Louis Verhaard | 1a92f78 | 2021-02-09 16:08:26 +0100 | [diff] [blame] | 1 | # Copyright (C) 2020-2021 Arm Limited or its affiliates. All rights reserved. |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 2 | # |
| 3 | # SPDX-License-Identifier: Apache-2.0 |
| 4 | # |
| 5 | # Licensed under the Apache License, Version 2.0 (the License); you may |
| 6 | # not use this file except in compliance with the License. |
| 7 | # You may obtain a copy of the License at |
| 8 | # |
| 9 | # www.apache.org/licenses/LICENSE-2.0 |
| 10 | # |
| 11 | # Unless required by applicable law or agreed to in writing, software |
| 12 | # distributed under the License is distributed on an AS IS BASIS, WITHOUT |
| 13 | # WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | # See the License for the specific language governing permissions and |
| 15 | # limitations under the License. |
| 16 | # |
| 17 | # Description: |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 18 | # Unit tests for tflite_graph_optimiser |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 19 | import numpy as np |
Louis Verhaard | ebf4af6 | 2021-01-27 15:57:57 +0100 | [diff] [blame] | 20 | import pytest |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 21 | |
Louis Verhaard | ae2d553 | 2020-12-11 17:19:54 +0100 | [diff] [blame] | 22 | from ethosu.vela.data_type import DataType |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 23 | from ethosu.vela.graph_optimiser import optimise_graph |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 24 | from ethosu.vela.nn_graph import NetworkType |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 25 | from ethosu.vela.operation import Op |
Louis Verhaard | ae2d553 | 2020-12-11 17:19:54 +0100 | [diff] [blame] | 26 | from ethosu.vela.operation import Padding |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 27 | from ethosu.vela.rewrite_graph import verify_graph_health |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 28 | from ethosu.vela.tensor import create_const_tensor |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 29 | from ethosu.vela.tensor import Shape4D |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 30 | from ethosu.vela.tensor import Tensor |
| 31 | from ethosu.vela.test import testutil |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 32 | from ethosu.vela.tflite_graph_optimiser import calc_explicit_padding |
| 33 | from ethosu.vela.tflite_graph_optimiser import convert_batched_fc_shape |
| 34 | from ethosu.vela.tflite_graph_optimiser import replace_pad_by_hw_pad |
| 35 | from ethosu.vela.tflite_graph_optimiser import rewrite_fully_connected_input |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 36 | |
| 37 | |
| 38 | def test_convert_batched_fc(): |
| 39 | """Tests shape conversion of batched fully connected""" |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 40 | ifm_shape = [4, 8] |
| 41 | ifm = create_const_tensor("test_in", ifm_shape, np.uint8, np.zeros(ifm_shape)) |
| 42 | w_shape = [8, 4] |
| 43 | weights = create_const_tensor("weight_in", w_shape, np.uint8, np.zeros(w_shape)) |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 44 | ofm = Tensor(ifm.shape, np.uint8, "test_out") |
| 45 | op = testutil.create_op(Op.FullyConnected, [ifm, weights], ofm) |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 46 | |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 47 | ifm.consumer_list.append(op) |
| 48 | |
| 49 | prev_op = op.clone() |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 50 | prev_op.ifm_shapes = op.ifm_shapes.copy() |
| 51 | prev_op.ofm_shapes = op.ofm_shapes.copy() |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 52 | |
Patrik Gustavsson | 2c2522d | 2021-01-29 11:51:31 +0100 | [diff] [blame] | 53 | rewrite_fully_connected_input(op, None, None) |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 54 | conv_op = convert_batched_fc_shape(op, None, None) |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 55 | assert conv_op.ifm == prev_op.ifm |
| 56 | assert conv_op.ofm == prev_op.ofm |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 57 | assert op.ifm_shapes[0] == Shape4D([1, 2, 2, 8]) |
| 58 | assert op.ofm_shapes[0] == Shape4D([1, 2, 2, 8]) |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 59 | assert conv_op.type == Op.FullyConnected |
| 60 | assert len(conv_op.ifm.shape) == 2 |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 61 | assert len(conv_op.ofm.shape) == 2 |
| 62 | assert conv_op.ifm.shape == conv_op.ofm.shape |
| 63 | |
| 64 | ifm.shape = [1, 8] |
| 65 | weights.shape = [8, 1] |
| 66 | ofm.shape = [1, 8] |
| 67 | op = testutil.create_op(Op.FullyConnected, [ifm, weights], ofm) |
| 68 | ifm.consumer_list.append(op) |
| 69 | |
| 70 | prev_op = op.clone() |
| 71 | prev_op.ifm_shapes = op.ifm_shapes.copy() |
| 72 | prev_op.ofm_shapes = op.ofm_shapes.copy() |
| 73 | |
Patrik Gustavsson | 2c2522d | 2021-01-29 11:51:31 +0100 | [diff] [blame] | 74 | rewrite_fully_connected_input(op, None, None) |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 75 | conv_op = convert_batched_fc_shape(op, None, None) |
| 76 | |
| 77 | assert conv_op.ifm == prev_op.ifm |
| 78 | assert conv_op.ofm == prev_op.ofm |
| 79 | assert op.ifm_shapes[0] == prev_op.ifm_shapes[0] |
| 80 | assert op.ofm_shapes[0] == prev_op.ofm_shapes[0] |
| 81 | assert conv_op.type == Op.FullyConnected |
| 82 | assert len(conv_op.ifm.shape) == 2 |
| 83 | assert len(conv_op.ofm.shape) == 2 |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 84 | assert conv_op.ifm.shape == conv_op.ofm.shape |
Louis Verhaard | ae2d553 | 2020-12-11 17:19:54 +0100 | [diff] [blame] | 85 | |
| 86 | |
Louis Verhaard | ebf4af6 | 2021-01-27 15:57:57 +0100 | [diff] [blame] | 87 | explicit_padding_test_data = [ |
| 88 | # Kernel size 2 |
| 89 | [(17, 1, 2, 1, 1), (1, 1)], |
| 90 | [(18, 1, 2, 0, 1), (0, 1)], |
| 91 | [(18, 1, 2, 1, 0), (1, 0)], |
| 92 | # Kernel size 3 |
| 93 | [(18, 2, 3, 1, 1), (1, 0)], |
| 94 | [(25, 2, 3, 1, 1), (1, 1)], |
| 95 | # Kernel size 4 |
| 96 | [(18, 1, 4, 1, 2), (1, 2)], |
| 97 | [(18, 1, 4, 2, 1), (2, 1)], |
| 98 | [(19, 1, 4, 2, 2), (2, 2)], |
| 99 | # Kernel size 5 |
| 100 | [(19, 1, 5, 1, 2), (1, 2)], |
| 101 | [(19, 1, 5, 0, 2), (0, 2)], |
| 102 | [(19, 1, 5, 1, 0), (1, 0)], |
| 103 | # Kernel size 21 |
| 104 | [(41, 2, 21, 8, 10), (8, 10)], |
| 105 | [(41, 3, 21, 10, 10), (10, 9)], |
| 106 | [(42, 3, 21, 10, 10), (10, 8)], |
| 107 | [(42, 3, 21, 9, 10), (9, 9)], |
| 108 | [(41, 3, 21, 10, 6), (10, 6)], |
| 109 | ] |
| 110 | |
| 111 | |
| 112 | @pytest.mark.parametrize("test_input, expected_result", explicit_padding_test_data) |
| 113 | def test_calc_explicit_padding(test_input, expected_result): |
| 114 | input_size, stride, filter_size, explicit_pad_before, explicit_pad_after = test_input |
| 115 | before, after = calc_explicit_padding(input_size, stride, filter_size, explicit_pad_before, explicit_pad_after) |
| 116 | assert (before, after) == expected_result |
| 117 | |
| 118 | |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 119 | def create_pad_and_conv2d( |
| 120 | in_shape, |
| 121 | out_shape, |
| 122 | padding, |
| 123 | in_dtype=DataType.int8, |
| 124 | out_dtype=DataType.int8, |
| 125 | pad_dtype=DataType.int32, |
| 126 | pad_setting=Padding.VALID, |
| 127 | kernel_size=3, |
| 128 | ): |
| 129 | """Creates Pad operator followed by a conv2d operator""" |
| 130 | qp = testutil.default_quant_params() |
| 131 | in0 = Tensor(in_shape, in_dtype, "in") |
| 132 | in0.quantization = qp |
| 133 | pad_tensor = create_const_tensor(name="pad", shape=list(np.shape(padding)), values=padding, dtype=pad_dtype) |
| 134 | out = Tensor(out_shape, out_dtype, "out") |
| 135 | out.quantization = qp.clone() |
| 136 | op = testutil.create_op(Op.Pad, [in0, pad_tensor], out) |
| 137 | op.run_on_npu = True |
| 138 | conv_out_tens = Tensor(in_shape, in_dtype, "output") |
| 139 | conv_out_tens.quantization = qp.clone() |
| 140 | weight_tens = Tensor([kernel_size, kernel_size, in_shape[-1], out_shape[-1]], in_dtype, "weights") |
James Peet | 7519d50 | 2021-07-19 16:47:58 +0100 | [diff] [blame] | 141 | weight_tens.values = np.zeros(weight_tens.shape, in_dtype.as_numpy_type()) |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 142 | weight_tens.quantization = qp.clone() |
| 143 | bias_tens = Tensor(out_shape, pad_dtype, "biases") |
| 144 | attrs = {"padding": pad_setting, "stride_w": 2, "stride_h": 2, "dilation_w_factor": 1, "dilation_h_factor": 1} |
| 145 | attrs["strides"] = (1, attrs["stride_h"], attrs["stride_w"], 1) |
| 146 | conv2d_op = testutil.create_op(Op.Conv2DBias, [out, weight_tens, bias_tens], conv_out_tens, attrs) |
| 147 | conv2d_op.add_input_tensor(out) |
| 148 | conv2d_op.run_on_npu = True |
| 149 | return op, conv2d_op |
| 150 | |
| 151 | |
| 152 | def test_pad_followed_by_conv_is_removed(): |
Louis Verhaard | ae2d553 | 2020-12-11 17:19:54 +0100 | [diff] [blame] | 153 | """ |
| 154 | Tests that the PAD operator is bypassed when followed by a convolution operator, |
| 155 | and that the padding of the convolution operation is correctly updated |
| 156 | """ |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 157 | pad_op, conv2d_op = create_pad_and_conv2d( |
| 158 | in_shape=[1, 76, 75, 64], out_shape=[1, 76, 75, 64], padding=[[0, 0], [2, 1], [1, 1], [0, 0]], kernel_size=4 |
| 159 | ) |
| 160 | nng = testutil.create_graph([pad_op, conv2d_op]) |
Louis Verhaard | ae2d553 | 2020-12-11 17:19:54 +0100 | [diff] [blame] | 161 | arch = testutil.create_arch() |
| 162 | |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 163 | replace_pad_by_hw_pad(conv2d_op, nng, arch) |
Louis Verhaard | ae2d553 | 2020-12-11 17:19:54 +0100 | [diff] [blame] | 164 | |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 165 | op = nng.subgraphs[0].output_tensors[0].ops[0] |
| 166 | assert op.type == Op.Conv2DBias |
Louis Verhaard | ae2d553 | 2020-12-11 17:19:54 +0100 | [diff] [blame] | 167 | assert op.attrs["padding"] == Padding.EXPLICIT |
| 168 | assert op.attrs["explicit_padding"] == (2, 1, 1, 1) |
| 169 | assert op.ifm.shape == [1, 76, 75, 64] |
| 170 | assert pad_op not in op.ifm.ops |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 171 | |
| 172 | |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 173 | leading_pad_test_data = [ |
| 174 | (2, 2, 11, True), |
| 175 | (1, 2, 11, False), |
| 176 | (2, 1, 11, False), |
| 177 | (5, 2, 11, True), |
| 178 | ] |
| 179 | |
| 180 | |
| 181 | @pytest.mark.parametrize("top, left, kernel_size, expect_pad_removed", leading_pad_test_data) |
| 182 | def test_leading_pad_size(top, left, kernel_size, expect_pad_removed): |
| 183 | # Tests PAD operator with big kernel size; top and left pad must be multiple of stride |
| 184 | out_shape = [1, 11 + left, 11 + top, 1] |
| 185 | padding = [[0, 0], [top, 0], [left, 0], [0, 0]] |
| 186 | pad_op, conv2d_op = create_pad_and_conv2d( |
| 187 | in_shape=[1, 11, 11, 1], out_shape=out_shape, padding=padding, kernel_size=kernel_size |
| 188 | ) |
| 189 | nng = testutil.create_graph([pad_op, conv2d_op]) |
| 190 | arch = testutil.create_arch() |
| 191 | replace_pad_by_hw_pad(conv2d_op, nng, arch) |
| 192 | op = nng.subgraphs[0].output_tensors[0].ops[0] |
| 193 | if expect_pad_removed: |
| 194 | assert op.attrs["padding"] == Padding.EXPLICIT |
| 195 | assert "explicit_padding" in op.attrs |
| 196 | assert op.ifm.shape == op.ofm.shape |
| 197 | assert pad_op not in op.ifm.ops |
| 198 | else: |
| 199 | assert pad_op in op.ifm.ops |
| 200 | assert op.attrs["padding"] == Padding.VALID |
| 201 | assert "explicit_padding" not in op.attrs |
| 202 | |
| 203 | |
Louis Verhaard | 1a92f78 | 2021-02-09 16:08:26 +0100 | [diff] [blame] | 204 | def test_optimise_pad_followed_by_avg_pool(): |
| 205 | """ |
| 206 | Tests that the PAD operator is bypassed when followed by a average pool operator, |
| 207 | and that the average pool is converted to a depthwise |
| 208 | """ |
| 209 | # Create Pad operation followed by AvgPool |
| 210 | quant = testutil.default_quant_params() |
| 211 | in_tens = Tensor([1, 76, 75, 64], DataType.uint8, "input") |
| 212 | in_tens.quantization = quant |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 213 | # Test with 3x2 input tensor |
| 214 | pad_input = create_const_tensor("pad_input", [3, 2], DataType.int32, [[2, 2], [1, 1], [0, 0]]) |
Louis Verhaard | 1a92f78 | 2021-02-09 16:08:26 +0100 | [diff] [blame] | 215 | temp_tens = Tensor([1, 79, 77, 64], DataType.uint8, "pad_out") |
| 216 | temp_tens.quantization = quant.clone() |
| 217 | out_tens = Tensor([1, 76, 75, 64], DataType.uint8, "output") |
| 218 | out_tens.quantization = quant.clone() |
| 219 | |
| 220 | pad_op = testutil.create_op(Op.Pad, [in_tens, pad_input], temp_tens) |
| 221 | attrs = { |
| 222 | "padding": Padding.VALID, |
| 223 | "ksize": [1, 5, 3, 1], |
| 224 | "stride_w": 2, |
| 225 | "stride_h": 2, |
| 226 | "dilation_w_factor": 1, |
| 227 | "dilation_h_factor": 1, |
| 228 | } |
| 229 | attrs["strides"] = (1, attrs["stride_h"], attrs["stride_w"], 1) |
| 230 | pad_op.run_on_npu = True |
| 231 | conv2d_op = testutil.create_op(Op.AvgPool, [temp_tens], out_tens, attrs) |
| 232 | conv2d_op.run_on_npu = True |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 233 | nng = testutil.create_graph([pad_op, conv2d_op]) |
Louis Verhaard | 1a92f78 | 2021-02-09 16:08:26 +0100 | [diff] [blame] | 234 | arch = testutil.create_arch() |
| 235 | |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 236 | replace_pad_by_hw_pad(conv2d_op, nng, arch) |
Louis Verhaard | 1a92f78 | 2021-02-09 16:08:26 +0100 | [diff] [blame] | 237 | |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 238 | op = nng.subgraphs[0].output_tensors[0].ops[0] |
Louis Verhaard | 1a92f78 | 2021-02-09 16:08:26 +0100 | [diff] [blame] | 239 | assert op.type == Op.DepthwiseConv2DBias |
| 240 | assert op.attrs["padding"] == Padding.EXPLICIT |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 241 | assert op.attrs["explicit_padding"] == (2, 1, 2, 1) |
Louis Verhaard | 1a92f78 | 2021-02-09 16:08:26 +0100 | [diff] [blame] | 242 | assert op.ifm.shape == [1, 76, 75, 64] |
| 243 | assert pad_op not in op.ifm.ops |
| 244 | # Check that bias and weight tensors have been added |
| 245 | assert op.bias.shape == [64] |
Louis Verhaard | 1a92f78 | 2021-02-09 16:08:26 +0100 | [diff] [blame] | 246 | assert op.weights.shape == [5, 3, 1, 64] |
| 247 | |
| 248 | |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 249 | pad_avg_pool_test_data = [ |
| 250 | ((3, 3), (1, 1, 1, 1), True), |
| 251 | ((3, 3), (2, 1, 1, 1), False), |
| 252 | ((3, 3), (1, 2, 1, 1), False), |
| 253 | ((3, 3), (1, 1, 2, 1), False), |
| 254 | ((3, 3), (1, 1, 1, 2), False), |
| 255 | ((2, 4), (1, 2, 1, 2), True), |
| 256 | ((5, 3), (2, 1, 2, 1), True), |
| 257 | ((5, 3), (0, 1, 2, 1), True), |
| 258 | ((5, 3), (2, 0, 2, 1), True), |
| 259 | ((5, 3), (2, 1, 0, 1), True), |
| 260 | ((5, 3), (2, 1, 0, 1), True), |
| 261 | ((4, 4), (2, 2, 2, 2), True), |
| 262 | ((4, 4), (1, 2, 2, 2), False), |
| 263 | ((4, 4), (2, 1, 2, 2), False), |
| 264 | ((4, 4), (2, 2, 1, 2), False), |
| 265 | ((4, 4), (2, 2, 2, 1), False), |
| 266 | ] |
| 267 | |
| 268 | |
| 269 | @pytest.mark.parametrize("k_size, padding, expect_pad_removed", pad_avg_pool_test_data) |
| 270 | def test_pad_followed_by_avg_pool(k_size, padding, expect_pad_removed): |
| 271 | # Tests PAD followed by AvgPool |
| 272 | k_w, k_h = k_size |
| 273 | top, left, bottom, right = padding |
| 274 | pad_values = [[0, 0], [top, bottom], [left, right], [0, 0]] |
| 275 | dtype = DataType.int8 |
| 276 | qp = testutil.default_quant_params() |
| 277 | in_shape = [1, 15, 17, 8] |
| 278 | out_shape = [1, in_shape[1] + top + bottom, in_shape[2] + left + right, in_shape[3]] |
| 279 | in0 = Tensor(in_shape, dtype, "in") |
| 280 | in0.quantization = qp |
| 281 | pad_tensor = create_const_tensor( |
| 282 | name="pad", shape=list(np.shape(pad_values)), values=pad_values, dtype=DataType.int32 |
| 283 | ) |
| 284 | out = Tensor(out_shape, dtype, "out") |
| 285 | out.quantization = qp.clone() |
| 286 | pad_op = testutil.create_op(Op.Pad, [in0, pad_tensor], out) |
| 287 | pool_out_tens = Tensor(in_shape, dtype, "output") |
| 288 | pool_out_tens.quantization = qp.clone() |
| 289 | attrs = { |
| 290 | "padding": Padding.VALID, |
| 291 | "ksize": [1, k_w, k_h, 1], |
| 292 | "stride_w": 1, |
| 293 | "stride_h": 1, |
| 294 | "dilation_w_factor": 1, |
| 295 | "dilation_h_factor": 1, |
| 296 | } |
| 297 | pool_op = testutil.create_op(Op.AvgPool, [out], pool_out_tens, attrs) |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 298 | pad_op.run_on_npu = True |
| 299 | pool_op.run_on_npu = True |
| 300 | nng = testutil.create_graph([pad_op, pool_op]) |
| 301 | arch = testutil.create_arch() |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 302 | nng = optimise_graph(nng, arch, NetworkType.TFLite) |
Louis Verhaard | c822d62 | 2021-03-11 14:59:06 +0100 | [diff] [blame] | 303 | sg = nng.subgraphs[0] |
| 304 | all_ops = sg.get_all_ops() |
| 305 | print("all_ops: ", all_ops) |
| 306 | # Pad should not be in the graph anymore, it should either have been removed or rewritten |
| 307 | assert not any(op.type == Op.Pad for op in all_ops) |
| 308 | op = nng.subgraphs[0].output_tensors[0].ops[0] |
| 309 | if expect_pad_removed: |
| 310 | # Expect rewrite to depthwise, PAD is removed |
| 311 | assert op.type == Op.DepthwiseConv2DBias |
| 312 | assert op.attrs["padding"] == Padding.EXPLICIT |
| 313 | assert any(pad > 0 for pad in op.attrs["explicit_padding"]) |
| 314 | assert op.ifm.shape == op.ofm.shape |
| 315 | # Check that bias and weight tensors have been added |
| 316 | assert len(op.bias.shape) > 0 |
| 317 | assert op.weights.shape is not None |
| 318 | else: |
| 319 | # Pad should have been rewritten to a number of average pool operations |
| 320 | assert all(op.type in (Op.AvgPool, Op.Const) for op in all_ops) |
| 321 | assert pool_op.type == Op.AvgPool |
| 322 | assert pool_op.attrs["padding"] == Padding.VALID |
| 323 | |
| 324 | |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 325 | def test_remove_reshape(): |
| 326 | """ |
| 327 | Test that the expected reshape are removed in graph_optimisation |
| 328 | """ |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 329 | |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 330 | # Create tensors and operators Test1 |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 331 | quant = testutil.default_quant_params() |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 332 | |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 333 | # create reshape1 op |
| 334 | ifm_shape = [64, 16] |
| 335 | reshape1_ofm_shape = [1, 4, 16, 16] |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 336 | reshape1_ifm = create_const_tensor("reshape1_in", ifm_shape, DataType.uint8, np.zeros(ifm_shape)) |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 337 | reshape1_ifm.quantization = quant |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 338 | reshape1_ofm = create_const_tensor("reshape1_out", reshape1_ofm_shape, DataType.uint8, np.zeros(reshape1_ofm_shape)) |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 339 | reshape1_ofm.quantization = quant |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 340 | shape_tens = create_const_tensor("reshape1_shape", [1], DataType.int32, reshape1_ofm_shape) |
| 341 | reshape1_op = testutil.create_op(Op.Reshape, [reshape1_ifm, shape_tens], reshape1_ofm, set_ifm_ofm_shapes=False) |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 342 | reshape1_op.attrs["new_shape"] = reshape1_ofm_shape |
| 343 | reshape1_op.run_on_npu = True |
| 344 | |
| 345 | # create conv op |
| 346 | conv_ofm = Tensor([1, 8, 8, 16], DataType.uint8, "output") |
| 347 | conv_ofm.quantization = quant.clone() |
| 348 | weight_tens = Tensor([1, 1, 16, 16], DataType.uint8, "weights") |
| 349 | weight_tens.values = np.zeros(weight_tens.shape, np.uint8) |
| 350 | weight_tens.quantization = quant.clone() |
| 351 | bias_tens = Tensor([16], DataType.int32, "biases") |
| 352 | |
| 353 | attrs = {"padding": Padding.SAME, "stride_w": 1, "stride_h": 1, "dilation_w_factor": 1, "dilation_h_factor": 1} |
| 354 | attrs["strides"] = (1, attrs["stride_h"], attrs["stride_w"], 1) |
| 355 | |
| 356 | conv2d_op = testutil.create_op( |
| 357 | Op.Conv2D, [reshape1_ofm, weight_tens, bias_tens], conv_ofm, attrs=attrs, set_ifm_ofm_shapes=False |
| 358 | ) |
| 359 | conv2d_op.run_on_npu = True |
| 360 | |
| 361 | # create reshape2 op |
| 362 | ofm_shape = [8, 8, 16] |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 363 | reshape2_ofm = create_const_tensor("reshape2_out", ofm_shape, DataType.uint8, np.zeros(ofm_shape)) |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 364 | reshape2_ofm.quantization = quant |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 365 | shape_tens = create_const_tensor("reshape2_shape", [1], DataType.int32, ofm_shape) |
| 366 | reshape2_op = testutil.create_op(Op.Reshape, [conv_ofm, shape_tens], reshape2_ofm, set_ifm_ofm_shapes=False) |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 367 | reshape2_op.attrs["new_shape"] = ofm_shape |
| 368 | reshape2_op.run_on_npu = True |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 369 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 370 | # Test1 no Reshape op is expected to remain in the NPU subgrapgh |
| 371 | # but first one will be put on CPU |
Patrik Gustavsson | 138d47f | 2021-02-08 10:13:48 +0100 | [diff] [blame] | 372 | # Network is Reshape-Conv-Reshape |
| 373 | # Result is Conv |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 374 | nng = testutil.create_graph([reshape1_op, conv2d_op, reshape2_op]) |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 375 | arch = testutil.create_arch() |
| 376 | assert verify_graph_health(nng) |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 377 | nng = optimise_graph(nng, arch, NetworkType.TFLite, True) |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 378 | assert verify_graph_health(nng) |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 379 | |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 380 | # Create tensors and operator Test2 |
| 381 | # create reshape op |
| 382 | reshape_ifm = create_const_tensor("reshape_in", ifm_shape, DataType.uint8, np.zeros(ifm_shape)) |
| 383 | reshape_ifm.quantization = quant |
| 384 | reshape_ofm = create_const_tensor("reshape1_out", reshape1_ofm_shape, DataType.uint8, np.zeros(reshape1_ofm_shape)) |
| 385 | reshape_ofm.quantization = quant |
| 386 | shape_tens = create_const_tensor("reshape1_shape", [1], DataType.int32, reshape1_ofm_shape) |
| 387 | reshape_op = testutil.create_op(Op.Reshape, [reshape_ifm, shape_tens], reshape_ofm, set_ifm_ofm_shapes=False) |
| 388 | reshape_op.attrs["new_shape"] = reshape1_ofm_shape |
| 389 | reshape_op.run_on_npu = True |
| 390 | |
| 391 | # Test2 Reshape ifm/ofm is sg input/output. |
| 392 | # Reshape op is expected to be replaced by a AvgPool 'NOP'. |
| 393 | # |
| 394 | # Network is Reshape |
| 395 | # expected is AvgPool |
| 396 | nng = testutil.create_graph([reshape_op]) |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 397 | assert verify_graph_health(nng) |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 398 | nng = optimise_graph(nng, arch, NetworkType.TFLite, True) |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 399 | assert verify_graph_health(nng) |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 400 | |
| 401 | |
| 402 | def test_remove_squeeze(): |
| 403 | """ |
| 404 | Tests that the expected squeeze are removed in graph_optimisation |
| 405 | """ |
| 406 | |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 407 | # Create tensors and operators Test1 |
| 408 | quant = testutil.default_quant_params() |
| 409 | |
| 410 | # create conv op |
| 411 | ifm_shape = [1, 1, 1, 1024] |
| 412 | conv_ifm = create_const_tensor("conv_in", ifm_shape, DataType.uint8, np.zeros(ifm_shape)) |
| 413 | conv_ifm.quantization = quant |
| 414 | conv_ofm = Tensor([1, 1, 1, 1001], DataType.uint8, "output") |
| 415 | conv_ofm.quantization = quant.clone() |
| 416 | weight_tens = Tensor([1, 1, 1024, 1001], DataType.uint8, "weights") |
| 417 | weight_tens.values = np.zeros(weight_tens.shape, np.uint8) |
| 418 | weight_tens.quantization = quant.clone() |
| 419 | bias_tens = Tensor([1001], DataType.int32, "biases") |
| 420 | |
| 421 | attrs = {"padding": Padding.SAME, "stride_w": 1, "stride_h": 1, "dilation_w_factor": 1, "dilation_h_factor": 1} |
| 422 | attrs["strides"] = (1, attrs["stride_h"], attrs["stride_w"], 1) |
| 423 | |
| 424 | conv2d_op = testutil.create_op( |
| 425 | Op.Conv2D, [conv_ifm, weight_tens, bias_tens], conv_ofm, attrs=attrs, set_ifm_ofm_shapes=False |
| 426 | ) |
| 427 | conv2d_op.run_on_npu = True |
| 428 | |
| 429 | # create squeeze op |
| 430 | ofm_shape = [1, 1001] |
| 431 | squeeze_ofm = create_const_tensor("squeeze_out", ofm_shape, DataType.uint8, np.zeros(ofm_shape)) |
| 432 | squeeze_ofm.quantization = quant.clone() |
| 433 | squeeze_op = testutil.create_op(Op.Squeeze, [conv_ofm], squeeze_ofm, set_ifm_ofm_shapes=False) |
| 434 | squeeze_op.attrs["squeeze_dims"] = [1, 2] |
| 435 | squeeze_op.run_on_npu = True |
| 436 | |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 437 | # Test1 no Squeeze op is expected to remain in the NPU subgrapgh |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 438 | # |
| 439 | # Network is Conv-Squeeze |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 440 | # Result is Conv |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 441 | nng = testutil.create_graph([conv2d_op, squeeze_op]) |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 442 | arch = testutil.create_arch() |
| 443 | assert verify_graph_health(nng) |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 444 | nng = optimise_graph(nng, arch, NetworkType.TFLite, True) |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 445 | assert verify_graph_health(nng) |
| 446 | |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 447 | # Create tensors and operator Test2 |
| 448 | # create squeeze op |
| 449 | ifm_shape = [1, 1, 1, 1001] |
| 450 | squeeze_ifm = create_const_tensor("squeeze_in", ifm_shape, DataType.uint8, np.zeros(ifm_shape)) |
| 451 | squeeze_ifm.quantization = quant |
| 452 | squeeze_ofm = create_const_tensor("squeeze_out", ofm_shape, DataType.uint8, np.zeros(ofm_shape)) |
| 453 | squeeze_ofm.quantization = quant.clone() |
| 454 | squeeze_op = testutil.create_op(Op.Squeeze, [squeeze_ifm], squeeze_ofm, set_ifm_ofm_shapes=False) |
| 455 | squeeze_op.attrs["squeeze_dims"] = [1, 2] |
| 456 | squeeze_op.run_on_npu = True |
| 457 | |
| 458 | # Test2 Squeeze ifm/ofm is sg input/output. |
| 459 | # Squeeze op is expected to be replaced by a AvgPool 'NOP'. |
| 460 | # |
| 461 | # Network is Squeeze |
| 462 | # expected is AvgPool |
| 463 | nng = testutil.create_graph([squeeze_op]) |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 464 | assert verify_graph_health(nng) |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame^] | 465 | nng = optimise_graph(nng, arch, NetworkType.TFLite, True) |
| 466 | assert verify_graph_health(nng) |
| 467 | |
| 468 | |
| 469 | def test_remove_expand_dims(): |
| 470 | """ |
| 471 | Tests that the expected ExpandDims are removed in graph_optimisation |
| 472 | """ |
| 473 | |
| 474 | # Create tensors and operators Test1 |
| 475 | quant = testutil.default_quant_params() |
| 476 | |
| 477 | # create ExpandDims op |
| 478 | ifm_shape = [4, 16, 16] |
| 479 | ofm_shape = [1, 4, 16, 16] |
| 480 | expand_dims_ifm = create_const_tensor("expand_dims_in", ifm_shape, DataType.uint8, np.zeros(ifm_shape)) |
| 481 | expand_dims_ifm.quantization = quant |
| 482 | expand_dims_ofm = create_const_tensor("expand_dims_out", ofm_shape, DataType.uint8, np.zeros(ofm_shape)) |
| 483 | expand_dims_ofm.quantization = quant.clone() |
| 484 | dim_tens = create_const_tensor("dim_tens", [], DataType.uint8, 1) |
| 485 | expand_dims_op = testutil.create_op( |
| 486 | Op.ExpandDims, [expand_dims_ifm, dim_tens], expand_dims_ofm, set_ifm_ofm_shapes=False |
| 487 | ) |
| 488 | expand_dims_op.run_on_npu = True |
| 489 | |
| 490 | # create conv op |
| 491 | conv_ofm = Tensor([1, 8, 8, 16], DataType.uint8, "output") |
| 492 | conv_ofm.quantization = quant.clone() |
| 493 | weight_tens = Tensor([1, 1, 16, 16], DataType.uint8, "weights") |
| 494 | weight_tens.values = np.zeros(weight_tens.shape, np.uint8) |
| 495 | weight_tens.quantization = quant.clone() |
| 496 | bias_tens = Tensor([16], DataType.int32, "biases") |
| 497 | |
| 498 | attrs = {"padding": Padding.SAME, "stride_w": 1, "stride_h": 1, "dilation_w_factor": 1, "dilation_h_factor": 1} |
| 499 | attrs["strides"] = (1, attrs["stride_h"], attrs["stride_w"], 1) |
| 500 | |
| 501 | conv2d_op = testutil.create_op( |
| 502 | Op.Conv2D, [expand_dims_ofm, weight_tens, bias_tens], conv_ofm, attrs=attrs, set_ifm_ofm_shapes=False |
| 503 | ) |
| 504 | conv2d_op.run_on_npu = True |
| 505 | |
| 506 | # Test1 no ExpandDims op is expected to remain in the NPU subgrapgh |
| 507 | # |
| 508 | # Network is ExpandDims-Conv |
| 509 | # Result is Conv |
| 510 | nng = testutil.create_graph([expand_dims_op, conv2d_op]) |
| 511 | arch = testutil.create_arch() |
| 512 | assert verify_graph_health(nng) |
| 513 | nng = optimise_graph(nng, arch, NetworkType.TFLite, True) |
| 514 | assert verify_graph_health(nng) |
| 515 | |
| 516 | # create ExpandDims op |
| 517 | expand_dims_ifm = create_const_tensor("expand_dims_in", ifm_shape, DataType.uint8, np.zeros(ifm_shape)) |
| 518 | expand_dims_ifm.quantization = quant |
| 519 | expand_dims_ofm = create_const_tensor("expand_dims_out", ofm_shape, DataType.uint8, np.zeros(ofm_shape)) |
| 520 | expand_dims_ofm.quantization = quant.clone() |
| 521 | dim_tens = create_const_tensor("dim_tens", [], DataType.uint8, 1) |
| 522 | expand_dims_op = testutil.create_op( |
| 523 | Op.ExpandDims, [expand_dims_ifm, dim_tens], expand_dims_ofm, set_ifm_ofm_shapes=False |
| 524 | ) |
| 525 | expand_dims_op.run_on_npu = True |
| 526 | |
| 527 | # Test2 ExpandDims ifm/ofm is sg input/output. |
| 528 | # ExpandDims op is expected to be replaced by a AvgPool 'NOP'. |
| 529 | # |
| 530 | # Network is ExpandDims |
| 531 | # expected is AvgPool |
| 532 | nng = testutil.create_graph([expand_dims_op]) |
| 533 | assert verify_graph_health(nng) |
| 534 | nng = optimise_graph(nng, arch, NetworkType.TFLite, True) |
Jonas Ohlsson | fbfd96e | 2021-08-25 11:38:03 +0200 | [diff] [blame] | 535 | assert verify_graph_health(nng) |