I am using the following code to perform progressive resizing for a semantic segmentation problem using UNET.
Problem:
I don’t think the resizing part actually works because the time that each epoch takes to train for learn_224 is exactly the same as that of learn_112. When create a learn_224 using unet_learn() each epoch takes much longer to train.
Am I doing this incorrectly?
size = 460 // 2
half= (224 //2, 224 // 2)
dblock_112 = DataBlock(blocks=(ImageBlock, MaskBlock(codes)),
get_items=get_relevant_images,
splitter=splitter,
get_y=get_mask,
item_tfms=Resize((size,size)),
batch_tfms=[*aug_transforms(size=half), Normalize.from_stats(*imagenet_stats)])
dls_112 = dblock_2.dataloaders(path/'images', bs=16)
learn_112 = unet_learner(dls_112, resnet34,
cbs=callbacks,
self_attention=True,
metrics=[Dice()]
)
learn_112.fine_tune(50)
# Let's trinf using double the size (224)
size = 460
half= (224, 224 )
dblock_2 = DataBlock(blocks=(ImageBlock, MaskBlock(codes)),
get_items=get_relevant_images,
splitter=splitter,
get_y=get_mask,
item_tfms=Resize((size,size)),
batch_tfms=[*aug_transforms(size=half), Normalize.from_stats(*imagenet_stats)])
dls_224 = dblock_2.dataloaders(path/'images', bs=16)
learn_224 = learn_112
learn_224.data = dls_224
x = next(iter(learn_224.data.train))
x[0].size() # prints (as expected): torch.Size([16, 3, 224, 224])
learn_224.unfreeze()
learn_224.fit_one_cycle(30, lr_max=slice(1e-4, 1e-3))
```