model-optimizer/mo/ops/pooling.py

   1 """
   2  Copyright (c) 2018 Intel Corporation
   3
   4  Licensed under the Apache License, Version 2.0 (the "License");
   5  you may not use this file except in compliance with the License.
   6  You may obtain a copy of the License at
   7
   8       http://www.apache.org/licenses/LICENSE-2.0
   9
  10  Unless required by applicable law or agreed to in writing, software
  11  distributed under the License is distributed on an "AS IS" BASIS,
  12  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13  See the License for the specific language governing permissions and
  14  limitations under the License.
  15 """
  16
  17 import networkx as nx
  18 import numpy as np
  19
  20 from mo.front.common.partial_infer.utils import tf_window_op_pad_infer
  21 from mo.front.extractor import attr_getter
  22 # from mo.front.common.partial_infer.pooling import pool_explicit_padding_infer
  23 from mo.front.extractor import spatial_getter
  24 from mo.front.onnx.extractors.utils import get_backend_pad
  25 from mo.graph.graph import Node
  26 from mo.ops.op import Op, PermuteAttrs
  27
  28
  29 class Pooling(Op):
  30     op = 'Pooling'
  31
  32     def __init__(self, graph: nx.MultiDiGraph, attrs: dict):
  33         super().__init__(graph, {
  34             'kind': 'op',
  35             'type': __class__.op,
  36             'op': __class__.op,
  37             'infer': __class__.infer,
  38         }, attrs)
  39
  40     def backend_attrs(self):
  41         return [
  42             ('strides', lambda node: ','.join(map(str, node['stride'][node.spatial_dims]))),
  43             ('kernel', lambda node: ','.join(map(str, node['window'][node.spatial_dims]))),
  44
  45             ('pads_begin', lambda node: ','.join(map(str, get_backend_pad(node.pad, node.spatial_dims, 0)))),
  46             ('pads_end', lambda node: ','.join(map(str, get_backend_pad(node.pad, node.spatial_dims, 1)))),
  47
  48             ('pool-method', 'pool_method'),
  49             ('exclude-pad', 'exclude_pad'),
  50
  51             'rounding_type',
  52             'auto_pad',
  53         ]
  54
  55     def backend_attrs_v2(self):
  56         return [
  57             ('stride', lambda node: attr_getter(node, 'stride')),
  58
  59             spatial_getter('stride-x', 'stride', 1),
  60             spatial_getter('stride-y', 'stride', 0),
  61             spatial_getter('kernel-x', 'window', 1),
  62             spatial_getter('kernel-y', 'window', 0),
  63             spatial_getter('pad-x', 'pad', 1, lambda x: x[0]),
  64             spatial_getter('pad-y', 'pad', 0, lambda x: x[0]),
  65
  66             ('pool-method', 'pool_method'),
  67             ('exclude-pad', 'exclude_pad'),
  68
  69             'rounding_type',
  70             'auto_pad',
  71         ]
  72
  73     @staticmethod
  74     def infer(node: Node):
  75         assert (len(node.in_nodes()) == 1)
  76         input_shape = node.in_node(0).shape
  77         if input_shape is None:
  78             return
  79
  80         if not node.has_valid('spatial_dims'):
  81             node['spatial_dims'] = np.delete([x for x in range(len(input_shape))],
  82                                              [node.batch_dims[0], node.channel_dims[0]])
  83
  84         input_spatial_shape = input_shape[node.spatial_dims]
  85
  86         # Setting default pad and stride attrs in case of None specified
  87         if not node.has_valid('pad'):
  88             node['pad'] = np.array([[0, 0] for x in range(len(input_shape))], dtype=np.int64)
  89         if not node.has_valid('pad_spatial_shape'):
  90             node['pad_spatial_shape'] = node.pad[node.spatial_dims]
  91         if not node.has_valid('stride'):
  92             node['stride'] = np.array([1 for x in range(len(input_shape))], dtype=np.int64)
  93
  94         if node.has_and_set('global_pool'):
  95             node.window[node.spatial_dims] = input_spatial_shape
  96         window_spatial_shape = node.window[node.spatial_dims]
  97         stride_spatial = node.stride[node.spatial_dims]
  98         assert any(stride_spatial), 'Stride can not be zero in node {}'.format(node.id)
  99
 100         if node.has_valid('auto_pad'):
 101             node.pad_spatial_shape, node.output_spatial_shape = tf_window_op_pad_infer(input_spatial_shape,
 102                                                                                        window_spatial_shape,
 103                                                                                        stride_spatial, node.auto_pad)
 104             pad = np.zeros((len(input_shape), 2), dtype=np.int64)
 105             pad[node.spatial_dims] = node.pad_spatial_shape
 106             node.pad = pad
 107         else:
 108
 109             pad_spatial_shape = np.add.reduce(node.pad_spatial_shape, axis=1)
 110
 111             rounding = np.floor
 112             if node.has_valid('pooling_convention') and node.pooling_convention == 'full':
 113                 rounding = np.ceil
 114             output_spatial_shape = np.array(rounding(
 115                 np.array(input_spatial_shape + pad_spatial_shape - window_spatial_shape,
 116                          dtype=np.float) / stride_spatial),
 117                 dtype=np.int64) + 1
 118
 119             original_pads = np.array([i[1] for i in node.pad_spatial_shape])
 120
 121             for i in range(len(input_spatial_shape)):
 122                 if original_pads[i] and (output_spatial_shape[i] - 1) * stride_spatial[i] >= \
 123                         input_spatial_shape[i] + original_pads[i]:
 124                     output_spatial_shape[i] -= 1
 125
 126             node['output_spatial_shape'] = output_spatial_shape
 127
 128         output_shape = input_shape.copy()
 129         output_shape[node.spatial_dims] = node.output_spatial_shape
 130         node.out_node().shape = output_shape
 131
 132         # Add permute_attrs
 133         PermuteAttrs.create_permute_attrs(node, attrs=[('pad', 'input:0'),
 134                                                        ('stride', 'input:0'),
 135                                                        ('window', 'input:0'),
 136                                                        ('spatial_dims', 'input:0')])