From eb45a9a1293193027e9032fc932493b34e1823a1 Mon Sep 17 00:00:00 2001 From: wanghaoshuang Date: Fri, 1 Nov 2019 14:28:00 +0800 Subject: [PATCH] Add simple prune API and unitest. --- paddleslim/common/__init__.py | 4 + paddleslim/common/graph_wrapper.py | 401 +++++++++++++++++++++ paddleslim/prune/__init__.py | 1 + paddleslim/prune/pruner.py | 551 +++++++++++++++++++++++++++++ setup.py | 8 +- tests/layers.py | 31 ++ tests/test_prune.py | 59 +++ 7 files changed, 1053 insertions(+), 2 deletions(-) create mode 100644 paddleslim/common/graph_wrapper.py create mode 100644 paddleslim/prune/pruner.py create mode 100644 tests/layers.py create mode 100644 tests/test_prune.py diff --git a/paddleslim/common/__init__.py b/paddleslim/common/__init__.py index 9d053150..f6ad753d 100644 --- a/paddleslim/common/__init__.py +++ b/paddleslim/common/__init__.py @@ -11,3 +11,7 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from . import graph_wrapper +from .graph_wrapper import * +__all__ = graph_wrapper.__all__ diff --git a/paddleslim/common/graph_wrapper.py b/paddleslim/common/graph_wrapper.py new file mode 100644 index 00000000..04d8b077 --- /dev/null +++ b/paddleslim/common/graph_wrapper.py @@ -0,0 +1,401 @@ +# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +import copy +import pickle +import numpy as np +from collections import OrderedDict +from collections import Iterable +from paddle.fluid.framework import Program, program_guard, Parameter, Variable + +__all__ = ['GraphWrapper', 'VarWrapper', 'OpWrapper'] + +OPTIMIZER_OPS = [ + 'momentum', + 'lars_momentum', + 'adagrad', + 'adam', + 'adamax', + 'dpsgd', + 'decayed_adagrad', + 'adadelta', + 'rmsprop', +] + + +class VarWrapper(object): + def __init__(self, var, graph): + assert isinstance(var, Variable) + assert isinstance(graph, GraphWrapper) + self._var = var + self._graph = graph + + def __eq__(self, v): + """ + Overwrite this function for ...in... syntax in python. + """ + return self._var.name == v._var.name + + def name(self): + """ + Get the name of the variable. + """ + return self._var.name + + def shape(self): + """ + Get the shape of the varibale. + """ + return self._var.shape + + def set_shape(self, shape): + """ + Set the shape of the variable. + """ + self._var.desc.set_shape(shape) + + def inputs(self): + """ + Get all the operators that use this variable as output. + Returns: + list: A list of operators. + """ + ops = [] + for op in self._graph.ops(): + if self in op.all_outputs(): + ops.append(op) + return ops + + def outputs(self): + """ + Get all the operators that use this variable as input. + Returns: + list: A list of operators. + """ + ops = [] + for op in self._graph.ops(): + if self in op.all_inputs(): + ops.append(op) + return ops + + +class OpWrapper(object): + def __init__(self, op, graph): + assert isinstance(graph, GraphWrapper) + self._op = op + self._graph = graph + + def __eq__(self, op): + """ + Overwrite this function for ...in... syntax in python. + """ + return self.idx() == op.idx() + + def all_inputs(self): + """ + Get all the input variables of this operator. + """ + return [ + self._graph.var(var_name) for var_name in self._op.input_arg_names + ] + + def all_outputs(self): + """ + Get all the output variables of this operator. + """ + return [ + self._graph.var(var_name) for var_name in self._op.output_arg_names + ] + + def idx(self): + """ + Get the id of this operator. + """ + return self._op.idx + + def type(self): + """ + Get the type of this operator. + """ + return self._op.type + + def is_bwd_op(self): + """ + Whether this operator is backward op. + """ + return self.type().endswith('_grad') + + def is_opt_op(self): + """ + Whether this operator is optimizer op. + """ + return self.type() in OPTIMIZER_OPS + + def inputs(self, name): + """ + Get all the varibales by the input name. + """ + return [self._graph.var(var_name) for var_name in self._op.input(name)] + + def outputs(self, name): + """ + Get all the varibales by the output name. + """ + return [ + self._graph.var(var_name) for var_name in self._op.output(name) + ] + + def set_attr(self, key, value): + """ + Set the value of attribute by attribute's name. + + Args: + key(str): the attribute name. + value(bool|int|str|float|list): the value of the attribute. + """ + self._op._set_attr(key, value) + + def attr(self, name): + """ + Get the attribute by name. + + Args: + name(str): the attribute name. + + Returns: + bool|int|str|float|list: The attribute value. The return value + can be any valid attribute type. + """ + return self._op.attr(name) + + +class GraphWrapper(object): + """ + It is a wrapper of paddle.fluid.framework.IrGraph with some special functions + for paddle slim framework. + """ + + def __init__(self, program=None, in_nodes=[], out_nodes=[]): + """ + Args: + program(framework.Program): A program with + in_nodes(dict): A dict to indicate the input nodes of the graph. + The key is user-defined and human-readable name. + The value is the name of Variable. + out_nodes(dict): A dict to indicate the input nodes of the graph. + The key is user-defined and human-readable name. + The value is the name of Variable. + """ + super(GraphWrapper, self).__init__() + self.program = Program() if program is None else program + self.persistables = {} + self.teacher_persistables = {} + for var in self.program.list_vars(): + if var.persistable: + self.persistables[var.name] = var + self.compiled_graph = None + in_nodes = [] if in_nodes is None else in_nodes + out_nodes = [] if out_nodes is None else out_nodes + self.in_nodes = OrderedDict(in_nodes) + self.out_nodes = OrderedDict(out_nodes) + self._attrs = OrderedDict() + + def all_parameters(self): + """ + Get all the parameters in this graph. + Returns: + list: A list of VarWrapper instances. + """ + params = [] + for block in self.program.blocks: + for param in block.all_parameters(): + params.append(VarWrapper(param, self)) + return params + + def is_parameter(self, var): + """ + Whether the given variable is parameter. + Args: + var(VarWrapper): The given varibale. + """ + return isinstance(var._var, Parameter) + + def is_persistable(self, var): + """ + Whether the given variable is persistable. + Args: + var(VarWrapper): The given varibale. + """ + return var._var.persistable + + def ops(self): + """ + Return all operator nodes included in the graph as a set. + """ + ops = [] + for block in self.program.blocks: + for op in block.ops: + ops.append(OpWrapper(op, self)) + return ops + + def vars(self): + """ + Get all the variables. + """ + return [VarWrapper(var, self) for var in self.program.list_vars()] + + def var(self, name): + """ + Get the variable by variable name. + """ + return VarWrapper(self.program.global_block().var(name), self) + + def clone(self, for_test=False): + """ + Clone a new graph from current graph. + Returns: + (GraphWrapper): The wrapper of a new graph. + """ + return GraphWrapper( + self.program.clone(for_test), + copy.deepcopy(self.in_nodes), copy.deepcopy(self.out_nodes)) + + def program(self): + """ + Get the program in current wrapper. + """ + return self.program + + def pre_ops(self, op): + """ + Get all the previous operators of target operator. + Args: + op(OpWrapper): Target operator.. + Returns: + list: A list of operators. + """ + ops = [] + for p in self.ops(): + for in_var in op.all_inputs(): + if in_var in p.all_outputs(): + ops.append(p) + return ops + + def next_ops(self, op): + """ + Get all the next operators of target operator. + Args: + op(OpWrapper): Target operator.. + Returns: + list: A list of operators. + """ + ops = [] + for p in self.ops(): + for out_var in op.all_outputs(): + if out_var in p.all_inputs(): + ops.append(p) + return ops + + def get_param_by_op(self, op): + """ + Get the parameters used by target operator. + """ + assert isinstance(op, OpWrapper) + params = [] + for var in op.all_inputs(): + if isinstance(var._var, Parameter): + params.append(var) + assert len(params) > 0 + return params + + def numel_params(self): + """ + Get the number of elements in all parameters. + """ + ret = 0 + for param in self.all_parameters(): + ret += np.product(param.shape()) + return ret + + def flops(self, only_conv=False): + """ + Get the flops of current graph. + Args: + only_conv: Only calculating the conv layers. default: False. + Returns: + int: The flops of current graph. + """ + flops = 0 + for op in self.ops(): + if op.type() in ['conv2d', 'depthwise_conv2d']: + filter_shape = op.inputs("Filter")[0].shape() + input_shape = op.inputs("Input")[0].shape() + output_shape = op.outputs("Output")[0].shape() + c_out, c_in, k_h, k_w = filter_shape + _, _, h_out, w_out = output_shape + groups = op.attr("groups") + kernel_ops = k_h * k_w * (c_in / groups) + if len(op.inputs("Bias")) > 0: + with_bias = 1 + else: + with_bias = 0 + flops += 2 * h_out * w_out * c_out * (kernel_ops + with_bias) + elif op.type() == 'pool2d' and not only_conv: + input_shape = op.inputs("X")[0].shape() + output_shape = op.outputs("Out")[0].shape() + _, c_out, h_out, w_out = output_shape + k_size = op.attr("ksize") + flops += h_out * w_out * c_out * (k_size[0]**2) + + elif op.type() == 'mul' and not only_conv: + x_shape = list(op.inputs("X")[0].shape()) + y_shape = op.inputs("Y")[0].shape() + if x_shape[0] == -1: + x_shape[0] = 1 + flops += 2 * x_shape[0] * x_shape[1] * y_shape[1] + + elif op.type() in ['relu', 'sigmoid', 'batch_norm' + ] and not only_conv: + input_shape = list(op.inputs("X")[0].shape()) + if input_shape[0] == -1: + input_shape[0] = 1 + flops += np.product(input_shape) + + return flops + + def update_param_shape(self, scope): + """ + Update the shape of parameters in the graph according to tensors in scope. + It is used after loading pruned parameters from file. + """ + for param in self.all_parameters(): + tensor_shape = np.array( + scope.find_var(param.name()).get_tensor()).shape + param.set_shape(tensor_shape) + + def infer_shape(self): + """ + Update the groups of convolution layer according to current filters. + It is used after loading pruned parameters from file. + """ + for op in self.ops(): + if op.type() != 'conditional_block': + op._op.desc.infer_shape(op._op.block.desc) + + def update_groups_of_conv(self): + for op in self.ops(): + if op.type() == 'depthwise_conv2d' or op.type( + ) == 'depthwise_conv2d_grad': + op.set_attr('groups', op.inputs('Filter')[0].shape()[0]) diff --git a/paddleslim/prune/__init__.py b/paddleslim/prune/__init__.py index 9d053150..926586c6 100644 --- a/paddleslim/prune/__init__.py +++ b/paddleslim/prune/__init__.py @@ -11,3 +11,4 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. +from pruner import Pruner diff --git a/paddleslim/prune/pruner.py b/paddleslim/prune/pruner.py new file mode 100644 index 00000000..4921e460 --- /dev/null +++ b/paddleslim/prune/pruner.py @@ -0,0 +1,551 @@ +# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License" +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np +import paddle.fluid as fluid +from ..common import VarWrapper, OpWrapper, GraphWrapper + +__all__ = ["prune"] + + +class Pruner(): + def __init__(self, criterion="l1_norm"): + """ + Args: + criterion(str): the criterion used to sort channels for pruning. + It only supports 'l1_norm' currently. + """ + self.criterion = criterion + + def prune(self, + program, + scope, + params, + ratios, + place=None, + lazy=False, + only_graph=False, + param_backup=None, + param_shape_backup=None): + """ + Pruning the given parameters. + Args: + program(fluid.Program): The program to be pruned. + scope(fluid.Scope): The scope storing paramaters to be pruned. + params(list): A list of parameter names to be pruned. + ratios(list): A list of ratios to be used to pruning parameters. + place(fluid.Place): The device place of filter parameters. Defalut: None. + lazy(bool): True means setting the pruned elements to zero. + False means cutting down the pruned elements. Default: False. + only_graph(bool): True means only modifying the graph. + False means modifying graph and variables in scope. Default: False. + param_backup(dict): A dict to backup the values of parameters. Default: None. + param_shape_backup(dict): A dict to backup the shapes of parameters. Default: None. + """ + + self.pruned_list = [] + graph = GraphWrapper(program.clone()) + self._prune_parameters( + graph, + scope, + params, + ratios, + place, + lazy=False, + only_graph=False, + param_backup=None, + param_shape_backup=None) + return graph.program + + def _prune_filters_by_ratio(self, + scope, + params, + ratio, + place, + lazy=False, + only_graph=False, + param_shape_backup=None, + param_backup=None): + """ + Pruning filters by given ratio. + Args: + scope(fluid.core.Scope): The scope used to pruning filters. + params(list): A list of filter parameters. + ratio(float): The ratio to be pruned. + place(fluid.Place): The device place of filter parameters. + lazy(bool): True means setting the pruned elements to zero. + False means cutting down the pruned elements. + only_graph(bool): True means only modifying the graph. + False means modifying graph and variables in scope. + """ + if params[0].name() in self.pruned_list[0]: + return + param_t = scope.find_var(params[0].name()).get_tensor() + pruned_idx = self._cal_pruned_idx( + params[0].name(), np.array(param_t), ratio, axis=0) + for param in params: + assert isinstance(param, VarWrapper) + param_t = scope.find_var(param.name()).get_tensor() + if param_backup is not None and (param.name() not in param_backup): + param_backup[param.name()] = copy.deepcopy(np.array(param_t)) + pruned_param = self._prune_tensor( + np.array(param_t), pruned_idx, pruned_axis=0, lazy=lazy) + if not only_graph: + param_t.set(pruned_param, place) + ori_shape = param.shape() + if param_shape_backup is not None and ( + param.name() not in param_shape_backup): + param_shape_backup[param.name()] = copy.deepcopy(param.shape()) + new_shape = list(param.shape()) + new_shape[0] = pruned_param.shape[0] + param.set_shape(new_shape) + self.pruned_list[0].append(param.name()) + return pruned_idx + + def _prune_parameter_by_idx(self, + scope, + params, + pruned_idx, + pruned_axis, + place, + lazy=False, + only_graph=False, + param_shape_backup=None, + param_backup=None): + """ + Pruning parameters in given axis. + Args: + scope(fluid.core.Scope): The scope storing paramaters to be pruned. + params(VarWrapper): The parameter to be pruned. + pruned_idx(list): The index of elements to be pruned. + pruned_axis(int): The pruning axis. + place(fluid.Place): The device place of filter parameters. + lazy(bool): True means setting the pruned elements to zero. + False means cutting down the pruned elements. + only_graph(bool): True means only modifying the graph. + False means modifying graph and variables in scope. + """ + if params[0].name() in self.pruned_list[pruned_axis]: + return + for param in params: + assert isinstance(param, VarWrapper) + param_t = scope.find_var(param.name()).get_tensor() + if param_backup is not None and (param.name() not in param_backup): + param_backup[param.name()] = copy.deepcopy(np.array(param_t)) + pruned_param = self._prune_tensor( + np.array(param_t), pruned_idx, pruned_axis, lazy=lazy) + if not only_graph: + param_t.set(pruned_param, place) + ori_shape = param.shape() + + if param_shape_backup is not None and ( + param.name() not in param_shape_backup): + param_shape_backup[param.name()] = copy.deepcopy(param.shape()) + new_shape = list(param.shape()) + new_shape[pruned_axis] = pruned_param.shape[pruned_axis] + param.set_shape(new_shape) + self.pruned_list[pruned_axis].append(param.name()) + + def _forward_search_related_op(self, graph, param): + """ + Forward search operators that will be affected by pruning of param. + Args: + graph(GraphWrapper): The graph to be searched. + param(VarWrapper): The current pruned parameter. + Returns: + list: A list of operators. + """ + assert isinstance(param, VarWrapper) + visited = {} + for op in graph.ops(): + visited[op.idx()] = False + stack = [] + for op in graph.ops(): + if (not op.is_bwd_op()) and (param in op.all_inputs()): + stack.append(op) + visit_path = [] + while len(stack) > 0: + top_op = stack[len(stack) - 1] + if visited[top_op.idx()] == False: + visit_path.append(top_op) + visited[top_op.idx()] = True + next_ops = None + if top_op.type() == "conv2d" and param not in top_op.all_inputs(): + next_ops = None + elif top_op.type() == "mul": + next_ops = None + else: + next_ops = self._get_next_unvisited_op(graph, visited, top_op) + if next_ops == None: + stack.pop() + else: + stack += next_ops + return visit_path + + def _get_next_unvisited_op(self, graph, visited, top_op): + """ + Get next unvisited adjacent operators of given operators. + Args: + graph(GraphWrapper): The graph used to search. + visited(list): The ids of operators that has been visited. + top_op: The given operator. + Returns: + list: A list of operators. + """ + assert isinstance(top_op, OpWrapper) + next_ops = [] + for op in graph.next_ops(top_op): + if (visited[op.idx()] == False) and (not op.is_bwd_op()): + next_ops.append(op) + return next_ops if len(next_ops) > 0 else None + + def _get_accumulator(self, graph, param): + """ + Get accumulators of given parameter. The accumulator was created by optimizer. + Args: + graph(GraphWrapper): The graph used to search. + param(VarWrapper): The given parameter. + Returns: + list: A list of accumulators which are variables. + """ + assert isinstance(param, VarWrapper) + params = [] + for op in param.outputs(): + if op.is_opt_op(): + for out_var in op.all_outputs(): + if graph.is_persistable(out_var) and out_var.name( + ) != param.name(): + params.append(out_var) + return params + + def _forward_pruning_ralated_params(self, + graph, + scope, + param, + place, + ratio=None, + pruned_idxs=None, + lazy=False, + only_graph=False, + param_backup=None, + param_shape_backup=None): + """ + Pruning all the parameters affected by the pruning of given parameter. + Args: + graph(GraphWrapper): The graph to be searched. + scope(fluid.core.Scope): The scope storing paramaters to be pruned. + param(VarWrapper): The given parameter. + place(fluid.Place): The device place of filter parameters. + ratio(float): The target ratio to be pruned. + pruned_idx(list): The index of elements to be pruned. + lazy(bool): True means setting the pruned elements to zero. + False means cutting down the pruned elements. + only_graph(bool): True means only modifying the graph. + False means modifying graph and variables in scope. + """ + assert isinstance( + graph, + GraphWrapper), "graph must be instance of slim.core.GraphWrapper" + assert isinstance( + param, + VarWrapper), "param must be instance of slim.core.VarWrapper" + + if param.name() in self.pruned_list[0]: + return + related_ops = self._forward_search_related_op(graph, param) + + if ratio is None: + assert pruned_idxs is not None + self._prune_parameter_by_idx( + scope, [param] + self._get_accumulator(graph, param), + pruned_idxs, + pruned_axis=0, + place=place, + lazy=lazy, + only_graph=only_graph, + param_backup=param_backup, + param_shape_backup=param_shape_backup) + + else: + pruned_idxs = self._prune_filters_by_ratio( + scope, [param] + self._get_accumulator(graph, param), + ratio, + place, + lazy=lazy, + only_graph=only_graph, + param_backup=param_backup, + param_shape_backup=param_shape_backup) + corrected_idxs = pruned_idxs[:] + + for idx, op in enumerate(related_ops): + if op.type() == "conv2d" and (param not in op.all_inputs()): + for in_var in op.all_inputs(): + if graph.is_parameter(in_var): + conv_param = in_var + self._prune_parameter_by_idx( + scope, [conv_param] + self._get_accumulator( + graph, conv_param), + corrected_idxs, + pruned_axis=1, + place=place, + lazy=lazy, + only_graph=only_graph, + param_backup=param_backup, + param_shape_backup=param_shape_backup) + if op.type() == "depthwise_conv2d": + for in_var in op.all_inputs(): + if graph.is_parameter(in_var): + conv_param = in_var + self._prune_parameter_by_idx( + scope, [conv_param] + self._get_accumulator( + graph, conv_param), + corrected_idxs, + pruned_axis=0, + place=place, + lazy=lazy, + only_graph=only_graph, + param_backup=param_backup, + param_shape_backup=param_shape_backup) + elif op.type() == "elementwise_add": + # pruning bias + for in_var in op.all_inputs(): + if graph.is_parameter(in_var): + bias_param = in_var + self._prune_parameter_by_idx( + scope, [bias_param] + self._get_accumulator( + graph, bias_param), + pruned_idxs, + pruned_axis=0, + place=place, + lazy=lazy, + only_graph=only_graph, + param_backup=param_backup, + param_shape_backup=param_shape_backup) + elif op.type() == "mul": # pruning fc layer + fc_input = None + fc_param = None + for in_var in op.all_inputs(): + if graph.is_parameter(in_var): + fc_param = in_var + else: + fc_input = in_var + + idx = [] + feature_map_size = fc_input.shape()[2] * fc_input.shape()[3] + range_idx = np.array(range(feature_map_size)) + for i in corrected_idxs: + idx += list(range_idx + i * feature_map_size) + corrected_idxs = idx + self._prune_parameter_by_idx( + scope, [fc_param] + self._get_accumulator(graph, fc_param), + corrected_idxs, + pruned_axis=0, + place=place, + lazy=lazy, + only_graph=only_graph, + param_backup=param_backup, + param_shape_backup=param_shape_backup) + + elif op.type() == "concat": + concat_inputs = op.all_inputs() + last_op = related_ops[idx - 1] + for out_var in last_op.all_outputs(): + if out_var in concat_inputs: + concat_idx = concat_inputs.index(out_var) + offset = 0 + for ci in range(concat_idx): + offset += concat_inputs[ci].shape()[1] + corrected_idxs = [x + offset for x in pruned_idxs] + elif op.type() == "batch_norm": + bn_inputs = op.all_inputs() + mean = bn_inputs[2] + variance = bn_inputs[3] + alpha = bn_inputs[0] + beta = bn_inputs[1] + self._prune_parameter_by_idx( + scope, [mean] + self._get_accumulator(graph, mean), + corrected_idxs, + pruned_axis=0, + place=place, + lazy=lazy, + only_graph=only_graph, + param_backup=param_backup, + param_shape_backup=param_shape_backup) + self._prune_parameter_by_idx( + scope, [variance] + self._get_accumulator(graph, variance), + corrected_idxs, + pruned_axis=0, + place=place, + lazy=lazy, + only_graph=only_graph, + param_backup=param_backup, + param_shape_backup=param_shape_backup) + self._prune_parameter_by_idx( + scope, [alpha] + self._get_accumulator(graph, alpha), + corrected_idxs, + pruned_axis=0, + place=place, + lazy=lazy, + only_graph=only_graph, + param_backup=param_backup, + param_shape_backup=param_shape_backup) + self._prune_parameter_by_idx( + scope, [beta] + self._get_accumulator(graph, beta), + corrected_idxs, + pruned_axis=0, + place=place, + lazy=lazy, + only_graph=only_graph, + param_backup=param_backup, + param_shape_backup=param_shape_backup) + + def _prune_parameters(self, + graph, + scope, + params, + ratios, + place, + lazy=False, + only_graph=False, + param_backup=None, + param_shape_backup=None): + """ + Pruning the given parameters. + Args: + graph(GraphWrapper): The graph to be searched. + scope(fluid.core.Scope): The scope storing paramaters to be pruned. + params(list): A list of parameter names to be pruned. + ratios(list): A list of ratios to be used to pruning parameters. + place(fluid.Place): The device place of filter parameters. + pruned_idx(list): The index of elements to be pruned. + lazy(bool): True means setting the pruned elements to zero. + False means cutting down the pruned elements. + only_graph(bool): True means only modifying the graph. + False means modifying graph and variables in scope. + """ + assert len(params) == len(ratios) + self.pruned_list = [[], []] + for param, ratio in zip(params, ratios): + assert isinstance(param, str) or isinstance(param, unicode) + param = graph.var(param) + self._forward_pruning_ralated_params( + graph, + scope, + param, + place, + ratio=ratio, + lazy=lazy, + only_graph=only_graph, + param_backup=param_backup, + param_shape_backup=param_shape_backup) + ops = param.outputs() + for op in ops: + if op.type() == 'conv2d': + brother_ops = self._search_brother_ops(graph, op) + for broher in brother_ops: + for p in graph.get_param_by_op(broher): + self._forward_pruning_ralated_params( + graph, + scope, + p, + place, + ratio=ratio, + lazy=lazy, + only_graph=only_graph, + param_backup=param_backup, + param_shape_backup=param_shape_backup) + + def _search_brother_ops(self, graph, op_node): + """ + Search brother operators that was affected by pruning of given operator. + Args: + graph(GraphWrapper): The graph to be searched. + op_node(OpWrapper): The start node for searching. + Returns: + list: A list of operators. + """ + visited = [op_node.idx()] + stack = [] + brothers = [] + for op in graph.next_ops(op_node): + if (op.type() != 'conv2d') and (op.type() != 'fc') and ( + not op.is_bwd_op()): + stack.append(op) + visited.append(op.idx()) + while len(stack) > 0: + top_op = stack.pop() + for parent in graph.pre_ops(top_op): + if parent.idx() not in visited and (not parent.is_bwd_op()): + if ((parent.type() == 'conv2d') or + (parent.type() == 'fc')): + brothers.append(parent) + else: + stack.append(parent) + visited.append(parent.idx()) + + for child in graph.next_ops(top_op): + if (child.type() != 'conv2d') and (child.type() != 'fc') and ( + child.idx() not in visited) and ( + not child.is_bwd_op()): + stack.append(child) + visited.append(child.idx()) + return brothers + + def _cal_pruned_idx(self, name, param, ratio, axis): + """ + Calculate the index to be pruned on axis by given pruning ratio. + Args: + name(str): The name of parameter to be pruned. + param(np.array): The data of parameter to be pruned. + ratio(float): The ratio to be pruned. + axis(int): The axis to be used for pruning given parameter. + If it is None, the value in self.pruning_axis will be used. + default: None. + Returns: + list: The indexes to be pruned on axis. + """ + prune_num = int(round(param.shape[axis] * ratio)) + reduce_dims = [i for i in range(len(param.shape)) if i != axis] + if self.criterion == 'l1_norm': + criterions = np.sum(np.abs(param), axis=tuple(reduce_dims)) + pruned_idx = criterions.argsort()[:prune_num] + return pruned_idx + + def _prune_tensor(self, tensor, pruned_idx, pruned_axis, lazy=False): + """ + Pruning a array by indexes on given axis. + Args: + tensor(numpy.array): The target array to be pruned. + pruned_idx(list): The indexes to be pruned. + pruned_axis(int): The axis of given array to be pruned on. + lazy(bool): True means setting the pruned elements to zero. + False means remove the pruned elements from memory. + default: False. + Returns: + numpy.array: The pruned array. + """ + mask = np.zeros(tensor.shape[pruned_axis], dtype=bool) + mask[pruned_idx] = True + + def func(data): + return data[~mask] + + def lazy_func(data): + data[mask] = 0 + return data + + if lazy: + return np.apply_along_axis(lazy_func, pruned_axis, tensor) + else: + return np.apply_along_axis(func, pruned_axis, tensor) diff --git a/setup.py b/setup.py index d79620c5..86421878 100644 --- a/setup.py +++ b/setup.py @@ -33,8 +33,12 @@ with open('./requirements.txt') as f: setup_requires = f.read().splitlines() packages = [ - 'paddleslim', 'paddleslim.prune', 'paddleslim.dist', 'paddleslim.nas', - 'paddleslim.analysis', 'paddleslim.quant' + 'paddleslim', + 'paddleslim.prune', + 'paddleslim.dist', + 'paddleslim.nas', + 'paddleslim.analysis', + 'paddleslim.quant', ] setup( diff --git a/tests/layers.py b/tests/layers.py new file mode 100644 index 00000000..78ef8395 --- /dev/null +++ b/tests/layers.py @@ -0,0 +1,31 @@ +import paddle.fluid as fluid +from paddle.fluid.param_attr import ParamAttr + + +def conv_bn_layer(input, + num_filters, + filter_size, + name, + stride=1, + groups=1, + act=None): + conv = fluid.layers.conv2d( + input=input, + num_filters=num_filters, + filter_size=filter_size, + stride=stride, + padding=(filter_size - 1) // 2, + groups=groups, + act=None, + param_attr=ParamAttr(name=name + "_weights"), + bias_attr=False, + name=name + "_out") + bn_name = name + "_bn" + return fluid.layers.batch_norm( + input=conv, + act=act, + name=bn_name + '_output', + param_attr=ParamAttr(name=bn_name + '_scale'), + bias_attr=ParamAttr(bn_name + '_offset'), + moving_mean_name=bn_name + '_mean', + moving_variance_name=bn_name + '_variance', ) diff --git a/tests/test_prune.py b/tests/test_prune.py new file mode 100644 index 00000000..f86bfe6d --- /dev/null +++ b/tests/test_prune.py @@ -0,0 +1,59 @@ +import sys +sys.path.append("../") +import unittest +import paddle.fluid as fluid +from paddleslim.prune import Pruner +from layers import conv_bn_layer + + +class TestPrune(unittest.TestCase): + def test_prune(self): + main_program = fluid.Program() + startup_program = fluid.Program() + with fluid.program_guard(main_program, startup_program): + input = fluid.data(name="image", shape=[None, 3, 16, 16]) + conv1 = conv_bn_layer(input, 8, 3, "conv1") + conv2 = conv_bn_layer(conv1, 8, 3, "conv2") + sum1 = conv1 + conv2 + conv3 = conv_bn_layer(sum1, 8, 3, "conv3") + conv4 = conv_bn_layer(conv3, 8, 3, "conv4") + sum2 = conv4 + sum1 + conv5 = conv_bn_layer(sum2, 8, 3, "conv5") + conv6 = conv_bn_layer(conv5, 8, 3, "conv6") + + shapes = {} + for param in main_program.global_block().all_parameters(): + shapes[param.name] = param.shape + + place = fluid.CPUPlace() + exe = fluid.Executor(place) + scope = fluid.Scope() + exe.run(startup_program, scope=scope) + pruner = Pruner() + main_program = pruner.prune( + main_program, + scope, + params=["conv4_weights"], + ratios=[0.5], + place=place, + lazy=False, + only_graph=False, + param_backup=None, + param_shape_backup=None) + + shapes = { + "conv5_weights": (8L, 4L, 3L, 3L), + "conv1_weights": (4L, 3L, 3L, 3L), + "conv6_weights": (8L, 8L, 3L, 3L), + "conv3_weights": (8L, 4L, 3L, 3L), + "conv2_weights": (4L, 4L, 3L, 3L), + "conv4_weights": (4L, 8L, 3L, 3L) + } + + for param in main_program.global_block().all_parameters(): + if "weights" in param.name: + self.assertTrue(param.shape == shapes[param.name]) + + +if __name__ == '__main__': + unittest.main() -- GitLab