Results layout

[priyankaankolekar]

Hello, I suppose the results in the Model Zoo for Classification are the NCHW format. How do you convert them to the NHWC format (also mentioned in the README in the Classification folder)? Can this conversion be done while running the training i.e. in the train_imagenet.py script itself? Thanks in advance for your help.

[aojunzz]

@priyankaankolekar Hi,

the default format is NHWC, you can refer to Line 76 https://github.com/NM-sparsity/NM-sparsity/blob/main/devkit/sparse_ops/sparse_ops.py#:~:text=return%20Sparse_NHWC.apply(self.weight%2C%20self.N%2C%20self.M)

[priyankaankolekar]

So are the results of Model Zoo here (https://github.com/NM-sparsity/NM-sparsity/tree/main/classification, Table titled "Results and Model Zoo") are in NHWC format? What is the difference between the tables "Converting the the NCHW format in Pytorch to NHWC in ASP" and "Results and Model Zoo"? Thanks.

[aojunzz]

@priyankaankolekar yes, the results in model zoo are NHWC format, "the results and model zoo" table is original paper results, after the paper published, the nvidia asp released source code, we noticed that the ASP method prune dense model using NCHW format, so we used our method to prune the dense model with NCHW format, the result shown in "Converting the the NCHW format in Pytorch to NHWC in ASP". according to our experiments, the format NCHW and NHWC had similar performance.

Thanks,

[priyankaankolekar]

Thank you so much for explaining this.

Priyanka.

[791136190]

@priyankaankolekar yes, the results in model zoo are NHWC format, "the results and model zoo" table is original paper results, after the paper published, the nvidia asp released source code, we noticed that the ASP method prune dense model using NCHW format, so we used our method to prune the dense model with NCHW format, the result shown in "Converting the the NCHW format in Pytorch to NHWC in ASP". according to our experiments, the format NCHW and NHWC had similar performance.

Thanks,

There is such an implementation in nvida's asp code：

# convs
t = t.permute(2,3,0,1).contiguous().view(shape[2]*shape[3]*shape[0], shape[1])
func = getattr(sys.modules[__name__], pattern, None)
mask = func(t, density)
mask = mask.view(shape[2], shape[3], shape[0], shape[1]).permute(2,3,0,1).contiguous()      
return mask.view(shape).type(ttype)

My question is : for pytorch weights tensor layout default is oihw，so nvidia's layout should be hwoi ？