How can I get unet.h5 model about chestx-ray-14 HOT 12 CLOSED

Hi @lbc402, I just upload the model to Kaggle, you can check out here

from chestx-ray-14.

@lbc402, did you load the pretrained model? the code to load is here

The model I uploaded is intended for the web demo, so it is not loaded by default.

from chestx-ray-14.

Thank to your reply. I'm not sure the link you provided is publicly accessible. In fact, I can't seem to open it.

from chestx-ray-14.

Hi @lbc402, I just make it publicly available, you can download the model now.

from chestx-ray-14.

Sorry @dattran2346 ，When I execute the code in cropped lung.ipynb, the returned py variable is a zero matrix.

d = {}
distorted_lung = []
for image_name in tqdm(image_names):
    image = Image.open(PATH/IMAGE_DN/image_name).convert('RGB')
    py = unet(V(tfm(image)[None]))  

I confirm that there is no problem with the input image. So I want to confirm with you whether the unet.h5 you provided is correct.

from chestx-ray-14.

I got it. I'm sure the model is loaded. In fact, the chexnet model can run successfully.

from chestx-ray-14.

The problem is to execute such a piece of code

unet = Unet(trained=True, model_name=unet_model).cuda()
unet.eval();
……
image = Image.open(PATH/IMAGE_DN/image_name).convert('RGB')
unet(V(tfm(image)[None])) 

And then we get this result

tensor([[[[0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
          [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
          [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
          ...,
          [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
          [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
          [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000]]]],
       device='cuda:0', grad_fn=<ReluBackward>)

from chestx-ray-14.

@lbc402, does the result from unet model are all zeros? Is the tfm working correctly, i.e, the image is properly normalized?

from chestx-ray-14.

@dattran2346 Yes, I got all zero results. I used the initialization parameters you defined, which seems to be no problem.

normalize = transforms.Normalize(IMAGENET_MEAN, IMAGENET_STD)
toTensor = transforms.ToTensor()
tfm = transforms.Compose([
    transforms.Resize((256, 256)),
    toTensor,
    normalize
])

I would like to ask you to confirm that this Unet model is the best model?

from chestx-ray-14.

@lbc402 I only keep the best single model, so yes, it is. But you're getting all zeros, which is quite interesting.

from chestx-ray-14.

@lbc402, I just checked the model, it's not all zeros. You should check the min and max value of the mask

Here is the working code

image = Image.open('/mnt/DATA/kaggle/siim/images/test/1.2.276.0.7230010.3.1.4.8323329.6249.1517875197.290574.jpg').convert('RGB')
image = Variable(tfm(image)[None])
py = torch.sigmoid(unet(image))
py = (py[0].cpu() > 0.5).type(torch.FloatTensor)
plt.imshow(py.numpy()[0])

from chestx-ray-14.

@dattran2346 Thank you very much. What a naive question I have.

from chestx-ray-14.