-
Notifications
You must be signed in to change notification settings - Fork 177
/
Pool.py
320 lines (290 loc) · 16 KB
/
Pool.py
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
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
# -------------------------------------------------------------------------
# Copyright (c) Microsoft Corporation. All rights reserved.
# Licensed under the MIT License. See License.txt in the project root for
# license information.
# --------------------------------------------------------------------------
import math
import numpy as np
from distutils.version import StrictVersion
from ....common._apply_operation import apply_mul, apply_div, apply_pad
from ....common._registration import register_converter
def calculate_legacy_pad_amount(H_in, pad_h, k_h, s_h):
'''
This function calculate padding amount along H-axis. It can be applied to other axes. It should be only used with
pooling conversion.
:param H_in: input dimension along H-axis
:param pad_h: padding amount at H-axis
:param k_h: kernel's H-axis dimension
:param s_h: stride along H-axis
:return: (top_padding_amount, bottom_padding_amount)
'''
# Calculate a common variable
H_temp = H_in + 2 * pad_h - k_h
# Pooling output shape under CoerML IncludeLastPixel padding mode
H_include_last_pad_out = math.ceil(H_temp / s_h) + 1
# Pooling output shape under valid padding mode
H_valid_pad_out = math.floor(H_temp / s_h) + 1
# Amount of values padded at top boundary. For max pooling, the padded value should be "-inf."
# For average pooling, we should pad zeros.
pad_t = pad_h
# Amount of values padded at bottom boundary (add extra pixels so that H_include_last_pad_out = floor( (H_adjusted_out - k_h) / stride) + 1)
if H_include_last_pad_out > H_valid_pad_out:
pad_b = pad_h + (s_h - H_temp % s_h)
else:
pad_b = pad_h
# Intermediate result with pad_t values at top and pad_b valules at bottom of the original input
H_adjusted_out = H_in + pad_t + pad_b
# Adjust padded result if the original pooling wants to cut off the last output pixel.
if (H_include_last_pad_out - 1) * s_h >= H_in + pad_h:
if H_adjusted_out % s_h == 0:
H_adjusted_out -= s_h
else:
H_adjusted_out -= H_adjusted_out % s_h
return (pad_t, H_adjusted_out - pad_t - H_in)
def create_legacy_pad(scope, input_name, output_name, H_in, W_in, k_h, k_w,
s_h, s_w, p_h, p_w, padded_value, container):
'''
This function adds one Pad operator into its last argument, which is a Container object. By feeding the output of
the created Pad operator into Pool operator under valid padding mode, we can achieve the same functionality of
CoreML' pooling under IncludeLastPixel padding mode.
:param scope:
:param input_name:
:param output_name:
:param H_in: input dimension along H-axis
:param W_in: input dimension along W-axis
:param k_h: kernel's H-axis dimension
:param k_w: kernel's W-axis dimension
:param s_h: stride along H-axis
:param s_w: stride along W-axis
:param p_h: padding amount at the beginning and the end of H-axis
:param p_w: padding amount at the beginning and the end of W-axis
:param padded_value: value used to fill padded area
:param container: Container object
'''
# Add a Pad operator to pre-process 4-D tensor
pad_t, pad_b = calculate_legacy_pad_amount(H_in, p_h, k_h, s_h)
pad_l, pad_r = calculate_legacy_pad_amount(W_in, p_w, k_w, s_w)
# CoreML pooling operator pads only their H- and W-axes. Here we assume the shape of the tensor to be padded
# is [N, C, H, W], so we have 8 padding amounts
# pads = [N_begin_index, C_begin_index, H_begin_index, W_begin_index,
# N_end_index, C_end_index, H_end_index, W_end_index]
# Because only H- and W-axes are padded in CoreML, we leave padding amounts of N- and C-axes zeros.
pads = [0, 0, pad_t, pad_l, 0, 0, pad_b, pad_r]
apply_pad(scope, input_name, output_name, container, pads=pads, value=padded_value)
# The conversion of pooling has several possible outcomes. Let's first define some symbols and then discuss their
# ONNX computational graphs case-by-case.
#
# Symbols:
# X: input 4-D tensor. It should have a shape [N, C, H, W] following CoreML's pooling definition.
# Y: output tensor identical to CorML's pooling. Its shapes depends on the pooling type applied.
#
# Case 1: global pooling
# X ---> ONNX Global Pooling ---> Y
# In this case, ONNX's pooling implementation should directly match CoreML's implementation, so it's just a naive
# translation.
#
# Case 2: local max/L2 pooling with same or valid padding
#
# X ---> ONNX Local Max/L2 Pooling ---> Y
#
# In this case, ONNX's pooling implementation should directly match CoreML's implementation, so it's just a naive
# translation.
#
# Case 3: local max/L2 pooling under CoreML's IncludeLastPixel padding
#
# X ---> Pad --> X' ---> ONNX Local Max/L2 Pooling ---> Y
#
# CoreML's IncludeLastPixel padding mode is not supported in ONNX's pooling. We combine a Pad
# operator and a pooling to simulate CoreML's behavior. In this case, the Pad takes all padding-related
# parameters from CoreML's pooling while ONNX's pooling is working under valid padding mode.
#
# Case 4: local average pooling with same or valid padding. exclude_pad_area is on.
#
# X ---> ONNX Local Average Pooling ---> Y
#
# Current ONNX pooling operator just follows Caffe2, so the padded area is naturally excluded when calculating the
# numerator and denumerator of the pixel average covered by the kernel. That is, the translation from CoreML to ONNX
# is trivial.
#
# Case 5: local average pooling with same or valid padding. exclude_pad_area is off.
#
# X ---> ONNX Local Average Pooling ---> Y' ------------> Mul ---> Y
# | ^
# | |
# '---> Scaler ---> Z ---> ONNX L1-norm Pooling ---> Z' ---'
#
# The Scaler has "alpha=0" and its "beta" is a constant. If "beta=1", the output of the L1-norm pooling, Z', is the
# effective kernel size applied at each pixel when padded area is excluded. Here we use "beta=1/kerenel_size" so
# that one value in Z' stands for
# (the kernel size without padded area) / (the kernel size with padded area)
# at a pixel. The output Y' is computed with exclude_pad_area=on, so the element-wise multiplication of Y' and Z'
# is Y.
#
# Case 6: local average pooling with IncludeLastPixel padding. exclude_pad_area is on.
#
# X ---> Pad ---> X' ---> ONNX Local Average Pooling ---> Y' ------> Div ---> Y
# | ^
# | |
# '---> Scaler ---> Z ---> Pad ---> Z' ---> ONNX L1-norm Pooling ---> Z''
#
# The Scaler has "alpha=0" and its "beta" is a constant. If "beta=1", the output of the L1-norm pooling, Z'', is
# the effective kernel size applied at each pixel when padded area is excluded (since Pad fills the
# padded area with zeros so those padded pixels are not counted by the L1-norm pooling). Here we use
# "beta=1/kerenel_size" so that one value in Z' stands for
# (the kernel size without padded area) / (the kernel size with padded area)
# at a pixel. The output Y' is computed as if exclude_pad_area=on, so the element-wise division of Y' and Z'' is Y.
#
# Case 7: local average pooling with IncludeLastPixel padding. exclude_pad_area is off.
#
# X ---> Pad --> X' ---> ONNX Local Average Pooling ---> Y
#
# Since Pad operators add zeros to X's margin and the local pooling here is working under valid padding, it's
# equivalent to the situation of exclude_pad_area=off.
def convert_pooling(scope, operator, container):
from coremltools.proto.NeuralNetwork_pb2 import PoolingLayerParams as Params
from coremltools.proto.NeuralNetwork_pb2 import SamePadding
params = operator.raw_operator.pooling
# The input of Pool
inputs = [variable.full_name for variable in operator.inputs]
# The output of Pool
outputs = [variable.full_name for variable in operator.outputs]
# Handle global pooling mode. This case if much simpler than the conversion of local pooling.
attrs = {'name': operator.full_name}
if params.globalPooling:
pooling_table = {params.MAX: 'GlobalMaxPool',
Params.AVERAGE: 'GlobalAveragePool',
Params.L2: 'GlobalLpPool'}
if params.type not in pooling_table:
raise ValueError('Unsupported pooling type: {}'.format(params.type))
op_type = pooling_table[params.type]
if params.type == Params.L2:
attrs['p'] = 2
container.add_node(op_type, inputs, outputs, op_version=2, **attrs)
else:
container.add_node(op_type, inputs, outputs, **attrs)
return
# From here to the end of this function, we will handle local pooling mode
if params.type == Params.MAX:
op_type = 'MaxPool'
op_version = 1
elif params.type == Params.AVERAGE:
op_type = 'AveragePool'
if operator.targeted_onnx_version < StrictVersion('1.2'):
op_version = 1
else:
op_version = 7
elif params.type == Params.L2:
op_type = 'LpPool'
attrs['p'] = 2
op_version = 2
else:
raise ValueError('Unsupported pooling type: {}'.format(params.type))
# CoreML default v.s. non-default parameters
kernel_shape = [3, 3] if len(params.kernelSize) <= 0 else params.kernelSize
strides = [1, 1] if len(params.stride) <= 0 else params.stride
attrs['kernel_shape'] = kernel_shape
attrs['strides'] = strides
# Set up padding attributes
pads = None
auto_pad = None
pad_type = params.WhichOneof('PoolingPaddingType')
if pad_type == 'valid':
if len(params.valid.paddingAmounts.borderAmounts) > 0:
pads = [0, 0, 0, 0]
pads[0] = params.valid.paddingAmounts.borderAmounts[0].startEdgeSize
pads[1] = params.valid.paddingAmounts.borderAmounts[1].startEdgeSize
pads[2] = params.valid.paddingAmounts.borderAmounts[0].endEdgeSize
pads[3] = params.valid.paddingAmounts.borderAmounts[1].endEdgeSize
# If padding amounts are all zero, there should be no padding list.
if all(pad == 0 for pad in pads):
pads = None
auto_pad = 'VALID'
else:
auto_pad = 'VALID'
elif pad_type == 'same':
if params.same.asymmetryMode == SamePadding.BOTTOM_RIGHT_HEAVY:
# In CoreML, BOTTOM_RIGHT_HEAVY means that the extra pixel (when not dividable by 2) will be added to the
# end of an axis. This behavior matches ONNX's SAME_UPPER mode.
# Reference: https://apple.github.io/coremltools/coremlspecification/sections/NeuralNetwork.html#samepadding
# https://github.com/onnx/onnx/blob/rel-1.2.1/docs/Operators.md#AveragePool
auto_pad = 'SAME_UPPER'
elif params.same.asymmetryMode == SamePadding.TOP_LEFT_HEAVY:
# In CoreML, TOP_LEFT_HEAVY means that the extra pixel (when not dividable by 2) will be added to the
# beginning of an axis. This behavior matches ONNX's SAME_LOWER mode.
# Reference: https://apple.github.io/coremltools/coremlspecification/sections/NeuralNetwork.html#samepadding
# https://github.com/onnx/onnx/blob/rel-1.2.1/docs/Operators.md#AveragePool
auto_pad = 'SAME_LOWER'
else:
raise ValueError('Unknown asymmetric mode: {}'.format(params.same.asymmetryMode))
elif pad_type == 'includeLastPixel':
# Here we use a Pad operator to mimic the behavior of this CoreML padding.
auto_pad = 'VALID'
H = operator.inputs[0].type.shape[2]
W = operator.inputs[0].type.shape[3]
pad_h = params.includeLastPixel.paddingAmounts[0]
pad_w = params.includeLastPixel.paddingAmounts[1]
legacy_padded_tensor_name = scope.get_unique_variable_name('legacy_padded_tensor')
padded_value = 0. if params.type != Params.MAX else 1 + np.finfo(np.float32).min
# Create a sub-graph of cases 3, 6, 7: X ---> Pad ---> X'
create_legacy_pad(scope, inputs[0], legacy_padded_tensor_name, H, W, kernel_shape[0], kernel_shape[1],
strides[0], strides[1], pad_h, pad_w, padded_value, container)
# Set the first input name to the output of Pad so that the following Pool operator won't access the
# original input.
inputs[0] = legacy_padded_tensor_name
else:
raise ValueError('Unsupported padding mode: {}'.format(pad_type))
if pads is not None:
attrs['pads'] = pads
if auto_pad is not None:
attrs['auto_pad'] = auto_pad
if (params.type == Params.AVERAGE and params.avgPoolExcludePadding and pad_type == 'includeLastPixel') or \
(params.type == Params.AVERAGE and not params.avgPoolExcludePadding and pad_type != 'includeLastPixel'):
# Case 5 & 6. See comment above.
# X --> Affine --> Z
X_name = operator.inputs[0].full_name
Y_name = operator.outputs[0].full_name
Z_name = scope.get_unique_variable_name('Z')
container.add_node('Affine', X_name, Z_name, name=scope.get_unique_operator_name('Affine'),
alpha=0., beta=1. / (kernel_shape[0] * kernel_shape[1]))
Z_prime_name = scope.get_unique_variable_name('Z_prime')
Y_prime_name = scope.get_unique_variable_name('Y_prime')
if pad_type != 'includeLastPixel':
# Create the major Pool operator.
# Associated sub-graph of case 5: X ---> Pool ---> Y'
container.add_node(op_type, inputs, Y_prime_name, **attrs)
# Create operators to calculate correction coefficients
# Associated sub-graph of case 5: Z ---> L1Pool ---> Z'
lp_pool_attrs = {'name': scope.get_unique_operator_name('LpPool'), 'kernel_shape': kernel_shape,
'strides': strides, 'p': 1}
if pads is not None:
lp_pool_attrs['pads'] = pads
if auto_pad is not None:
lp_pool_attrs['auto_pad'] = auto_pad
container.add_node('LpPool', Z_name, Z_prime_name, op_version=2, **lp_pool_attrs)
# Element-wisely apply adjustment coefficients and create the expected CoreML output
# Associated sub-graph of case 5: Y', Z' ---> Mul ---> Y
apply_mul(scope, [Y_prime_name, Z_prime_name], Y_name, container, broadcast=0)
else:
# Create the major Pool operator
# Associated sub-graph of case 6: X' ---> Pool ---> Y'
container.add_node(op_type, inputs, Y_prime_name, **attrs)
# Create operators to correct Pool's output
Y_name = operator.outputs[0].full_name
Z_prime_prime_name = scope.get_unique_variable_name('Z_prime_prime')
# Pad the constant tensor.
# Associated sub-graph of case 6: Z ---> Pad ---> Z'
create_legacy_pad(scope, Z_name, Z_prime_name, operator.inputs[0].type.shape[2],
operator.inputs[0].type.shape[3], kernel_shape[0], kernel_shape[1],
strides[0], strides[1], params.includeLastPixel.paddingAmounts[0],
params.includeLastPixel.paddingAmounts[1], 0., container)
# Associated sub-graph of case 6: Z' ---> L1Pool ---> Z''
lp_pool_attrs = {'name': scope.get_unique_operator_name('LpPool'), 'kernel_shape': kernel_shape,
'strides': strides, 'p': 1, 'audo_pad': 'VALID'}
container.add_node('LpPool', Z_prime_name, Z_prime_prime_name, op_version=2, **lp_pool_attrs)
# Element-wisely apply adjustment coefficients and create the expected CoreML output
# Associated sub-graph of case 6: Y', Z'' ---> Div ---> Y
apply_div(scope, [Y_prime_name, Z_prime_prime_name], Y_name, container, broadcast=0)
else:
# Create the major Pool operator
container.add_node(op_type, inputs, outputs, op_version=op_version, **attrs)
register_converter('pooling', convert_pooling)