How many iterations will a training run last?

$$\text{iterations}=\{\begin{array}{llc}\lfloor \frac{|\mathcal{D}|}{\text{batch size}}\rfloor \cdot \text{epochs}& & \text{if }\text{drop\_last=True}\\ & & \\ \lceil \frac{|\mathcal{D}|}{\text{batch size}}\rceil \cdot \text{epochs}& & \text{if }\text{drop\_last=False}\end{array}$$

where $|\mathcal{D}|$ is the dataset cardinality (size in terms of number of samples) and drop_last corresponds to the PyTorch DataLoader initialisation argument that determines if the final samples are discarded (True) or retained (False) if they are insufficient to populate a whole batch.

• Note the term $\frac{|\mathcal{D}|}{\text{batch size}}$ represents the number of iterations per epoch

def compute_iterations_per_epoch(dataset_size: int, batch_size: int, drop_last: bool) -> int:
    return dataset_size // batch_size if drop_last else -(-dataset_size // batch_size)
 
 
def compute_n_epochs(iterations: int, dataset_size: int, batch_size: int, drop_last: bool) -> int:
    iterations_per_epoch = compute_iterations_per_epoch(dataset_size, batch_size, drop_last)
    return -(-iterations // iterations_per_epoch)