test_pass_fold_scale_axis.py 23.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements.  See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership.  The ASF licenses this file
# to you 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.
17
import numpy as np
18

19
import tvm
20
from tvm import te
Zhi committed
21 22 23
from tvm import relay
from tvm.relay import transform

24 25 26
def _get_positive_scale(size):
    return np.random.uniform(0.5, 1, size=size).astype('float32')

27

Zhi committed
28 29
def run_opt_pass(expr, opt_pass):
    assert isinstance(opt_pass, transform.Pass)
30
    mod = tvm.IRModule.from_expr(expr)
Zhi committed
31
    mod = opt_pass(mod)
32
    entry = mod["main"]
Zhi committed
33 34 35
    return entry if isinstance(expr, relay.Function) else entry.body


36 37 38
def test_fold_fwd_simple():
    """Simple testcase."""
    def before(x, conv_weight, in_bias, in_scale, channels):
39
        args = [x, conv_weight, in_bias]
40 41 42 43 44 45 46 47
        in_bias = relay.expand_dims(in_bias, axis=1, num_newaxis=2)
        x = relay.multiply(x, in_scale)
        x = relay.nn.relu(x)
        x = relay.add(x, in_bias)
        y = relay.nn.conv2d(x, conv_weight,
                            channels=channels,
                            kernel_size=(3, 3),
                            padding=(1, 1))
48

49 50 51 52
        return relay.Function(args, y)

    def expected(x, conv_weight, in_bias, in_scale, channels):
        # use a fixed order of args so alpha equal check can pass
53
        args = [x, conv_weight, in_bias]
54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71
        in_bias = relay.expand_dims(in_bias, axis=1, num_newaxis=2)
        squeezed_scale = relay.squeeze(in_scale, axis=[1,2])
        x = relay.nn.relu(x)
        in_bias = relay.divide(in_bias, relay.expand_dims(squeezed_scale, axis=1, num_newaxis=2))
        x = relay.add(x, in_bias)
        conv_weight = relay.multiply(
            conv_weight , relay.expand_dims(squeezed_scale, axis=1, num_newaxis=2))
        y = relay.nn.conv2d(x, conv_weight,
                            channels=channels,
                            kernel_size=(3, 3),
                            padding=(1, 1))
        return relay.Function(args, y)

    def check(shape, channels):
        x =  relay.var("x", shape=shape)
        in_channels = shape[1]
        weight = relay.var("weight")
        in_bias = relay.var("in_bias", shape=(in_channels,))
72
        in_scale = relay.const(_get_positive_scale((in_channels, 1, 1)))
73
        y1 = before(x, weight, in_bias, in_scale, channels)
Zhi committed
74
        y1 = run_opt_pass(y1, transform.InferType())
75 76
        type_dict = {x.name_hint:x.checked_type for x in y1.params}
        weight = relay.var("weight", type_dict["weight"])
Zhi committed
77
        y1_folded = run_opt_pass(y1, transform.ForwardFoldScaleAxis())
78
        y1_expected = expected(x, weight, in_bias, in_scale, channels)
79

Zhi committed
80 81 82
        y1_folded = run_opt_pass(y1_folded, transform.InferType())
        y1_expected = run_opt_pass(y1_expected, transform.InferType())
        assert relay.analysis.alpha_equal(y1_folded, y1_expected)
83 84 85 86 87 88 89

    check((2, 4, 10, 10), 2)


def test_fold_fwd_dual_path():
    """scale axis being consumed by two consumers"""
    def before(x, conv_weight, in_bias, in_scale, channels):
90
        args = [x, conv_weight, in_bias]
91 92 93 94 95 96 97
        x = relay.multiply(in_scale, x)
        x = relay.nn.relu(x)
        x = relay.subtract(x, in_bias)
        y1 = relay.nn.conv2d(x, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             data_layout="NHWC",
98
                             kernel_layout="HWIO",
99 100 101 102 103 104
                             groups=channels,
                             padding=(1, 1))
        y2 = relay.nn.conv2d(x, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             data_layout="NHWC",
105
                             kernel_layout="HWIO",
106 107 108 109 110 111
                             groups=channels,
                             padding=(1, 1))
        z = relay.add(y1, y2)
        return relay.Function(args, z)

    def expected(x, conv_weight, in_bias, in_scale, channels):
112
        args = [x, conv_weight, in_bias]
113 114 115 116 117 118 119 120
        x = relay.nn.relu(x)
        in_bias = relay.divide(in_bias, in_scale)
        x = relay.subtract(x, in_bias)
        y1 = relay.nn.conv2d(x,
                             relay.multiply(conv_weight, in_scale),
                             channels=channels,
                             kernel_size=(3, 3),
                             data_layout="NHWC",
121
                             kernel_layout="HWIO",
122 123 124 125 126 127 128
                             groups=channels,
                             padding=(1, 1))
        y2 = relay.nn.conv2d(x,
                             relay.multiply(conv_weight, in_scale),
                             channels=channels,
                             kernel_size=(3, 3),
                             data_layout="NHWC",
129
                             kernel_layout="HWIO",
130 131 132 133 134
                             groups=channels,
                             padding=(1, 1))
        z = relay.add(y1, y2)
        return relay.Function(args, z)

135 136 137
    def check(dshape, channels):
        x =  relay.var("x", shape=dshape)
        in_channels = dshape[-1]
138 139
        # test depthwise
        assert in_channels == channels
140 141
        wshape = (3, 3, 1, channels) # HWIO
        weight = relay.var("weight", shape=wshape)
142
        in_bias = relay.var("in_bias", shape=(in_channels,))
143
        in_scale = relay.const(_get_positive_scale(in_channels,))
144
        y1 = before(x, weight, in_bias, in_scale, channels)
Zhi committed
145 146
        y1 = run_opt_pass(y1, transform.InferType())
        y1_folded = run_opt_pass(y1, transform.ForwardFoldScaleAxis())
147 148 149
        type_dict = {x.name_hint:x.checked_type for x in y1.params}
        weight = relay.var("weight", type_dict["weight"])
        y1_expected = expected(x, weight, in_bias, in_scale, channels)
Zhi committed
150 151
        y1_expected = run_opt_pass(y1_expected, transform.InferType())
        assert relay.analysis.alpha_equal(y1_folded, y1_expected)
152 153 154 155 156 157 158 159 160 161 162 163 164 165 166

    check((2, 4, 10, 3), 3)


def test_fold_fwd_fail():
    """testcase where we canont fold"""
    def before(x, conv_weight, in_bias, in_scale, channels):
        x = relay.multiply(x, in_scale)
        xx = relay.nn.leaky_relu(x, alpha=0.1)
        y1 = relay.nn.conv2d(xx, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             data_layout="NHWC",
                             padding=(1, 1))
        z = relay.add(y1, x)
Zhi committed
167
        return relay.Function(relay.analysis.free_vars(z), z)
168 169 170 171 172 173 174 175

    def check(shape, channels):
        x =  relay.var("x", shape=shape)
        in_channels = shape[-1]
        # test depthwise
        assert in_channels == channels
        weight = relay.var("weight")
        in_bias = relay.var("in_bias", shape=(in_channels,))
176
        in_scale = relay.const(_get_positive_scale(size=(in_channels,)))
177
        y1 = before(x, weight, in_bias, in_scale, channels)
Zhi committed
178 179 180
        y1 = run_opt_pass(y1, transform.InferType())
        y1_folded = run_opt_pass(y1, transform.ForwardFoldScaleAxis())
        assert relay.analysis.alpha_equal(y1, y1_folded)
181 182 183 184

    check((2, 11, 10, 4), 4)


185 186 187 188 189 190 191 192 193 194 195
def test_fold_fwd_relu_fail():
    """testcase where we canont fold because scale can not pass relu"""
    def before(x, conv_weight, in_bias, in_scale, channels):
        x = relay.multiply(x, in_scale)
        xx = relay.nn.relu(x)
        y1 = relay.nn.conv2d(xx, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             data_layout="NHWC",
                             padding=(1, 1))
        z = relay.add(y1, x)
Zhi committed
196
        return relay.Function(relay.analysis.free_vars(z), z)
197 198 199 200 201 202 203 204 205

    def check(shape, channels, in_scale):
        x =  relay.var("x", shape=shape)
        in_channels = shape[-1]
        # test depthwise
        assert in_channels == channels
        weight = relay.var("weight")
        in_bias = relay.var("in_bias", shape=(in_channels,))
        y1 = before(x, weight, in_bias, in_scale, channels)
Zhi committed
206 207 208
        y1 = run_opt_pass(y1, transform.InferType())
        y1_folded = run_opt_pass(y1, transform.ForwardFoldScaleAxis())
        assert relay.analysis.alpha_equal(y1, y1_folded)
209 210 211

    in_scale = relay.var("in_scale", shape=(4,))
    check((2, 11, 10, 4), 4, in_scale)
212
    in_scale = relay.const(-_get_positive_scale((4,)))
213 214 215
    check((2, 11, 10, 4), 4, in_scale)


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
def test_fold_fwd_negative_scale():
    """Testcase of folding negative scale"""
    def before(x, conv_weight, in_scale, channels):
        args = [x, conv_weight]
        x = relay.multiply(x, in_scale)
        y = relay.nn.conv2d(x, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             padding=(1, 1))
        return relay.Function(args, y)

    def expected(x, conv_weight, in_scale, channels):
        # use a fixed order of args so alpha equal check can pass
        args = [x, conv_weight]
        squeezed_scale = relay.squeeze(in_scale, axis=[1,2])
        conv_weight = relay.multiply(
            conv_weight , relay.expand_dims(squeezed_scale, axis=1, num_newaxis=2))
        y = relay.nn.conv2d(x,
                             conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             padding=(1, 1))
        return relay.Function(args, y)

    def check(shape, channels):
        x =  relay.var("x", shape=shape)
        in_channels = shape[1]
        in_scale = relay.const(-_get_positive_scale((in_channels, 1, 1)))
        weight = relay.var("weight")
        y1 = before(x, weight, in_scale, channels)
Zhi committed
246
        y1 = run_opt_pass(y1, transform.InferType())
247 248
        type_dict = {x.name_hint:x.checked_type for x in y1.params}
        weight = relay.var("weight", type_dict["weight"])
Zhi committed
249
        y1_folded = run_opt_pass(y1, transform.ForwardFoldScaleAxis())
250
        y1_expected = expected(x, weight, in_scale, channels)
Zhi committed
251 252
        y1_expected = run_opt_pass(y1_expected, transform.InferType())
        assert relay.analysis.alpha_equal(y1_folded, y1_expected)
253 254 255 256

    check((2, 4, 10, 10), 4)


257 258 259
def test_fold_bwd_simple():
    """Simple testcase."""
    def before(x, conv_weight, out_bias, out_scale, channels):
260
        args = [x, conv_weight, out_bias]
261 262 263 264 265 266 267 268 269 270 271 272
        out_bias = relay.expand_dims(out_bias, axis=1, num_newaxis=2)
        y = relay.nn.conv2d(x, conv_weight,
                            channels=channels,
                            kernel_size=(3, 3),
                            padding=(1, 1))
        y = relay.add(y, out_bias)
        y = relay.nn.relu(y)
        y = relay.multiply(y, out_scale)
        return relay.Function(args, y)

    def expected(x, conv_weight, out_bias, out_scale, channels):
        # use a fixed order of args so alpha equal check can pass
273
        args = [x, conv_weight, out_bias]
274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293
        out_bias = relay.expand_dims(out_bias, axis=1, num_newaxis=2)
        squeezed_scale = relay.squeeze(out_scale, axis=[1,2])
        conv_weight = relay.multiply(
            conv_weight , relay.expand_dims(squeezed_scale, axis=1, num_newaxis=3))

        y = relay.nn.conv2d(x, conv_weight,
                            channels=channels,
                            kernel_size=(3, 3),
                            padding=(1, 1))
        out_bias = relay.multiply(out_bias,
                                  relay.expand_dims(squeezed_scale, axis=1, num_newaxis=2))
        y = relay.add(y, out_bias)
        y = relay.nn.relu(y)
        return relay.Function(args, y)

    def check(shape, channels):
        x =  relay.var("x", shape=shape)
        in_channels = shape[1]
        weight = relay.var("weight")
        out_bias = relay.var("out_bias", shape=(channels,))
294
        out_scale = relay.const(_get_positive_scale((channels, 1, 1)))
295 296

        y1 = before(x, weight, out_bias, out_scale, channels)
Zhi committed
297
        y1 = run_opt_pass(y1, transform.InferType())
298 299
        type_dict = {x.name_hint:x.checked_type for x in y1.params}
        weight = relay.var("weight", type_dict["weight"])
Zhi committed
300
        y1_folded = run_opt_pass(y1, transform.BackwardFoldScaleAxis())
301
        y1_expected = expected(x, weight, out_bias, out_scale, channels)
Zhi committed
302 303
        y1_expected = run_opt_pass(y1_expected, transform.InferType())
        assert relay.analysis.alpha_equal(y1_folded, y1_expected)
304 305 306 307 308 309 310

    check((2, 4, 10, 10), 8)


def test_fold_bwd_dual_path():
    """Dual path testcase."""
    def before(x, conv_weight, out_bias, out_scale, channels):
311
        args = [x, conv_weight, out_bias]
312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327
        y1 = relay.nn.conv2d(x, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             padding=(1, 1))
        y1 = relay.nn.relu(y1)
        y2 = relay.nn.conv2d(x, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             padding=(1, 1))
        y2 = relay.nn.relu(y2)
        y = relay.add(y1, y2)
        y = relay.multiply(y, out_scale)
        return relay.Function(args, y)

    def expected(x, conv_weight, out_bias, out_scale, channels):
        # use a fixed order of args so alpha equal check can pass
328
        args = [x, conv_weight, out_bias]
329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352
        out_bias = relay.expand_dims(out_bias, axis=1, num_newaxis=2)
        squeezed_scale = relay.squeeze(out_scale, axis=[1,2])
        def fold_conv_weight():
            return  relay.multiply(
                conv_weight ,
                relay.expand_dims(squeezed_scale, axis=1, num_newaxis=3))
        y1 = relay.nn.conv2d(x, fold_conv_weight(),
                            channels=channels,
                            kernel_size=(3, 3),
                            padding=(1, 1))
        y1 = relay.nn.relu(y1)
        y2 = relay.nn.conv2d(x, fold_conv_weight(),
                            channels=channels,
                            kernel_size=(3, 3),
                            padding=(1, 1))
        y2 = relay.nn.relu(y2)
        y = relay.add(y1, y2)
        return relay.Function(args, y)

    def check(shape, channels):
        x =  relay.var("x", shape=shape)
        in_channels = shape[1]
        weight = relay.var("weight")
        out_bias = relay.var("out_bias", shape=(channels,))
353
        out_scale = relay.const(_get_positive_scale((channels, 1, 1)))
354 355

        y1 = before(x, weight, out_bias, out_scale, channels)
Zhi committed
356
        y1 = run_opt_pass(y1, transform.InferType())
357 358
        type_dict = {x.name_hint:x.checked_type for x in y1.params}
        weight = relay.var("weight", type_dict["weight"])
Zhi committed
359
        y1_folded = run_opt_pass(y1, transform.BackwardFoldScaleAxis())
360
        y1_expected = expected(x, weight, out_bias, out_scale, channels)
Zhi committed
361 362
        y1_expected = run_opt_pass(y1_expected, transform.InferType())
        assert relay.analysis.alpha_equal(y1_folded, y1_expected)
363 364 365 366

    check((2, 4, 10, 10), 8)


367 368
def test_fold_bwd_dual_consumer():
    def before(x, conv_weight, out_bias, out_scale, channels):
369
        args = [x, conv_weight, out_bias]
370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395
        y0 = relay.nn.conv2d(x, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             padding=(1, 1))
        y0 = relay.multiply(y0, out_scale)
        y0 = relay.nn.relu(y0)

        y1 = relay.nn.conv2d(y0, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             padding=(1, 1))
        y1 = relay.multiply(y1, out_scale)
        y1 = relay.nn.relu(y1)

        y2 = relay.nn.conv2d(y0, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             padding=(1, 1))
        y2 = relay.multiply(y2, out_scale)
        y2 = relay.nn.relu(y2)

        y = relay.add(y1, y2)
        return relay.Function(args, y)

    def expected(x, conv_weight, out_bias, out_scale, channels):
        # use a fixed order of args so alpha equal check can pass
396
        args = [x, conv_weight, out_bias]
397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424
        def fold_conv_weight():
            squeezed_scale = relay.squeeze(out_scale, axis=[1,2])
            return  relay.multiply(
                conv_weight ,
                relay.expand_dims(squeezed_scale, axis=1, num_newaxis=3))
        y0 = relay.nn.conv2d(x, fold_conv_weight(),
                            channels=channels,
                            kernel_size=(3, 3),
                            padding=(1, 1))
        y0 = relay.nn.relu(y0)
        y1 = relay.nn.conv2d(y0, fold_conv_weight(),
                            channels=channels,
                            kernel_size=(3, 3),
                            padding=(1, 1))
        y1 = relay.nn.relu(y1)
        y2 = relay.nn.conv2d(y0, fold_conv_weight(),
                            channels=channels,
                            kernel_size=(3, 3),
                            padding=(1, 1))
        y2 = relay.nn.relu(y2)
        y = relay.add(y1, y2)
        return relay.Function(args, y)

    def check(shape, channels):
        x =  relay.var("x", shape=shape)
        in_channels = shape[1]
        weight = relay.var("weight")
        out_bias = relay.var("out_bias", shape=(channels,))
425
        out_scale = relay.const(_get_positive_scale((channels,1, 1)))
426 427

        y1 = before(x, weight, out_bias, out_scale, channels)
Zhi committed
428
        y1 = run_opt_pass(y1, transform.InferType())
429 430
        type_dict = {x.name_hint:x.checked_type for x in y1.params}
        weight = relay.var("weight", type_dict["weight"])
Zhi committed
431
        y1_folded = run_opt_pass(y1, transform.BackwardFoldScaleAxis())
432
        y1_expected = expected(x, weight, out_bias, out_scale, channels)
Zhi committed
433 434
        y1_expected = run_opt_pass(y1_expected, transform.InferType())
        assert relay.analysis.alpha_equal(y1_folded, y1_expected)
435 436 437 438

    check((2, 4, 10, 10), 4)


439 440 441
def test_fold_bwd_fail():
    """Dual path testcase."""
    def fail1(x, conv_weight, out_bias, out_scale, channels):
442
        args = [x, conv_weight, out_bias]
443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461
        out_bias = relay.expand_dims(out_bias, axis=1, num_newaxis=2)
        y1 = relay.nn.conv2d(x, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             padding=(1, 1))
        y1 = relay.nn.relu(y1)
        y2 = relay.nn.conv2d(x, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             padding=(1, 1),
                             out_layout="CNHW")
        # fold will fail because the axis from two path
        # differs from each other.
        y2 = relay.nn.relu(y2)
        y = relay.add(y1, y2)
        y = relay.multiply(y, out_scale)
        return relay.Function(args, y)

    def fail2(x, conv_weight, out_bias, out_scale, channels):
462
        args = [x, conv_weight, out_bias]
463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478
        out_bias = relay.expand_dims(out_bias, axis=1, num_newaxis=2)
        y1 = relay.nn.conv2d(x, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             padding=(1, 1))
        y2 = relay.nn.relu(y1)
        # fold will fail because y1 is referred also by y2
        y1 = relay.multiply(y1, out_scale)
        y = relay.add(y1, y2)
        return relay.Function(args, y)

    def check(shape, channels, fbefore):
        x =  relay.var("x", shape=shape)
        in_channels = shape[1]
        weight = relay.var("weight")
        out_bias = relay.var("out_bias", shape=(channels,))
479
        out_scale = relay.const(_get_positive_scale((channels, 1, 1)))
480
        y1 = fbefore(x, weight, out_bias, out_scale, channels)
Zhi committed
481 482 483
        y1 = run_opt_pass(y1, transform.InferType())
        y1_folded = run_opt_pass(y1, transform.BackwardFoldScaleAxis())
        assert relay.analysis.alpha_equal(y1_folded, y1)
484 485 486 487 488

    check((4, 4, 10, 10), 4, fail1)
    check((4, 4, 10, 10), 4, fail2)


489 490 491 492 493 494 495 496 497 498
def test_fold_bwd_relu_fail():
    """testcase where we canont fold because scale can not pass relu"""
    def before(x, conv_weight, out_scale, channels):
        y = relay.nn.conv2d(x, conv_weight,
                             channels=channels,
                             kernel_size=(3, 3),
                             data_layout="NCHW",
                             padding=(1, 1))
        y = relay.nn.relu(y)
        y = relay.multiply(x, out_scale)
Zhi committed
499
        return relay.Function(relay.analysis.free_vars(y), y)
500 501 502 503 504 505

    def check(shape, channels, out_scale):
        x =  relay.var("x", shape=shape)
        in_channels = shape[1]
        weight = relay.var("weight")
        y1 = before(x, weight, out_scale, channels)
Zhi committed
506 507 508
        y1 = run_opt_pass(y1, transform.InferType())
        y1_folded = run_opt_pass(y1, transform.BackwardFoldScaleAxis())
        assert relay.analysis.alpha_equal(y1, y1_folded)
509 510 511 512 513 514 515

    out_scale = relay.var("in_scale", shape=(4, 1, 1))
    check((4, 4, 10, 10), 4, out_scale)
    out_scale = relay.const(np.random.uniform(size=(4, 1, 1), low=-1.0, high=0.0)).astype("float32")
    check((4, 4, 10, 10), 4, out_scale)


516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543
def test_fold_bwd_negative_scale():
    """Testcase of folding negative scale"""
    def before(x, conv_weight, out_scale, channels):
        args = [x, conv_weight]
        y = relay.nn.conv2d(x, conv_weight,
                            channels=channels,
                            kernel_size=(3, 3),
                            padding=(1, 1))
        y = relay.multiply(y, out_scale)
        return relay.Function(args, y)

    def expected(x, conv_weight, out_scale, channels):
        # use a fixed order of args so alpha equal check can pass
        args = [x, conv_weight]
        squeezed_scale = relay.squeeze(out_scale, axis=[1,2])
        conv_weight = relay.multiply(
            conv_weight , relay.expand_dims(squeezed_scale, axis=1, num_newaxis=3))
        y = relay.nn.conv2d(x, conv_weight,
                            channels=channels,
                            kernel_size=(3, 3),
                            padding=(1, 1))
        return relay.Function(args, y)

    def check(shape, channels):
        x =  relay.var("x", shape=shape)
        weight = relay.var("weight")
        out_scale = relay.const(-_get_positive_scale((channels, 1, 1)))
        y1 = before(x, weight, out_scale, channels)
Zhi committed
544
        y1 = run_opt_pass(y1, transform.InferType())
545 546
        type_dict = {x.name_hint:x.checked_type for x in y1.params}
        weight = relay.var("weight", type_dict["weight"])
Zhi committed
547
        y1_folded = run_opt_pass(y1, transform.BackwardFoldScaleAxis())
548
        y1_expected = expected(x, weight, out_scale, channels)
Zhi committed
549 550
        y1_expected = run_opt_pass(y1_expected, transform.InferType())
        assert relay.analysis.alpha_equal(y1_folded, y1_expected)
551 552 553 554

    check((2, 4, 10, 10), 8)


555 556 557 558
if __name__ == "__main__":
    test_fold_fwd_simple()
    test_fold_fwd_dual_path()
    test_fold_fwd_fail()
559
    test_fold_fwd_relu_fail()
560
    test_fold_fwd_negative_scale()
561 562
    test_fold_bwd_simple()
    test_fold_bwd_dual_path()
563
    test_fold_bwd_dual_consumer()
564
    test_fold_bwd_fail()
565
    test_fold_bwd_relu_fail()
566
    test_fold_bwd_negative_scale()