Conv2dRFSearchOp

class mmcv.cnn.Conv2dRFSearchOp(op_layer: torch.nn.modules.module.Module, global_config: dict, verbose: bool = True)

Enable Conv2d with receptive field searching ability.

Parameters

• op_layer (nn.Module) – pytorch module, e,g, Conv2d

• global_config (dict) –

  config dict. Defaults to None. By default this must include:

  • ”init_alphas”: The value for initializing weights of each branch.

  • ”num_branches”: The controller of the size of search space (the number of branches).

  • ”exp_rate”: The controller of the sparsity of search space.

  • ”mmin”: The minimum dilation rate.

  • ”mmax”: The maximum dilation rate.

  Extra keys may exist, but are used by RFSearchHook, e.g., “step”, “max_step”, “search_interval”, and “skip_layer”.

• verbose (bool) – Determines whether to print rf-next related logging messages. Defaults to True.

estimate_rates() → None

Estimate new dilation rate based on trained branch_weights.

expand_rates() → None

Expand dilation rate.

forward(input: torch.Tensor) → torch.Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.