CamVid Demo

This is a small demo showing that the functionality of misas is not at all limited to the domain of medical imaging. However, the main use case for the authors of misas is currently medical imaging, therefore this is just a proof of concept. It should give you an idea on how to work with a new dataset. It can be extended if users need it. The example builds on the camvid example in the fastai1 documentation.

Please reach out via GitHub if you have problems using misas on your own dataset.

from fastai.vision.all import *

path = untar_data(URLs.CAMVID_TINY)

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

array(['Animal', 'Archway', 'Bicyclist', 'Bridge', 'Building', 'Car',
       'CartLuggagePram', 'Child', 'Column_Pole', 'Fence', 'LaneMkgsDriv',
       'LaneMkgsNonDriv', 'Misc_Text', 'MotorcycleScooter', 'OtherMoving',
       'ParkingBlock', 'Pedestrian', 'Road', 'RoadShoulder', 'Sidewalk',
       'SignSymbol', 'Sky', 'SUVPickupTruck', 'TrafficCone',
       'TrafficLight', 'Train', 'Tree', 'Truck_Bus', 'Tunnel',
       'VegetationMisc', 'Void', 'Wall'], dtype='<U17')

fnames = get_image_files(path/"images")

def label_func(fn): return path/"labels"/f"{fn.stem}_P{fn.suffix}"

cam_fn = fnames[0]
mask_fn = label_func(fnames[0])

cam_img = lambda: Image.open(cam_fn).convert("RGB")
mask = lambda: Image.open(mask_fn).convert("I")

dls = SegmentationDataLoaders.from_label_func(
    path, bs=8, fnames = fnames, label_func = label_func, codes = codes
)

dls.show_batch(max_n=6)

learn = unet_learner(dls, resnet34)

Downloading: "https://download.pytorch.org/models/resnet34-b627a593.pth" to /home/markus/.cache/torch/hub/checkpoints/resnet34-b627a593.pth

learn.fine_tune(6)

epoch	train_loss	valid_loss	time
0	3.004354	2.548569	00:30

epoch	train_loss	valid_loss	time
0	2.048129	1.562803	00:37
1	1.704228	1.354810	00:33
2	1.483399	1.031199	00:30
3	1.315127	0.921900	00:27
4	1.164001	0.827559	00:33
5	1.051317	0.808568	00:32

learn.show_results(max_n=6, figsize=(7,8))

interp = SegmentationInterpretation.from_learner(learn)
interp.plot_top_losses(k=3)

camvid = DataBlock(blocks=(ImageBlock, MaskBlock(codes)),
                   get_items = get_image_files,
                   get_y = label_func,
                   splitter=RandomSplitter(),
                   batch_tfms=aug_transforms(size=(120,160)))

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

dls.show_batch(max_n=6)

# learn.save('mini_train')

plt.imshow(cam_img().resize((128,128)))

<matplotlib.image.AxesImage>

from misas.core import *
from misas.core import default_cmap

learn.prepareSize = lambda item: item.resize((128,128))

class Fastai_model:
    def __init__(self, learner):
        self.trainedModel = learner
        self.resize128 = lambda x: x.resize ((128,128))
        self.trainedModel.remove_cbs(ProgressCallback)

        
    def prepareSize(self, item):
        return self.resize128(item)
    
    def predict(self, image):
        image = PILImage.create(np.array(image))
        output = self.trainedModel.predict(image)
        output = PILImage.create(output [0])
        output = Image.fromarray(np.array(output)) #mode="I"
        return output

Cam_vid = Fastai_model(learn)

plot_series(get_rotation_series(cam_img(), Cam_vid), vmax=31, vmin=0)

plot_series(get_zoom_series(cam_img(), Cam_vid), vmax=31, vmin=0, nrow=2)

found_classes = np.unique(np.array(Cam_vid.predict(cam_img())))

codes[found_classes]

array(['Building', 'Car', 'LaneMkgsDriv', 'Road', 'Sidewalk', 'Sky',
       'Tree', 'Void', 'Wall'], dtype='<U17')

result = eval_rotation_series(cam_img(),mask(),Cam_vid,components=codes)

When plotting the evaluation series, it makes sense to only plot classes that actually occur.

plot_eval_series(result[np.append("deg",codes[found_classes])])