I use Data Augmentation
\nrandom_transforms = Compose([\n RandCropByLabelClassesd(\n keys=[\"image\", \"label\"],\n label_key=\"label\",\n spatial_size=[128, 128, 128],\n num_classes=7,\n num_samples=my_num_samples\n ),\n RandRotate90d(\n keys=[\"image\", \"label\"],\n prob=0.3,\n spatial_axes=[0, 1]\n ),\n RandRotate90d(\n keys=[\"image\", \"label\"],\n prob=0.2,\n spatial_axes=[1, 2]\n ),\n RandFlipd(\n keys=[\"image\", \"label\"],\n prob=0.3,\n spatial_axis=0\n ),\n RandFlipd(\n keys=[\"image\", \"label\"],\n prob=0.3,\n spatial_axis=1\n ),\n # Optionally, you can also flip along the third axis:\n # RandFlipd(\n # keys=[\"image\", \"label\"],\n # prob=0.3,\n # spatial_axis=2\n # ),\n RandAffined(\n keys=[\"image\", \"label\"],\n prob=0.5,\n rotate_range=(0.17, 0.17, 0.17),\n scale_range=(0.05, 0.05, 0.05),\n mode=(\"bilinear\", \"nearest\"),\n padding_mode=\"zeros\"\n ),\n RandScaleIntensityd(\n keys=\"image\",\n prob=0.2,\n factors=0.1\n ),\n RandShiftIntensityd(\n keys=\"image\",\n prob=0.2,\n offsets=0.1\n ),\n RandAdjustContrastd(\n keys=\"image\",\n prob=0.2,\n gamma=(0.9, 1.1)\n ),\n RandHistogramShiftd(\n keys=\"image\",\n prob=0.2,\n num_control_points=10\n )\n])\n
Model Architecture
\nMy model is based on the MONAI UNet and uses the following configurations:
Channels: (64, 128, 256, 256)
\nStrides Pattern: (2, 2, 1)
\nNumber of Residual Units: 1
\nDue to limited GPU resources, I utilized an L4 GPU.
Performance Results
\nPure 3D UNet: Achieved a public score of 0.722 and a private score of 0.726.
\nEnsemble with YOLO: Reached a public score of 0.757 and a private score of 0.755.
Additional Enhancements
\nI also incorporated a filtering mechanism that ignores any cluster where the standard deviation of each coordinate exceeds 40% of the particle radius.
For further details, please refer to the notebook.
\nhttps://www.kaggle.com/code/junhanzangai/czii-cryo-s
\nhttps://www.kaggle.com/code/junhanzangai/submission-test\\
And i attach my code, too.
\nThank you for giving me this opportunity.
", "messages": [ { "id": "3116419", "postDate": "02/06/2025 00:21:54", "content": "I use Data Augmentation
\nrandom_transforms = Compose([\n RandCropByLabelClassesd(\n keys=[\"image\", \"label\"],\n label_key=\"label\",\n spatial_size=[128, 128, 128],\n num_classes=7,\n num_samples=my_num_samples\n ),\n RandRotate90d(\n keys=[\"image\", \"label\"],\n prob=0.3,\n spatial_axes=[0, 1]\n ),\n RandRotate90d(\n keys=[\"image\", \"label\"],\n prob=0.2,\n spatial_axes=[1, 2]\n ),\n RandFlipd(\n keys=[\"image\", \"label\"],\n prob=0.3,\n spatial_axis=0\n ),\n RandFlipd(\n keys=[\"image\", \"label\"],\n prob=0.3,\n spatial_axis=1\n ),\n # Optionally, you can also flip along the third axis:\n # RandFlipd(\n # keys=[\"image\", \"label\"],\n # prob=0.3,\n # spatial_axis=2\n # ),\n RandAffined(\n keys=[\"image\", \"label\"],\n prob=0.5,\n rotate_range=(0.17, 0.17, 0.17),\n scale_range=(0.05, 0.05, 0.05),\n mode=(\"bilinear\", \"nearest\"),\n padding_mode=\"zeros\"\n ),\n RandScaleIntensityd(\n keys=\"image\",\n prob=0.2,\n factors=0.1\n ),\n RandShiftIntensityd(\n keys=\"image\",\n prob=0.2,\n offsets=0.1\n ),\n RandAdjustContrastd(\n keys=\"image\",\n prob=0.2,\n gamma=(0.9, 1.1)\n ),\n RandHistogramShiftd(\n keys=\"image\",\n prob=0.2,\n num_control_points=10\n )\n])\n
Model Architecture
\nMy model is based on the MONAI UNet and uses the following configurations:
Channels: (64, 128, 256, 256)
\nStrides Pattern: (2, 2, 1)
\nNumber of Residual Units: 1
\nDue to limited GPU resources, I utilized an L4 GPU.
Performance Results
\nPure 3D UNet: Achieved a public score of 0.722 and a private score of 0.726.
\nEnsemble with YOLO: Reached a public score of 0.757 and a private score of 0.755.
Additional Enhancements
\nI also incorporated a filtering mechanism that ignores any cluster where the standard deviation of each coordinate exceeds 40% of the particle radius.
For further details, please refer to the notebook.
\nhttps://www.kaggle.com/code/junhanzangai/czii-cryo-s
\nhttps://www.kaggle.com/code/junhanzangai/submission-test\\
And i attach my code, too.
\nThank you for giving me this opportunity.
", "rawMarkdown": "I use Data Augmentation\n\n`random_transforms = Compose([\n RandCropByLabelClassesd(\n keys=[\"image\", \"label\"],\n label_key=\"label\",\n spatial_size=[128, 128, 128],\n num_classes=7,\n num_samples=my_num_samples\n ),\n RandRotate90d(\n keys=[\"image\", \"label\"],\n prob=0.3,\n spatial_axes=[0, 1]\n ),\n RandRotate90d(\n keys=[\"image\", \"label\"],\n prob=0.2,\n spatial_axes=[1, 2]\n ),\n RandFlipd(\n keys=[\"image\", \"label\"],\n prob=0.3,\n spatial_axis=0\n ),\n RandFlipd(\n keys=[\"image\", \"label\"],\n prob=0.3,\n spatial_axis=1\n ),\n # Optionally, you can also flip along the third axis:\n # RandFlipd(\n # keys=[\"image\", \"label\"],\n # prob=0.3,\n # spatial_axis=2\n # ),\n RandAffined(\n keys=[\"image\", \"label\"],\n prob=0.5,\n rotate_range=(0.17, 0.17, 0.17),\n scale_range=(0.05, 0.05, 0.05),\n mode=(\"bilinear\", \"nearest\"),\n padding_mode=\"zeros\"\n ),\n RandScaleIntensityd(\n keys=\"image\",\n prob=0.2,\n factors=0.1\n ),\n RandShiftIntensityd(\n keys=\"image\",\n prob=0.2,\n offsets=0.1\n ),\n RandAdjustContrastd(\n keys=\"image\",\n prob=0.2,\n gamma=(0.9, 1.1)\n ),\n RandHistogramShiftd(\n keys=\"image\",\n prob=0.2,\n num_control_points=10\n )\n])\n`\n\n**Model Architecture**\nMy model is based on the MONAI UNet and uses the following configurations:\n\nChannels: (64, 128, 256, 256)\nStrides Pattern: (2, 2, 1)\nNumber of Residual Units: 1\nDue to limited GPU resources, I utilized an L4 GPU.\n\n**Performance Results**\nPure 3D UNet: Achieved a public score of 0.722 and a private score of 0.726.\nEnsemble with YOLO: Reached a public score of 0.757 and a private score of 0.755.\n\n**Additional Enhancements**\nI also incorporated a filtering mechanism that ignores any cluster where the standard deviation of each coordinate exceeds 40% of the particle radius.\n\nFor further details, please refer to the notebook.\nhttps://www.kaggle.com/code/junhanzangai/czii-cryo-s\nhttps://www.kaggle.com/code/junhanzangai/submission-test\\\n\nAnd i attach my code, too.\n\nThank you for giving me this opportunity.", "votes": null }, { "id": "3116421", "postDate": "02/06/2025 00:25:22", "content": "Oh, wow! I really like that filter idea!
", "rawMarkdown": "Oh, wow! I really like that filter idea!", "votes": null }, { "id": "3116425", "postDate": "02/06/2025 00:29:45", "content": "Congratulations! Wow, I didn't expect YOLO to boost the scores that much. That's amazing!
", "rawMarkdown": "Congratulations! Wow, I didn't expect YOLO to boost the scores that much. That's amazing!", "votes": null }, { "id": "3116442", "postDate": "02/06/2025 00:56:18", "content": "In my case, filtering id is main for scoring up.
", "rawMarkdown": "In my case, filtering id is main for scoring up.", "votes": null }, { "id": "3116443", "postDate": "02/06/2025 00:58:02", "content": "In fact, as I working with 3D, I realized that 2.5D could have a good impact.
", "rawMarkdown": "In fact, as I working with 3D, I realized that 2.5D could have a good impact.", "votes": null }, { "id": "3116449", "postDate": "02/06/2025 01:01:15", "content": "so if i understand correctly was that mostly getting rid of non-spherical particles? or rather particle candidates i suppose.
", "rawMarkdown": "so if i understand correctly was that mostly getting rid of non-spherical particles? or rather particle candidates i suppose.", "votes": null }, { "id": "3116458", "postDate": "02/06/2025 01:14:56", "content": "Your understanding is partly correct but there's a nuance. The filtering mechanism isn't intended solely to remove non-spherical particles. Instead, it's designed to ignore any candidate cluster where the standard deviation along any coordinate exceeds 40% of the expected particle radius. This serves as a proxy for detecting clusters with overly dispersed predictions—which likely indicates unstable or false-positive detections—rather than a direct assessment of the particle’s inherent shape. In other words, it's more about ensuring the predicted candidate clusters are spatially compact and consistent with what you'd expect for a true particle detection.
\nAdditionally, I used the following code to filter out as many genuine particles as possible, but the score actually dropped.
\n`BLOB_THRESHOLD = 250
\nclasses = [1, 2, 3, 4, 5, 6]
\ntotal_start = time.time()
with torch.no_grad():
\n location_df = []
run root.runs:\n run_start = time.time() \n\n (run)\n\n \n load_start = time.time()\n tomo = run.get_voxel_spacing()\n tomo_arr = tomo.get_tomogram(tomo_type).numpy() \n load_end = time.time()\n ()\n\n \n prep_start = time.time()\n data_dict = [{: tomo_arr}]\n tomo_ds = CacheDataset(data=data_dict, transform=inference_transforms, cache_rate=)\n volume_tensor = tomo_ds[][].unsqueeze().to() \n prep_end = time.time()\n ()\n\n \n \n infer_start = time.time()\n out_logits = sliding_window_inference(\n inputs=volume_tensor,\n roi_size=(, , ),\n sw_batch_size=,\n predictor=ensemble_tta_predictor, \n overlap=,\n mode=\n )\n infer_end = time.time()\n ()\n\n \n post_start = time.time()\n out_probs = torch.softmax(out_logits, dim=) \n out_probs_np = out_probs[].cpu().numpy() \n reconstructed_mask = np.argmax(out_probs_np, axis=) \n post_end = time.time()\n ()\n\n \n aspect_ratio_thresholds = {\n : ,\n : ,\n : ,\n : ,\n : \n }\n\n cc_start = time.time()\n location = {}\n\n \n c classes:\n \n cc = cc3d.connected_components(reconstructed_mask == c)\n stats = cc3d.statistics(cc)\n\n \n centroids = np.array(stats[][:]) \n bbox_list = stats[][:] \n voxel_counts = np.array(stats[][:])\n\n valid_indices = []\n \n class_name = id_to_name[c]\n ar_threshold = aspect_ratio_thresholds.get(class_name, )\n\n i, bbox (bbox_list):\n \n \n slice_z, slice_y, slice_x = bbox \n depth = slice_z.stop - slice_z.start\n height = slice_y.stop - slice_y.start\n width = slice_x.stop - slice_x.start\n\n \n aspect_ratio = (width, height, depth) / (width, height, depth)\n\n \n voxel_counts[i] > BLOB_THRESHOLD aspect_ratio >= ar_threshold:\n valid_indices.append(i)\n\n (valid_indices) > :\n \n valid_centroids = centroids[valid_indices]\n \n valid_centroids = valid_centroids * \n \n valid_xyz = np.ascontiguousarray(valid_centroids[:, ::-])\n location[class_name] = valid_xyz\n\n cc_end = time.time()\n ()\n\n \n df = dict_to_df(location, run.name)\n location_df.append(df)\n\n run_end = time.time()\n ()\n\n\nlocation_df = pd.concat(location_df)\ntotal_end = time.time()
\n`
Yeah, that makes sense. Thanks!
", "rawMarkdown": "Yeah, that makes sense. Thanks!", "votes": null }, { "id": "3116945", "postDate": "02/06/2025 13:03:35", "content": "congratulation!
", "rawMarkdown": "congratulation!", "votes": null }, { "id": "3117064", "postDate": "02/06/2025 15:47:06", "content": "Congratulations! I also used 3DUnet model, but it didn`t do so well. Would you mind telling tips for hyperparameter tuning?
", "rawMarkdown": "Congratulations! I also used 3DUnet model, but it didn`t do so well. Would you mind telling tips for hyperparameter tuning?", "votes": null }, { "id": "3117080", "postDate": "02/06/2025 16:06:58", "content": "Similar to #8(https://www.kaggle.com/competitions/czii-cryo-et-object-identification/discussion/561515), i found dropout, batchnorm and channels to be key. The increase in channels affected my model, so I spent a lot of time optimizing them. If you look at my notes, you'll see that there is learning for many different channels.
\nhttps://www.kaggle.com/code/junhanzangai/submission-test/
", "rawMarkdown": "Similar to #8(https://www.kaggle.com/competitions/czii-cryo-et-object-identification/discussion/561515), i found dropout, batchnorm and channels to be key. The increase in channels affected my model, so I spent a lot of time optimizing them. If you look at my notes, you'll see that there is learning for many different channels.\n\nhttps://www.kaggle.com/code/junhanzangai/submission-test/", "votes": null }, { "id": "3117387", "postDate": "02/07/2025 00:03:21", "content": "thank you for great sharing!
", "rawMarkdown": "thank you for great sharing!", "votes": null }, { "id": "3179351", "postDate": "04/15/2025 10:50:27", "content": "Hello, Thank you for very informative sharing!!!
\nI'm studying YOLO now. If possible, could you share the train code of YOLO?
Oh, wow! I really like that filter idea!
", "votes": null, "replies": [ { "id": 3116442, "author_name": "junhanzangai", "author_url": "", "post_date": "02/06/2025 00:56:18", "content": "In my case, filtering id is main for scoring up.
", "votes": null, "replies": [ { "id": 3116449, "author_name": "davidlist", "author_url": "", "post_date": "02/06/2025 01:01:15", "content": "so if i understand correctly was that mostly getting rid of non-spherical particles? or rather particle candidates i suppose.
", "votes": null, "replies": [ { "id": 3116458, "author_name": "junhanzangai", "author_url": "", "post_date": "02/06/2025 01:14:56", "content": "Your understanding is partly correct but there's a nuance. The filtering mechanism isn't intended solely to remove non-spherical particles. Instead, it's designed to ignore any candidate cluster where the standard deviation along any coordinate exceeds 40% of the expected particle radius. This serves as a proxy for detecting clusters with overly dispersed predictions—which likely indicates unstable or false-positive detections—rather than a direct assessment of the particle’s inherent shape. In other words, it's more about ensuring the predicted candidate clusters are spatially compact and consistent with what you'd expect for a true particle detection.
\nAdditionally, I used the following code to filter out as many genuine particles as possible, but the score actually dropped.
\n`BLOB_THRESHOLD = 250
\nclasses = [1, 2, 3, 4, 5, 6]
\ntotal_start = time.time()
with torch.no_grad():
\n location_df = []
run root.runs:\n run_start = time.time() \n\n (run)\n\n \n load_start = time.time()\n tomo = run.get_voxel_spacing()\n tomo_arr = tomo.get_tomogram(tomo_type).numpy() \n load_end = time.time()\n ()\n\n \n prep_start = time.time()\n data_dict = [{: tomo_arr}]\n tomo_ds = CacheDataset(data=data_dict, transform=inference_transforms, cache_rate=)\n volume_tensor = tomo_ds[][].unsqueeze().to() \n prep_end = time.time()\n ()\n\n \n \n infer_start = time.time()\n out_logits = sliding_window_inference(\n inputs=volume_tensor,\n roi_size=(, , ),\n sw_batch_size=,\n predictor=ensemble_tta_predictor, \n overlap=,\n mode=\n )\n infer_end = time.time()\n ()\n\n \n post_start = time.time()\n out_probs = torch.softmax(out_logits, dim=) \n out_probs_np = out_probs[].cpu().numpy() \n reconstructed_mask = np.argmax(out_probs_np, axis=) \n post_end = time.time()\n ()\n\n \n aspect_ratio_thresholds = {\n : ,\n : ,\n : ,\n : ,\n : \n }\n\n cc_start = time.time()\n location = {}\n\n \n c classes:\n \n cc = cc3d.connected_components(reconstructed_mask == c)\n stats = cc3d.statistics(cc)\n\n \n centroids = np.array(stats[][:]) \n bbox_list = stats[][:] \n voxel_counts = np.array(stats[][:])\n\n valid_indices = []\n \n class_name = id_to_name[c]\n ar_threshold = aspect_ratio_thresholds.get(class_name, )\n\n i, bbox (bbox_list):\n \n \n slice_z, slice_y, slice_x = bbox \n depth = slice_z.stop - slice_z.start\n height = slice_y.stop - slice_y.start\n width = slice_x.stop - slice_x.start\n\n \n aspect_ratio = (width, height, depth) / (width, height, depth)\n\n \n voxel_counts[i] > BLOB_THRESHOLD aspect_ratio >= ar_threshold:\n valid_indices.append(i)\n\n (valid_indices) > :\n \n valid_centroids = centroids[valid_indices]\n \n valid_centroids = valid_centroids * \n \n valid_xyz = np.ascontiguousarray(valid_centroids[:, ::-])\n location[class_name] = valid_xyz\n\n cc_end = time.time()\n ()\n\n \n df = dict_to_df(location, run.name)\n location_df.append(df)\n\n run_end = time.time()\n ()\n\n\nlocation_df = pd.concat(location_df)\ntotal_end = time.time()
\n`
Yeah, that makes sense. Thanks!
", "votes": null, "replies": [] } ] } ] } ] } ] }, { "id": 3116425, "author_name": "snnclsr", "author_url": "", "post_date": "02/06/2025 00:29:45", "content": "Congratulations! Wow, I didn't expect YOLO to boost the scores that much. That's amazing!
", "votes": null, "replies": [ { "id": 3116443, "author_name": "junhanzangai", "author_url": "", "post_date": "02/06/2025 00:58:02", "content": "In fact, as I working with 3D, I realized that 2.5D could have a good impact.
", "votes": null, "replies": [] } ] }, { "id": 3116945, "author_name": "", "author_url": "", "post_date": "02/06/2025 13:03:35", "content": "congratulation!
", "votes": null, "replies": [] }, { "id": 3117064, "author_name": "ayaha0619", "author_url": "", "post_date": "02/06/2025 15:47:06", "content": "Congratulations! I also used 3DUnet model, but it didn`t do so well. Would you mind telling tips for hyperparameter tuning?
", "votes": null, "replies": [ { "id": 3117080, "author_name": "junhanzangai", "author_url": "", "post_date": "02/06/2025 16:06:58", "content": "Similar to #8(https://www.kaggle.com/competitions/czii-cryo-et-object-identification/discussion/561515), i found dropout, batchnorm and channels to be key. The increase in channels affected my model, so I spent a lot of time optimizing them. If you look at my notes, you'll see that there is learning for many different channels.
\nhttps://www.kaggle.com/code/junhanzangai/submission-test/
", "votes": null, "replies": [ { "id": 3117387, "author_name": "ayaha0619", "author_url": "", "post_date": "02/07/2025 00:03:21", "content": "thank you for great sharing!
", "votes": null, "replies": [] } ] } ] }, { "id": 3179351, "author_name": "taiki2", "author_url": "", "post_date": "04/15/2025 10:50:27", "content": "Hello, Thank you for very informative sharing!!!
\nI'm studying YOLO now. If possible, could you share the train code of YOLO?