Duplicate VGG19 models

[lutzroeder]

There are two different VGG19 models in the model zoo. Can the files be consolidated or renamed so users referring to "VGG19 in the model zoo" will be less ambiguous.

https://github.com/onnx/models/tree/master/vgg19
https://github.com/onnx/models/tree/master/models/image_classification/vgg

[snnn]

"The VGG19 in the model zoo generated from caffe2"
"The VGG19 in the model zoo generated from mxnet"

That would be enough.

[houseroad]

@snnn thanks for the explanation

[prasanthpul]

That's pretty confusing for the users of the zoo to have 2 different versions of the model. Which one should they use and why?

[houseroad]

probably we can have some evaluation. And only keep one.

[prasanthpul]

we don't even link to both from the TOC so one of the models is "hidden".

[crahrig]

Adding on here that this is also captured by #82

[ankkhedia]

@prasanthpul @crahrig The models listed #82 were in the model zoo before the need for all the artifacts was standardized. Previous models :
https://github.com/onnx/models/tree/master/squeezenet
https://github.com/onnx/models/tree/master/resnet50
https://github.com/onnx/models/tree/master/vgg19
They do not meet all the requirements of reporting pre-processing, post-processing and accuracy numbers.
However, since the models are generated from different frameworks, they are still coexisting.

[snnn]

Hi @ankkhedia ,

Is there any update on this bug: #91 ? Does it only affect the resnet100 model?

Thanks.

[ankkhedia]

@snnn got a little busy and was not able to look into the issue #91, I will try to address the issue by next week. Sorry for the inconvenience.

[snnn]

Thank you @ankkhedia

[mx-iao]

@snnn @ankkhedia @prasanthpul @houseroad have we reached consensus on this? either 1) merge and keep only one model or 2) keep both, group in same folder, and differentiate based on training framework used ?

[snnn]

The VGG19 model from mxnet seems having some problems. The input of a GEMM op is not a 2-D tensor.

[vandanavk]

The VGG19 model from mxnet seems having some problems. The input of a GEMM op is not a 2-D tensor.

This should be fixed with PR apache/mxnet#13356.

[snnn]

The VGG19 model from mxnet seems having some problems. The input of a GEMM op is not a 2-D tensor.

This should be fixed with PR apache/incubator-mxnet#13356.

So, someone will update the models here?

[ankkhedia]

@snnn I will update the models after generating them with this fix.

[lutzroeder]

The actual issue seems to be there are duplicate VGG19 models in the zoo (see above). How does "updating" the files resolve this issue? Shouldn't one set be removed or the README.md files be updated to clarify?

[ankkhedia]

@lutzroeder This is about @snnn comment about issues with MXNet generated models.
Regarding duplicate models, the models have been generated from different frameworks and there is no clear consensus as to remove the models or keep both in same folder and differentiate on basis of names

@prasanthpul any thoughts on the same?

[ankkhedia]

@snnn The models has been uploaded after the fix. Please feel free to verify and let us know in case of any issues