I wonder what was the approach for the teams who got 60% and above
", "messages": [ { "id": 1617443, "postDate": "2021-12-14T05:03:34.063Z", "content": "I wonder what was the approach for the teams who got 60% and above
", "rawMarkdown": "I wonder what was the approach for the teams who got 60% and above", "votes": 5 }, { "id": 1619351, "postDate": "2021-12-15T21:09:19.277Z", "content": "Yep, we also used the pre-trained CLIP model without significant tuning. I'd be very curious others approaches, especially if you didn't use contrastive learning - we were not able to get much better than 25-30% using non-DL methods.
\nHere's a condensed script of how we produced the results for our top leaderboard submission if anyone is curious: https://gist.github.com/aspfohl/ad69768755703a4890ce373814fbf09f
", "rawMarkdown": "Yep, we also used the pre-trained CLIP model without significant tuning. I'd be very curious others approaches, especially if you didn't use contrastive learning - we were not able to get much better than 25-30% using non-DL methods.\n\nHere's a condensed script of how we produced the results for our top leaderboard submission if anyone is curious: https://gist.github.com/aspfohl/ad69768755703a4890ce373814fbf09f", "votes": 1 }, { "id": 1620556, "postDate": "2021-12-16T21:57:46.657Z", "content": "I also used a clip based model. In the end I was essentially just using the pre-trained weights. I tried to fine tune it on the data, but it wasn't going particularly well so I had to give up on it. I also did everything on my laptop and training the model meant I could use it for anything really, so I only had the option to train at night. And in the end I ran out of nights where I could train before I could properly tune things.
\nThe one thing I did do that I think was different, and what ultimately put me a couple points above some of the other teams clustered in the upper 0.6 range was process the descriptions a little differently. The original CLIP authors mention that more context is better and that appending statements to the front of descriptions can improve performance. I did that and it turned out to be the case.
", "rawMarkdown": "I also used a clip based model. In the end I was essentially just using the pre-trained weights. I tried to fine tune it on the data, but it wasn't going particularly well so I had to give up on it. I also did everything on my laptop and training the model meant I could use it for anything really, so I only had the option to train at night. And in the end I ran out of nights where I could train before I could properly tune things. \n\nThe one thing I did do that I think was different, and what ultimately put me a couple points above some of the other teams clustered in the upper 0.6 range was process the descriptions a little differently. The original CLIP authors mention that more context is better and that appending statements to the front of descriptions can improve performance. I did that and it turned out to be the case. ", "votes": 2 }, { "id": 1619313, "postDate": "2021-12-15T19:58:22.720Z", "content": "Hi Khaled, we (team TILAPIA, rank 4th, score 0.681) used CLIP from OpenAI. CLIP is a non-generative neural network designed to learn image features through text supervision. The idea is to do contrastive learning. CLIP can be applied with a variety of NLP (like BERT) and CV model architectures (like Resnet-50 or ViT-B/32). You can find their github here. You may find this paper and this youtube tutorial helpful.
\nCLIP is the best model we found and got us 0.681. We believe more finetuning will further boost the performance but we didn't have time to do it. We are the lowest team above 0.60 so we'd love to know what other teams did. I know that @anna462's team also employed CLIP with probably a better implementation so feel free to comment any insights 😊
\nAt the same time, it would be especially helpful if @jamesan2020 can share your idea when you have time!
", "rawMarkdown": "Hi Khaled, we (team TILAPIA, rank 4th, score 0.681) used CLIP from OpenAI. CLIP is a non-generative neural network designed to learn image features through text supervision. The idea is to do contrastive learning. CLIP can be applied with a variety of NLP (like BERT) and CV model architectures (like Resnet-50 or ViT-B/32). You can find their [github](https://github.com/openai/CLIP) here. You may find this [paper](https://arxiv.org/abs/2103.00020) and this [youtube tutorial](https://www.youtube.com/watch?v=T9XSU0pKX2E) helpful. \n\nCLIP is the best model we found and got us 0.681. We believe more finetuning will further boost the performance but we didn't have time to do it. We are the lowest team above 0.60 so we'd love to know what other teams did. I know that @anna462's team also employed CLIP with probably a better implementation so feel free to comment any insights 😊\n\nAt the same time, it would be especially helpful if @jamesan2020 can share your idea when you have time!", "votes": 2 }, { "id": 1620489, "postDate": "2021-12-16T20:39:04.407Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 1619351, "author_name": "Anna462", "author_url": "", "post_date": "2021-12-15T21:09:19.277000", "content": "Yep, we also used the pre-trained CLIP model without significant tuning. I'd be very curious others approaches, especially if you didn't use contrastive learning - we were not able to get much better than 25-30% using non-DL methods.
\nHere's a condensed script of how we produced the results for our top leaderboard submission if anyone is curious: https://gist.github.com/aspfohl/ad69768755703a4890ce373814fbf09f
", "votes": 1, "replies": [] }, { "id": 1620556, "author_name": "Elliot", "author_url": "", "post_date": "2021-12-16T21:57:46.657000", "content": "I also used a clip based model. In the end I was essentially just using the pre-trained weights. I tried to fine tune it on the data, but it wasn't going particularly well so I had to give up on it. I also did everything on my laptop and training the model meant I could use it for anything really, so I only had the option to train at night. And in the end I ran out of nights where I could train before I could properly tune things.
\nThe one thing I did do that I think was different, and what ultimately put me a couple points above some of the other teams clustered in the upper 0.6 range was process the descriptions a little differently. The original CLIP authors mention that more context is better and that appending statements to the front of descriptions can improve performance. I did that and it turned out to be the case.
", "votes": 2, "replies": [] }, { "id": 1619313, "author_name": "JiaZZZZZ", "author_url": "", "post_date": "2021-12-15T19:58:22.720000", "content": "Hi Khaled, we (team TILAPIA, rank 4th, score 0.681) used CLIP from OpenAI. CLIP is a non-generative neural network designed to learn image features through text supervision. The idea is to do contrastive learning. CLIP can be applied with a variety of NLP (like BERT) and CV model architectures (like Resnet-50 or ViT-B/32). You can find their github here. You may find this paper and this youtube tutorial helpful.
\nCLIP is the best model we found and got us 0.681. We believe more finetuning will further boost the performance but we didn't have time to do it. We are the lowest team above 0.60 so we'd love to know what other teams did. I know that @anna462's team also employed CLIP with probably a better implementation so feel free to comment any insights 😊
\nAt the same time, it would be especially helpful if @jamesan2020 can share your idea when you have time!
", "votes": 2, "replies": [] }, { "id": 1620489, "author_name": "", "author_url": "", "post_date": "2021-12-16T20:39:04.407000", "content": "", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "1617443": "I wonder what was the approach for the teams who got 60% and above", "1619351": "Yep, we also used the pre-trained CLIP model without significant tuning. I'd be very curious others approaches, especially if you didn't use contrastive learning - we were not able to get much better than 25-30% using non-DL methods.\n\nHere's a condensed script of how we produced the results for our top leaderboard submission if anyone is curious: https://gist.github.com/aspfohl/ad69768755703a4890ce373814fbf09f", "1620556": "I also used a clip based model. In the end I was essentially just using the pre-trained weights. I tried to fine tune it on the data, but it wasn't going particularly well so I had to give up on it. I also did everything on my laptop and training the model meant I could use it for anything really, so I only had the option to train at night. And in the end I ran out of nights where I could train before I could properly tune things. \n\nThe one thing I did do that I think was different, and what ultimately put me a couple points above some of the other teams clustered in the upper 0.6 range was process the descriptions a little differently. The original CLIP authors mention that more context is better and that appending statements to the front of descriptions can improve performance. I did that and it turned out to be the case. ", "1619313": "Hi Khaled, we (team TILAPIA, rank 4th, score 0.681) used CLIP from OpenAI. CLIP is a non-generative neural network designed to learn image features through text supervision. The idea is to do contrastive learning. CLIP can be applied with a variety of NLP (like BERT) and CV model architectures (like Resnet-50 or ViT-B/32). You can find their [github](https://github.com/openai/CLIP) here. You may find this [paper](https://arxiv.org/abs/2103.00020) and this [youtube tutorial](https://www.youtube.com/watch?v=T9XSU0pKX2E) helpful. \n\nCLIP is the best model we found and got us 0.681. We believe more finetuning will further boost the performance but we didn't have time to do it. We are the lowest team above 0.60 so we'd love to know what other teams did. I know that @anna462's team also employed CLIP with probably a better implementation so feel free to comment any insights 😊\n\nAt the same time, it would be especially helpful if @jamesan2020 can share your idea when you have time!", "1620489": "" } }