have you reproduced the results in original paper?

[iamhankai]

have you reproduced the results in original paper?

[JunrQ]

No

[iamhankai]

No

Me too.. How much difference now?

[JunrQ]

No

Me too.. How much difference now?

Actually, I didn't run experiments in original paper. The dataset, block type, arch I used for doing experiments are different.

[QiuPaul]

As you said，you only use the method of fbnet to search another task network?

[JunrQ]

As you said，you only use the method of fbnet to search another task network?

I run the ImageNet search as the paper, and it converges well.

But when i do the classification task with the sampled model, it failed to converge.

I don't know why... may be different initialization? bn? random seed??

[QiuPaul]

As you said，you only use the method of fbnet to search another task network?

I run the ImageNet search as the paper, and it converges well.

But when i do the classification task with the sampled model, it failed to converge.

I don't know why... may be different initialization? bn? random seed??

Ah, you mean the training of the final searched network failed to converge, or the search in another classification task? i guess the search in another classification task failed to converge, but i also wonder how it happens, actually,i recently work on fbnet and snas...

[JunrQ]

As you said，you only use the method of fbnet to search another task network?

I run the ImageNet search as the paper, and it converges well.
But when i do the classification task with the sampled model, it failed to converge.
I don't know why... may be different initialization? bn? random seed??

Ah, you mean the training of the final searched network failed to converge, or the search in another classification task? i guess the search in another classification task failed to converge, but i also wonder how it happens, actually,i recently work on fbnet and snas...

The search part converges ( Same settings as original paper ).

But the network (choose one from eight blocks for every TBS(to be search) layer) failed to converge on a very simple classification task.

One possible reason: too many layers (22 * 3 ? I don't remember the exact number), but no batchnorm.

[iamhankai]

One possible reason: too many layers (22 * 3 ? I don't remember the exact number), but no batchnorm.

Have u tried the model provided in the paper?

[JunrQ]

One possible reason: too many layers (22 * 3 ? I don't remember the exact number), but no batchnorm.

Have u tried the model provided in the paper?

Not yet.

[brisker]

@JunrQ
Python version?
PyTorch version?

[JunrQ]

@JunrQ
Python version?
PyTorch version?

Python 2.7
PyTorch 1.0

[kbhardwa]

Anything new on this? I was wondering if this code reaches about 97-98% accuracy on CIFAR-10. Please let me know. Thanks a lot.