{ "id": 355114, "title": "How to write competition metric in tensorflow", "url": "/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/355114", "author_name": "", "post_date": "2022-09-25T13:25:19.558059600Z", "votes": 2, "comment_count": 4, "views": 0, "content": "

I would like to create custom metrics in tensorflow to match the metrics of this competition as shown in the following discussion, but I don't know how to write the code. If anyone has written the code, I would appreciate it if you could let me know.
\nhttps://www.kaggle.com/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/341854

", "messages": [ { "id": "1954806", "postDate": "09/25/2022 13:25:19", "content": "

I would like to create custom metrics in tensorflow to match the metrics of this competition as shown in the following discussion, but I don't know how to write the code. If anyone has written the code, I would appreciate it if you could let me know.
\nhttps://www.kaggle.com/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/341854

", "rawMarkdown": "I would like to create custom metrics in tensorflow to match the metrics of this competition as shown in the following discussion, but I don't know how to write the code. If anyone has written the code, I would appreciate it if you could let me know.\nhttps://www.kaggle.com/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/341854", "votes": null }, { "id": "1955048", "postDate": "09/25/2022 17:11:44", "content": "

You should only need to change the loss function. Try this:

\n

loss_fn = tf.losses.BinaryCrossEntropy(from_logits=True, reduction=tf.keras.losses.Reduction.NONE)

", "rawMarkdown": "You should only need to change the loss function. Try this:\n\n`loss_fn = tf.losses.BinaryCrossEntropy(from_logits=True, reduction=tf.keras.losses.Reduction.NONE)`", "votes": null }, { "id": "1955620", "postDate": "09/26/2022 04:33:16", "content": "

Thank you for your reply.
\nI also tried to use BinaryCrossEntropy, but I felt that alone it would not take into account the weights specific to this competition, as shown below.

\n

competition_weights = {\n '-' : torch.tensor([7, 1, 1, 1, 1, 1, 1, 1], dtype=torch.float, device=device),\n '+' : torch.tensor([14, 2, 2, 2, 2, 2, 2, 2], dtype=torch.float, device=device),\n}

", "rawMarkdown": "Thank you for your reply.\nI also tried to use BinaryCrossEntropy, but I felt that alone it would not take into account the weights specific to this competition, as shown below.\n\n`competition_weights = {\n '-' : torch.tensor([7, 1, 1, 1, 1, 1, 1, 1], dtype=torch.float, device=device),\n '+' : torch.tensor([14, 2, 2, 2, 2, 2, 2, 2], dtype=torch.float, device=device),\n}`", "votes": null }, { "id": "1959972", "postDate": "09/28/2022 12:25:57", "content": "

Look at this notebook, the final function is the correct one, I believe
\nhttps://www.kaggle.com/code/solverworld/rsna2022-comp-metric

", "rawMarkdown": "Look at this notebook, the final function is the correct one, I believe\nhttps://www.kaggle.com/code/solverworld/rsna2022-comp-metric", "votes": null }, { "id": "1961897", "postDate": "09/29/2022 12:08:03", "content": "

Thanks for the reply.
\nI would like to know the metrics implemented in sensorflow, not in pytorch.

", "rawMarkdown": "Thanks for the reply.\nI would like to know the metrics implemented in sensorflow, not in pytorch.", "votes": null } ], "comments": [ { "id": 1955048, "author_name": "samuelcortinhas", "author_url": "", "post_date": "09/25/2022 17:11:44", "content": "

You should only need to change the loss function. Try this:

\n

loss_fn = tf.losses.BinaryCrossEntropy(from_logits=True, reduction=tf.keras.losses.Reduction.NONE)

", "votes": null, "replies": [ { "id": 1955620, "author_name": "kun4qi", "author_url": "", "post_date": "09/26/2022 04:33:16", "content": "

Thank you for your reply.
\nI also tried to use BinaryCrossEntropy, but I felt that alone it would not take into account the weights specific to this competition, as shown below.

\n

competition_weights = {\n '-' : torch.tensor([7, 1, 1, 1, 1, 1, 1, 1], dtype=torch.float, device=device),\n '+' : torch.tensor([14, 2, 2, 2, 2, 2, 2, 2], dtype=torch.float, device=device),\n}

", "votes": null, "replies": [] } ] }, { "id": 1959972, "author_name": "solverworld", "author_url": "", "post_date": "09/28/2022 12:25:57", "content": "

Look at this notebook, the final function is the correct one, I believe
\nhttps://www.kaggle.com/code/solverworld/rsna2022-comp-metric

", "votes": null, "replies": [ { "id": 1961897, "author_name": "kun4qi", "author_url": "", "post_date": "09/29/2022 12:08:03", "content": "

Thanks for the reply.
\nI would like to know the metrics implemented in sensorflow, not in pytorch.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1954806": "I would like to create custom metrics in tensorflow to match the metrics of this competition as shown in the following discussion, but I don't know how to write the code. If anyone has written the code, I would appreciate it if you could let me know.\nhttps://www.kaggle.com/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/341854", "1955048": "You should only need to change the loss function. Try this:\n\n`loss_fn = tf.losses.BinaryCrossEntropy(from_logits=True, reduction=tf.keras.losses.Reduction.NONE)`", "1955620": "Thank you for your reply.\nI also tried to use BinaryCrossEntropy, but I felt that alone it would not take into account the weights specific to this competition, as shown below.\n\n`competition_weights = {\n '-' : torch.tensor([7, 1, 1, 1, 1, 1, 1, 1], dtype=torch.float, device=device),\n '+' : torch.tensor([14, 2, 2, 2, 2, 2, 2, 2], dtype=torch.float, device=device),\n}`", "1959972": "Look at this notebook, the final function is the correct one, I believe\nhttps://www.kaggle.com/code/solverworld/rsna2022-comp-metric", "1961897": "Thanks for the reply.\nI would like to know the metrics implemented in sensorflow, not in pytorch." }, "source": "meta" }