Design Doc¶

Overview¶

The model compression framework has two main components: pruner and module wrapper.

pruner¶

A pruner is responsible for :

provide a cal_mask method that calculates masks for weight and bias.
replace the module with module wrapper based on config.
modify the optimizer so that the cal_mask method is called every time the step method is called.

module wrapper¶

A module wrapper is a module containing :

the origin module
some buffers used by cal_mask
a new forward method that applies masks before running the original forward method.

the reasons to use module wrapper :

some buffers are needed by cal_mask to calculate masks and these buffers should be registered in module wrapper so that the original modules are not contaminated.
a new forward method is needed to apply masks to weight before calling the real forward method.

How it works¶

A basic pruner usage:

configure_list = [{
    'sparsity': 0.7,
    'op_types': ['BatchNorm2d'],
}]

optimizer = torch.optim.SGD(model.parameters(), lr=0.001, momentum=0.9, weight_decay=1e-4)
pruner = SlimPruner(model, configure_list, optimizer)
model = pruner.compress()

A pruner receive model, config and optimizer as arguments. In the __init__ method, the step method of the optimizer is replaced with a new step method that calls cal_mask. Also, all modules are checked if they need to be pruned based on config. If a module needs to be pruned, then this module is replaced by a module wrapper. Afterward, the new model and new optimizer are returned, which can be trained as before. compress method will calculate the default masks.

Implement a new pruning algorithm¶

Implementing a new pruning algorithm requires implementing a new pruner class, which should subclass Pruner and override the cal_mask method. The cal_mask is called byoptimizer.step method. The Pruner base class provided basic functionality listed above, for example, replacing modules and patching optimizer.

A basic pruner look likes this:

class NewPruner(Pruner):
    def __init__(self, model, config_list, optimizer)
        super().__init__(model, config_list, optimizer)
        # do some initialization

    def calc_mask(self, wrapper, **kwargs):
        # do something to calculate weight_mask
        wrapper.weight_mask = weight_mask

Set wrapper attribute¶

Sometimes cal_mask must save some state data, therefore users can use set_wrappers_attribute API to register attribute just like how buffers are registered in PyTorch modules. These buffers will be registered to module wrapper. Users can access these buffers through module wrapper.

class NewPruner(Pruner):
    def __init__(self, model, config_list, optimizer):
        super().__init__(model, config_list, optimizer)
        self.set_wrappers_attribute("if_calculated", False)
    
    def calc_mask(self, wrapper):
        # do something to calculate weight_mask
        if wrapper.if_calculated:
            pass
        else:
            wrapper.if_calculated = True
            # update masks

Collect data during forward¶

Sometimes users want to collect some data during the modules’ forward method, for example, the mean value of the activation. Therefore user can add a customized collector to module.

class ActivationRankFilterPruner(Pruner):
    def __init__(self, model, config_list, optimizer, activation='relu', statistics_batch_num=1):
        super().__init__(model, config_list, optimizer)
        self.set_wrappers_attribute("if_calculated", False)
        self.set_wrappers_attribute("collected_activation", [])
        self.statistics_batch_num = statistics_batch_num

        def collector(module_, input_, output):
            if len(module_.collected_activation) < self.statistics_batch_num:
                module_.collected_activation.append(self.activation(output.detach().cpu()))
        self.add_activation_collector(collector)
        assert activation in ['relu', 'relu6']
        if activation == 'relu':
            self.activation = torch.nn.functional.relu
        elif activation == 'relu6':
            self.activation = torch.nn.functional.relu6
        else:
            self.activation = None

The collector function will be called each time the forward method runs.

Users can also remove this collector like this:

collector_id = self.add_activation_collector(collector)
# ...
self.remove_activation_collector(collector_id)

Multi-GPU support¶

On multi-GPU training, buffers and parameters are copied to multiple GPU every time the forward method runs on multiple GPU. If buffers and parameters are updated in the forward method, an in-place update is needed to ensure the update is effective. Since cal_mask is called in the optimizer.step method, which happens after the forward method and happens only on one GPU, it supports multi-GPU naturally.