Newbie here! This might be a dumb question but
\nCan someone tell me how to average the Cv i am using 5 fold Cv during training but i am stuck at how do i ensemble or may i say average them? Do we need to average the model weights? Or should i ensemble during inference? Or is there something else? Would be a great help if someone could tell me what is the best option and how to do it properly
Newbie here! This might be a dumb question but
\nCan someone tell me how to average the Cv i am using 5 fold Cv during training but i am stuck at how do i ensemble or may i say average them? Do we need to average the model weights? Or should i ensemble during inference? Or is there something else? Would be a great help if someone could tell me what is the best option and how to do it properly
like this
\ndef (model_class, checkpoint_path, device):\n model = ()\n model.(torch.(checkpoint_path, map_location=device))\n model.(device)\n model.()\n return model\n\ndef (model_class, checkpoint_paths, dataloader, device=):\n models = [(model_class, ckpt, device) for ckpt in checkpoint_paths]\n\n all_preds = []\n\n with torch.():\n for batch in (dataloader, desc=):\n inputs = batch.(device)\n\n fold_preds = []\n for model in models:\n preds = (inputs)\n fold_preds.(preds)\n\n # Stack along new dimension and average\n fold_preds = torch.(fold_preds, dim=) # (num_folds, batch_size, ...)\n avg_preds = fold_preds.(dim=) # (batch_size, ...)\n\n all_preds.(avg_preds.())\n\n all_preds = torch.(all_preds, dim=)\n return all_preds\n@seeingtimes Thanks a lot! I suppose this snippet should be put in the inference notebook but will it not increase the inference time?
", "rawMarkdown": "@seeingtimes Thanks a lot! I suppose this snippet should be put in the inference notebook but will it not increase the inference time?", "replies": [ { "id": 3188150, "postDate": "2025-04-27T05:41:30.740Z", "content": "It will increase the inference time since you are inferring five models instead of one.
", "rawMarkdown": "It will increase the inference time since you are inferring five models instead of one." }, { "id": 3188185, "postDate": "2025-04-27T07:23:35.703Z", "rawMarkdown": "", "isDeleted": true } ] } ] } ], "comments": [ { "id": 3187723, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-04-26T12:39:29.647000", "content": "like this
\ndef (model_class, checkpoint_path, device):\n model = ()\n model.(torch.(checkpoint_path, map_location=device))\n model.(device)\n model.()\n return model\n\ndef (model_class, checkpoint_paths, dataloader, device=):\n models = [(model_class, ckpt, device) for ckpt in checkpoint_paths]\n\n all_preds = []\n\n with torch.():\n for batch in (dataloader, desc=):\n inputs = batch.(device)\n\n fold_preds = []\n for model in models:\n preds = (inputs)\n fold_preds.(preds)\n\n # Stack along new dimension and average\n fold_preds = torch.(fold_preds, dim=) # (num_folds, batch_size, ...)\n avg_preds = fold_preds.(dim=) # (batch_size, ...)\n\n all_preds.(avg_preds.())\n\n all_preds = torch.(all_preds, dim=)\n return all_preds\n@seeingtimes Thanks a lot! I suppose this snippet should be put in the inference notebook but will it not increase the inference time?
", "votes": 0, "replies": [ { "id": 3188150, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-04-27T05:41:30.740000", "content": "It will increase the inference time since you are inferring five models instead of one.
", "votes": 0, "replies": [] }, { "id": 3188185, "author_name": "", "author_url": "", "post_date": "2025-04-27T07:23:35.703000", "content": "", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3187650": "Newbie here! This might be a dumb question but \nCan someone tell me how to average the Cv i am using 5 fold Cv during training but i am stuck at how do i ensemble or may i say average them? Do we need to average the model weights? Or should i ensemble during inference? Or is there something else? Would be a great help if someone could tell me what is the best option and how to do it properly", "3187723": "like this \n```\ndef load_model(model_class, checkpoint_path, device):\n model = model_class()\n model.load_state_dict(torch.load(checkpoint_path, map_location=device))\n model.to(device)\n model.eval()\n return model\n\ndef fold_ensemble_predict(model_class, checkpoint_paths, dataloader, device='cuda'):\n models = [load_model(model_class, ckpt, device) for ckpt in checkpoint_paths]\n \n all_preds = []\n \n with torch.no_grad():\n for batch in tqdm(dataloader, desc=\"Ensemble Inference\"):\n inputs = batch.to(device)\n\n fold_preds = []\n for model in models:\n preds = model(inputs)\n fold_preds.append(preds)\n\n # Stack along new dimension and average\n fold_preds = torch.stack(fold_preds, dim=0) # (num_folds, batch_size, ...)\n avg_preds = fold_preds.mean(dim=0) # (batch_size, ...)\n \n all_preds.append(avg_preds.cpu())\n \n all_preds = torch.cat(all_preds, dim=0)\n return all_preds\n```" } }