joeylitalien / noise2noise-pytorch Goto Github PK

Hello, I would like to ask, why the test image I input is 1024 * 1024, but the output is indeed 256 * 256, and I did not see where the picture was cropped, I look forward to your reply, thank you.

I am curious about the differences from the paper model

1. Why did you choose to use ConvTranspose instead of Upsampling
2. You used RELU in all parts and Leaky RELU after the last layer. But from the paper I think the author meant Leaky RELU everywhere except the last layer. And In the last layer only linear activation

Unrelated question
3) Did you use any pre-processing for the images e.g. means subtraction, normalization etc. I think that would be needed since we don't have BN layers

I am trying to implement the model in Tensorflow and having the problem of INF loss that starts in the 2nd epoch. So I hope your answers would help me. Thank you in advance!

Update: I seem to have solved the problem by adding Batch Norm layers to the UNET model. But still I am puzzled how the authors managed to get a stable training without Batch Norm

Traceback (most recent call last):
File "train.py", line 7, in
from datasets import load_dataset
File "/Users/snehagathani/Desktop/noise/src/datasets.py", line 9, in
from utils import load_hdr_as_tensor
File "/Users/snehagathani/Desktop/noise/src/utils.py", line 12, in
import OpenEXR
ImportError: dlopen(/anaconda3/lib/python3.6/site-packages/OpenEXR.cpython-36m-darwin.so, 2): Symbol not found: __ZN7Imf_2_314TypedAttributeISsE13readValueFromERNS_7IStreamEii
Referenced from: /anaconda3/lib/python3.6/site-packages/OpenEXR.cpython-36m-darwin.so
Expected in: flat namespace
in /anaconda3/lib/python3.6/site-packages/OpenEXR.cpython-36m-darwin.so

This error comes for all the training files.

Hi ,

Can you please share the results of Monte Carlo Renderings N2N ?

Regards
Muzahid

Hello,there may be problems with this place
raise ValueError('Invalid noise type: {}'.format(noise_type))
should be
raise ValueError('Invalid noise type: {}'.format(self.noise_type))

Hi I checked the code feel confused with the code in unet.py:

    def forward(self, x):
        """Through encoder, then decoder by adding U-skip connections. """

        # Encoder
        pool1 = self._block1(x)
        pool2 = self._block2(pool1)
        pool3 = self._block2(pool2)
        pool4 = self._block2(pool3)
        pool5 = self._block2(pool4)

pool2-pool5 is computed by self._block2, So it means pool2-pool5 re-use the same conv weight and bias. Does it accepted? I think the u-net should use different conv weight of different layer.

Dear Authors,

There are no instructions on how to use the code for my own noisy-noisy pairs. The current code takes a dataset of images as input and applies two independent noises on each instance, leading to pairs of noisy-noisy train set. But what if I already have my own pairs of noisy-noisy images and want to train the network on them? Is this possible with current architecture?

Best,
Ali.

i used --cuda on google colab but gpu is not being used how to fix it

Hi Authors,

I found the official implementations of adding poission noise use the following methods:
https://github.com/NVlabs/noise2noise/blob/c40a0481198bb524d0b70c2cc452f21bd7aec85c/train.py#L47

    return np.random.poisson(chi*(x+0.5))/chi - 0.5

Do you think we can utilize this method?

hello, i want to know if you have codes about image reconstruction or if the codes can be used for image reconstruction.

I train the model on val2017 dataset and my own dataset, but the valid result is about 21db, have you tested them ?Thank you!