{ "id": 180056, "title": "TF Record Dataset and Own Baseline", "url": "/competitions/landmark-recognition-2020/discussion/180056", "author_name": "Martin Kovacevic Buvinic", "post_date": "2020-09-03T17:26:15.612000", "votes": 26, "comment_count": 6, "views": 0, "content": "

I am training models and it does not take that much time with 60% of the data. Im getting interesting results with a really basic model, (not using arcface, cosine-softmax etc….).

\n

Using 1 fold validation with 20% of the data give me a validation accuracy of 0.79 and a gap of 0.76.
\nThe public leadearboad is 0.1188. Im not using retrieval inference so that could be a reason why the score is so low. Onother point to mention is that my inference pipeline gives the same score as my model validation so i confirm it is correct. At last, the inference pipeline is very fast this way, it predicts the private test set in 30 minutes (because i am using a standard multi classification model, not retrieval approach. Im pretty sure there is a huge terrain of improvement.

\n

Question to consider:

\n\n

Here are the datasets i am using, the images are resize to 384 x 384 and have 100% quality

\n

https://www.kaggle.com/ragnar123/landmark-tfrecords-384

\n

https://www.kaggle.com/ragnar123/landmark-tfrecords-384-2

\n

Their are 2 because the 50 tf records dont fit in one dataset (more than 100GB). The 50 tf records are stratified by the class and also they are encoded.

\n

Here is a link where you will find how to map the tf records.

\n

https://www.kaggle.com/ragnar123/landmark-image-train

\n

Here their is a column names landmark_id_encode. This is the target label encoded so that you can map the final results of the model to predict the private test.

\n

Finally, the dataset now has 80% of the data, in the following days i will upload the rest to have the 50 tf records (this is 100% of the training data).

\n

I hope this is helpfull for training your own models. If you want i can make a notebook with the complete pipeline. Cheers and have fun

", "messages": [ { "id": 997020, "postDate": "2020-09-03T17:26:15.613Z", "content": "

I am training models and it does not take that much time with 60% of the data. Im getting interesting results with a really basic model, (not using arcface, cosine-softmax etc….).

\n

Using 1 fold validation with 20% of the data give me a validation accuracy of 0.79 and a gap of 0.76.
\nThe public leadearboad is 0.1188. Im not using retrieval inference so that could be a reason why the score is so low. Onother point to mention is that my inference pipeline gives the same score as my model validation so i confirm it is correct. At last, the inference pipeline is very fast this way, it predicts the private test set in 30 minutes (because i am using a standard multi classification model, not retrieval approach. Im pretty sure there is a huge terrain of improvement.

\n

Question to consider:

\n\n

Here are the datasets i am using, the images are resize to 384 x 384 and have 100% quality

\n

https://www.kaggle.com/ragnar123/landmark-tfrecords-384

\n

https://www.kaggle.com/ragnar123/landmark-tfrecords-384-2

\n

Their are 2 because the 50 tf records dont fit in one dataset (more than 100GB). The 50 tf records are stratified by the class and also they are encoded.

\n

Here is a link where you will find how to map the tf records.

\n

https://www.kaggle.com/ragnar123/landmark-image-train

\n

Here their is a column names landmark_id_encode. This is the target label encoded so that you can map the final results of the model to predict the private test.

\n

Finally, the dataset now has 80% of the data, in the following days i will upload the rest to have the 50 tf records (this is 100% of the training data).

\n

I hope this is helpfull for training your own models. If you want i can make a notebook with the complete pipeline. Cheers and have fun

", "rawMarkdown": "I am training models and it does not take that much time with 60% of the data. Im getting interesting results with a really basic model, (not using arcface, cosine-softmax etc....).\n\nUsing 1 fold validation with 20% of the data give me a validation accuracy of 0.79 and a gap of 0.76.\nThe public leadearboad is 0.1188. Im not using retrieval inference so that could be a reason why the score is so low. Onother point to mention is that my inference pipeline gives the same score as my model validation so i confirm it is correct. At last, the inference pipeline is very fast this way, it predicts the private test set in 30 minutes (because i am using a standard multi classification model, not retrieval approach. Im pretty sure there is a huge terrain of improvement.\n\nQuestion to consider: \n\n- The distribution of the target is the same for the training set and the private test set? I believe it is not the same\n- How can we resolve the non landmark problem?, we want to predict empty or a really low confidence score for that cases\n\nHere are the datasets i am using, the images are resize to 384 x 384 and have 100% quality\n\nhttps://www.kaggle.com/ragnar123/landmark-tfrecords-384\n\nhttps://www.kaggle.com/ragnar123/landmark-tfrecords-384-2\n\nTheir are 2 because the 50 tf records dont fit in one dataset (more than 100GB). The 50 tf records are stratified by the class and also they are encoded. \n\nHere is a link where you will find how to map the tf records.\n\nhttps://www.kaggle.com/ragnar123/landmark-image-train\n\nHere their is a column names landmark_id_encode. This is the target label encoded so that you can map the final results of the model to predict the private test.\n\nFinally, the dataset now has 80% of the data, in the following days i will upload the rest to have the 50 tf records (this is 100% of the training data).\n\nI hope this is helpfull for training your own models. If you want i can make a notebook with the complete pipeline. Cheers and have fun\n\n", "votes": 26 }, { "id": 1013295, "postDate": "2020-09-16T16:11:53.673Z", "content": "

Thank you for the dataset. Unfortunately, I was not able to download this one: https://www.kaggle.com/ragnar123/landmark-tfrecords-384 using this Kaggle API command. It returned an error message \"404 - Not Found\". Is it because of storage limitations in your Kaggle dataset?

", "rawMarkdown": "Thank you for the dataset. Unfortunately, I was not able to download this one: https://www.kaggle.com/ragnar123/landmark-tfrecords-384 using this Kaggle API command. It returned an error message \"404 - Not Found\". Is it because of storage limitations in your Kaggle dataset?" }, { "id": 1012062, "postDate": "2020-09-15T22:14:10.183Z", "content": "

Who is there to be team with with me i have more great approach than this

", "rawMarkdown": "Who is there to be team with with me i have more great approach than this" }, { "id": 1012060, "postDate": "2020-09-15T22:11:57.213Z", "content": "

So Helpful
\n@ragnar123

", "rawMarkdown": "So Helpful\n@ragnar123 " }, { "id": 1011138, "postDate": "2020-09-15T09:03:38.420Z", "content": "

I see there is a limit of 20GB on datasets. How were you able to create ~75GB sized datasets?? Thanks

", "rawMarkdown": "I see there is a limit of 20GB on datasets. How were you able to create ~75GB sized datasets?? Thanks" }, { "id": 1005700, "postDate": "2020-09-10T17:08:09.040Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1000999, "postDate": "2020-09-07T02:07:35.137Z", "content": "

Thanks, this is very useful！

", "rawMarkdown": "Thanks, this is very useful！", "votes": 1 } ], "comments": [ { "id": 1013295, "author_name": "FGPC", "author_url": "", "post_date": "2020-09-16T16:11:53.673000", "content": "

Thank you for the dataset. Unfortunately, I was not able to download this one: https://www.kaggle.com/ragnar123/landmark-tfrecords-384 using this Kaggle API command. It returned an error message \"404 - Not Found\". Is it because of storage limitations in your Kaggle dataset?

", "votes": 0, "replies": [] }, { "id": 1012062, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2020-09-15T22:14:10.183000", "content": "

Who is there to be team with with me i have more great approach than this

", "votes": 0, "replies": [] }, { "id": 1012060, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2020-09-15T22:11:57.213000", "content": "

So Helpful
\n@ragnar123

", "votes": 0, "replies": [] }, { "id": 1011138, "author_name": "JkReddy", "author_url": "", "post_date": "2020-09-15T09:03:38.420000", "content": "

I see there is a limit of 20GB on datasets. How were you able to create ~75GB sized datasets?? Thanks

", "votes": 0, "replies": [] }, { "id": 1005700, "author_name": "", "author_url": "", "post_date": "2020-09-10T17:08:09.040000", "content": "", "votes": 0, "replies": [] }, { "id": 1000999, "author_name": "xckkcxxck", "author_url": "", "post_date": "2020-09-07T02:07:35.137000", "content": "

Thanks, this is very useful！

", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "997020": "I am training models and it does not take that much time with 60% of the data. Im getting interesting results with a really basic model, (not using arcface, cosine-softmax etc....).\n\nUsing 1 fold validation with 20% of the data give me a validation accuracy of 0.79 and a gap of 0.76.\nThe public leadearboad is 0.1188. Im not using retrieval inference so that could be a reason why the score is so low. Onother point to mention is that my inference pipeline gives the same score as my model validation so i confirm it is correct. At last, the inference pipeline is very fast this way, it predicts the private test set in 30 minutes (because i am using a standard multi classification model, not retrieval approach. Im pretty sure there is a huge terrain of improvement.\n\nQuestion to consider: \n\n- The distribution of the target is the same for the training set and the private test set? I believe it is not the same\n- How can we resolve the non landmark problem?, we want to predict empty or a really low confidence score for that cases\n\nHere are the datasets i am using, the images are resize to 384 x 384 and have 100% quality\n\nhttps://www.kaggle.com/ragnar123/landmark-tfrecords-384\n\nhttps://www.kaggle.com/ragnar123/landmark-tfrecords-384-2\n\nTheir are 2 because the 50 tf records dont fit in one dataset (more than 100GB). The 50 tf records are stratified by the class and also they are encoded. \n\nHere is a link where you will find how to map the tf records.\n\nhttps://www.kaggle.com/ragnar123/landmark-image-train\n\nHere their is a column names landmark_id_encode. This is the target label encoded so that you can map the final results of the model to predict the private test.\n\nFinally, the dataset now has 80% of the data, in the following days i will upload the rest to have the 50 tf records (this is 100% of the training data).\n\nI hope this is helpfull for training your own models. If you want i can make a notebook with the complete pipeline. Cheers and have fun\n\n", "1013295": "Thank you for the dataset. Unfortunately, I was not able to download this one: https://www.kaggle.com/ragnar123/landmark-tfrecords-384 using this Kaggle API command. It returned an error message \"404 - Not Found\". Is it because of storage limitations in your Kaggle dataset?", "1012062": "Who is there to be team with with me i have more great approach than this", "1012060": "So Helpful\n@ragnar123 ", "1011138": "I see there is a limit of 20GB on datasets. How were you able to create ~75GB sized datasets?? Thanks", "1005700": "", "1000999": "Thanks, this is very useful！" } }