Just wondering if anyone use TF and have standing in a medal zone.
", "messages": [ { "id": "1670829", "postDate": "01/31/2022 22:26:58", "content": "Just wondering if anyone use TF and have standing in a medal zone.
", "rawMarkdown": "Just wondering if anyone use TF and have standing in a medal zone.", "votes": null }, { "id": "1673149", "postDate": "02/02/2022 14:38:50", "content": "No one does? Seems that this needs to be fixed…
", "rawMarkdown": "No one does? Seems that this needs to be fixed...", "votes": null }, { "id": "1673555", "postDate": "02/02/2022 20:06:05", "content": "It is becoming harder and harder to get out TF2.0 models since 1. their repo is very barebones and doesn't have many of the \"luxuries\" like in yolov5 (strong data augmentation, optimizations, model selection (mmtracking/mmdetection), and it is harder to work with since even for some of the \"medium\"-sized models, I couldn't even get to run on my 3090). And 2. now with TPUs out of the question I have pretty much given up on the TF prize. There may be some things you can do the lstm video research on their GitHub but I don't see many people getting the prize.
", "rawMarkdown": "It is becoming harder and harder to get out TF2.0 models since 1. their repo is very barebones and doesn't have many of the \"luxuries\" like in yolov5 (strong data augmentation, optimizations, model selection (mmtracking/mmdetection), and it is harder to work with since even for some of the \"medium\"-sized models, I couldn't even get to run on my 3090). And 2. now with [TPUs out of the question](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/291541#1671823) I have pretty much given up on the TF prize. There may be some things you can do the lstm video research on their GitHub but I don't see many people getting the prize.", "votes": null }, { "id": "1673635", "postDate": "02/02/2022 21:19:13", "content": "Thanks for answering! As for \"lstm video research\" - do you mean this https://github.com/tensorflow/models/tree/master/research/lstm_object_detection ?
", "rawMarkdown": "Thanks for answering! As for \"lstm video research\" - do you mean this https://github.com/tensorflow/models/tree/master/research/lstm_object_detection ?", "votes": null }, { "id": "1673700", "postDate": "02/02/2022 22:26:32", "content": "Yeah, I mainly mentioned it because temporal information seems to be an important aspect that almost all public solutions haven't explored.
", "rawMarkdown": "Yeah, I mainly mentioned it because temporal information seems to be an important aspect that almost all public solutions haven't explored.", "votes": null }, { "id": "1677563", "postDate": "02/05/2022 20:09:31", "content": "I tried to use it on Kaggle, and here's my attempt https://www.kaggle.com/atamazian/tf1-lstm-object-detection-training-debug. Unfortunately, I don't know how to write pipeline file.
", "rawMarkdown": "I tried to use it on Kaggle, and here's my attempt https://www.kaggle.com/atamazian/tf1-lstm-object-detection-training-debug. Unfortunately, I don't know how to write pipeline file.", "votes": null }, { "id": "1677580", "postDate": "02/05/2022 20:25:24", "content": "Yeah, It is hard to work with, I'd keep trying tf2.0 OD (not lstm) and try to improve the score. I am focusing on yolov5 and mmtracking and I will try tf2.0 again in the last week.
", "rawMarkdown": "Yeah, It is hard to work with, I'd keep trying tf2.0 OD (not lstm) and try to improve the score. I am focusing on yolov5 and mmtracking and I will try tf2.0 again in the last week.", "votes": null }, { "id": "1677615", "postDate": "02/05/2022 20:59:21", "content": "I think LSTM may be better, but they even have no examples, no docs on pretrained models… it's really hard.
", "rawMarkdown": "I think LSTM may be better, but they even have no examples, no docs on pretrained models... it's really hard.", "votes": null }, { "id": "1682012", "postDate": "02/08/2022 21:11:56", "content": "Yeah, I agree it is an absolute nightmare. I spent three full weeks trying to learn TF OD API, and once I thought I handled it, I found half of the models provided in their zoo not working. I actually don't know why they call it API. It might be much easier to implement the models from scrach than using TF OD API.
", "rawMarkdown": "Yeah, I agree it is an absolute nightmare. I spent three full weeks trying to learn TF OD API, and once I thought I handled it, I found half of the models provided in their zoo not working. I actually don't know why they call it API. It might be much easier to implement the models from scrach than using TF OD API.", "votes": null } ], "comments": [ { "id": 1673149, "author_name": "atamazian", "author_url": "", "post_date": "02/02/2022 14:38:50", "content": "No one does? Seems that this needs to be fixed…
", "votes": null, "replies": [] }, { "id": 1673555, "author_name": "outwrest", "author_url": "", "post_date": "02/02/2022 20:06:05", "content": "It is becoming harder and harder to get out TF2.0 models since 1. their repo is very barebones and doesn't have many of the \"luxuries\" like in yolov5 (strong data augmentation, optimizations, model selection (mmtracking/mmdetection), and it is harder to work with since even for some of the \"medium\"-sized models, I couldn't even get to run on my 3090). And 2. now with TPUs out of the question I have pretty much given up on the TF prize. There may be some things you can do the lstm video research on their GitHub but I don't see many people getting the prize.
", "votes": null, "replies": [ { "id": 1673635, "author_name": "atamazian", "author_url": "", "post_date": "02/02/2022 21:19:13", "content": "Thanks for answering! As for \"lstm video research\" - do you mean this https://github.com/tensorflow/models/tree/master/research/lstm_object_detection ?
", "votes": null, "replies": [] }, { "id": 1673700, "author_name": "outwrest", "author_url": "", "post_date": "02/02/2022 22:26:32", "content": "Yeah, I mainly mentioned it because temporal information seems to be an important aspect that almost all public solutions haven't explored.
", "votes": null, "replies": [] }, { "id": 1677563, "author_name": "atamazian", "author_url": "", "post_date": "02/05/2022 20:09:31", "content": "I tried to use it on Kaggle, and here's my attempt https://www.kaggle.com/atamazian/tf1-lstm-object-detection-training-debug. Unfortunately, I don't know how to write pipeline file.
", "votes": null, "replies": [] }, { "id": 1677580, "author_name": "outwrest", "author_url": "", "post_date": "02/05/2022 20:25:24", "content": "Yeah, It is hard to work with, I'd keep trying tf2.0 OD (not lstm) and try to improve the score. I am focusing on yolov5 and mmtracking and I will try tf2.0 again in the last week.
", "votes": null, "replies": [] }, { "id": 1677615, "author_name": "atamazian", "author_url": "", "post_date": "02/05/2022 20:59:21", "content": "I think LSTM may be better, but they even have no examples, no docs on pretrained models… it's really hard.
", "votes": null, "replies": [] }, { "id": 1682012, "author_name": "azzamradman", "author_url": "", "post_date": "02/08/2022 21:11:56", "content": "Yeah, I agree it is an absolute nightmare. I spent three full weeks trying to learn TF OD API, and once I thought I handled it, I found half of the models provided in their zoo not working. I actually don't know why they call it API. It might be much easier to implement the models from scrach than using TF OD API.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1670829": "Just wondering if anyone use TF and have standing in a medal zone.", "1673149": "No one does? Seems that this needs to be fixed...", "1673555": "It is becoming harder and harder to get out TF2.0 models since 1. their repo is very barebones and doesn't have many of the \"luxuries\" like in yolov5 (strong data augmentation, optimizations, model selection (mmtracking/mmdetection), and it is harder to work with since even for some of the \"medium\"-sized models, I couldn't even get to run on my 3090). And 2. now with [TPUs out of the question](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/291541#1671823) I have pretty much given up on the TF prize. There may be some things you can do the lstm video research on their GitHub but I don't see many people getting the prize.", "1673635": "Thanks for answering! As for \"lstm video research\" - do you mean this https://github.com/tensorflow/models/tree/master/research/lstm_object_detection ?", "1673700": "Yeah, I mainly mentioned it because temporal information seems to be an important aspect that almost all public solutions haven't explored.", "1677563": "I tried to use it on Kaggle, and here's my attempt https://www.kaggle.com/atamazian/tf1-lstm-object-detection-training-debug. Unfortunately, I don't know how to write pipeline file.", "1677580": "Yeah, It is hard to work with, I'd keep trying tf2.0 OD (not lstm) and try to improve the score. I am focusing on yolov5 and mmtracking and I will try tf2.0 again in the last week.", "1677615": "I think LSTM may be better, but they even have no examples, no docs on pretrained models... it's really hard.", "1682012": "Yeah, I agree it is an absolute nightmare. I spent three full weeks trying to learn TF OD API, and once I thought I handled it, I found half of the models provided in their zoo not working. I actually don't know why they call it API. It might be much easier to implement the models from scrach than using TF OD API." }, "source": "meta" }