Hello,

Due to computational limitations, I have reduced the n_classes to 5 and n_samples = 2. But in the function pdist() in the file models. py, it is showing the above mentioned error: RuntimeError: t() expects a tensor with <= 2 dimensions, but self is 4D during training time.
pool_out = self.avg(conv_out).squeeze()
This gives pool_out as 4D-tensor.

Can you please check this?

Thanks,

Hello, I have been trying to reproduce the benchmark results reported in your paper, more specifically the ones where you train a ResNet-101 backbone with an API-net on top on the CUB-200-2011 dataset.

My first attempt was on a 16G GPU, so I could only set --n_classes 9, and --n_samples 3. I reached 87.0% validation accuracy, whereas in the paper a 88.6% was reported. The obvious explanation for this was that reducing the number of classes and samples, reduced API-net's performance, which makes sense.
So I have tried some experiments on a larger GPU, so that I could use more classes and samples, but this didn't give me the expected improvement.

When I set --n_classes 25, --n_samples 3, I only got 85.9% validation accuracy. (Notice that this is even lower than my first 9x3 experiment!) I even tried using Adam instead of SGD, but no improvement there.
So, I tried --n_classes 19, --n_samples 4, but got similar results.
I noticed that the training curves flatten early: already around epoch 30 training accuracy is 100%, and the loss and validation accuracy hardly change anymore.

Any ideas about what could go wrong or how to improve my results would be very welcome, thanks!

Hi, I have tried your code on my dataset, but the softmax loss converge slowly compared to other methods. do you know the reason? thank you.

I train the two classes classification. It can work in training time.But it can not work in validation

What GPU do you use ? and how much memory this project need ?

Congratulations to the great improvement on your paper，but could you update a version which can have a testing to reproduce the great improvement

Hi,

I recently did some research on fine-grained classification, and I found that there is a 'val.txt' in your code. Is the 'val.txt' in the code actually the test set in datasets like CUB? Whether should I split trainset into train and val set when I do some experiments on datasets like CUB_200_2011, since I also didn't find any description about validation set in other existing papers?

Thank you.

Kind regards.

Hello, I have read your paper and your code. Thanks for you great job.
I want to use your model in image matching task. My dream model is: input is 2 images, and output is their match score, just as your training phase.
I review your code and your paper, and find that: this model is required the one image and output score is <Pooling feat + FC> while testing.
My question is: If I use an image pair(2 images) to get the final score from your model, would this model work find in the fine-grained dataset?
Thank you.

If you use class size 30.sample images 4. The batch number is 120. It is so big.