Wondering if anyone has found any useful augmentations that they would be willing to share. Only 1 augmentation has worked for me so far, but there must be more!
\nWorked
\nDidnt Work
\nAny and all ideas are welcome π
", "messages": [ { "id": 3209493, "postDate": "2025-05-25T23:24:35.160Z", "content": "Wondering if anyone has found any useful augmentations that they would be willing to share. Only 1 augmentation has worked for me so far, but there must be more!
\nWorked
\nDidnt Work
\nAny and all ideas are welcome π
", "rawMarkdown": "Wondering if anyone has found any useful augmentations that they would be willing to share. Only 1 augmentation has worked for me so far, but there must be more! \n\n**Worked**\n- Sensor flipping\n\n**Didnt Work**\n- Mixup (within class)\n- Mixup (between class)\n- Sensor shifting (w/ masked loss)\n- CoarseDropout\n- GaussianNoise\n\n\nAny and all ideas are welcome π\n", "votes": 18 }, { "id": 3209624, "postDate": "2025-05-26T05:19:48.737Z", "content": "Great discussion! For geophysical data, have you tried physics-informed augmentations?
\nTime-domain approaches:
\nFrequency-domain:
\nThe key insight from @Doomsday is spot-on - this isn't typical image data, so physics-based augmentations might work better than standard computer vision techniques. Worth testing small velocity perturbations (Β±5-10%) as they're geologically realistic! π
", "rawMarkdown": "Great discussion! For geophysical data, have you tried **physics-informed augmentations**? \n\n**Time-domain approaches:**\n- Velocity scaling (simulates different subsurface velocities)\n- Amplitude scaling with frequency-dependent attenuation\n- Phase shifts to simulate different source positions\n\n**Frequency-domain:**\n- Bandpass filtering to simulate different acquisition parameters\n- Spectral whitening for noise robustness\n\nThe key insight from @Doomsday is spot-on - this isn't typical image data, so physics-based augmentations might work better than standard computer vision techniques. Worth testing small velocity perturbations (Β±5-10%) as they're geologically realistic! π", "votes": 3 }, { "id": 3210615, "postDate": "2025-05-27T13:01:48.443Z", "content": "Not really considered augmentation, but have has anyone tried deep-supervision?
\nOn most vision segmentation models it does tend to significantly improve CV scores and generalisation, but like Doomsday mentioned, this is not really a classic vision problem, so unsure if it would work.
I applied some pruning and intensity augmentation , but based on the current results, it seems that this might be detrimental to training (though I haven't run many epochs yet). As for mixup, I haven't conducted any controlled experiments; I directly integrated it into my pipeline.
", "rawMarkdown": "I applied some pruning and intensity augmentation , but based on the current results, it seems that this might be detrimental to training (though I haven't run many epochs yet). As for mixup, I haven't conducted any controlled experiments; I directly integrated it into my pipeline. ", "votes": 1 }, { "id": 3221855, "postDate": "2025-06-11T14:24:09.280Z", "content": "The source locations are indexed as 0, 17, 34, 52, 69 along the x-direction, but simply reversing these indices would result in 69, 52, 35, 17, 0, creating a discrepancy between indices 35 and 34. To correct this, instead of simply reversing the indices, you could perform a forward simulation on the reversed velocity map using methods like those implemented in https://www.kaggle.com/code/jaewook704/waveform-inversion-vel-to-seis. While data generation for this takes considerable time, it is significantly more accurate than inference processes in Deepwave.
", "rawMarkdown": "The source locations are indexed as 0, 17, 34, 52, 69 along the x-direction, but simply reversing these indices would result in 69, 52, 35, 17, 0, creating a discrepancy between indices 35 and 34. To correct this, instead of simply reversing the indices, you could perform a forward simulation on the reversed velocity map using methods like those implemented in https://www.kaggle.com/code/jaewook704/waveform-inversion-vel-to-seis. While data generation for this takes considerable time, it is significantly more accurate than inference processes in Deepwave.\n", "votes": 2 }, { "id": 3210600, "postDate": "2025-05-27T12:40:07.137Z", "content": "I tried summing the sensor inputs into one channel and concatenating to the original input so instead of shape (B, 5, 1000, 70), I had (B, 6, 1000, 70). The extra channel is supposed to be the simulation of all sources firing at the same time. I think it's physically plausible to add them like that since it's in time domain, I'm not super sure.
Anyways, it didn't make much of a difference on initial experiments with CurveFault_B-only training.
", "rawMarkdown": "I tried summing the sensor inputs into one channel and concatenating to the original input so instead of shape `(B, 5, 1000, 70)`, I had `(B, 6, 1000, 70)`. The extra channel is supposed to be the simulation of all sources firing at the same time. I think it's physically plausible to add them like that since it's in time domain, I'm not super sure.\n\nAnyways, it didn't make much of a difference on initial experiments with CurveFault_B-only training.", "votes": 2 }, { "id": 3209701, "postDate": "2025-05-26T07:31:41.413Z", "content": "Worked
\nDidn't work
\nLike what exactly are you referring to with sensor flipping ?? , are we interchanging two sensors or receivers ??
", "rawMarkdown": "Like what exactly are you referring to with sensor flipping ?? , are we interchanging two sensors or receivers ?? ", "isDeleted": true } ] }, { "id": 3209569, "postDate": "2025-05-26T03:49:25.853Z", "content": "Yes tried all of them .. didn't help for full run with 72x72 and didn't help with few epoch run with full size.
\nIs it because it's not like usual segmentation problem where the more variety of images you show ,it becomes robust to overfitting .. here it's mostly like a a transformation based on equation ?
", "rawMarkdown": "Yes tried all of them .. didn't help for full run with 72x72 and didn't help with few epoch run with full size. \n\nIs it because it's not like usual segmentation problem where the more variety of images you show ,it becomes robust to overfitting .. here it's mostly like a a transformation based on equation ?", "votes": 2, "replies": [ { "id": 3209582, "postDate": "2025-05-26T04:06:52.507Z", "content": "Good point. Maybe we need to get a bit more creative to make something work here π€
", "rawMarkdown": "Good point. Maybe we need to get a bit more creative to make something work here π€", "votes": 2, "replies": [ { "id": 3209600, "postDate": "2025-05-26T04:29:51.393Z", "content": "I used a very low probability of mixup +cutmix .. and the more I increased the probability the worser it got . However ,one disclaimer , mixup and cut mix needsheavy models and sometimes it shows improvement when you have a long run to get better minima .. I have not tested it for that long run.
\nDisclaimer 2: earlier comp mixup /cutmix /Aug helped when the training loss becomes lower and validation loss not so much.. which means overfitting zone .. here I have seen train and validation loss almost go hand in hand ..I have not reached a overfitting zone where let's say train loss is 10 and valid loss is 20 . So not sure if this Augs would help with the current state of my models and training
", "rawMarkdown": "I used a very low probability of mixup +cutmix .. and the more I increased the probability the worser it got . However ,one disclaimer , mixup and cut mix needsheavy models and sometimes it shows improvement when you have a long run to get better minima .. I have not tested it for that long run. \n\nDisclaimer 2: earlier comp mixup /cutmix /Aug helped when the training loss becomes lower and validation loss not so much.. which means overfitting zone .. here I have seen train and validation loss almost go hand in hand ..I have not reached a overfitting zone where let's say train loss is 10 and valid loss is 20 . So not sure if this Augs would help with the current state of my models and training ", "votes": 1 } ] } ] }, { "id": 3209503, "postDate": "2025-05-26T00:01:58.893Z", "content": "How much flipping help?
\nAs a TTA it did not help me at all.
It looks to be helpful in training against overfitting.
", "rawMarkdown": "It looks to be helpful in training against overfitting.", "votes": 1, "replies": [ { "id": 3209509, "postDate": "2025-05-26T00:21:46.373Z", "content": "You did ablation study and have a score with/without to compare?
\nWell I'll probably just have to try it myself and seeβ¦
Here are the validation scores with and without TTA. The models evaluated here were trained with sensor flipping.
\n| TTA | \nValidation MAE | \n
|---|---|
| Yes | \n28.95 | \n
| No | \n29.38 | \n
@brendanartley thx but this does not answer how much it help compared to training without flipping. It's obvious flipping TTA would help if the model was also trained with flipping.
", "rawMarkdown": "@brendanartley thx but this does not answer how much it help compared to training without flipping. It's obvious flipping TTA would help if the model was also trained with flipping.", "votes": -2 }, { "id": 3209714, "postDate": "2025-05-26T07:49:16.580Z", "content": "@shlomoron
\nI think I did an experiment a while back
\nTraining with flip and validating with flip-TTA helps (obviously as you pointed out)
\nTraining without flip and using flip-TTA in validation hurts performance (around -0.01)
Flipping help me too. Both training&tta.
", "rawMarkdown": "Flipping help me too. Both training&tta.", "votes": 2 } ] } ], "comments": [ { "id": 3209624, "author_name": "Harshith Vaddiparthy", "author_url": "", "post_date": "2025-05-26T05:19:48.737000", "content": "Great discussion! For geophysical data, have you tried physics-informed augmentations?
\nTime-domain approaches:
\nFrequency-domain:
\nThe key insight from @Doomsday is spot-on - this isn't typical image data, so physics-based augmentations might work better than standard computer vision techniques. Worth testing small velocity perturbations (Β±5-10%) as they're geologically realistic! π
", "votes": 3, "replies": [] }, { "id": 3210615, "author_name": "NMVR", "author_url": "", "post_date": "2025-05-27T13:01:48.443000", "content": "Not really considered augmentation, but have has anyone tried deep-supervision?
\nOn most vision segmentation models it does tend to significantly improve CV scores and generalisation, but like Doomsday mentioned, this is not really a classic vision problem, so unsure if it would work.
I applied some pruning and intensity augmentation , but based on the current results, it seems that this might be detrimental to training (though I haven't run many epochs yet). As for mixup, I haven't conducted any controlled experiments; I directly integrated it into my pipeline.
", "votes": 1, "replies": [] }, { "id": 3221855, "author_name": "Harui-ig", "author_url": "", "post_date": "2025-06-11T14:24:09.280000", "content": "The source locations are indexed as 0, 17, 34, 52, 69 along the x-direction, but simply reversing these indices would result in 69, 52, 35, 17, 0, creating a discrepancy between indices 35 and 34. To correct this, instead of simply reversing the indices, you could perform a forward simulation on the reversed velocity map using methods like those implemented in https://www.kaggle.com/code/jaewook704/waveform-inversion-vel-to-seis. While data generation for this takes considerable time, it is significantly more accurate than inference processes in Deepwave.
", "votes": 2, "replies": [] }, { "id": 3210600, "author_name": "MGΓΆksu", "author_url": "", "post_date": "2025-05-27T12:40:07.137000", "content": "I tried summing the sensor inputs into one channel and concatenating to the original input so instead of shape (B, 5, 1000, 70), I had (B, 6, 1000, 70). The extra channel is supposed to be the simulation of all sources firing at the same time. I think it's physically plausible to add them like that since it's in time domain, I'm not super sure.
Anyways, it didn't make much of a difference on initial experiments with CurveFault_B-only training.
", "votes": 2, "replies": [] }, { "id": 3209701, "author_name": "SeshuRaju π§ββοΈ", "author_url": "", "post_date": "2025-05-26T07:31:41.413000", "content": "Worked
\nDidn't work
\nLike what exactly are you referring to with sensor flipping ?? , are we interchanging two sensors or receivers ??
", "votes": 0, "replies": [] } ] }, { "id": 3209569, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2025-05-26T03:49:25.853000", "content": "Yes tried all of them .. didn't help for full run with 72x72 and didn't help with few epoch run with full size.
\nIs it because it's not like usual segmentation problem where the more variety of images you show ,it becomes robust to overfitting .. here it's mostly like a a transformation based on equation ?
", "votes": 2, "replies": [ { "id": 3209582, "author_name": "Bartley", "author_url": "", "post_date": "2025-05-26T04:06:52.507000", "content": "Good point. Maybe we need to get a bit more creative to make something work here π€
", "votes": 2, "replies": [ { "id": 3209600, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2025-05-26T04:29:51.393000", "content": "I used a very low probability of mixup +cutmix .. and the more I increased the probability the worser it got . However ,one disclaimer , mixup and cut mix needsheavy models and sometimes it shows improvement when you have a long run to get better minima .. I have not tested it for that long run.
\nDisclaimer 2: earlier comp mixup /cutmix /Aug helped when the training loss becomes lower and validation loss not so much.. which means overfitting zone .. here I have seen train and validation loss almost go hand in hand ..I have not reached a overfitting zone where let's say train loss is 10 and valid loss is 20 . So not sure if this Augs would help with the current state of my models and training
", "votes": 1, "replies": [] } ] } ] }, { "id": 3209503, "author_name": "greySnow", "author_url": "", "post_date": "2025-05-26T00:01:58.893000", "content": "How much flipping help?
\nAs a TTA it did not help me at all.
It looks to be helpful in training against overfitting.
", "votes": 1, "replies": [ { "id": 3209509, "author_name": "greySnow", "author_url": "", "post_date": "2025-05-26T00:21:46.373000", "content": "You did ablation study and have a score with/without to compare?
\nWell I'll probably just have to try it myself and seeβ¦
Here are the validation scores with and without TTA. The models evaluated here were trained with sensor flipping.
\n| TTA | \nValidation MAE | \n
|---|---|
| Yes | \n28.95 | \n
| No | \n29.38 | \n
@brendanartley thx but this does not answer how much it help compared to training without flipping. It's obvious flipping TTA would help if the model was also trained with flipping.
", "votes": -2, "replies": [] }, { "id": 3209714, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2025-05-26T07:49:16.580000", "content": "@shlomoron
\nI think I did an experiment a while back
\nTraining with flip and validating with flip-TTA helps (obviously as you pointed out)
\nTraining without flip and using flip-TTA in validation hurts performance (around -0.01)
Flipping help me too. Both training&tta.
", "votes": 2, "replies": [] } ] } ], "raw_markdown_by_id": { "3209493": "Wondering if anyone has found any useful augmentations that they would be willing to share. Only 1 augmentation has worked for me so far, but there must be more! \n\n**Worked**\n- Sensor flipping\n\n**Didnt Work**\n- Mixup (within class)\n- Mixup (between class)\n- Sensor shifting (w/ masked loss)\n- CoarseDropout\n- GaussianNoise\n\n\nAny and all ideas are welcome π\n", "3209624": "Great discussion! For geophysical data, have you tried **physics-informed augmentations**? \n\n**Time-domain approaches:**\n- Velocity scaling (simulates different subsurface velocities)\n- Amplitude scaling with frequency-dependent attenuation\n- Phase shifts to simulate different source positions\n\n**Frequency-domain:**\n- Bandpass filtering to simulate different acquisition parameters\n- Spectral whitening for noise robustness\n\nThe key insight from @Doomsday is spot-on - this isn't typical image data, so physics-based augmentations might work better than standard computer vision techniques. Worth testing small velocity perturbations (Β±5-10%) as they're geologically realistic! π", "3210615": "Not really considered augmentation, but have has anyone tried deep-supervision?\nOn most vision segmentation models it does tend to significantly improve CV scores and generalisation, but like Doomsday mentioned, this is not really a classic vision problem, so unsure if it would work.", "3209958": "I applied some pruning and intensity augmentation , but based on the current results, it seems that this might be detrimental to training (though I haven't run many epochs yet). As for mixup, I haven't conducted any controlled experiments; I directly integrated it into my pipeline. ", "3221855": "The source locations are indexed as 0, 17, 34, 52, 69 along the x-direction, but simply reversing these indices would result in 69, 52, 35, 17, 0, creating a discrepancy between indices 35 and 34. To correct this, instead of simply reversing the indices, you could perform a forward simulation on the reversed velocity map using methods like those implemented in https://www.kaggle.com/code/jaewook704/waveform-inversion-vel-to-seis. While data generation for this takes considerable time, it is significantly more accurate than inference processes in Deepwave.\n", "3210600": "I tried summing the sensor inputs into one channel and concatenating to the original input so instead of shape `(B, 5, 1000, 70)`, I had `(B, 6, 1000, 70)`. The extra channel is supposed to be the simulation of all sources firing at the same time. I think it's physically plausible to add them like that since it's in time domain, I'm not super sure.\n\nAnyways, it didn't make much of a difference on initial experiments with CurveFault_B-only training.", "3209701": "**Worked**\n- Sensor Flipping ( Horizontal )\n\n**Didn't work**\n- GaussianNoise", "3209569": "Yes tried all of them .. didn't help for full run with 72x72 and didn't help with few epoch run with full size. \n\nIs it because it's not like usual segmentation problem where the more variety of images you show ,it becomes robust to overfitting .. here it's mostly like a a transformation based on equation ?", "3209503": "How much flipping help?\nAs a TTA it did not help me at all." } }