I wonder if i could get some help with my own RGB input about fast-depth

[fangchangma]

Hi. To test with your own images, simply read them in as a 4-dimensional pytorch floating point tensor of size 1x224x224x3. The raw rgb values [0,255] should be divided by a constant factor 255.0, such that all pixel values fall in the range of [0, 1].

[GustavoCamargoRL]

I tried doing this but i've got this error "RuntimeError: Given groups=1, weight of size 32 3 3 3, expected input[1, 244, 244, 3] to have 3 channels, but got 244 channels instead". My image is 244x244 and i'm giving the right format, as you can see here:

import matplotlib.pyplot as plt
import numpy as np
img = plt.imread("img.jpg")/255.
img.shape
(244, 244, 3)
img = np.expand_dims(img, axis=0)
img.shape
(1, 244, 244, 3)
i = torch.from_numpy(img)
Traceback (most recent call last):
File "", line 1, in
NameError: name 'torch' is not defined
import torch
i = torch.from_numpy(img)
i.shape
torch.Size([1, 244, 244, 3])

So i don't know why this error is occurring, if i'm giving the exact same format.

[dwofk]

The PyTorch conv2d function assumes inputs to be in 'NCHW' format, meaning that the tensor you feed into the network should be of shape [1, 3, 224, 224]. From your code snippet, you may be using a 'NHWC' format -- try permuting the tensor dimensions to change to 'NCHW'.

[fangchangma]

My image is 244x244 and i'm giving the right format

Also, the correct image size is 224 x 224, not 244 x 244

[GustavoCamargoRL]

Oh thanks! it worked, just one more problem, the results were these :

I placed my input code in the "args.evaluate" if condition, and then saved my results in a ply file, so my question is if there is any pos processing missing for the correct prediction of the depth map that i forgoted to do, or it just didn't work for this image.

[fangchangma]

Have you divided the input RGB values by 255.0, as in this line?

fast-depth/dataloaders/nyu.py

Line 56 in b1266da

rgb_np = np.asfarray(rgb_np, dtype='float') / 255

[GustavoCamargoRL]

Not exactly like this.
This is my input code :

    img = plt.imread("img.jpg")/255.  #normalization
    img = np.reshape(img, (3, 224, 224))
    img = np.expand_dims(img, axis=0)
    print(img.shape)
    with torch.no_grad():
        pred = model(torch.from_numpy(img).float().cuda())
        np.save('pred.npy', pred.cpu())
    
    print(pred)
    import sys
    sys.exit(0)

[fangchangma]

img = np.reshape(img, (3, 224, 224))

I believe it should be permutation of dimensions here, rather than reshaping (which breaks the data ordering). Please try img = np.transpose(img, (2,0,1)) and see if it makes a difference.

[GustavoCamargoRL]

It worked much better!
I will try better results with other images. Thanks for the help!

[mathmax12]

Thanks for the work @dwofk @fangchangma.
I am trying the same thing as @GustavoCamargoRL did.

   while True:  
        image_cuda = torch.from_numpy(img).float().cuda()
        pred = 0
        print(pred)
        with torch.no_grad():
            pred = model(image_cuda)
            #np.save('pred.npy', pred.cpu())
        print(pred)

The output from the first iteration looks good. But at each iteration, the output is different from the output of other iterations even with the same input image (See below pic).
If I kill the thread and execute the code each time at the first iteration I will get the same output.

I print the pred values and find that it does differ from the previous iteration even with the same input image and the same model.

Is there anything I missed for using the model?

[mathmax12]

@GustavoCamargoRL Do you have the same issue?

[LulaSan]

@mathmax12 Have you done this by using tvm apache?

[mathmax12]

@LulaSan It turns out that this caused by the tvm . the latest tvm solved this

[LulaSan]

@mathmax12 Ok thank you, can I ask you how do you visualize the results? By using their code visualize.py?

[mathmax12]

You can save the results as https://github.com/dwofk/fast-depth/blob/master/main.py#L98
or using cv2.imshow() to display