boyuanjiang / matching-networks-pytorch

Matching Networks for one shot learning

matching-networks-pytorch's Issues

Pytorch version requirement 0.3.1

Since the update of Pytorch has changed its API for some functions, thus the version >= 0.4.0 is not compatible with this code.
You need to create a new conda environment with python3 and Pytorch 0.3.1 to run.

Otherwise, you will find it stack with problem:"TypeError iteration over a 0-d tensor"

That's all.

Final accuracy

Do you know what the final accuracies of this model trained on Omniglot are?

The results from the paper are:

5-way/1-shot  = 98.1%
5-way/5-shot  = 98.9%
20-way/1-shot = 93.8%
20-way/5-shot = 98.5%

Thanks
Ben

TypeError: iteration over a 0-d tensor

training objective

The training objective in the original paper is theta, should the loss value be theta? Why is loss in your program a cross entropy loss?

Seems fce are not correctly used

In the forward function of matching_net.py, output of self.g is passed to self.lstm, but the output of self.lstm is not used.

Note the 'output' becomes 'outputs' in 'outputs = self.lstm(output)'

code line 198

        # use fce?
        if self.fce:
            outputs = self.lstm(output)

        # get similarities between support set embeddings and target
        similarites = self.dn(support_set=output[:-1], input_image=output[-1])

ValueError: source code string cannot contain null bytes

File "mainOmniglot.py", line 1, in
from data_loader import OmniglotNShotDataset
ValueError: source code string cannot contain null bytes
Can any help?Thank you...

How to run this model with newest pytorch?

You need to change code:
line 133 def repackage_hidden(self,h): """Wraps hidden states in new Variables, to detach them from their history.""" if type(h) == Variable: return Variable(h.data) else: return tuple(self.repackage_hidden(v) for v in h)
to
def repackage_hidden(self, h): """Wraps hidden states in new Variables, to detach them from their history.""" if isinstance(h, torch.Tensor): return h.detach() else: return tuple(self.repackage_hidden(v) for v in h)
and all acc.data[0], loss.data[0] should be changed to acc.item(),loss.item().

then you can run the code with newest pytorch.

It is shameful to plagiarize others

The source project：https://github.com/gitabcworld/MatchingNetworks

About the testing

I have a little question about the testing.

In the testing part, the query image always comes from the 5 support classes. But, in practice, there is a totally unknown query coming, how we choose support set (because we don't know its class).

Actually I have tested this query with all possible support sets, but I got many high confidences.

This does not make any sense, we can NOT recognise the query by the output confidence.

That makes me confused about how to use matching net in real practice.

Is my understanding all right?

Cosine distance calculation problem

In the source code, the author calculates the cosine distance as follows.

    sum_support = torch.sum(torch.pow(support_image, 2), 1) 
    support_manitude = sum_support.clamp(eps, float("inf")).rsqrt() 
    dot_product = input_image.unsqueeze(1).bmm(support_image.unsqueeze(2)).squeeze()
    cosine_similarity = dot_product * support_manitude * input_manitude
    similarities.append(cosine_similarity)

But in my opinion, the right the cosine distance should be calculated as follows.

    sum_support = torch.sum(torch.pow(support_image, 2), 1) 
    support_manitude = sum_support.clamp(eps, float("inf")).rsqrt() 
    sum_input = torch.sum(torch.pow(input_image, 2), 1)
    input_manitude = sum_input.clamp(eps, float("inf")).rsqrt()
    dot_product = input_image.unsqueeze(1).bmm(support_image.unsqueeze(2)).squeeze()
    cosine_similarity = dot_product * support_manitude * input_manitude
    similarities.append(cosine_similarity)

Am i right? If not, what is the mistake?