I notice that the computation resource is limited in submissions (only CPU notebooks and <= 2 hours). Therefore, faster inference may be necessary to apply some complex strategies.
\nIn this notebook, ONNX is employed to speed up the inference (about x2-3 faster than the original version). Hope it is helpful for you. By the way, ONNX and OpenVINO are also popular in BirdCLEF 2023.
", "messages": [ { "id": 2750338, "postDate": "2024-04-13T15:36:36.163Z", "content": "I notice that the computation resource is limited in submissions (only CPU notebooks and <= 2 hours). Therefore, faster inference may be necessary to apply some complex strategies.
\nIn this notebook, ONNX is employed to speed up the inference (about x2-3 faster than the original version). Hope it is helpful for you. By the way, ONNX and OpenVINO are also popular in BirdCLEF 2023.
", "rawMarkdown": "I notice that the computation resource is limited in submissions (only CPU notebooks and <= 2 hours). Therefore, faster inference may be necessary to apply some complex strategies. \n\nIn this [notebook](https://www.kaggle.com/code/zijiangyang1116/birdclef-24-inference-with-onnx), ONNX is employed to speed up the inference (about x2-3 faster than the original version). Hope it is helpful for you. By the way, ONNX and OpenVINO are also popular in BirdCLEF 2023.", "votes": 20 }, { "id": 2824821, "postDate": "2024-05-20T04:33:40.527Z", "content": "Hey @zijiangyang1116, thanks for the resource. I tried using the same notebook with the efficientnet_b5 model but it was giving an OOM error while submission.
", "rawMarkdown": "Hey @zijiangyang1116, thanks for the resource. I tried using the same notebook with the efficientnet_b5 model but it was giving an OOM error while submission. ", "replies": [ { "id": 2838738, "postDate": "2024-05-27T07:38:24.330Z", "content": "The BATCH_SIZE is set to 64. Therefore, reducing it to 32 or lower may be necessary for efficientnet_b5.
Thank you for sharing your insights! I am trying to convert a TensorFlow model to ONNX, but I am experiencing errors during the conversion process (tf2onnx.convert.from_keras) due to incompatible library versions. Here are the versions I tried that led to errors:
\nTensorFlow (tf): 2.15.0
\nTensorFlow IO (tfio): 0.35.0
\ntf2onnx: 1.16.1
\nONNX: 1.16.0
\nONNX Runtime: 1.17.3
Yes. The training is shared in this notebook.
", "rawMarkdown": "Yes. The training is shared in [this notebook](https://www.kaggle.com/code/zijiangyang1116/birdclef-24-efficientnetb0-pytorch-train)." } ] }, { "id": 2750429, "postDate": "2024-04-13T17:02:47.103Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 2824821, "author_name": "Rishav dash", "author_url": "", "post_date": "2024-05-20T04:33:40.527000", "content": "Hey @zijiangyang1116, thanks for the resource. I tried using the same notebook with the efficientnet_b5 model but it was giving an OOM error while submission.
", "votes": 0, "replies": [ { "id": 2838738, "author_name": "Koolo", "author_url": "", "post_date": "2024-05-27T07:38:24.330000", "content": "The BATCH_SIZE is set to 64. Therefore, reducing it to 32 or lower may be necessary for efficientnet_b5.
Thank you for sharing your insights! I am trying to convert a TensorFlow model to ONNX, but I am experiencing errors during the conversion process (tf2onnx.convert.from_keras) due to incompatible library versions. Here are the versions I tried that led to errors:
\nTensorFlow (tf): 2.15.0
\nTensorFlow IO (tfio): 0.35.0
\ntf2onnx: 1.16.1
\nONNX: 1.16.0
\nONNX Runtime: 1.17.3
Yes. The training is shared in this notebook.
", "votes": 0, "replies": [] } ] }, { "id": 2750429, "author_name": "", "author_url": "", "post_date": "2024-04-13T17:02:47.103000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2750338": "I notice that the computation resource is limited in submissions (only CPU notebooks and <= 2 hours). Therefore, faster inference may be necessary to apply some complex strategies. \n\nIn this [notebook](https://www.kaggle.com/code/zijiangyang1116/birdclef-24-inference-with-onnx), ONNX is employed to speed up the inference (about x2-3 faster than the original version). Hope it is helpful for you. By the way, ONNX and OpenVINO are also popular in BirdCLEF 2023.", "2824821": "Hey @zijiangyang1116, thanks for the resource. I tried using the same notebook with the efficientnet_b5 model but it was giving an OOM error while submission. ", "2752476": "Thank you for sharing your insights! I am trying to convert a TensorFlow model to ONNX, but I am experiencing errors during the conversion process (tf2onnx.convert.from_keras) due to incompatible library versions. Here are the versions I tried that led to errors:\n\nTensorFlow (tf): 2.15.0\nTensorFlow IO (tfio): 0.35.0\ntf2onnx: 1.16.1\nONNX: 1.16.0\nONNX Runtime: 1.17.3", "2847726": "", "2750429": "" } }