{ "id": 179975, "title": "Pretrained model in baseline", "url": "/competitions/landmark-recognition-2020/discussion/179975", "author_name": "", "post_date": "2020-09-03T13:20:03.531980300Z", "votes": 2, "comment_count": 13, "views": 0, "content": "

Anyone who knows is the pre-trained model provided in the baseline file trained on v1 or cleaned v2 dataset?

", "messages": [ { "id": "996667", "postDate": "09/03/2020 13:20:03", "content": "

Anyone who knows is the pre-trained model provided in the baseline file trained on v1 or cleaned v2 dataset?

", "rawMarkdown": "Anyone who knows is the pre-trained model provided in the baseline file trained on v1 or cleaned v2 dataset?", "votes": null }, { "id": "996927", "postDate": "09/03/2020 16:36:25", "content": "

Cleaned v2 I reckon

", "rawMarkdown": "Cleaned v2 I reckon", "votes": null }, { "id": "997126", "postDate": "09/03/2020 19:00:37", "content": "

For the baseline in this challenge, it's actually pretrained on GLDv1, taken from the current version of the DELG paper.

", "rawMarkdown": "For the baseline in this challenge, it's actually pretrained on GLDv1, taken from the current version of the DELG paper.", "votes": null }, { "id": "997140", "postDate": "09/03/2020 19:18:08", "content": "

@andrefaraujo I am a bit confused as I also saw in the paper that you use GLDv1, but there is not too much overlap in classes, or am I missing something? Do you mind linking the exact dataset you are using here?

", "rawMarkdown": "andrefaraujo I am a bit confused as I also saw in the paper that you use GLDv1, but there is not too much overlap in classes, or am I missing something? Do you mind linking the exact dataset you are using here?", "votes": null }, { "id": "997145", "postDate": "09/03/2020 19:24:57", "content": "

The model is used to extract features that are matched at inference time to determine the label for each test image. It's not really using the classifier at test time (I think this was probably your confusion?). The features trained on GLDv1 tend to work well in GLDv2 as well since both are datasets based on landmarks.

", "rawMarkdown": "The model is used to extract features that are matched at inference time to determine the label for each test image. It's not really using the classifier at test time (I think this was probably your confusion?). The features trained on GLDv1 tend to work well in GLDv2 as well since both are datasets based on landmarks.", "votes": null }, { "id": "997152", "postDate": "09/03/2020 19:30:55", "content": "

Thank you for this message! I tried to replace the pre-trained model in the baseline by 'R101-DELG' from DELF repository, but it seems that this model is different from the one in the baseline since there is no attribute of \"input_global_scales_ind\" in the baseline. Did you do any modifications on the model?

", "rawMarkdown": "Thank you for this message! I tried to replace the pre-trained model in the baseline by 'R101-DELG' from DELF repository, but it seems that this model is different from the one in the baseline since there is no attribute of \"input_global_scales_ind\" in the baseline. Did you do any modifications on the model?", "votes": null }, { "id": "997158", "postDate": "09/03/2020 19:33:08", "content": "

The host said it is trained on v1

", "rawMarkdown": "The host said it is trained on v1", "votes": null }, { "id": "997172", "postDate": "09/03/2020 19:51:12", "content": "

The one they provided in the Retrieval competition was trained on GLDv2 clean (https://www.kaggle.com/camaskew/baseline-landmark-retrieval-model). I thought it was the same here

", "rawMarkdown": "The one they provided in the Retrieval competition was trained on GLDv2 clean (https://www.kaggle.com/camaskew/baseline-landmark-retrieval-model). I thought it was the same here", "votes": null }, { "id": "997205", "postDate": "09/03/2020 20:14:29", "content": "

indeed, the retrieval one was trained on GLDv2-clean

", "rawMarkdown": "indeed, the retrieval one was trained on GLDv2-clean", "votes": null }, { "id": "997208", "postDate": "09/03/2020 20:17:53", "content": "

yes, the recent ones we uploaded to the DELF repository have a few changes indeed, and they would indeed fail on the baseline notebook provided for this competition. It's just that they have the new \"input_global_scales_ind\" endpoint, so you can add it in the baseline notebook.

\n

References:

\n", "rawMarkdown": "yes, the recent ones we uploaded to the DELF repository have a few changes indeed, and they would indeed fail on the baseline notebook provided for this competition. It's just that they have the new \"input_global_scales_ind\" endpoint, so you can add it in the baseline notebook.\n\nReferences:\n- [Performing inference with the new model format, pointer showing the use of this new endpoint](https://github.com/tensorflow/models/blob/master/research/delf/delf/python/examples/extractor.py#L106)\n- [Explanation of this new endpoint](https://github.com/tensorflow/models/blob/master/research/delf/delf/protos/delf_config.proto#L68)", "votes": null }, { "id": "997260", "postDate": "09/03/2020 21:19:42", "content": "

Thank you very much!

", "rawMarkdown": "Thank you very much!", "votes": null }, { "id": "997980", "postDate": "09/04/2020 11:48:56", "content": "

Hi @andrefaraujo
\nI had tried to remake and train r101 model which was used in the delg paper, but I cannot find r101 code.(r50 code was easily found.) Did you upload elsewhere?

", "rawMarkdown": "Hi @andrefaraujo \nI had tried to remake and train r101 model which was used in [the delg paper](https://arxiv.org/pdf/2001.05027.pdf), but I cannot find r101 code.(r50 code was easily found.) Did you upload elsewhere?", "votes": null }, { "id": "1003121", "postDate": "09/08/2020 17:11:19", "content": "

currently the open-source DELG codebase only supports training R50, but our plan is to extend to R101 and other architectures. (note that we do have the pretrained R101 released model though)

", "rawMarkdown": "currently the open-source DELG codebase only supports training R50, but our plan is to extend to R101 and other architectures. (note that we do have the pretrained R101 released model though)", "votes": null }, { "id": "1013818", "postDate": "09/17/2020 02:15:57", "content": "

I'm looking forward to it.
\nThanks

", "rawMarkdown": "I'm looking forward to it.\nThanks", "votes": null } ], "comments": [ { "id": 996927, "author_name": "arc144", "author_url": "", "post_date": "09/03/2020 16:36:25", "content": "

Cleaned v2 I reckon

", "votes": null, "replies": [ { "id": 997158, "author_name": "hao0214", "author_url": "", "post_date": "09/03/2020 19:33:08", "content": "

The host said it is trained on v1

", "votes": null, "replies": [] }, { "id": 997172, "author_name": "arc144", "author_url": "", "post_date": "09/03/2020 19:51:12", "content": "

The one they provided in the Retrieval competition was trained on GLDv2 clean (https://www.kaggle.com/camaskew/baseline-landmark-retrieval-model). I thought it was the same here

", "votes": null, "replies": [] }, { "id": 997205, "author_name": "andrefaraujo", "author_url": "", "post_date": "09/03/2020 20:14:29", "content": "

indeed, the retrieval one was trained on GLDv2-clean

", "votes": null, "replies": [] } ] }, { "id": 997126, "author_name": "andrefaraujo", "author_url": "", "post_date": "09/03/2020 19:00:37", "content": "

For the baseline in this challenge, it's actually pretrained on GLDv1, taken from the current version of the DELG paper.

", "votes": null, "replies": [ { "id": 997140, "author_name": "philippsinger", "author_url": "", "post_date": "09/03/2020 19:18:08", "content": "

@andrefaraujo I am a bit confused as I also saw in the paper that you use GLDv1, but there is not too much overlap in classes, or am I missing something? Do you mind linking the exact dataset you are using here?

", "votes": null, "replies": [] }, { "id": 997145, "author_name": "andrefaraujo", "author_url": "", "post_date": "09/03/2020 19:24:57", "content": "

The model is used to extract features that are matched at inference time to determine the label for each test image. It's not really using the classifier at test time (I think this was probably your confusion?). The features trained on GLDv1 tend to work well in GLDv2 as well since both are datasets based on landmarks.

", "votes": null, "replies": [] }, { "id": 997152, "author_name": "hao0214", "author_url": "", "post_date": "09/03/2020 19:30:55", "content": "

Thank you for this message! I tried to replace the pre-trained model in the baseline by 'R101-DELG' from DELF repository, but it seems that this model is different from the one in the baseline since there is no attribute of \"input_global_scales_ind\" in the baseline. Did you do any modifications on the model?

", "votes": null, "replies": [] }, { "id": 997208, "author_name": "andrefaraujo", "author_url": "", "post_date": "09/03/2020 20:17:53", "content": "

yes, the recent ones we uploaded to the DELF repository have a few changes indeed, and they would indeed fail on the baseline notebook provided for this competition. It's just that they have the new \"input_global_scales_ind\" endpoint, so you can add it in the baseline notebook.

\n

References:

\n", "votes": null, "replies": [] }, { "id": 997260, "author_name": "hao0214", "author_url": "", "post_date": "09/03/2020 21:19:42", "content": "

Thank you very much!

", "votes": null, "replies": [] }, { "id": 997980, "author_name": "yasagure", "author_url": "", "post_date": "09/04/2020 11:48:56", "content": "

Hi @andrefaraujo
\nI had tried to remake and train r101 model which was used in the delg paper, but I cannot find r101 code.(r50 code was easily found.) Did you upload elsewhere?

", "votes": null, "replies": [] }, { "id": 1003121, "author_name": "andrefaraujo", "author_url": "", "post_date": "09/08/2020 17:11:19", "content": "

currently the open-source DELG codebase only supports training R50, but our plan is to extend to R101 and other architectures. (note that we do have the pretrained R101 released model though)

", "votes": null, "replies": [] }, { "id": 1013818, "author_name": "yasagure", "author_url": "", "post_date": "09/17/2020 02:15:57", "content": "

I'm looking forward to it.
\nThanks

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "996667": "Anyone who knows is the pre-trained model provided in the baseline file trained on v1 or cleaned v2 dataset?", "996927": "Cleaned v2 I reckon", "997126": "For the baseline in this challenge, it's actually pretrained on GLDv1, taken from the current version of the DELG paper.", "997140": "andrefaraujo I am a bit confused as I also saw in the paper that you use GLDv1, but there is not too much overlap in classes, or am I missing something? Do you mind linking the exact dataset you are using here?", "997145": "The model is used to extract features that are matched at inference time to determine the label for each test image. It's not really using the classifier at test time (I think this was probably your confusion?). The features trained on GLDv1 tend to work well in GLDv2 as well since both are datasets based on landmarks.", "997152": "Thank you for this message! I tried to replace the pre-trained model in the baseline by 'R101-DELG' from DELF repository, but it seems that this model is different from the one in the baseline since there is no attribute of \"input_global_scales_ind\" in the baseline. Did you do any modifications on the model?", "997158": "The host said it is trained on v1", "997172": "The one they provided in the Retrieval competition was trained on GLDv2 clean (https://www.kaggle.com/camaskew/baseline-landmark-retrieval-model). I thought it was the same here", "997205": "indeed, the retrieval one was trained on GLDv2-clean", "997208": "yes, the recent ones we uploaded to the DELF repository have a few changes indeed, and they would indeed fail on the baseline notebook provided for this competition. It's just that they have the new \"input_global_scales_ind\" endpoint, so you can add it in the baseline notebook.\n\nReferences:\n- [Performing inference with the new model format, pointer showing the use of this new endpoint](https://github.com/tensorflow/models/blob/master/research/delf/delf/python/examples/extractor.py#L106)\n- [Explanation of this new endpoint](https://github.com/tensorflow/models/blob/master/research/delf/delf/protos/delf_config.proto#L68)", "997260": "Thank you very much!", "997980": "Hi @andrefaraujo \nI had tried to remake and train r101 model which was used in [the delg paper](https://arxiv.org/pdf/2001.05027.pdf), but I cannot find r101 code.(r50 code was easily found.) Did you upload elsewhere?", "1003121": "currently the open-source DELG codebase only supports training R50, but our plan is to extend to R101 and other architectures. (note that we do have the pretrained R101 released model though)", "1013818": "I'm looking forward to it.\nThanks" }, "source": "meta" }