uclaopt / provable_plug_and_play Goto Github PK

Hi, I found this is a very interesting theoretical paper with good practical justification.

I was wondering whether you could help me resolve the question about the experiment setting of CSMRI:
In Section 6 (Compressed sensing MRI.), the paper said the CSMRI results are evaluated in 30% sample with additive Gaussian noise \sigma=15/255. However, I cannot understand the meaning of noise intensity here, as the value range in Fourier space is indeed no longer 0-1 (or 0-255).
In your implementation, the value range of input image is [0-1], while the intensity of added noise is 15 rather than 15/255 described in the paper (note both 15 and 15/255 have no clear physical meaning in Fourier space)

Thank you very much!

Hello, when I reproduce your code of single photon imaging part, I don't quite understand the update part of x.

def inverse_step(u, v, K1, K, rho):
xtilde = v - u
x = np.copy(xtilde)
K0 = np.square(K) - K1

indices_0 = (K1 == 0)
x[indices_0] = xtilde[indices_0] - K0[indices_0] / rho

func = lambda y: K1 / (np.exp(y) - 1) - rho*y - K0 + rho*xtilde
indices_1 = np.logical_not(indices_0)

# binary search?
bmin = 1e-5 * np.ones_like(x, dtype=np.float64)
bmax = 100  * np.ones_like(x, dtype=np.float64)
bave = (bmin + bmax) / 2.0
for i in range(30):
    tmp = func(bave)
    indices_pos = np.logical_and(tmp > 0, indices_1)
    indices_neg = np.logical_and(tmp < 0, indices_1)
    indices_zero = np.logical_and(tmp == 0, indices_1)
    indices_0 = np.logical_or(indices_0, indices_zero)
    indices_1 = np.logical_not(indices_0)

    bmin[indices_pos] = bave[indices_pos]
    bmax[indices_neg] = bave[indices_neg]
    bave[indices_1] = (bmin[indices_1] + bmax[indices_1]) / 2.0

x[K1 != 0] = bave[K1 != 0]
return np.clip(x, 0.0, 1.0)

Does it use dichotomy to find the minimum value? What's its advantage over making the derivative equal to zero? Looking forward to your reply

Hi, really amazing work.
I have some questions about the convergence:

1. Why set sigma = 0 in BN layer：layers.append(bn_layer(features, 0.0)) # bn layer. Is it because that BN is not recommended for SN? if so, how to ensure the denoiser satisfies a certain Lipschitz condition after adding the BN layer?
2. Is convergence only related to network weight？Since during the ADMM process, the inputs to the projection network do not usually follow a Gaussian distribution, if I change the noise type and the loss function (choose L1 rather than L2) during the training process, can convergence be guaranteed?

Many thanks. Looking forward to your reply！

Hello,

Thank you for uploading the training codes of the paper, and well done for you work!

I am trying to training the RealSN-DnCNN on coloured images (3 channels) with a Lips of 1 but I can see that the codes are not adapted for that. I made some changes myself for example I added

    if module.weight.shape[0] == 3:
        C_out = 3

The NN seems to be trained however when I print the spectral values σ(Κ) after a point they go much higher than 1. This leads ADMM to a strange behaviour even if the NN is denoising really well.

It seems that by introducing the 3 channels, there is a bag in the code that it gets activated but not when you require only 1 channel. Could I kindly ask what changes would you make to the codes to be trained on coloured images? It would be so helpful.

Thank you very much again,
Savvas

hello,

I view we haven't the same PSNR. the deep learning algorithm use the false PSNR.
I propose to delete psnr in utils and put :

from skimage.metrics import peak_signal_noise_ratio

peak_signal_noise_ratio(img1,img2, data_range=255)
and for x_init, because it doesn't work if it's not modified :
`x_init = np.fft.ifft2(y).real # .real modification

Can you provide the original training code? Which loss function is used to train the network? Thanks!