scut-mingqinchen / self2self Goto Github PK

Hi, I notice that in Punet.py you added some poisson noise to the masked image.

if is_realnoisy:
    response = tf.squeeze(tf.random_poisson(25 * response, [1]) / 25, 0)

Can you explain why this operation is necessary?

Thanks.

Hi,

I did reproduce your result Set9(25, 50).
And it shows similar PSNR score.

But, SSIM score(8.71, 8.11) is worse than the paper(9.56, 9.28).

I did also make BSD68 result, it shows similar PSNR and SSIM.
I think the way to calculate color SSIM is wrong.

I tried three variance of color SSIM.

1. calculate SSIM with mean image of color channel
2. calculate SSIM with cv2.cvt BGR to GRAY
3. calculate SSIM with LUMA, which is Y in YCbCr space.
4. calculate SSIM separately for color and average the score.

Could you tell me how to get your SSIM score in Set9.

I noticed that EBs are using partial convolution without dropouts, while DBs are using standard convolution with dropouts. My understanding of this is that the current DB setting is a tradeoff between a) mimicking the behavior of partial convolutional layers, and b) not using the inaccurate downsampled masks in upsample blocks. But I'm not sure if this is the case. Would you mind explaining the reasoning behind this?

If this is the case, probably a soft-mask, which generated by Conv2d and ConvTrans2d instead of maxpool, can be an alternative to the current settings?

Thanks in advance

When I training the realnoisy , the picture will be blacker , and the loss will be bigger.And the 1000 training steps result seems good ,but when training times get bigger, the result becomes worser and the noise appears

Hello,

Thank you for publishing the code! Can the model be trained on a dataset to make it more generalizable similar to dataset based models like Noise2Fast or Noise2Noise?

Kind regards

Hi.

I'm wondering if you have codes of pytorch version!

Thx.

I want to test this model in my own images, do you have a plane to share the trained model?

In Punet.py, the mask is scaled by a factor that is equal to dropout rate (0.7), is it a bug?

mask_tensor = tf.nn.dropout(mask_tensor, 0.7) * 0.7

According to TF doc, the tensor being dropouted has expectation of 1.0/(1-rate), which would be 3.3333 in this setting. So the scaling factor should be 0.3 instead of 0.7?

mask_tensor = tf.nn.dropout(mask_tensor, 0.7) * 0.3

default dropout_rate (p in your code) is 0.7
and Bernoulli sample value is also 0.7.
Is the dropout value equal to the Bernoulli sample value?
if i want to make the result more sharp. which is the best value of the Bernoulli sample value and dropout_rate?

tensorflow/core/common_runtime/executor.cc:641] Executor failed to create kernel. Invalid argument: Default MaxPoolingOp only supports NHWC on device type CPU
[[{{node MaxPool_1}}]]
Traceback (most recent call last):
File "/root/py3env/lib64/python3.6/site-packages/tensorflow/python/client/session.py", line 1356, in _do_call
return fn(*args)
File "/root/py3env/lib64/python3.6/site-packages/tensorflow/python/client/session.py", line 1341, in _run_fn
options, feed_dict, fetch_list, target_list, run_metadata)
File "/root/py3env/lib64/python3.6/site-packages/tensorflow/python/client/session.py", line 1429, in _call_tf_sessionrun
run_metadata)
tensorflow.python.framework.errors_impl.InvalidArgumentError: Default MaxPoolingOp only supports NHWC on device type CPU
[[{{node MaxPool_1}}]]

Hi.

I executed demo_denoising.pu without any touching codes, but i got ValueError.

ValueError: Layer p_conv2d_1 was called with an input that isn't a symbolic tensor. Received type: <class 'tensorflow.python.framework.ops.Tensor'>. Full input: [<tf.Tensor 'enc_conv0/MirrorPad:0' shape=(1, 1, 323, 483) dtype=float32>, <tf.Tensor 'enc_conv0/PadV2:0' shape=(1, 1, 323, 483) dtype=float32>]. All inputs to the layer should be tensors.

What's the problem?