Hi. Everyone.\nI am sharing a method to train this huge dataset on TPU using @iafoss approach in Tensorflow.\nInference kernal is available here. \nTraining kernal is available. \nI trained 30 epochs with TPU in less than 2 hours with Batch Size of 64.\nIt is a starter code, so we can make it a lot better. I am a noob so I will welcome improvements and suggestions. \n@akensert helped a lot through public conversation to implement this. \nIf anyone could help, I would like to get suggestions about my validation.\nLike validation Loss is low in training. But when I predict my validation set separately then there is a huge gap. It happened to me before but that was due to BatchNormalization convergence. In this case I am not sure. If anyone could help that would be great.\nHowever I created dataset of 48x256x256x3 here based on @raghaw 's kernal\nYou kagglers are great. I am learning a lot and this is fun.\nThanks. Enjoy :)
", "messages": [ { "id": 905382, "postDate": "2020-06-28T14:01:28.760Z", "content": "Hi. Everyone.\nI am sharing a method to train this huge dataset on TPU using @iafoss approach in Tensorflow.\nInference kernal is available here. \nTraining kernal is available. \nI trained 30 epochs with TPU in less than 2 hours with Batch Size of 64.\nIt is a starter code, so we can make it a lot better. I am a noob so I will welcome improvements and suggestions. \n@akensert helped a lot through public conversation to implement this. \nIf anyone could help, I would like to get suggestions about my validation.\nLike validation Loss is low in training. But when I predict my validation set separately then there is a huge gap. It happened to me before but that was due to BatchNormalization convergence. In this case I am not sure. If anyone could help that would be great.\nHowever I created dataset of 48x256x256x3 here based on @raghaw 's kernal\nYou kagglers are great. I am learning a lot and this is fun.\nThanks. Enjoy :)
", "rawMarkdown": "Hi. Everyone.\nI am sharing a method to train this huge dataset on TPU using @iafoss approach in Tensorflow.\nInference [kernal](https://www.kaggle.com/micheomaano/pandas-42x256x256x3-inference) is available here. \nTraining [kernal](https://www.kaggle.com/micheomaano/tpu-training-tensorflow-iafoos-method-42x256x256x3/) is available. \nI trained 30 epochs with TPU in less than 2 hours with Batch Size of 64.\nIt is a starter code, so we can make it a lot better. I am a noob so I will welcome improvements and suggestions. \n@akensert helped a lot through public conversation to implement this. \nIf anyone could help, I would like to get suggestions about my validation.\nLike validation Loss is low in training. But when I predict my validation set separately then there is a huge gap. It happened to me before but that was due to BatchNormalization convergence. In this case I am not sure. If anyone could help that would be great.\nHowever I created dataset of 48x256x256x3 [here](https://www.kaggle.com/micheomaano/medium-resolution-dataset-48x256x256) based on @raghaw 's [kernal](https://www.kaggle.com/c/prostate-cancer-grade-assessment/discussion/161500)\nYou kagglers are great. I am learning a lot and this is fun.\nThanks. Enjoy :)", "votes": 15 }, { "id": 926296, "postDate": "2020-07-12T16:27:28.863Z", "content": "Hi Salman,
\n\nGood work on the kernel. Some feedback/questions
\n\nThanks for sharing TPU approach.
", "rawMarkdown": "Thanks for sharing TPU approach.", "votes": 2 } ], "comments": [ { "id": 926296, "author_name": "Vee", "author_url": "", "post_date": "2020-07-12T16:27:28.863000", "content": "Hi Salman,
\n\nGood work on the kernel. Some feedback/questions
\n\nThanks for sharing TPU approach.
", "votes": 2, "replies": [] } ], "raw_markdown_by_id": { "905382": "Hi. Everyone.\nI am sharing a method to train this huge dataset on TPU using @iafoss approach in Tensorflow.\nInference [kernal](https://www.kaggle.com/micheomaano/pandas-42x256x256x3-inference) is available here. \nTraining [kernal](https://www.kaggle.com/micheomaano/tpu-training-tensorflow-iafoos-method-42x256x256x3/) is available. \nI trained 30 epochs with TPU in less than 2 hours with Batch Size of 64.\nIt is a starter code, so we can make it a lot better. I am a noob so I will welcome improvements and suggestions. \n@akensert helped a lot through public conversation to implement this. \nIf anyone could help, I would like to get suggestions about my validation.\nLike validation Loss is low in training. But when I predict my validation set separately then there is a huge gap. It happened to me before but that was due to BatchNormalization convergence. In this case I am not sure. If anyone could help that would be great.\nHowever I created dataset of 48x256x256x3 [here](https://www.kaggle.com/micheomaano/medium-resolution-dataset-48x256x256) based on @raghaw 's [kernal](https://www.kaggle.com/c/prostate-cancer-grade-assessment/discussion/161500)\nYou kagglers are great. I am learning a lot and this is fun.\nThanks. Enjoy :)", "926296": "Hi Salman,\n\nGood work on the kernel. Some feedback/questions\n\n1. Why are you using a dropout in the inference kernel ? I feel its not required, unless I am missing something\n2. In the trainning&inference kernels, why do you subclass a Model instead of using the simpler functional API. Any particular reason?", "916323": "", "906106": "Thanks for sharing TPU approach." } }