luvata / cs330-stanford-deep-multi-task-and-meta-learning Goto Github PK

Hello,

I tried to implement the assignment of HW1 by myself, and I was getting a flat loss and a flat accuracy around 20% in the default case, i.e: num_classes=5, num_classes=1, meta_batch_size=16.

Now I used your tensorflow implementation and got the same result.

Is this expected ?

Thanks.

shuffle at load_data ?

Some options:

1. shuffle(encoding) + shuffle(retrieval)? (shuffle separately, then concatenate)
2. shuffle(encoding + retrieval) (shuffle both together)
3. Others idea ?

Currently I'm using (2), there can be chance that retrieval samples labels are duplicated or not appear in encoding samples. IMO I think the prediction will not be biased since the model must learn to deal with duplication or novelty class in retrieval phase
Did I miss something ?

[Edit 1] Newest commit is currently using options (2)
[Edit 2]

On paper MANN

Interestingly, the MANN displayed better than random guessing on the first instance within a class. Seemingly, it employed a strategy of educated guessing; if a particular sample produced a key that was a poor match to any of the bindings stored in external memory, then the network was less likely to choose the class labels associated with these stored bindings, and hence increased its probability of correctly guessing this new class on the first instance.

So it seems (1) is still correct

hello!
In the matrix 'all_image_batches', why don't all shots in the same class come from the same class?
For instance, if I print 'images[0, 0, 0].reshape((28,28))' and 'images[0, 1, 0].reshape((28,28))', I do expect them to be the similar picture form same class.Thanks!

Hi @Luvata

Is the code for HW2 also complete, if not will it be possible for you to upload the completed code.It will be of great help.

Thank You.

Hey,

I tried solving this in PyTorch, but my training loss is not improving.

I tried your PyTorch solution (copy and paste), same problem.

Am I missing something?