ValueError: negative dimensions are not allowed in object_recognition.py about mann

[wzell]

hallo redhat12345,

the Alexnet implementation is from https://github.com/heuritech/convnets-keras. There, you can find your issue: heuritech/convnets-keras#18.

Seems that this issue has to do with the "image_dim_ordering" parameter in the keras.json file.

Please let me know if that helps.

[wzell]

Thank you :-).

The data augmentation is only for comparability reasons to other approaches that use data augmentation (Domain-adversarial neural networks, the mmd family of papers, DAN etc. and deep-CORAL).

I did not analyse the number of images on the accuracy. In the paper you can see that there are two approaches (mann and robust-mann in the code). One needs the augmented images (mann) and it is only for sake of comparability with other approaches that use data augmentation. The other approach (robust-mann) does not need the data augmentation. Please see the paper for details.

If you are not interested in running the full experiment (needs ~1 week), I suggest to uncomment everything that uses data augmentation (CNN network+mann) and just run the robust-mann implementation with only one repetition. This will need ~10min for the whole dataset. With increasing repetition number, the results should approach the table in the paper.

[wzell]

Done. Hope that helps.

[redhat12345]

I have tried with both theano 0.8.1 and theano 0.9.0. Got the same error. I am not sure what happened ?

[redhat12345]

By editing ~/.keras/keras.json file

{
"image_dim_ordering": "th",
"epsilon": 1e-07,
"floatx": "float32",
"backend": "theano"
}
It worked for me. Thank you so much for your kind reply. By the way you did a great job.

[redhat12345]

It creates a lot of augmented images (100000) per dataset. Is there any effect of numbers of augmented images in accuracy?

[wzell]

I close this issue since the error is solved and I change its name. Please open a new one if further errors occur.

[redhat12345]

Thank you so much. It would be really appreciable if you publish the commented version of the code which will give the results quickly. I am sorry that in close issue, I made a comment.