I have been trying to figure out why my submission code is so slow for the last whole day but have not had any success. I did some profiling and found that the inference code, literally the one line output = model(inputs) is the part slowing everything down. DataLoading speeds and everything else seem fine but this part is almost 3-4 times slower than most public notebooks' code taking 10-12+ seconds to run inference on one 4 minute clip compared to most public notebooks which can run at 5 seconds per 4 minute clip. I tried using onnx but it did not help much and even if it did i think my code is unusually slow even before onnx speedup so I want to figure out the issue. Would anyone be willing to look at my code and help? this is the link to my submission code https://www.kaggle.com/code/snehalverma10/birdclef-submission
I have been trying to figure out why my submission code is so slow for the last whole day but have not had any success. I did some profiling and found that the inference code, literally the one line output = model(inputs) is the part slowing everything down. DataLoading speeds and everything else seem fine but this part is almost 3-4 times slower than most public notebooks' code taking 10-12+ seconds to run inference on one 4 minute clip compared to most public notebooks which can run at 5 seconds per 4 minute clip. I tried using onnx but it did not help much and even if it did i think my code is unusually slow even before onnx speedup so I want to figure out the issue. Would anyone be willing to look at my code and help? this is the link to my submission code https://www.kaggle.com/code/snehalverma10/birdclef-submission
I was originally using timm tf_efficientnet_b0_ns. Then i switched to torchvision.models.efficientnetb0 to see if it helped and it did not.
", "rawMarkdown": "I was originally using timm tf_efficientnet_b0_ns. Then i switched to torchvision.models.efficientnetb0 to see if it helped and it did not.", "votes": null }, { "id": "2762771", "postDate": "04/20/2024 04:53:06", "content": "How about this topic?
\nhttps://www.kaggle.com/competitions/birdclef-2024/discussion/492649
", "rawMarkdown": "How about this topic?\n\nhttps://www.kaggle.com/competitions/birdclef-2024/discussion/492649", "votes": null }, { "id": "2765593", "postDate": "04/21/2024 08:14:14", "content": "so - can't tell too much without seeing the model…. I will say that some of my earlier efforts involved adding additional layers to an efficientnet - and adding the layers caused a larger-than-intuitive performance drop.
\nIf you are adding a bunch of layers - I might look at that first.
", "rawMarkdown": "so - can't tell too much without seeing the model.... I will say that some of my earlier efforts involved adding additional layers to an efficientnet - and adding the layers caused a larger-than-intuitive performance drop.\n\nIf you are adding a bunch of layers - I might look at that first.", "votes": null } ], "comments": [ { "id": 2760218, "author_name": "snehalverma10", "author_url": "", "post_date": "04/19/2024 06:52:14", "content": "I was originally using timm tf_efficientnet_b0_ns. Then i switched to torchvision.models.efficientnetb0 to see if it helped and it did not.
", "votes": null, "replies": [] }, { "id": 2762771, "author_name": "goyashuto", "author_url": "", "post_date": "04/20/2024 04:53:06", "content": "How about this topic?
\nhttps://www.kaggle.com/competitions/birdclef-2024/discussion/492649
", "votes": null, "replies": [] }, { "id": 2765593, "author_name": "richolson", "author_url": "", "post_date": "04/21/2024 08:14:14", "content": "so - can't tell too much without seeing the model…. I will say that some of my earlier efforts involved adding additional layers to an efficientnet - and adding the layers caused a larger-than-intuitive performance drop.
\nIf you are adding a bunch of layers - I might look at that first.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2760217": "I have been trying to figure out why my submission code is so slow for the last whole day but have not had any success. I did some profiling and found that the inference code, literally the one line `output = model(inputs)` is the part slowing everything down. DataLoading speeds and everything else seem fine but this part is almost 3-4 times slower than most public notebooks' code taking 10-12+ seconds to run inference on one 4 minute clip compared to most public notebooks which can run at 5 seconds per 4 minute clip. I tried using onnx but it did not help much and even if it did i think my code is unusually slow even before onnx speedup so I want to figure out the issue. Would anyone be willing to look at my code and help? this is the link to my submission code https://www.kaggle.com/code/snehalverma10/birdclef-submission", "2760218": "I was originally using timm tf_efficientnet_b0_ns. Then i switched to torchvision.models.efficientnetb0 to see if it helped and it did not.", "2762771": "How about this topic?\n\nhttps://www.kaggle.com/competitions/birdclef-2024/discussion/492649", "2765593": "so - can't tell too much without seeing the model.... I will say that some of my earlier efforts involved adding additional layers to an efficientnet - and adding the layers caused a larger-than-intuitive performance drop.\n\nIf you are adding a bunch of layers - I might look at that first." }, "source": "meta" }