minibatchqueue without shuffling by default

mbq = minibatchqueue(ads,MiniBatchSize=miniBatchSize,MiniBatchFormat="BC");

If without mbq.shuffle, mbq retains the original sequence in ads? That is, minibatchqueue just cuts ads into length(ads)/miniBatchSize*portions?

e.g. ads=arrayDatastore(linspace(0,2,1000),IterationDimension=1) and if miniBatchSize=100, then the 1st batch of mbq is just [0, ..., 0.2]?

[robinho robinho]

?

Answer 1

[Ben]

I think you need to use IterationDimension=2 to read out [0, ..., 0.2] in one mini-batch:

ads=arrayDatastore(linspace(0,2,1000),IterationDimension=2);

miniBatchSize=100;

mbq = minibatchqueue(ads,MiniBatchSize=miniBatchSize,MiniBatchFormat="BC");

batch = next(mbq);

This will read out a batch of 100 observations, first [0, ..., 0.1982], then [0.2002,...,0.3984] etc. These will be output as 1x100 dlarray with "CB" labels - i.e. a batch of 100 observations.

[robinho robinho]

thank you so much Ben! Sorry for a typo of my OP, but I meant to post:

ads=arrayDatastore(linspace(0,2,1000)',IterationDimension=1) and if miniBatchSize=100, then the 1st batch of mbq is just [0, ..., 0.1982]?

So can we confirm that without mbq.shuffle, mbq retains the original sequence in ads and next() doesn't contain any randomness in it?

[Ben]

Yes, there should be no randomness, and if the miniBatchSize exactly divides the number of observations then minibatchqueue will just read through the data in chunks of size miniBatchSize (the exception is that if miniBatchSize doesn't exactly divide the number of observations then your last minibatch will be smaller - the remainder of numObservations / miniBatchSize).

Once the minibatchqueue runs out of data use reset to loop back to the start:

ads=arrayDatastore(linspace(0,2,1000)',IterationDimension=1);

mbq = minibatchqueue(ads,MiniBatchSize=100);

while hasdata(mbq)

batch = next(mbq);

end

reset(mbq);

[robinho robinho]

thank you. What difference will it make to the training if mbq.shuffle isn't used?

[Ben]

The data won't be shuffled at the end of each epoch. This could impact the model performance.

[robinho robinho]

right, but what impact would no-shuffling make?

[robinho robinho]

?