RuntimeError: Expected object of scalar type Byte but got scalar type Float for argument #3 'mat1' in call to _th_addmm_

vision
1

Hi, I am quite new to PyTorch and FastAI, and I am getting this error while using fastai’s unet_learner and according to the traceback, the error is generated at learn.lr_find(). Can you please let me know how I can fix this? Thanks in advance.

Here’s a part of my code:

class UNIZDataset(Dataset):
 
  def __init__(self, folderlist):

    self.c = 5
    count = 0
    for folder in tqdm(os.listdir(data_path)):
        images, labels = self.extract_data(os.path.join(data_path,folder))
        tensor_x = torch.from_numpy(np.array(images))
        tensor_y = torch.from_numpy(np.array(labels))
        if count == 0:
            self.x = tensor_x
            self.y = tensor_y
            count+=1
        else:
            self.x = torch.cat((self.x,tensor_x), dim = 0)
            self.y = torch.cat((self.y,tensor_y), dim = 0)

 def extract_data(self, path):
    x = []
    y = []
    data_x = os.path.join(path,'DATA_X')
    data_y = os.path.join(path,'DATA_Y')
    fileList = glob.glob(os.path.join(data_x,'*.tif'))
    for file in sorted(fileList):
        x.append(cv2.imread(os.path.join(data_x,file)))
        y.append(cv2.imread(os.path.join(data_y,file)))
    return (x,y)

def __getitem__(self, index):
    image = self.x[index,:,:]
    image = image.permute(2,0,1)
    label = self.y[index,:,:]
    label = label.permute(2,0,1)
    return (image,label)

def __len__(self):
    return self.x.size()[0]

train_dataset = UNIZDataset(train_files)
valid_dataset = UNIZDataset(valid_files)

databnch = DataBunch.create(train_dataset,valid_dataset, bs = 18)
learn = unet_learner(data=databnch,  arch=models.resnet34)
learn.lr_find()
1 Like
2

Maybe you need to convert the images to float32…can you replace

        tensor_x = torch.from_numpy(np.array(images))

with

        tensor_x = torch.from_numpy(np.array(images)).float()
2 Likes
3

Yes, that worked! Thank you :smile:

4

You might want to divide by 255 like this

        tensor_x = torch.from_numpy(np.array(images)).float() / 255.

otherwise the model may take long time to train

5

Thanks! I will do that :grin:. I am also running into another issue “IndexError: Target 20 is out of bounds.” when it starts to train. Can you please tell me what’s causing this error? Thanks again in advance!