PyTorch Certified Associate (PTCA) free question 10 / 10 · Domain 4: Data Handling
A dataset has 10 samples and a DataLoader uses batch_size=3 with drop_last=True. How many batches does one epoch yield, and what happens to the last sample?
- 4 batches; the last batch is smaller and holds one sample
- 4 batches; the last sample is repeated to fill the batch
- 3 batches; the incomplete last batch is dropped
- 3 batches; the last sample is moved to the start of the next epoch
Show the answer and explanation
Answer: C
Why this is correct
drop_last controls what happens when the dataset size is not divisible by the batch size. The page says that, when set to True, the last incomplete batch is dropped; if False, the last batch will be smaller. Ten samples with batch_size=3 form three full batches and one leftover sample. The BatchSampler example on the page shows the same data both ways: with drop_last=False it yields [[0, 1, 2], [3, 4, 5], [6, 7, 8], [9]], and with drop_last=True it yields [[0, 1, 2], [3, 4, 5], [6, 7, 8]].
Why the other options are wrong
"4 batches; the last batch is smaller and holds one sample" … That is the result with drop_last=False. The page says that if drop_last is False and the dataset size is not divisible by the batch size, the last batch will be smaller.
"4 batches; the last sample is repeated to fill the batch" … The page describes two outcomes only: the incomplete last batch is dropped, or it stays smaller. It does not describe repeating samples to fill a batch.
"3 batches; the last sample is moved to the start of the next epoch" … The page says drop_last=True drops the last incomplete batch. The BatchSampler example with drop_last=True yields three batches of three indices and leaves index 9 out.
Key points
drop_last=True makes every batch the same size, at the price of skipping the leftover samples of the epoch.