transpose.py 12.9 KB
Newer Older
J
jiangjiajun 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261
import copy
import sys


class TransposeOpt:
    def __init__(self):
        self.image_layers = [
            'fluid.layers.conv2d', 'fluid.layers.batch_norm',
            'fluid.layers.conv2d_transpose', 'fluid.layers.resize_nearest',
            'fluid.layers.resize_bilinear', 'fluid.layers.pool2d',
            'fluid.layers.pad2d'
        ]
        self.direct_layers = [
            'fluid.layers.relu', 'fluid.layers.relu6', 'fluid.layers.abs',
            'fluid.layers.sigmoid', 'fluid.layers.exp', 'fluid.layers.rsqrt',
            'fluid.layers.swish_f32', 'fluid.layers.tanh',
            'fluid.layers.softplus', 'fluid.layers.leaky_relu',
            'fluid.layers.floor', 'fluid.layers.erf', 'fluid.layers.swish'
        ]
        self.elementwise_layers = [
            'fluid.layers.elementwise_add', 'fluid.layers.elementwise_sub',
            'fluid.layers.elementwise_mul', 'fluid.layers.elementwise_div'
        ]
        #        self.reduce_layers = []
        self.reduce_layers = [
            'fluid.layers.reduce_mean', 'fluid.layers.reduce_all',
            'fluid.layers.reduce_max', 'fluid.layers.reduce_any',
            'fluid.layers.reduce_sum', 'fluid.layers.reduce_prod'
        ]

    def get_transpose_num(self, graph):
        count = 0
        for layer_id, layer in graph.layers.items():
            if layer.kernel == "fluid.layers.transpose":
                count += 1
        return count

    def run(self, graph):
        total_layer_num = len(graph.layers)
        scanned_layers = set()
        optimized_transpose_layers = list()
        optimized_reduce_layers = list()
        optimized_concat_layers = list()
        optimized_elementwise_layers = list()

        def strip_transpose(_graph):
            layers = copy.deepcopy(_graph.layers)
            for layer_id, layer in layers.items():
                if layer_id in scanned_layers:
                    continue
                scanned_layers.add(layer_id)
                percent = round(len(scanned_layers) / total_layer_num * 100, 2)
                sys.stderr.write("\rOptimize Transpose Layers...{}%".format(
                    percent))

                if layer.kernel != "fluid.layers.transpose":
                    continue
                if layer.attrs["perm"] != [0, 2, 3, 1]:
                    continue
                transpose_layers = list()
                propagate_layers = list()
                reduce_layers = list()
                concat_layers = list()
                # 此elementwise_layers专用于存储shape(4) + shape(1)的形式layer
                elementwise_layers = list()
                can_be_optimized = True
                for out in _graph.edges_out.get(layer_id, []):
                    if _graph.layers[out].kernel == "fluid.layers.transpose":
                        if _graph.layers[out].attrs["perm"] != [0, 3, 1, 2]:
                            can_be_optimized = False
                            break
                        transpose_layers.append(out)
                    elif _graph.layers[out].kernel in self.elementwise_layers:
                        propagate_layers.append(out)
                    elif _graph.layers[out].kernel in self.direct_layers:
                        if _graph.layers[out].outputs[0] in _graph.outputs:
                            can_be_optimized = False
                            break
                        propagate_layers.append(out)
                    elif _graph.layers[out].kernel in self.reduce_layers:
                        if _graph.layers[out].outputs[0] in _graph.outputs:
                            can_be_optimized = False
                            break
                        if not _graph.layers[out].attrs.get('keep_dim', False):
                            can_be_optimized = False
                            break
                        propagate_layers.append(out)
                        reduce_layers.append(out)
                    elif _graph.layers[out].kernel == "fluid.layers.concat":
                        if _graph.layers[out].outputs[0] in _graph.outputs:
                            can_be_optimized = False
                            break
                        propagate_layers.append(out)
                        concat_layers.append(out)
                    else:
                        can_be_optimized = False
                        break

                visited_layers = set()
                while len(propagate_layers) > 0 and can_be_optimized:
                    current_id = propagate_layers.pop(0)
                    visited_layers.add(current_id)
                    for out in _graph.edges_out.get(current_id, []):
                        if _graph.layers[
                                out].kernel == "fluid.layers.transpose":
                            if _graph.layers[out].attrs["perm"] != [0, 3, 1, 2]:
                                can_be_optimized = False
                                break
                            transpose_layers.append(out)
                        elif _graph.layers[
                                out].kernel in self.elementwise_layers:
                            if _graph.layers[out].outputs[0] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if out not in visited_layers:
                                propagate_layers.append(out)
                        elif _graph.layers[out].kernel in self.direct_layers:
                            if _graph.layers[out].outputs[0] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if out not in visited_layers:
                                propagate_layers.append(out)
                        elif _graph.layers[out].kernel in self.reduce_layers:
                            if _graph.layers[out].outputs[0] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if not _graph.layers[out].attrs.get('keep_dim',
                                                                False):
                                can_be_optimized = False
                                break
                            if out not in visited_layers:
                                propagate_layers.append(out)
                                reduce_layers.append(out)
                        elif _graph.layers[out].kernel == "fluid.layers.concat":
                            if _graph.layers[out].outputs[0] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if out not in visited_layers:
                                propagate_layers.append(out)
                                concat_layers.append(out)
                        else:
                            can_be_optimized = False
                            break
                    for ipt in _graph.edges_in.get(current_id, []):
                        if _graph.layers[
                                current_id].kernel in self.elementwise_layers:
                            try:
                                x_shape = _graph.layers[
                                    current_id].input_shapes['x']
                                y_shape = _graph.layers[
                                    current_id].input_shapes['y']
                                if _graph.layers[ipt].outputs[
                                        0] == _graph.layers[current_id].inputs[
                                            'x']:
                                    if len(x_shape) <= 1:
                                        elementwise_layers.append(current_id)
                                        continue
                                elif _graph.layers[ipt].outputs[
                                        0] == _graph.layers[current_id].inputs[
                                            'y']:
                                    if len(y_shape) <= 1:
                                        elementwise_layers.append(current_id)
                                        continue
                                else:
                                    raise Exception(
                                        "Unexcepted situation happend while optimizing transpose"
                                    )
                            except Exception as e:
                                can_be_optimized = False
                                break
                        if _graph.layers[
                                ipt].kernel == "fluid.layers.transpose":
                            if _graph.layers[ipt].attrs["perm"] != [0, 2, 3, 1]:
                                can_be_optimized = False
                                break
                            if ipt not in visited_layers:
                                transpose_layers.append(ipt)
                        elif _graph.layers[
                                ipt].kernel in self.elementwise_layers:
                            if _graph.layers[ipt].outputs[0] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if ipt not in visited_layers:
                                propagate_layers.append(ipt)
                        elif _graph.layers[ipt].kernel in self.direct_layers:
                            if _graph.layers[ipt].outputs[0] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if ipt not in visited_layers:
                                propagate_layers.append(ipt)
                        elif _graph.layers[ipt].kernel in self.reduce_layers:
                            if _graph.layers[ipt].outputs[0] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if not _graph.layers[ipt].attrs.get('keep_dim',
                                                                False):
                                can_be_optimized = False
                                break
                            if ipt not in visited_layers:
                                propagate_layers.append(ipt)
                                reduce_layers.append(ipt)
                        elif _graph.layers[ipt].kernel == "fluid.layers.concat":
                            if _graph.layers[ipt].outputs[0] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if ipt not in visited_layers:
                                propagate_layers.append(ipt)
                                concat_layers.append(ipt)
                        else:
                            can_be_optimized = False
                            break
                    if not can_be_optimized:
                        break
                if not can_be_optimized:
                    continue

                transpose_layers.append(layer_id)
                transpose_layers = list(set(transpose_layers))
                for l in transpose_layers:
                    if graph.layers[l].outputs[0] in graph.outputs:
                        can_be_optimized = False
                        break
                if not can_be_optimized:
                    continue

                for l in transpose_layers:
                    _graph.del_layer(l)

                optimized_transpose_layers.extend(transpose_layers)
                optimized_reduce_layers.extend(reduce_layers)
                optimized_concat_layers.extend(concat_layers)
                optimized_elementwise_layers.extend(elementwise_layers)
                return True
            return False

        before_transpose_num = self.get_transpose_num(graph)
        opt_graph = copy.deepcopy(graph)
        total_layer_num = len(opt_graph.layers)

        while strip_transpose(opt_graph):
            pass

        for layer_id in list(set(optimized_transpose_layers)):
            graph.del_layer(layer_id)
        for layer_id in list(set(optimized_reduce_layers)):
            dim = graph.layers[layer_id].attrs.get('dim', None)
            if dim is not None:
                for i in range(len(dim)):
                    dim[i] = [0, 2, 3, 1][dim[i]]
                graph.layers[layer_id].attrs['dim'] = dim
        for layer_id in list(set(optimized_concat_layers)):
            axis = graph.layers[layer_id].attrs.get('axis', 0)
            graph.layers[layer_id].attrs['axis'] = [0, 2, 3, 1][axis]
        for layer_id in list(set(optimized_elementwise_layers)):
            axis = graph.layers[layer_id].attrs.get('axis', -1)
            graph.layers[layer_id].attrs['axis'] = [0, 2, 3, 1][axis]

        current_transpose_num = self.get_transpose_num(graph)
        print(
            "\nTranspose layers optimized, before: transpose_num={}, after: transpose_num={}".
            format(before_transpose_num, current_transpose_num))