![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F5483160%2F478068740d14fb19934fb2f237a9deca%2F1.png?generation=1769161443528035&alt=media\")
I would like to express my sincere gratitude to Kaggle and the competition organizers for providing this invaluable opportunity. Special thanks to @hengck23 for sharing his strong baseline, which served as a crucial foundation for my work.
\nMy solution focuses on optimizing Stage 2 of the baseline provided by @hengck23. The key insight is to directly regress the lead signal, bypassing the traditional pipeline of \"segmentation followed by post-processing.\" By treating this as a direct regression task, the model learns the signal features more effectively and reduces the cumulative error often introduced during the post-processing stage.
\n
The training dataset contains ECG signals with diverse sampling rates, ranging from 2.5 kHz to 10 kHz. To ensure high-quality ground truth and maintain optimal Signal-to-Noise Ratio (SNR), a consistent resampling strategy is required. I implemented a benchmarking framework to evaluate the fidelity of various resampling algorithms—including polyphase, linear, cubic spline, and FFT-based methods. The performance was measured using a transformation:
The key observations is
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F5483160%2F2af8c00564301f404ea7b900bf715442%2F1.png?generation=1769138986791210&alt=media\")
Moreover, to accelerate image resampling and integrate it into the training process, I use resample_torch
def resample_torch(self, x, num, dim=-1):\n dim = (x.dim() + dim) if dim < 0 else dim\n X = torch.fft.fft(x, dim=dim)\n Nx = X.shape[dim]\n\n sl = [slice(None)] * X.ndim\n newshape = list(X.shape)\n newshape[dim] = num\n Y = torch.zeros(newshape, dtype=X.dtype, device=X.device)\n\n N = min(num, Nx)\n sl[dim] = slice(0, (N + 1) // 2)\n Y[sl] = X[sl]\n sl[dim] = slice(-(N - 1) // 2, None)\n Y[sl] = X[sl]\n\n if N % 2 == 0:\n if N < Nx:\n sl[dim] = slice(N//2, N//2+1)\n Y[sl] += X[sl]\n elif N < num:\n sl[dim] = slice(num-N//2, num-N//2+1)\n Y[sl] /= 2\n temp = Y[sl]\n sl[dim] = slice(N//2, N//2+1)\n Y[sl] = temp\n\n y = torch.fft.ifft(Y, dim=dim).real * (float(num) / float(Nx))\n return y\nThis module converts 2D feature embeddings into precise physical voltage values for ECG leads by using a Soft-Argmax mechanism to estimate vertical coordinates.
\n# Signal Regression Head\nclass MaskEmbeddingToLeadSignalSoftArgmax(nn.Module):\n def __init__(self, n_leads=4, embedding_dim=32, temperature=0.5):\n super().__init__()\n\n self.n_leads = n_leads\n self.temperature = temperature\n self.lead_y_logits = nn.Conv2d(embedding_dim, n_leads, kernel_size=1)\n\n self.register_buffer('zero_mv', torch.tensor([703.5, 987.5, 1271.5, 1531.5]).view(1, 4, 1))\n self.register_buffer('mv_to_pixel', torch.tensor(79.0))\n\n def forward(self, masked_feat):\n B, C, H, W = masked_feat.shape\n device = masked_feat.device\n dtype = masked_feat.dtype\n\n y_logits = self.lead_y_logits(masked_feat)\n prob = torch.softmax(y_logits / self.temperature, dim=2)\n\n y_coord = torch.arange(H, device=device, dtype=dtype).view(1,1,H,1)\n y_pixel = (prob * y_coord).sum(dim=2) # [B, L, W]\n\n pred_mv = (self.zero_mv - y_pixel) / self.mv_to_pixel\n return pred_mv, prob\n| Signal Length | \nTTA (hflip) | \nEpochs | \nLB Score | \n
|---|---|---|---|
| 2560 | \nFalse | \n60 | \n21.36 | \n
| 5120 | \nFalse | \n100 | \n22.20 | \n
| 5120 | \nTrue | \n100 | \n22.36 | \n
| 5120 | \nTrue | \n150 | \n22.43 | \n
I would like to express my sincere gratitude to Kaggle and the competition organizers for providing this invaluable opportunity. Special thanks to @hengck23 for sharing his strong baseline, which served as a crucial foundation for my work.
\nMy solution focuses on optimizing Stage 2 of the baseline provided by @hengck23. The key insight is to directly regress the lead signal, bypassing the traditional pipeline of \"segmentation followed by post-processing.\" By treating this as a direct regression task, the model learns the signal features more effectively and reduces the cumulative error often introduced during the post-processing stage.
\n
The training dataset contains ECG signals with diverse sampling rates, ranging from 2.5 kHz to 10 kHz. To ensure high-quality ground truth and maintain optimal Signal-to-Noise Ratio (SNR), a consistent resampling strategy is required. I implemented a benchmarking framework to evaluate the fidelity of various resampling algorithms—including polyphase, linear, cubic spline, and FFT-based methods. The performance was measured using a transformation:
The key observations is
\n
Moreover, to accelerate image resampling and integrate it into the training process, I use resample_torch
def resample_torch(self, x, num, dim=-1):\n dim = (x.dim() + dim) if dim < 0 else dim\n X = torch.fft.fft(x, dim=dim)\n Nx = X.shape[dim]\n\n sl = [slice(None)] * X.ndim\n newshape = list(X.shape)\n newshape[dim] = num\n Y = torch.zeros(newshape, dtype=X.dtype, device=X.device)\n\n N = min(num, Nx)\n sl[dim] = slice(0, (N + 1) // 2)\n Y[sl] = X[sl]\n sl[dim] = slice(-(N - 1) // 2, None)\n Y[sl] = X[sl]\n\n if N % 2 == 0:\n if N < Nx:\n sl[dim] = slice(N//2, N//2+1)\n Y[sl] += X[sl]\n elif N < num:\n sl[dim] = slice(num-N//2, num-N//2+1)\n Y[sl] /= 2\n temp = Y[sl]\n sl[dim] = slice(N//2, N//2+1)\n Y[sl] = temp\n\n y = torch.fft.ifft(Y, dim=dim).real * (float(num) / float(Nx))\n return y\nThis module converts 2D feature embeddings into precise physical voltage values for ECG leads by using a Soft-Argmax mechanism to estimate vertical coordinates.
\n# Signal Regression Head\nclass MaskEmbeddingToLeadSignalSoftArgmax(nn.Module):\n def __init__(self, n_leads=4, embedding_dim=32, temperature=0.5):\n super().__init__()\n\n self.n_leads = n_leads\n self.temperature = temperature\n self.lead_y_logits = nn.Conv2d(embedding_dim, n_leads, kernel_size=1)\n\n self.register_buffer('zero_mv', torch.tensor([703.5, 987.5, 1271.5, 1531.5]).view(1, 4, 1))\n self.register_buffer('mv_to_pixel', torch.tensor(79.0))\n\n def forward(self, masked_feat):\n B, C, H, W = masked_feat.shape\n device = masked_feat.device\n dtype = masked_feat.dtype\n\n y_logits = self.lead_y_logits(masked_feat)\n prob = torch.softmax(y_logits / self.temperature, dim=2)\n\n y_coord = torch.arange(H, device=device, dtype=dtype).view(1,1,H,1)\n y_pixel = (prob * y_coord).sum(dim=2) # [B, L, W]\n\n pred_mv = (self.zero_mv - y_pixel) / self.mv_to_pixel\n return pred_mv, prob\n| Signal Length | \nTTA (hflip) | \nEpochs | \nLB Score | \n
|---|---|---|---|
| 2560 | \nFalse | \n60 | \n21.36 | \n
| 5120 | \nFalse | \n100 | \n22.20 | \n
| 5120 | \nTrue | \n100 | \n22.36 | \n
| 5120 | \nTrue | \n150 | \n22.43 | \n
@hengck23 is the goat. he does a lot in simplifying complex stuff
", "rawMarkdown": "hengck23 is the goat. he does a lot in simplifying complex stuff", "votes": null }, { "id": "3395569", "postDate": "01/23/2026 08:41:09", "content": "If i may ask, how did you train on such a big dataset? Hardware config? and how long did it take?
", "rawMarkdown": "If i may ask, how did you train on such a big dataset? Hardware config? and how long did it take?", "votes": null }, { "id": "3395573", "postDate": "01/23/2026 08:45:10", "content": "It takes approximately 10 hours to complete full training (without validation) using 24 NVIDIA RTX 3090 GPUs (24 GB each).
", "rawMarkdown": "It takes approximately 10 hours to complete full training (without validation) using 24 NVIDIA RTX 3090 GPUs (24 GB each).", "votes": null }, { "id": "3395719", "postDate": "01/23/2026 14:42:02", "content": "SImple but effective, nice solution. Do you have a measure of how much scipy.signal.resample improved CV/LB over torch.nn.functional.interpolate?
Nice work! Do you mean you used 24 GPUs for parallel training?
", "rawMarkdown": "Nice work! Do you mean you used 24 GPUs for parallel training?", "votes": null } ], "comments": [ { "id": 3395503, "author_name": "samu2505", "author_url": "", "post_date": "01/23/2026 05:09:45", "content": "@hengck23 is the goat. he does a lot in simplifying complex stuff
", "votes": null, "replies": [] }, { "id": 3395569, "author_name": "jaiaaryan", "author_url": "", "post_date": "01/23/2026 08:41:09", "content": "If i may ask, how did you train on such a big dataset? Hardware config? and how long did it take?
", "votes": null, "replies": [ { "id": 3395573, "author_name": "sjtuwangshuo", "author_url": "", "post_date": "01/23/2026 08:45:10", "content": "It takes approximately 10 hours to complete full training (without validation) using 24 NVIDIA RTX 3090 GPUs (24 GB each).
", "votes": null, "replies": [ { "id": 3421680, "author_name": "mancyw", "author_url": "", "post_date": "03/16/2026 04:32:26", "content": "Nice work! Do you mean you used 24 GPUs for parallel training?
", "votes": null, "replies": [] } ] } ] }, { "id": 3395719, "author_name": "brendanartley", "author_url": "", "post_date": "01/23/2026 14:42:02", "content": "SImple but effective, nice solution. Do you have a measure of how much scipy.signal.resample improved CV/LB over torch.nn.functional.interpolate?