url": "", "post_date": "2021-12-12T15:10:46.960000", "content": "Thanks Addison,
\nJust a small question. Can you elaborate more on how performance will be calculated?
\nWill inference performance only be considered or both training time and inference?
\nAnd if so do we have to train the model in the submission or we can train it offline and send the model weights with the submission with the training code is not executed on the submission?
The time used for this is the time it takes your submission notebook to execute on the private dataset. You absolutely may train offline and upload your model as a dataset to only do inference in the submission.
", "votes": 2, "replies": [] }, { "id": 1646909, "author_name": "SAS", "author_url": "", "post_date": "2022-01-12T06:49:22.023000", "content": "Hi @wcukierski,
\nI hope, I understood your comment correctly that the speed is measured just for inference on the unseen (private) test dataset which makes sense. However, in the description of the above post \"TensorFlow in both training and inference is intended for this bonus prize\", I find it contradictory as it explicitly mentions training time. Am I missing something here?
\nThanks!
Thanks Addison,
\nCan we use model from tf1_detection_zoo?
Much thanks!
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1599844": "Hey all,\n\nWe had a question raised on the bonus prize and use of TensorFlow: \n\nThe rule stating \"Use TensorFlow 2.X on GPU\" as a qualification for the bonus prize has been clarified to note this means *training and inference* code.\n\nThe TensorFlow team is working with the CSIRO team to install a model based on the winning solution(s) on a TensorFlow-configured device to be used in future research and exploration. Therefore the use of TensorFlow in both training and inference is intended for this bonus prize. \n\nI've updated the language for avoidance of doubt.\n\nThere are no restrictions to any technologies or libraries used for the main leaderboard prize.", "1614719": "Thanks for clarification!\nHow about this case?\n- Port pytorch pretrained weight to TF model\n- Finetune TF model with this dataset\n- Inference with the TF model", "1687242": "Then can we use Deep Learning Keras as well", "1653276": "@addisonhoward Could you as a summary of this competition record and publish TOP5 (or paid zone) models prediction videos (test dataset) with GT and inference? 😄 In my opinion it would be great summary of competition solution quality.", "1629372": "Can we use TensorRT with TensorFlow? any updates on this plz", "1621013": "Can we use TensorRT with TensorFlow?", "1671687": "@addisonhoward \n\nAre models trained on TPU still eligible for bonus prize? I know it clearly states GPU but in terms of using Kaggle GPU, I am limited to specific models at a low resolution and batch size. The TF2 OD GitHub supports TPUs and would allow me to make stronger models and I'd still be using TensorFlow for training and inference. ", "1649442": "I run two epochs.how to use the file 'fasterrcnn_resnet50_fpn-e1.pth' on the output file?", "1631202": "Thanks Addison,Will inference performance only be considered or both training time and inference?", "1615686": "Thanks Addison,\nJust a small question. Can you elaborate more on how performance will be calculated? \nWill inference performance only be considered or both training time and inference?\nAnd if so do we have to train the model in the submission or we can train it offline and send the model weights with the submission with the training code is not executed on the submission?", "1646216": "Thanks Addison, \nCan we use model from tf1_detection_zoo?", "1618464": "", "1616989": "", "1688095": "Much thanks!" } }