def test_torchify():
# We take the fitst 8 blocks of 256 frames from video '01.00'.
# Video '01.00' has 2250 frames at a pixel resolution of 512x512.
x, y, meta = data.load_dataset('./data', n=8, frames=256, subset=['01.00'])
ds = data.Torchify(x, y)
assert len(ds) == 8
# We can successfully iterate over the dataset.
for i in range(len(ds)):
x, y = ds[i]
assert x.shape == (256, 512, 512) # 256 frames per datum
assert y.shape == (512, 512)
assert x.dtype == 'float64'
assert y.dtype == 'int64'
assert isinstance(x, np.ndarray)
assert isinstance(y, np.ndarray)
# We can successfully iterate over the dataset using a dataloader.
loader = torch.utils.data.DataLoader(ds, batch_size=4)
for batch in loader:
x, y = batch
assert x.shape == (4, 256, 512, 512
) # batch of 4, 256 frames per datum
assert y.shape == (4, 512, 512) # 2 classes, background and foreground
assert isinstance(x, torch.DoubleTensor)
assert isinstance(y, torch.LongTensor)
def property_nmf():
output = []
for vid in [TEST_SET]:
tempdict = {}
video, y, meta = data.load_dataset(base_path='./data', subset=[vid], preprocess=preprocess.ed_open, frames=1000)
rois = [{"coordinates": [list(k) for k in i]} for i in nmf.nmf_extraction(np.dstack(video[0]), k=10)]
tempdict["dataset"]=vid
tempdict["regions"]=rois
output.append(tempdict)
str_out = json.dumps(output, default=lambda x: int(x))
print(str_out)
def test_segmentation():
torch.set_default_tensor_type('torch.DoubleTensor')
net_ctor = lambda: UNet(5, 2, depth=3, size=32)
loss = F.cross_entropy
x, y, meta = load_dataset('./data', subset=['01.00'], n=5, stop=1)
x = [a.compute() for a in x]
model = TorchSegmenter(net_ctor, loss)
model.fit(x, y, epochs=1)
model.predict(x)
model.score(x, y)
def test_dataset():
# We take the fitst 8 blocks of 256 frames from video '01.00'.
# Video '01.00' has 2250 frames at a pixel resolution of 512x512.
x, y, meta = data.load_dataset('./data',
n=8,
frames=256,
skip=4,
subset=['01.00'])
assert len(x) == len(y) == len(meta) == 8
# We can successfully iterate over the dataset.
for i in range(8):
assert x[i].shape == (256, 512, 512) # 256 frames per datum
assert x[i].dtype == 'float64'
assert y[i].shape == (512, 512)
assert y[i].dtype == 'int64'
assert isinstance(x[i], da.Array)
assert isinstance(y[i], np.ndarray)
assert isinstance(meta[i], dict)
def test_gridsearch():
from sklearn.base import BaseEstimator
from sklearn.model_selection import GridSearchCV
# Grid search over a stub estimator with one dummy param
class StubEstimator(BaseEstimator):
def __init__(self, foo=1):
self.foo = foo
def score(self, x, y):
return self.foo
def fit(self, x, y):
assert isinstance(x[0], da.Array)
assert isinstance(y[0], np.ndarray)
return self
# The dataset should be compatible with GridSearchCV
x, y, meta = data.load_dataset('./data', n=8, frames=256, subset=['01.00'])
stub = StubEstimator(foo=1)
grid_stub = GridSearchCV(stub, {'foo': [1, 2, 3]})
grid_stub.fit(x, y)