cm[0][1]) + 50.0 * cm[1][1] / (
cm[1][1] + cm[1][0] + cm[0][1])
cm = {}
with torch.no_grad():
for town in miniworld.towns:
print(town)
cm[town] = np.zeros((3, 3), dtype=int)
for i in range(miniworld.data[town].nbImages):
imageraw, label = miniworld.data[town].getImageAndLabel(i)
if False:
# 100%
pred = label.copy()
label = dataloader.convertIn3classNP(label)
if False:
# 100%
pred = dataloader.convertIn3classNP(label).astype(int)
pred = np.abs(pred - 1)
pred = np.uint8(np.abs(1 - pred))
label = dataloader.convertIn3classNP(label)
if True:
# ONLY AROUND 90%
pred = dataloader.convertIn3classNP(label).astype(int)
pred = np.abs(pred - 1)
pred = np.uint8(np.abs(1 - pred))
assert label.shape == pred.shape
(image.shape[2], image.shape[3]))
power2resize = torch.nn.AdaptiveAvgPool2d(
((image.shape[2] // 64) * 64, (image.shape[3] // 64) * 64))
image = power2resize(image)
if image.shape[2] < 128 and image.shape[3] < 128:
print("can not handle such small image")
quit()
else:
pred = dataloader.largeforward(net, image, device)
pred = globalresize(pred)
_, pred = torch.max(pred[0], 0)
pred = pred.cpu().numpy()
label = dataloader.convertIn3classNP(label)
assert label.shape == pred.shape
cm[town] += confusion_matrix(label.flatten(),
pred.flatten(),
labels=[0, 1, 2])
if town in ["toulouse/test"]:
image = np.transpose(image[0].cpu().numpy(), axes=(1, 2, 0))
im = PIL.Image.fromarray(np.uint8(image))
im.save("build/" + town[0:-5] + "_" + str(i) + "_x.png")
labelim = PIL.Image.fromarray(np.uint8(label) * 125)
labelim.save("build/" + town[0:-5] + "_" + str(i) + "_y.png")
predim = PIL.Image.fromarray(np.uint8(pred) * 125)
predim.save("build/" + town[0:-5] + "_" + str(i) + "_z.png")