\nAbove copy (model weights from my public datasets )
\nI try simple idea .
\nRuntime in gpu only 102.3s(version 6),but still timeout.
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4184342%2F922aaaee976627b735c89487332deb8c%2Ftimeout.png?generation=1600775799447568&alt=media\")
\nI don't know what problem happened?
\nCan someone help me?
my notebook
\nAbove copy (model weights from my public datasets )
\nI try simple idea .
\nRuntime in gpu only 102.3s(version 6),but still timeout.
\n
\nI don't know what problem happened?
\nCan someone help me?
my notebook
\nAbove copy (model weights from my public datasets )
\nI try simple idea .
\nRuntime in gpu only 102.3s(version 6),but still timeout.
\n
\nI don't know what problem happened?
\nCan someone help me?
@chihantsai The 102 second run is only for the public test set of 5 patients, the actual run on the hidden private test that gives you a score contains about 200 patients. From my notebook runs using dicom images it takes around 2.5 hours for preprocessing all these images in the hidden test set even with multiprocessing.
\nThe bottleneck here is the CPU and since you are not preprocessing the images and caching them your dataloader pipeline is not efficient. So i suggest you to first preprocess all the images and cache them as numpy arrays and then load those arrays within your dataset loader. Just 1 last thing to remember your runtime is capped at 4 hours for GPU and 9 hours for CPU.
", "rawMarkdown": "chihantsai The 102 second run is only for the public test set of 5 patients, the actual run on the hidden private test that gives you a score contains about 200 patients. From my notebook runs using dicom images it takes around 2.5 hours for preprocessing all these images in the hidden test set even with multiprocessing.\n\nThe bottleneck here is the CPU and since you are not preprocessing the images and caching them your dataloader pipeline is not efficient. So i suggest you to first preprocess all the images and cache them as numpy arrays and then load those arrays within your dataset loader. Just 1 last thing to remember your runtime is capped at 4 hours for GPU and 9 hours for CPU.", "votes": null }, { "id": "1022476", "postDate": "09/22/2020 15:08:20", "content": "thank you very much.
", "rawMarkdown": "thank you very much.", "votes": null } ], "comments": [ { "id": 1022446, "author_name": "yovinyahathugoda", "author_url": "", "post_date": "09/22/2020 14:35:20", "content": "@chihantsai The 102 second run is only for the public test set of 5 patients, the actual run on the hidden private test that gives you a score contains about 200 patients. From my notebook runs using dicom images it takes around 2.5 hours for preprocessing all these images in the hidden test set even with multiprocessing.
\nThe bottleneck here is the CPU and since you are not preprocessing the images and caching them your dataloader pipeline is not efficient. So i suggest you to first preprocess all the images and cache them as numpy arrays and then load those arrays within your dataset loader. Just 1 last thing to remember your runtime is capped at 4 hours for GPU and 9 hours for CPU.
", "votes": null, "replies": [ { "id": 1022476, "author_name": "chihantsai", "author_url": "", "post_date": "09/22/2020 15:08:20", "content": "thank you very much.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1022235": "[my notebook](https://www.kaggle.com/chihantsai/image-1-model-b4-inference-timeout)\n[Above copy](https://www.kaggle.com/chihantsai/fork-of-image-1-model-b4-v6-inference/output) (model weights from my public datasets )\nI try simple idea .\nRuntime in gpu only 102.3s(version 6),but still timeout.\n\nI don't know what problem happened?\nCan someone help me?", "1022446": "chihantsai The 102 second run is only for the public test set of 5 patients, the actual run on the hidden private test that gives you a score contains about 200 patients. From my notebook runs using dicom images it takes around 2.5 hours for preprocessing all these images in the hidden test set even with multiprocessing.\n\nThe bottleneck here is the CPU and since you are not preprocessing the images and caching them your dataloader pipeline is not efficient. So i suggest you to first preprocess all the images and cache them as numpy arrays and then load those arrays within your dataset loader. Just 1 last thing to remember your runtime is capped at 4 hours for GPU and 9 hours for CPU.", "1022476": "thank you very much." }, "source": "meta" }