simet.metrics.downstream_task.sample.sample_downstream_task¶

simet.metrics.downstream_task.sample.sample_downstream_task ¶

SampleDownstreamTask ¶

Bases: DownstreamTask, ABC

Toy downstream task: train a small CNN and report test accuracy.

Trains a :class:SimpleCNN classifier for a few epochs on a provided training dataloader and evaluates on a test dataloader, returning the final accuracy as a scalar. Intended as a lightweight example of how to wire a downstream task; not optimized for performance.

Assumptions

Binary classification (num_classes=2) using CrossEntropyLoss.
Dataloaders yield (image_tensor, target) pairs where target is integer-encoded (0/1).
Uses CUDA if available, otherwise CPU.

Hyperparameters

Optimizer: Adam(lr=MY_OPTIM_LR) with MY_OPTIM_LR = 1e-3.
Epochs: MY_EPOCHS = 3.

Attributes:

Name	Type	Description
`train_set`	`DataLoader[VisionDataset]`	Set during `_compute`.
`test_set`	`DataLoader[VisionDataset]`	Set during `_compute`.
`device`	`device`	`"cuda"` if available else `"cpu"`.
`model`	`Module`	The classifier (initialized in `_compute`).
`criterion`	`Module`	Loss function (`CrossEntropyLoss`).
`optimizer`	`Optimizer`	Adam optimizer on `model` params.