Dataloader parameter collate_ Use of FN

In short , This parameter is used to set dataloader The last output batch Content ;dataloader One time from dataset obtain batch Size data , But the data itself is fragmented , Take the picture for example , For example, we batch by 8, Then we get 8 individual [3,256,256]（256 Shape for picture , Set it up casually ） The tensor of size , adopt collate_fn This parameter translates to a shape of [8,3,256,256] Tensor of as dataloader Output .

In general , This parameter does not need to be set , What is an unusual situation , For example, when the data length is different , The most obvious one is NLP The length of the sentence in the example , Every batch The sentences inside are different in length , If you use the default collate_fn Method , You may report a mistake , In this case, you need to customize collate_fn Parameters .

Rewriting this parameter is simple , Is to customize a function , Suppose we name this function here my_collate_fn(batch), Be careful , There can only be one input variable batch,batch Contains dataset Inside getitem All values returned .

take vqa Task as an example , such as dataset Of getitem Return the data of one picture at a time data,label, And corresponding question,answer, If we set dataloader Get... At one time 8 Size batch, Then the my_collate_fn The variable of is a list, This list contain 8 individual [data, label, question, answer], If you want to use the default collate_fn, Requirements , this 8 individual [data, label, question, answer] Inside , The shape of each variable is the same , That is to say 8 individual data In the same shape ,8 individual label In the same shape ,8 individual question The shape of the ,8 individual answer The shape of , In this way, the default collate_fn Parameters , Otherwise, an error will be reported . If not, you need to customize collate_fn Parameters

Examples are as follows , At this point, the customized vqa_collate_fn The input is list by 8 The variable of , Let's look at the third element of each of these variables individually , You can find , The first 0 And the 1 The length of the third element of the variables is different , In this case, if the default collate_fn You're going to report a mistake , And in here, , In our custom functions , We put the third element of each variable in a list It goes straight back to ：

The example comes from ALBEF