API Overview
This is a brief overview of the state of the NeuroBench API, the specifications NeuroBench components are expected to conform to, and any exceptions or errata that have come up during development.
Example Python snippets may be shown here and will be kept up-to-date.
Overview
Components |
---|
Data |
Dataset |
PreProcessor |
PostProcessor |
Model |
Metrics |
Benchmark |
Specifications
Data:
Format:
tensor: A PyTorch tensor of shape (batch, timesteps, features*), where features* can be any number of dimensions.
Note: See Known Errata below for special cases of data formatting.
Dataset:
Output:
(data, targets): A tuple of PyTorch tensors. The first dimension (batch) is expected to match.
Alternatively, if other metadata is required from the dataset, like in 1MP object detection, then there can be a 3-tuple with kwargs.
Output:
(data, targets, kwargs): kwargs is a dictionary of metadata.
Pre-processor:
Processing data / preprocessing.
Input:
(data, targets): A tuple of PyTorch tensors. The first dimension (batch) is expected to match.
Output:
(data, targets): A tuple of PyTorch tensors. The first dimension (batch) is expected to match.
class PreProcessor(NeuroBenchPreProcessor):
def __init__(self):
...
def __call__(self, dataset):
...
alg = PreProcessor()
new_dataset = alg(dataset) # dataset: (data, targets)
Post-processor:
Accumulating predictions / postprocessing.
Input:
preds: A PyTorch tensor.
Output:
results: A PyTorch tensor. Post-processors may be chained together. Final shape is expected to match the data targets for comparison.
class PostProcessor(NeuroBenchPostProcessor):
def __init__(self):
...
def __call__(self, preds):
...
alg = PostProcessor()
model = NeuroBenchModel(...)
preds = model(data) # data: (batch, timesteps, features*)
results = alg(preds)
Model:
Input:
data: A PyTorch tensor of shape (batch, timesteps, features*)
Output:
preds: A PyTorch tensor. Can either be the final shape to be compared with targets or an arbitrary shape to be postprocessed by Post-processors(s).
class SNNTorchModel(NeuroBenchModel):
def __init__(self, net):
...
def __call__(self, batch):
...
model = SNNTorchModel(net)
preds = model(batch)
Metrics:
There are two types of metrics: static and data. Static metrics can be computed using the model alone, while data metrics require the model predictions and the targets as well.
Static metrics are stateless functions.
Data metrics can also be stateless functions (in which case they are accumulated over batched evaluation via mean), or they can be stateful subclasses of AccumulatedMetric.
**Static Metrics:**
Input:
model: A NeuroBenchModel object.
Output:
result: Any type. The result of the metric.
**Data Metrics:**
Input:
model: A NeuroBenchModel object.
preds: A PyTorch tensor. To be compared with targets.
data: Tuple of (data, targets).
Output:
result: A float or int.
def static_metric(model):
...
def data_metric(model, preds, data):
# must return an int or float to be accumulated with mean
return compare(preds, data[1])
class data_metric_with_state(AccumulatedMetric):
def __init__(self):
...
def __call__(self, model, preds, data):
# accumulate state from this batch
return self.compute()
def compute():
# compute metric from accumulated state
Benchmark:
Input:
model: The NeuroBenchModel to be tested.
dataloader: A PyTorch DataLoader which loads the evaluation dataset.
pre-processors: A list of pre-processors.
post-processors: A list of post-processors.
metric_list: [[static_metrics], [data_metrics]], where each are strings. The names of the metric will be used to call it from the metrics file. User defined metrics should be discouraged.
Output:
results: A dict of {metric: result}.
model = TorchModel(net)
test_set = NeuroBenchDataset(...)
test_set_loader = DataLoader(test_set, batch_size=16, shuffle=False)
preprocessors = [PreProcessor1(), PreProcessor2()]
postprocessors = [PostProcessor1()]
static_metrics = ["footprint", "connection_sparsity"]
data_metrics = ["accuracy", "activation_sparsity"]
benchmark = Benchmark(
model,
test_set_loader,
preprocessors,
postprocessors,
[static_metrics, data_metrics]
)
results = benchmark.run()
Known Errata
Any anomalies that break the high-level API will be noted here but attempts will be made to keep this to a minimum.
Data formatting: For the sequence-to-sequence prediction tasks (MackeyGlass and PrimateReaching), the dataset is one time series, and it is presented as [num points, 1, features]. Each of the data points is considered as a separate inference task for the model, so it is bundled into the zero dimension. When using a DataLoader, ensure that shuffle=False if your model is sequential.