Image Segmentation with CAMVID

A basic segmentation example for CAMVID

import torch

from fastai.basics import *
from fastai.callback.all import *
from fastai.vision.all import *

path = untar_data(URLs.CAMVID)

valid_fnames = (path/'valid.txt').read_text().split('\n')

def ListSplitter(valid_items):
    def _inner(items):
        val_mask = tensor([o.name in valid_items for o in items])
        return [~val_mask,val_mask]
    return _inner

codes = np.loadtxt(path/'codes.txt', dtype=str)

camvid = DataBlock(blocks=(ImageBlock, MaskBlock(codes)),
                   get_items=get_image_files,
                   splitter=ListSplitter(valid_fnames),
                   get_y=lambda o: path/'labels'/f'{o.stem}_P{o.suffix}',
                   batch_tfms=[*aug_transforms(size=(360,480)), Normalize.from_stats(*imagenet_stats)])

dls = camvid.dataloaders(path/"images", bs=8)

/home/jhoward/miniconda3/lib/python3.8/site-packages/torch/_tensor.py:575: UserWarning: floor_divide is deprecated, and will be removed in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values.
To keep the current behavior, use torch.div(a, b, rounding_mode='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode='floor'). (Triggered internally at  /opt/conda/conda-bld/pytorch_1623448278899/work/aten/src/ATen/native/BinaryOps.cpp:467.)
  return torch.floor_divide(self, other)
/home/jhoward/miniconda3/lib/python3.8/site-packages/torch/_tensor.py:1023: UserWarning: torch.solve is deprecated in favor of torch.linalg.solveand will be removed in a future PyTorch release.
torch.linalg.solve has its arguments reversed and does not return the LU factorization.
To get the LU factorization see torch.lu, which can be used with torch.lu_solve or torch.lu_unpack.
X = torch.solve(B, A).solution
should be replaced with
X = torch.linalg.solve(A, B) (Triggered internally at  /opt/conda/conda-bld/pytorch_1623448278899/work/aten/src/ATen/native/BatchLinearAlgebra.cpp:760.)
  ret = func(*args, **kwargs)

dls = SegmentationDataLoaders.from_label_func(path, bs=8,
    fnames = get_image_files(path/"images"), 
    label_func = lambda o: path/'labels'/f'{o.stem}_P{o.suffix}',                                     
    codes = codes,                         
    batch_tfms=[*aug_transforms(size=(360,480)), Normalize.from_stats(*imagenet_stats)])

dls.show_batch(max_n=2, figsize=(20, 7))

codes = np.loadtxt(path/'codes.txt', dtype=str)
dls.vocab = codes

name2id = {v:k for k,v in enumerate(codes)}
void_code = name2id['Void']

def acc_camvid(input, target):
    target = target.squeeze(1)
    mask = target != void_code
    return (input.argmax(dim=1)[mask]==target[mask]).float().mean()

opt_func = partial(Adam, lr=3e-3, wd=0.01)#, eps=1e-8)

learn = unet_learner(dls, resnet34, loss_func=CrossEntropyLossFlat(axis=1), opt_func=opt_func, path=path, metrics=acc_camvid,
                     norm_type=None, wd_bn_bias=True)

/home/jhoward/miniconda3/lib/python3.8/site-packages/torch/nn/functional.py:718: UserWarning: Named tensors and all their associated APIs are an experimental feature and subject to change. Please do not use them for anything important until they are released as stable. (Triggered internally at  /opt/conda/conda-bld/pytorch_1623448278899/work/c10/core/TensorImpl.h:1156.)
  return torch.max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode)

get_c(dls)

learn.lr_find()

SuggestedLRs(valley=0.00010964782268274575)

lr= 3e-3
learn.freeze()

learn.fit_one_cycle(10, slice(lr), pct_start=0.9, wd=1e-2)

epoch	train_loss	valid_loss	acc_camvid	time
0	1.200469	0.869627	0.769983	00:57
1	0.840649	0.809244	0.776909	00:47
2	0.716685	0.638332	0.838415	00:47
3	0.670508	0.551083	0.851559	00:47
4	0.664709	0.588863	0.849711	00:47
5	0.603191	0.502482	0.867659	00:47
6	0.592773	0.507730	0.869631	00:47
7	0.541870	0.540163	0.863005	00:47
8	0.531527	0.429516	0.878525	00:47
9	0.463456	0.345390	0.900292	00:47

learn.show_results(max_n=2, rows=2, vmin=1, vmax=30, figsize=(14, 10))

learn.save('stage-1')

learn.load('stage-1')
learn.unfreeze()

#learn.opt.clear_state() #Not necessarily useful

lrs = slice(lr/400,lr/4)

learn.fit_one_cycle(12, lrs, pct_start=0.8, wd=1e-2)

epoch	train_loss	valid_loss	acc_camvid	time
0	0.415170	0.350871	0.897328	00:42
1	0.405012	0.341905	0.899924	00:42
2	0.400426	0.330662	0.904413	00:42
3	0.385431	0.329282	0.904444	00:42
4	0.372985	0.322414	0.912512	00:42
5	0.366623	0.306477	0.916740	00:42
6	0.362156	0.298581	0.913030	00:42
7	0.343045	0.290931	0.919178	00:42
8	0.327369	0.295092	0.921611	00:42
9	0.334783	0.280629	0.922483	00:42
10	0.295628	0.260418	0.929844	00:42
11	0.269825	0.260967	0.928652	00:42

learn.show_results(max_n=4, vmin=1, vmax=30, figsize=(15,6))