Utilities¶
-
class
condensa.util.AverageMeter¶ Computes and stores the average and current value
-
class
condensa.util.EventTimer¶ Simple timer class.
-
property
elapsed_seconds¶ Returns elapsed seconds.
-
reset()¶ Reset timer.
-
property
-
condensa.util.accuracy(output, target, topk=(1, ))¶ Computes the precision@k for the specified values of k
- Parameters
output (torch.Tensor) – Predicted output batch
target (torch.Tensor) – Actual output batch
topk (Tuple) – Top-k value
- Returns
Model accuracy
- Return type
float
-
condensa.util.cnn_statistics(model, criterion, dataloader)¶ Computes accuracy of given CNN model.
- Parameters
model (torch.nn.Module) – PyTorch model
criterion – Loss function
dataloader – Data loader to use
- Returns
Top-1 and Top-5 accuracies
- Return type
Tuple(top1, top5)
-
condensa.util.compressed_model_stats(w, wc)¶ Retrieve various statistics for compressed model.
- Parameters
w (torch.nn.Module) – Original model
wc (torch.nn.Module) – Compressed model
- Returns
Dictionary of compressed model statistics
- Return type
dict
-
condensa.util.empty_stat_fn(model, criterion, dataloader)¶ Empty model statistics function: returns loss.
- Parameters
model (torch.nn.Module) – PyTorch model
loss_fn – Loss function
dataloader – Data loader to use
- Returns
Tuple of loss, dictionary of statistics
- Return type
Tuple(float, dict)
-
condensa.util.is_leaf_node(module)¶ Checks if given module is a leaf module.
- Parameters
module (torch.nn.Module) – PyTorch module
- Returns
Boolean value representing whether module is a leaf.
- Return type
bool
-
condensa.util.loss(model, criterion, dataloader)¶ Computes loss on given dataset.
- Parameters
model (torch.nn.Module) – PyTorch model
loss_fn – Loss function
dataloader – Data loader to use
- Returns
Loss
- Return type
float
-
condensa.util.magnitude_threshold(module, density)¶ Computes a magnitude-based threshold for given module.
- Parameters
module (torch.nn.Module) – PyTorch module
density (float) – Desired ratio of nonzeros to total elements
- Returns
Magnitude threshold
- Return type
float
-
condensa.util.pretrain(epochs, model, trainloader, criterion, optimizer)¶ No-frills pre-training method.
- Parameters
epochs (int) – Number of epochs
model (torch.nn.Module) – PyTorch model
trainloader – Training dataloader
criterion – Loss criterion
optimizer – Optimizer to use