Hi Everyone,
\nI noticed an issue with the EffNetV2 weights dataset I had created and used in my recent EffNetV2 train and infer notebooks. I have fixed the issue and updated the dataset. So, please use the updated weights dataset for all of your work.
\nI have re-run and saved all the notebooks with the updated EffNetV2 weights dataset as well. The notebooks have no code change. The only change is to the dataset. The train notebook will get a CV AUC of over 0.812+ with the updated weights compared to the 0.805+ before.
\nI'm linking the dataset and the notebook which should be used, with the updated dataset version for clarification below:
\n🏷️ EffNetV2 TfHub Weights Dataset - Use v4 or above:
\n\n \n\n
🏷️ EffNetV2 Keras Study Level Train notebook - Use v6 or above:
\n\n \n\n
Thanks. Have a great day!
", "messages": [ { "id": "1396850", "postDate": "07/22/2021 14:39:05", "content": "Hi Everyone,
\nI noticed an issue with the EffNetV2 weights dataset I had created and used in my recent EffNetV2 train and infer notebooks. I have fixed the issue and updated the dataset. So, please use the updated weights dataset for all of your work.
\nI have re-run and saved all the notebooks with the updated EffNetV2 weights dataset as well. The notebooks have no code change. The only change is to the dataset. The train notebook will get a CV AUC of over 0.812+ with the updated weights compared to the 0.805+ before.
\nI'm linking the dataset and the notebook which should be used, with the updated dataset version for clarification below:
\n🏷️ EffNetV2 TfHub Weights Dataset - Use v4 or above:
\n\n \n\n
🏷️ EffNetV2 Keras Study Level Train notebook - Use v6 or above:
\n\n \n\n
Thanks. Have a great day!
", "rawMarkdown": "Hi Everyone,\n\nI noticed an issue with the **EffNetV2 weights dataset** I had created and used in my recent EffNetV2 train and infer notebooks. I have fixed the issue and updated the dataset. So, **please use the updated weights dataset for all of your work.**\n\n**I have re-run and saved all the notebooks with the updated EffNetV2 weights dataset as well. The notebooks have no code change**. The only change is to the dataset. **The train notebook will get a CV AUC of over 0.812+ with the updated weights** compared to the 0.805+ before.\n\nI'm linking the dataset and the notebook which should be used, with the updated dataset version for clarification below:\n\n🏷️ EffNetV2 TfHub Weights Dataset - Use v4 or above:
\n\n> [EfficientNetV2 TFHub Weight Files](https://www.kaggle.com/sreevishnudamodaran/efficientnetv2-tfhub-weight-files?select=tfhub_models)🏷️ EffNetV2 Keras Study Level Train notebook - Use v6 or above:
\n\n> [SIIM EffNetV2 Keras Study Train [TPU CV0.805+]🎏](https://www.kaggle.com/sreevishnudamodaran/siim-effnetv2-keras-study-train-tpu-cv0-805)Thank you for updating us with this issue !
\nWhere does this difference comes from ? Where did find the \"bad\" weights and where di you find the good weights ?
\nI'm trying to find a way to get those weights outside a tfhub layer in order to have more flexibility, do you know if that is possible ?
", "rawMarkdown": "Thank you for updating us with this issue ! \n\nWhere does this difference comes from ? Where did find the \"bad\" weights and where di you find the good weights ? \n\nI'm trying to find a way to get those weights outside a tfhub layer in order to have more flexibility, do you know if that is possible ?", "votes": null }, { "id": "1397746", "postDate": "07/23/2021 13:14:44", "content": "thank you so much. i think its good code. I'll use it.
", "rawMarkdown": "thank you so much. i think its good code. I'll use it.", "votes": null }, { "id": "1397834", "postDate": "07/23/2021 14:17:45", "content": "All the weights are from TensorFlow Hub (https://tfhub.dev/)
\nThe issue was with the code that I used to download the weights, rename them and put them in directories to create the dataset. Some weight files got interchanged. That was the only issue.
I saw the other trick you posted to load the weights in the notebook comments 👍👍
", "rawMarkdown": "All the weights are from TensorFlow Hub (https://tfhub.dev/)\nThe issue was with the code that I used to download the weights, rename them and put them in directories to create the dataset. Some weight files got interchanged. That was the only issue.\n\nI saw the other trick you posted to load the weights in the notebook comments 👍👍", "votes": null }, { "id": "1400984", "postDate": "07/26/2021 19:27:10", "content": "Thanks for sharing!
", "rawMarkdown": "Thanks for sharing!", "votes": null } ], "comments": [ { "id": 1397598, "author_name": "tomdarmon", "author_url": "", "post_date": "07/23/2021 10:24:19", "content": "Thank you for updating us with this issue !
\nWhere does this difference comes from ? Where did find the \"bad\" weights and where di you find the good weights ?
\nI'm trying to find a way to get those weights outside a tfhub layer in order to have more flexibility, do you know if that is possible ?
", "votes": null, "replies": [ { "id": 1397834, "author_name": "sreevishnudamodaran", "author_url": "", "post_date": "07/23/2021 14:17:45", "content": "All the weights are from TensorFlow Hub (https://tfhub.dev/)
\nThe issue was with the code that I used to download the weights, rename them and put them in directories to create the dataset. Some weight files got interchanged. That was the only issue.
I saw the other trick you posted to load the weights in the notebook comments 👍👍
", "votes": null, "replies": [] } ] }, { "id": 1397746, "author_name": "tensorchoko", "author_url": "", "post_date": "07/23/2021 13:14:44", "content": "thank you so much. i think its good code. I'll use it.
", "votes": null, "replies": [] }, { "id": 1400984, "author_name": "minhtien1405", "author_url": "", "post_date": "07/26/2021 19:27:10", "content": "Thanks for sharing!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1396850": "Hi Everyone,\n\nI noticed an issue with the **EffNetV2 weights dataset** I had created and used in my recent EffNetV2 train and infer notebooks. I have fixed the issue and updated the dataset. So, **please use the updated weights dataset for all of your work.**\n\n**I have re-run and saved all the notebooks with the updated EffNetV2 weights dataset as well. The notebooks have no code change**. The only change is to the dataset. **The train notebook will get a CV AUC of over 0.812+ with the updated weights** compared to the 0.805+ before.\n\nI'm linking the dataset and the notebook which should be used, with the updated dataset version for clarification below:\n\n🏷️ EffNetV2 TfHub Weights Dataset - Use v4 or above:
\n\n> [EfficientNetV2 TFHub Weight Files](https://www.kaggle.com/sreevishnudamodaran/efficientnetv2-tfhub-weight-files?select=tfhub_models)🏷️ EffNetV2 Keras Study Level Train notebook - Use v6 or above:
\n\n> [SIIM EffNetV2 Keras Study Train [TPU CV0.805+]🎏](https://www.kaggle.com/sreevishnudamodaran/siim-effnetv2-keras-study-train-tpu-cv0-805)