\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F7389764%2Fc59bfcfa058869e325edde03c04b4f21%2F2025-10-10%20014324.png?generation=1760028871691289&alt=media\")
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F7389764%2F30f245d657642f6a3f06022467eda067%2F2025-10-10%20014437.png?generation=1760028881846870&alt=media\")
I created several models, but for some reason, training converges after just one epoch.
\nI tried various approaches for the train data—simply slicing it, aggregating it with np.mean(), and so on.
This is a modified version of the code I used for local training. It only has 10 SeriesInstanceUIDs, but it should work:
\n[RSNA-IAD]train_swin_s.ipynb
The key points of this Notebook are as follows:
\nReference:
\nSummary:
\nThe graph of training error and validation error from training on the first fold looks like this:
\n
\n
Do you think this code is functioning correctly? Or am I doing something wrong?
", "messages": [ { "id": 3300156, "postDate": "2025-10-09T17:02:28.417Z", "content": "I created several models, but for some reason, training converges after just one epoch.
\nI tried various approaches for the train data—simply slicing it, aggregating it with np.mean(), and so on.
This is a modified version of the code I used for local training. It only has 10 SeriesInstanceUIDs, but it should work:
\n[RSNA-IAD]train_swin_s.ipynb
The key points of this Notebook are as follows:
\nReference:
\nSummary:
\nThe graph of training error and validation error from training on the first fold looks like this:
\n
\n
Do you think this code is functioning correctly? Or am I doing something wrong?
", "rawMarkdown": "I created several models, but for some reason, training converges after just one epoch.\nI tried various approaches for the train data—simply slicing it, aggregating it with np.mean(), and so on.\n\nThis is a modified version of the code I used for local training. It only has 10 SeriesInstanceUIDs, but it should work:\n[[RSNA-IAD]train_swin_s.ipynb](https://www.kaggle.com/code/ataracsia/rsna-iad-train-swin-s)\n\nThe key points of this Notebook are as follows:\n\n**Reference:**\n- RSNA-IAD | Swin Transformer| LB\n- Train 224x224 DICOM->PNGs EfficientNetV2S\n\n**Summary:**\n\n1. Convert .dcm files to 384x384 npy files, save them, and simultaneously save the patient's metadata to a .csv file\n2. Calculate mean, std, and kurtosis channel-wise for each SeriesInstanceUID and save them to npy files\n3. Save the path of each npy file to a .csv file\n4. Create a classification model using image data and meta data based on timm.createmodel(‘swin_small_patch4_window7_224’).\n5. Use a custom criterion.\n6. Split into 5 folds using Multilabel Stratified KFold and train for 20 epochs on each fold.\n\nThe graph of training error and validation error from training on the first fold looks like this:\n\n\n\nDo you think this code is functioning correctly? Or am I doing something wrong?" } ], "comments": [], "raw_markdown_by_id": { "3300156": "I created several models, but for some reason, training converges after just one epoch.\nI tried various approaches for the train data—simply slicing it, aggregating it with np.mean(), and so on.\n\nThis is a modified version of the code I used for local training. It only has 10 SeriesInstanceUIDs, but it should work:\n[[RSNA-IAD]train_swin_s.ipynb](https://www.kaggle.com/code/ataracsia/rsna-iad-train-swin-s)\n\nThe key points of this Notebook are as follows:\n\n**Reference:**\n- RSNA-IAD | Swin Transformer| LB\n- Train 224x224 DICOM->PNGs EfficientNetV2S\n\n**Summary:**\n\n1. Convert .dcm files to 384x384 npy files, save them, and simultaneously save the patient's metadata to a .csv file\n2. Calculate mean, std, and kurtosis channel-wise for each SeriesInstanceUID and save them to npy files\n3. Save the path of each npy file to a .csv file\n4. Create a classification model using image data and meta data based on timm.createmodel(‘swin_small_patch4_window7_224’).\n5. Use a custom criterion.\n6. Split into 5 folds using Multilabel Stratified KFold and train for 20 epochs on each fold.\n\nThe graph of training error and validation error from training on the first fold looks like this:\n\n\n\nDo you think this code is functioning correctly? Or am I doing something wrong?" } }