I wonder how this function, say MultimodalMinibatchLoaderCaption:next_batch, works on preparing the text data.
Seems that the tensor txt is created to store a batch of chosen sentences and finally returned as training data. But I'm confused on the assignment of txt in the for-loop in lines 83 to 101. It's NOT assigned directly like the img or lab (line 80 & 81).
So why txt is assigned in this way? How this works? I guess the for-loop is the flipping of sentence as the cmd:option -flip says and as my comprehension this operation only add some noise to the sentences data like image flipping?
p.s. I can run this code correctly on the CUB dataset but failed on wikipedia dataset, with NO augmentation on images as this paper say and for each image there is only 1 corresponding sentence. I preprocessed the data to keep the same shape with the data provided by the author, i.e., image of shape (#samples, 1024, 1) and text (#samples, 201, 1)
I wonder how this function, say
MultimodalMinibatchLoaderCaption:next_batch, works on preparing the text data. Seems that the tensortxtis created to store a batch of chosen sentences and finally returned as training data. But I'm confused on the assignment oftxtin the for-loop in lines 83 to 101. It's NOT assigned directly like theimgorlab(line 80 & 81). So whytxtis assigned in this way? How this works? I guess the for-loop is the flipping of sentence as thecmd:option -flipsays and as my comprehension this operation only add some noise to the sentences data like image flipping?p.s. I can run this code correctly on the CUB dataset but failed on wikipedia dataset, with NO augmentation on images as this paper say and for each image there is only
1corresponding sentence. I preprocessed the data to keep the same shape with the data provided by the author, i.e., image of shape(#samples, 1024, 1)and text(#samples, 201, 1)