Cats&Dogs model improvement #254
Conversation
ermanok
commented
Sep 19, 2023
•
ermanok
commented
- Dataloader changed to augment data randomly at each epoch
- Training & QAT policies updated for optimized model training
Augmentation approach fixed. Random augmentations added to be applied dynamically while loading the data samples.
Old dataloader removed. Linter errors fixed.
Linter fixes
policies/qat_policy_late_cd.yaml
Outdated
| @@ -0,0 +1,4 @@ | |||
| --- | |||
| start_epoch: 30 | |||
There was a problem hiding this comment.
It is not very late, considering the 200 epochs. We might want to drop "late" from the filename.
| album.ShiftScaleRotate(shift_limit=0.05, scale_limit=0.05, rotate_limit=15, p=0.5), | ||
| album.RandomCrop(height=resize_size[0], width=resize_size[1]), | ||
| album.HorizontalFlip(p=0.5), | ||
| album.Normalize(mean=(0.0, 0.0, 0.0), std=(1.0, 1.0, 1.0))]) |
There was a problem hiding this comment.
One question: Doesn't this normalize all images & all channels independently? It is different than dividing the entire dataset by 255, which we employ in the other vision examples. Whichever correct way is needed to be applied to all examples.
There was a problem hiding this comment.
Both approaches are very similar to each other. The previous appraoch divides the dataset by 255 while this one divides with the maximum value of the sample. Both should provide similar results and it requires to make extensive experiments to understand if one is correct.
There was a problem hiding this comment.
That's correct. I assumed that it divides each channel with its own maximum value. Doublechecked the documentation and it divides with max possible, which is 255.
| album.ShiftScaleRotate(shift_limit=0.05, scale_limit=0.05, rotate_limit=15, p=0.5), | ||
| album.RandomCrop(height=resize_size[0], width=resize_size[1]), | ||
| album.HorizontalFlip(p=0.5), | ||
| album.Normalize(mean=(0.0, 0.0, 0.0), std=(1.0, 1.0, 1.0))]) |
There was a problem hiding this comment.
That's correct. I assumed that it divides each channel with its own maximum value. Doublechecked the documentation and it divides with max possible, which is 255.
scripts/train_catsdogs.sh
Outdated
| @@ -1,2 +1,2 @@ | |||
| #!/bin/sh | |||
| python train.py --epochs 250 --optimizer Adam --lr 0.001 --wd 0 --deterministic --compress policies/schedule-catsdogs.yaml --model ai85cdnet --dataset cats_vs_dogs --confusion --param-hist --embedding --device MAX78000 "$@" | |||
| python train.py --epochs 200 --optimizer Adam --lr 0.001 --wd 0 --deterministic --compress policies/schedule-catsdogs.yaml --qat-policy policies/qat_policy_late_cd.yaml --model ai85cdnet --dataset cats_vs_dogs --confusion --param-hist --embedding --device MAX78000 "$@" | |||
There was a problem hiding this comment.
Seems that "qat_policy_late_cd.yaml" should be replaced with "qat_policy_cd.yaml".
There was a problem hiding this comment.
Correct and fixed...
fixes the script after changing qat policy filename changed
