Same code behave diffrently

vision
1

1111111
1111111994×293 113 KB

Hey guys, i am using this repo GitHub - hamdaan19/UNet-Multiclass: This repository contains code used to train U-Net on a multi-class segmentation dataset..

Evaluate function in this repo producing the second image, which is what i want.

But i need a real time function so i wrote this:

model = UNET(in_channels=3, classes=19)
model.state_dict(destination=torch.load("D:/ide/Saves/RedHorse/env/HorseNet/NN/Static/NNmodels/2_Segmentation_Model_Segmentation_model.pth", map_location="cpu"))
model.eval()

IMAGE_HEIGHT, IMAGE_WIDTH = 300, 600

x = cv2.imread("D:/ide/Saves/RedHorse/env/HorseNet/NN/extracted_data/Cityspaces.zip/images/val/frankfurt/frankfurt_000000_015389_leftImg8bit.png")

x = cv2.resize(x, (IMAGE_WIDTH,IMAGE_HEIGHT))
x = cv2.cvtColor(x, cv2.COLOR_BGR2RGB)

x = torch.from_numpy(x).float()
x = x.permute(2,0,1)
x = x.unsqueeze(0)

with torch.no_grad():
    y = model(x)

y = torch.nn.functional.softmax(y, dim=1)
y = torch.argmax(y, dim=1).float()

y.apply_(lambda x: t2l[x].id)

y = y.numpy()

y = y.transpose(1,2,0)

plt.imshow(y)
plt.show()

cv2.imwrite("D:/Downloads/UNet-Multiclass/saved_images/.png",y)

and output of this code is first image. I really don’t understand is this a problem or not. If its not what should i do to get the outputs like second one?

2

I guess you might be missing some preprocessing steps such as normalizing the image and I would generally try to compare the intermediate outputs between the original repository output and your “real time” code.

3

Thank you for yours answer but i already try that. Before this implementation i basically copy all the code that evaluate function in the repo. The only differences, in the evaluate function its using data loader, but i am using cv2.imread. Before you ask, yes i did check data loaders transforms to. I just don’t understand whats the issue.