collect_weight_statistics
python
collect_weight_statistics(model: nn.Module) -> list[LayerWeightStatistics]Measure every convolutional and linear layer's weights.
A purely structural read of the model, useful alongside activation statistics. A weight matrix whose effective rank is a small fraction of its size has many output units pointing the same way, so the layer is narrower than it looks; a spectral norm drifting far above its initial value is the usual sign of an optimiser pushing on a layer that has stopped contributing.
Parameters:
model(Module) – The model.
Returns:
list[LayerWeightStatistics]– One entry perConv2dorLinearlayer, in definition order.