Hi,
I have two questions regarding dimensions,
1. Lets for example focus on the Unary unit, as far as I can see from the paper the input for this unit is of the shape : where is the embeddind dimension (i.e for image or for text..) and is the number of samples and after doing the mathematics mentioned in the paper we get that the output dimension is of the shape : which a little bit confuses me am I missing something ? otherwise how can we look at the importance of each feature for each sample? (here we got only 1 sample output).

2. I tried to run the following code which simply tries to add an unary attention on word embeddings of dimension 300 and I noticed that the atten class expects a tensor input of 3 dimensions (I assume the first dimension is for the batches number? fix me if i'm missing something)
s_output, s_lengths = nn.utils.rnn.pad_packed_sequence(s) print(s_output.transpose(1,2).size()) # This will print : Torch.Size([48, 300, 128]) s_with_atten = self.attention(utils=[s_output.transpose(1,2).cuda()]) print(s_with_atten[0].size()) # This will print : Torch.Size([48, 300])
I defined self.attention as following :
self.attention = Atten(util_e = [params.word_embedding_dim] , high_order_utils=[], pairwise_flag = False , unary_flag=True , self_flag = False)
can you elaborate more on the output dimensions printed?

shouldn't size_flag=False be set to True instead?
Because you already passed a size list to Atten, and inside the constructor, there is a code for initializing an empty size list in case size_flag is set to false, but you already passed a size list and thus no need to init a new size list with Nones

Hello,

When I download AVSD annotations, it does not work. And it needs my account.

Hi,
I have a question about the following line of code :

Assuming we set pairwise_flag=false, and unary_flag = true, as far as I can see from the code this will also add a pairwise potential between the passed utility and itself + unary potential for the same utility.
my question is what's the point of adding a pairwise for the utility and itself since pairwise_flag=false and unary_flag=true

The provided link to download the dataset seems to be removed.

Which dataset and features should I download?