import copy import warnings import abc import torch from torch import nn from torch.nn.utils import parametrize from torch.nn.modules.container import ModuleDict, ModuleList from .parametrization import PruningParametrization, ZeroesParametrization, ActivationReconstruction, BiasHook from torch.ao.sparsity import BaseSparsifier, module_to_fqn, fqn_to_module from torch.ao.sparsity.sparsifier.utils import get_arg_info_from_tensor_fqn __all__ = ["BasePruner"] SUPPORTED_MODULES = { # added to config if None given nn.Linear, nn.Conv2d, nn.BatchNorm2d, # will need manual update to match conv2d } NEEDS_ZEROS = { # these layers should have pruned indices zero-ed, not removed nn.BatchNorm2d } class BasePruner(BaseSparsifier): r"""Base class for all pruners. Abstract methods that need to be implemented: - update_mask: Function to compute a new mask for all keys in the `groups` attribute. Args: - defaults [dict]: default configurations will be attached to the configuration. Only the keys that don't exist in the `config` will be updated. - also_prune_bias [bool]: whether to prune bias in addition to weights (to prune full output channel) or not; default=True. """ def __init__(self, defaults, also_prune_bias=True): super().__init__(defaults) self.prune_bias = also_prune_bias def _get_modules_and_tensor_names(self, config, use_path): modules = [] tensor_names = [] if use_path: if type(config['module']) is tuple: # (Conv2d, BN) for module_fqn, tensor_name in zip(config['module_fqn'], config['tensor_name']): module = fqn_to_module(self.model, module_fqn) modules.append(module) tensor_names.append(tensor_name) else: module = fqn_to_module(self.model, config['module_fqn']) modules.append(module) tensor_name = config['tensor_name'] tensor_names.append(tensor_name) else: if type(config['module']) is tuple: for module, tensor_name in zip(config['module'], config['tensor_name']): modules.append(module) tensor_names.append(tensor_name) else: module = config['module'] modules.append(module) tensor_name = config['tensor_name'] tensor_names.append(tensor_name) return modules, tensor_names def _prepare(self, use_path=False, *args, **kwargs): r"""Adds mask parametrization to the layer weight """ self.activation_handles = [] # store removable hook handles self.bias_handles = [] for config in self.groups: modules, tensor_names = self._get_modules_and_tensor_names(config, use_path) for module, tensor_name in zip(modules, tensor_names): if not isinstance(module, tuple(NEEDS_ZEROS)): # add pruning parametrization and forward hooks if getattr(module, 'mask', None) is None: module.register_buffer('mask', torch.tensor(getattr(module, tensor_name).shape[0])) param = config.get('parametrization', PruningParametrization) parametrize.register_parametrization(module, tensor_name, param(module.mask), unsafe=True) assert isinstance(module.parametrizations, ModuleDict) # make mypy happy assert isinstance(module.parametrizations.weight, ModuleList) if isinstance(module, tuple(SUPPORTED_MODULES)): self.activation_handles.append(module.register_forward_hook( ActivationReconstruction(getattr(module.parametrizations, tensor_name)[0]) )) else: raise NotImplementedError("This module type is not supported yet.") else: # needs zeros if getattr(module, 'mask', None) is None: module.register_buffer('mask', torch.tensor(getattr(module, tensor_name).shape[0])) param = config.get('parametrization', ZeroesParametrization) parametrize.register_parametrization(module, tensor_name, param(module.mask), unsafe=True) if module.bias is not None: module.register_parameter('_bias', nn.Parameter(module.bias.detach())) module.bias = None self.bias_handles.append(module.register_forward_hook(BiasHook(module.parametrizations.weight[0], self.prune_bias))) if len(modules) == 2: # (Conv2d, BN) # should have the same set of pruned outputs modules[1].parametrizations.weight[0].pruned_outputs = modules[0].parametrizations.weight[0].pruned_outputs def make_config_from_model(self, model, SUPPORTED_MODULES=SUPPORTED_MODULES, NEEDS_ZEROS=NEEDS_ZEROS): self.config = [] stack = [model] while stack: module = stack.pop() for name, child in module.named_children(): if type(child) in SUPPORTED_MODULES: child_fqn = module_to_fqn(model, child) assert isinstance(child_fqn, str) # for mypy self.config.append({'tensor_fqn': child_fqn + '.weight'}) else: if NEEDS_ZEROS is not None and type(child) in NEEDS_ZEROS and hasattr(self, "prune_bias") and self.prune_bias: # only useful for Pruner warnings.warn(f"Models with {type(child)} layers have config provided by user.") stack.append(child) def prepare(self, model, config): r"""Prepares a model, by adding the parametrizations and forward post-hooks. Note:: The model is modified inplace. If you need to preserve the original model, use copy.deepcopy. Args: - model [nn.Module]: model to configure. The model itself is not saved but used for the state_dict saving / loading. - config [list]: configuration elements could either be instances of tuples of dict maps or dict maps. The dicts must have a key 'tensor_fqn' with the value being the fqn of the tensor to be pruned. """ self.model = model # TODO: Need to figure out how to load without this. self.config = config # If no config -- try getting all the supported layers if self.config is None: # Add all models to the config self.make_config_from_model(self.model) for module_config in self.config: if type(module_config) is tuple: first_layer, next_layer = module_config assert isinstance(first_layer, nn.Conv2d) and isinstance(next_layer, nn.BatchNorm2d) assert isinstance(module_config, tuple) # for mypy module_config = {'module': module_config} local_args = copy.deepcopy(self.defaults) local_args.update(module_config) module_fqn_list = [] tensor_fqn_list = [] tensor_name_list = [] for module in local_args['module']: module_fqn = module_to_fqn(model, module) if module_fqn is None: module_fqn = '' if module_fqn and module_fqn[0] == '.': module_fqn = module_fqn[1:] module_fqn_list.append(module_fqn) tensor_fqn_list.append(module_fqn + '.weight') tensor_name_list.append('weight') local_args['module_fqn'] = module_fqn_list local_args['tensor_fqn'] = tensor_fqn_list local_args['tensor_name'] = tensor_name_list else: if isinstance(module_config, nn.Module): module_config = {'module': module_config} # type: ignore[dict-item] local_args = copy.deepcopy(self.defaults) local_args.update(module_config) # now that we're working with a dict, does it have the new format? if local_args.get('tensor_fqn', None) is not None: tensor_fqn = local_args.get('tensor_fqn') assert isinstance(tensor_fqn, str) # for mypy info_from_tensor_fqn = get_arg_info_from_tensor_fqn(model, tensor_fqn) for key in info_from_tensor_fqn.keys(): if key in local_args: # info_from_tensor_fqn will chop leading '.' from tensor_fqn so ignore that assert key == 'tensor_fqn' or info_from_tensor_fqn[key] == local_args[key], ( "Given both `{}` and `tensor_fqn`, it is expected them to " "agree!".format(key) ) local_args.update(info_from_tensor_fqn) else: module = local_args['module'] module_fqn = module_to_fqn(model, module) if module_fqn and module_fqn[0] == '.': module_fqn = module_fqn[1:] local_args['module_fqn'] = module_fqn local_args['tensor_name'] = "weight" assert isinstance(module_fqn, str) # for mypy local_args['tensor_fqn'] = module_fqn + ".weight" self.groups.append(local_args) self._prepare() def squash_mask(self, use_path=False, *args, **kwargs): for config in self.groups: modules, tensor_names = self._get_modules_and_tensor_names(config, use_path) for module, tensor_name in zip(modules, tensor_names): parametrize.remove_parametrizations(module, tensor_name, leave_parametrized=True) if getattr(module._parameters, 'mask', None): del module._parameters['mask'] elif getattr(module._buffers, 'mask', None): del module._buffers['mask'] delattr(module, 'mask') def get_module_pruned_outputs(self, module, tensor_name='weight'): r"""Returns the set of pruned indices of module""" assert parametrize.is_parametrized(module) # can only get pruned indices of pruned module return getattr(module.parametrizations, tensor_name)[0].pruned_outputs # assume only one parametrization attached def step(self, use_path=False): if not self.enable_mask_update: return with torch.no_grad(): for config in self.groups: modules, tensor_names = self._get_modules_and_tensor_names(config, use_path) untupled_args: dict = {} untupled_args.update() # only need to update the first module in modules if len(modules) > 1 # since they should share the same set of pruned outputs untupled_args['module'] = modules[0] untupled_args['tensor_name'] = tensor_names[0] self.update_mask(**config) @abc.abstractmethod def update_mask(self, module, tensor_name, **kwargs): pass