ironhead

make sure that the values of the image is in [0,1]

你好，请问一下你的average pool的得到f_b与f_f是如何实现的？是对Encoder的feature进行Masked Average吗，我同样尝试了本文的模型，最终L_rgb 出现 nan，疑似出现了梯度爆炸

> 是的，我这里解决了这个问题。之前的错误是由于引用论文的Adaptive 3D LUT的github公开实现似乎存在问题，python调用时是输入是[batch_size,3(rgb),width,height],但cuda cpp实现的是[3(rgb),batch_size,width,height]的输入。导致存在我调用时存在越界访问的情况。 另外对于f×m和f×(1-m)进行average pool意思是，直接对整个f的尺寸进行全局平均池化吗，我的实现类似是(fxm).sum()/m.sum(),(fx(1-m)).sum()/(1-m).summ()，缩小mask尺寸时使用的是AvgPool

@taoxinhao13 请问你们的backbone （ISSAM）是从头开始训练的还是使用了pretrained weights

我发现，c++ cuda中计算rgb时采用的位置是 index (假设这是r), index + batch * width * height (假设这是g), index + 2 * batch * width * height(假设这是b)，那么这里计算的是image[0][b_i][w_i][h_i] (r),image[1][b_i][w_i][h_i] (g),image[2][b_i][w_i][h_i] (b).那么调用时image的维度应当是[3][batch][width][height]才对？ 否则计算的rgb值都不一定在同一个pixel上和同一个通道上才对？ 是否是我理解有误了？我看python中调用时似乎维度使用的是[batch][channel(3)][width][height] 如果我的理解没有错误的话，如果batch_size > 1时，这里使用的就不是LUT了。

> Actually i dive in the cuda c++ file. Then i found that the input image Tensor should be in size [CHANNEL_SIZE(3), BATCH_SIZE,WIDTH,HEIGHT]. However, in the python code, the input...

same problem

@WEIIEW97 hey, sir. Actually i dive in the cuda c++ file. Then i found that the input `image` Tensor should be in size `[CHANNEL_SIZE(3), BATCH_SIZE,WIDTH,HEIGHT]`. However, in the python code...

@WEIIEW97 Thanks for your sharing. If It is a basic migration. I think there are some codes need to modify in the original training/evaluation python code. Some Modifications show as...

> > make sure that the values of the image is in [0,1] > > in the datasets.py, "TF.to_tensor" is to convert 0/255 to the 0/1 range, why is there...