You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2022/11/11 02:22:57 UTC
[GitHub] [tvm] wzh99 opened a new issue, #13359: [Bug] Inconsistent results of a Relay program at optimization level 3
wzh99 opened a new issue, #13359:
URL: https://github.com/apache/tvm/issues/13359
### Expected behavior
The following Relay program should produce consistent results between optimization level 0 and 3:
```
#[version = "0.0.5"]
def @main(%x0: Tensor[(1, 4, 3, 2), float32] /* ty=Tensor[(1, 4, 3, 2), float32] span=from_string:4:15 */, %x1: float32 /* ty=float32 span=from_string:4:20 */, %x2: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:7:31 */, %x3: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:7:36 */, %x4: Tensor[(4, 1, 1, 3, 2), float32] /* ty=Tensor[(4, 1, 1, 3, 2), float32] span=from_string:13:24 */, %x5: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:14:28 */, %x6: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:14:33 */, %x7: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:14:38 */, %x8: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:14:43 */, %x9: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:16:28 */, %x10: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:16:33 */, %x11: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:16:39 */, %x12: Tenso
r[(4), float32] /* ty=Tensor[(4), float32] span=from_string:16:45 */, %x13: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:22:33 */, %x14: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:22:39 */) {
%0 = nn.pad(%x0, %x1, pad_width=[[0, 0], [0, 0], [2, 2], [1, 0]]) /* ty=Tensor[(1, 4, 7, 3), float32] span=from_string:5:20 */;
%1 = expand_dims(%0, axis=3) /* ty=Tensor[(1, 4, 7, 1, 3), float32] span=from_string:19:13 */;
%2 = nn.adaptive_avg_pool3d(%1, output_size=[1, 1, 1]) /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:8:17 */;
%3 = nn.instance_norm(%2, %x2, %x3) /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:8:13 */;
%4 = (%3, %3, %2) /* ty=(Tensor[(1, 4, 1, 1, 1), float32], Tensor[(1, 4, 1, 1, 1), float32], Tensor[(1, 4, 1, 1, 1), float32]) span=from_string:9:20 */;
%5 = concatenate(%4, axis=1) /* ty=Tensor[(1, 12, 1, 1, 1), float32] span=from_string:10:12 */;
%6 = cos(%5) /* ty=Tensor[(1, 12, 1, 1, 1), float32] span=from_string:11:12 */;
%7 = sum(%6, axis=[1, 2, 3], keepdims=True) /* ty=Tensor[(1, 1, 1, 1, 1), float32] span=from_string:12:15 */;
%8 = round(%7) /* ty=Tensor[(1, 1, 1, 1, 1), float32] span=from_string:20:23 */;
%9 = nn.conv3d(%8, %x4, strides=[1, 2, 2], padding=[0, 1, 2, 0, 1, 0], dilation=[1, 1, 2], channels=4, kernel_size=[1, 3, 2]) /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:14:23 */;
%10 = nn.batch_norm(%9, %x5, %x6, %x7, %x8, center=False, scale=False) /* ty=(Tensor[(1, 4, 1, 1, 1), float32], Tensor[(4), float32], Tensor[(4), float32]) span=from_string:15:9 */;
%11 = %10.0 /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:16:23 */;
%12 = nn.batch_norm(%11, %x9, %x10, %x11, %x12, center=False, scale=False) /* ty=(Tensor[(1, 4, 1, 1, 1), float32], Tensor[(4), float32], Tensor[(4), float32]) span=from_string:17:9 */;
%13 = %12.0 /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:18:13 */;
%14 = sum(%1, axis=[2], keepdims=True) /* ty=Tensor[(1, 4, 1, 1, 3), float32] span=from_string:20:18 */;
%15 = exp(%13) /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:21:13 */;
%16 = maximum(%14, %8) /* ty=Tensor[(1, 4, 1, 1, 3), float32] span=from_string:21:18 */;
%17 = add(%15, %16) /* ty=Tensor[(1, 4, 1, 1, 3), float32] span=from_string:22:28 */;
%18 = nn.instance_norm(%17, %x13, %x14) /* ty=Tensor[(1, 4, 1, 1, 3), float32] span=from_string:23:4 */;
(%18,) /* ty=(Tensor[(1, 4, 1, 1, 3), float32],) span=from_string:4:3 */
}
```
### Actual behavior
The results are inconsistent:
```
Reference:
[[[[[0.0242452 0.0242452 0.0242452 ]]]
[[[0.40037617 0.40037617 0.40037617]]]
[[[0.78425235 0.78425235 0.78425235]]]
[[[0.75829774 0.75829774 0.75829774]]]]]
Optimized:
[[[[[0.0242452 0.0242452 0.0242452 ]]]
[[[0.40037617 0.40037617 0.40037617]]]
[[[0.78425235 0.78425235 0.78425235]]]
[[[0.19935817 0.19935817 0.19935817]]]]]
```
### Environment
Windows 11 22H2, TVM commit
### Steps to reproduce
Preferably a minimal script to cause the issue to occur.
```python
import numpy as np
import tvm
from tvm import parser, relay, transform, cpu
from tvm.contrib.graph_executor import GraphModule
code = """
#[version = "0.0.5"]
def @main(%x0: Tensor[(1, 4, 3, 2), float32] /* ty=Tensor[(1, 4, 3, 2), float32] span=from_string:4:15 */, %x1: float32 /* ty=float32 span=from_string:4:20 */, %x2: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:7:31 */, %x3: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:7:36 */, %x4: Tensor[(4, 1, 1, 3, 2), float32] /* ty=Tensor[(4, 1, 1, 3, 2), float32] span=from_string:13:24 */, %x5: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:14:28 */, %x6: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:14:33 */, %x7: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:14:38 */, %x8: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:14:43 */, %x9: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:16:28 */, %x10: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:16:33 */, %x11: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:16:39 */, %x12: Tenso
r[(4), float32] /* ty=Tensor[(4), float32] span=from_string:16:45 */, %x13: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:22:33 */, %x14: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:22:39 */) {
%0 = nn.pad(%x0, %x1, pad_width=[[0, 0], [0, 0], [2, 2], [1, 0]]) /* ty=Tensor[(1, 4, 7, 3), float32] span=from_string:5:20 */;
%1 = expand_dims(%0, axis=3) /* ty=Tensor[(1, 4, 7, 1, 3), float32] span=from_string:19:13 */;
%2 = nn.adaptive_avg_pool3d(%1, output_size=[1, 1, 1]) /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:8:17 */;
%3 = nn.instance_norm(%2, %x2, %x3) /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:8:13 */;
%4 = (%3, %3, %2) /* ty=(Tensor[(1, 4, 1, 1, 1), float32], Tensor[(1, 4, 1, 1, 1), float32], Tensor[(1, 4, 1, 1, 1), float32]) span=from_string:9:20 */;
%5 = concatenate(%4, axis=1) /* ty=Tensor[(1, 12, 1, 1, 1), float32] span=from_string:10:12 */;
%6 = cos(%5) /* ty=Tensor[(1, 12, 1, 1, 1), float32] span=from_string:11:12 */;
%7 = sum(%6, axis=[1, 2, 3], keepdims=True) /* ty=Tensor[(1, 1, 1, 1, 1), float32] span=from_string:12:15 */;
%8 = round(%7) /* ty=Tensor[(1, 1, 1, 1, 1), float32] span=from_string:20:23 */;
%9 = nn.conv3d(%8, %x4, strides=[1, 2, 2], padding=[0, 1, 2, 0, 1, 0], dilation=[1, 1, 2], channels=4, kernel_size=[1, 3, 2]) /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:14:23 */;
%10 = nn.batch_norm(%9, %x5, %x6, %x7, %x8, center=False, scale=False) /* ty=(Tensor[(1, 4, 1, 1, 1), float32], Tensor[(4), float32], Tensor[(4), float32]) span=from_string:15:9 */;
%11 = %10.0 /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:16:23 */;
%12 = nn.batch_norm(%11, %x9, %x10, %x11, %x12, center=False, scale=False) /* ty=(Tensor[(1, 4, 1, 1, 1), float32], Tensor[(4), float32], Tensor[(4), float32]) span=from_string:17:9 */;
%13 = %12.0 /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:18:13 */;
%14 = sum(%1, axis=[2], keepdims=True) /* ty=Tensor[(1, 4, 1, 1, 3), float32] span=from_string:20:18 */;
%15 = exp(%13) /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:21:13 */;
%16 = maximum(%14, %8) /* ty=Tensor[(1, 4, 1, 1, 3), float32] span=from_string:21:18 */;
%17 = add(%15, %16) /* ty=Tensor[(1, 4, 1, 1, 3), float32] span=from_string:22:28 */;
%18 = nn.instance_norm(%17, %x13, %x14) /* ty=Tensor[(1, 4, 1, 1, 3), float32] span=from_string:23:4 */;
(%18,) /* ty=(Tensor[(1, 4, 1, 1, 3), float32],) span=from_string:4:3 */
}
"""
mod = parser.parse(code)
inputs = np.load('inputs.npz')
params = np.load('params.npz')
with transform.PassContext(opt_level=0):
lib = relay.build(mod, params=dict(params), target='llvm')
gmod = GraphModule(lib['default'](cpu()))
gmod.run(**inputs)
ref_out = gmod.get_output(0).numpy()
with transform.PassContext(opt_level=3):
mod, params = relay.optimize(mod, params=params, target='llvm')
lib = relay.build(mod, params=params, target='llvm')
gmod = GraphModule(lib['default'](cpu()))
gmod.run(**inputs)
opt_out = gmod.get_output(0).numpy()
print('Reference:\n', ref_out)
print('Optimized:\n', opt_out)
```
File `inputs.npz` and `params.npz` are packed in [data.zip](https://github.com/apache/tvm/files/9985896/data.zip). A visualization of the computation graph in PDF format [graph.gv.pdf](https://github.com/apache/tvm/files/9985903/graph.gv.pdf) is also attached.
### Current Findings
1. This bug only occurs in 0.9 and all releases after it, but not in 0.8.
2. This bug seems to be related to `FoldScaleAxis` pass enabled at optimization level 3. If this pass is disabled, the inconsistency will not occur.
3. The test case presented above seems to be minimal to trigger this inconsistency. If any of the operators is removed (i.e., its uses are replaced with the input of the operator), the inconsistency no longer occurs.
4. The result at optimization level 3 seems to be equal to the one of the graph with `round` (`%8`) removed at any optimization level, but I am not sure if the inconsistency is related to this operator.
### Triage
* needs-triage
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [I] [Bug] Inconsistent results of a Relay program at optimization level 3 [tvm]
Posted by "wzh99 (via GitHub)" <gi...@apache.org>.
wzh99 commented on issue #13359:
URL: https://github.com/apache/tvm/issues/13359#issuecomment-1934190453
Thanks for digging into this issue. I think your localization of the error is reasonable. The `FoldScaleAxis` pass only introduces slight differences, but they are magnified by the `nn.instance_norm` operator and the final difference can be quite large. This is quite interesting.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [I] [Bug] Inconsistent results of a Relay program at optimization level 3 [tvm]
Posted by "hxzd5568 (via GitHub)" <gi...@apache.org>.
hxzd5568 commented on issue #13359:
URL: https://github.com/apache/tvm/issues/13359#issuecomment-1933639434
This bug comes from conv3d-batch_norm pattern under FoldScaleAxis pass.
It is a numerical errors because ehancing imprecision can decrease the error from $0.18$ to $10^{-5}$ .
It is also caused by conv3d-batch_norm pattern under FoldScaleAxis because breaking any of these conditions can cause the error to zero.
## It is a numerical error
```
import numpy as np
import tvm
from tvm import parser, relay, transform, cpu
from tvm.contrib.graph_executor import GraphModule
def MRE(y_true, y_pred,): #precision along with tf.keras.metrics.MeanRelativeError
if np.isinf(y_true).any()==1 or np.isnan(y_true).any()==1:
print('y_true have inf\\nan:locating...')
#self.locate_naninf('1')
return 0
if np.isinf(y_pred).any()==1 or np.isnan(y_pred).any()==1:
print('y_pred have inf\\nan:locating...')
return 0
#self.locate_naninf('5')
else:
pass
d = np.abs(y_true.astype(np.float64) - y_pred)
relative_error = np.average( d \
/ (np.abs(y_true).astype(np.float64) + 1e-8) * np.not_equal(y_true, 0)\
+ np.equal(y_true, 0)* d )
return relative_error
code2 = """
#[version = "0.0.5"]
def @main(%x0: Tensor[(1, 4, 3, 2), float64] /* ty=Tensor[(1, 4, 3, 2), float64] span=from_string:4:15 */, %x1: float64 /* ty=float64 span=from_string:4:20 */, %x2: Tensor[(4), float64] /* ty=Tensor[(4), float64] span=from_string:7:31 */, %x3: Tensor[(4), float64] /* ty=Tensor[(4), float64] span=from_string:7:36 */, %x4: Tensor[(4, 1, 1, 3, 2), float64] /* ty=Tensor[(4, 1, 1, 3, 2), float64] span=from_string:13:24 */, %x5: Tensor[(4), float64] /* ty=Tensor[(4), float64] span=from_string:14:28 */, %x6: Tensor[(4), float64] /* ty=Tensor[(4), float64] span=from_string:14:33 */, %x7: Tensor[(4), float64] /* ty=Tensor[(4), float64] span=from_string:14:38 */, %x8: Tensor[(4), float64] /* ty=Tensor[(4), float64] span=from_string:14:43 */, %x9: Tensor[(4), float64] /* ty=Tensor[(4), float64] span=from_string:16:28 */, %x10: Tensor[(4), float64] /* ty=Tensor[(4), float64] span=from_string:16:33 */, %x11: Tensor[(4), float64] /* ty=Tensor[(4), float64] span=from_string:16:39 */, %x12: Tenso
r[(4), float64] /* ty=Tensor[(4), float64] span=from_string:16:45 */, %x13: Tensor[(4), float64] /* ty=Tensor[(4), float64] span=from_string:22:33 */, %x14: Tensor[(4), float64] /* ty=Tensor[(4), float64] span=from_string:22:39 */) {
%0 = nn.pad(%x0, %x1, pad_width=[[0, 0], [0, 0], [2, 2], [1, 0]]) /* ty=Tensor[(1, 4, 7, 3), float64] span=from_string:5:20 */;
%1 = expand_dims(%0, axis=3) /* ty=Tensor[(1, 4, 7, 1, 3), float64] span=from_string:19:13 */;
%2 = nn.adaptive_avg_pool3d(%1, output_size=[1, 1, 1]) /* ty=Tensor[(1, 4, 1, 1, 1), float64] span=from_string:8:17 */;
%3 = nn.instance_norm(%2, %x2, %x3) /* ty=Tensor[(1, 4, 1, 1, 1), float64] span=from_string:8:13 */;
%4 = (%3, %3, %2) /* ty=(Tensor[(1, 4, 1, 1, 1), float64], Tensor[(1, 4, 1, 1, 1), float64], Tensor[(1, 4, 1, 1, 1), float64]) span=from_string:9:20 */;
%5 = concatenate(%4, axis=1) /* ty=Tensor[(1, 12, 1, 1, 1), float64] span=from_string:10:12 */;
%6 = cos(%5) /* ty=Tensor[(1, 12, 1, 1, 1), float64] span=from_string:11:12 */;
%7 = sum(%6, axis=[1, 2, 3], keepdims=True) /* ty=Tensor[(1, 1, 1, 1, 1), float64] span=from_string:12:15 */;
%8 = round(%7) /* ty=Tensor[(1, 1, 1, 1, 1), float64] span=from_string:20:23 */;
%9 = nn.conv3d(%8, %x4, strides=[1, 2, 2], padding=[0, 1, 2, 0, 1, 0], dilation=[1, 1, 2], channels=4, kernel_size=[1, 3, 2]) /* ty=Tensor[(1, 4, 1, 1, 1), float64] span=from_string:14:23 */;
%10 = nn.batch_norm(%9, %x5, %x6, %x7, %x8, center=False, scale=False) /* ty=(Tensor[(1, 4, 1, 1, 1), float64], Tensor[(4), float64], Tensor[(4), float64]) span=from_string:15:9 */;
%11 = %10.0 /* ty=Tensor[(1, 4, 1, 1, 1), float64] span=from_string:16:23 */;
%12 = nn.batch_norm(%11, %x9, %x10, %x11, %x12, center=False, scale=False) /* ty=(Tensor[(1, 4, 1, 1, 1), float64], Tensor[(4), float64], Tensor[(4), float64]) span=from_string:17:9 */;
%13 = %12.0 /* ty=Tensor[(1, 4, 1, 1, 1), float64] span=from_string:18:13 */;
%14 = sum(%1, axis=[2], keepdims=True) /* ty=Tensor[(1, 4, 1, 1, 3), float64] span=from_string:20:18 */;
%15 = exp(%13) /* ty=Tensor[(1, 4, 1, 1, 1), float64] span=from_string:21:13 */;
%16 = maximum(%14, %8) /* ty=Tensor[(1, 4, 1, 1, 3), float64] span=from_string:21:18 */;
%17 = add(%15, %16) /* ty=Tensor[(1, 4, 1, 1, 3), float64] span=from_string:22:28 */;
%18 = nn.instance_norm(%17, %x13, %x14) /* ty=Tensor[(1, 4, 1, 1, 3), float64] span=from_string:23:4 */;
(%18,) /* ty=(Tensor[(1, 4, 1, 1, 3), float64],) span=from_string:4:3 */
}
"""
#### ----------------
mod = parser.parse(code2)
inputs = np.load('inputs.npz')
params = np.load('params.npz')
inputs0 = dict()
params0 = dict()
for k, v in dict(params).items():
params0[k]=v.astype('float64')
for k, v in dict(inputs).items():
inputs0[k]=v.astype('float64')
with transform.PassContext(opt_level=0):
lib = relay.build(mod, params=params0, target='llvm')
gmod = GraphModule(lib['default'](cpu()))
gmod.run(**inputs0)
ref_out = gmod.get_output(0).numpy()
with transform.PassContext(opt_level=3):
mod, params = relay.optimize(mod, params=params0, target='llvm')
lib = relay.build(mod, params=params, target='llvm')
gmod = GraphModule(lib['default'](cpu()))
gmod.run(**inputs0)
opt_out = gmod.get_output(0).numpy()
# print('Reference:\n', ref_out)
# print('Optimized:\n', opt_out)
print(MRE(ref_out,opt_out))
```
## Breaking conv3d-batchnorm pattern makes the error to zero.
```
import numpy as np
import tvm
from tvm import parser, relay, transform, cpu
from tvm.contrib.graph_executor import GraphModule
def MRE(y_true, y_pred,): #precision along with tf.keras.metrics.MeanRelativeError
if np.isinf(y_true).any()==1 or np.isnan(y_true).any()==1:
print('y_true have inf\\nan:locating...')
#self.locate_naninf('1')
return 0
if np.isinf(y_pred).any()==1 or np.isnan(y_pred).any()==1:
print('y_pred have inf\\nan:locating...')
return 0
#self.locate_naninf('5')
else:
pass
d = np.abs(y_true.astype(np.float64) - y_pred)
relative_error = np.average( d \
/ (np.abs(y_true).astype(np.float64) + 1e-8) * np.not_equal(y_true, 0)\
+ np.equal(y_true, 0)* d )
return relative_error
code = """
#[version = "0.0.5"]
def @main(%x0: Tensor[(1, 4, 3, 2), float32] /* ty=Tensor[(1, 4, 3, 2), float32] span=from_string:4:15 */, %x1: float32 /* ty=float32 span=from_string:4:20 */, %x2: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:7:31 */, %x3: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:7:36 */, %x4: Tensor[(4, 1, 1, 3, 2), float32] /* ty=Tensor[(4, 1, 1, 3, 2), float32] span=from_string:13:24 */, %x5: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:14:28 */, %x6: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:14:33 */, %x7: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:14:38 */, %x8: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:14:43 */, %x9: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:16:28 */, %x10: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:16:33 */, %x11: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:16:39 */, %x12: Tenso
r[(4), float32] /* ty=Tensor[(4), float32] span=from_string:16:45 */, %x13: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:22:33 */, %x14: Tensor[(4), float32] /* ty=Tensor[(4), float32] span=from_string:22:39 */) {
%0 = nn.pad(%x0, %x1, pad_width=[[0, 0], [0, 0], [2, 2], [1, 0]]) /* ty=Tensor[(1, 4, 7, 3), float32] span=from_string:5:20 */;
%1 = expand_dims(%0, axis=3) /* ty=Tensor[(1, 4, 7, 1, 3), float32] span=from_string:19:13 */;
%2 = nn.adaptive_avg_pool3d(%1, output_size=[1, 1, 1]) /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:8:17 */;
%3 = nn.instance_norm(%2, %x2, %x3) /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:8:13 */;
%4 = (%3, %3, %2) /* ty=(Tensor[(1, 4, 1, 1, 1), float32], Tensor[(1, 4, 1, 1, 1), float32], Tensor[(1, 4, 1, 1, 1), float32]) span=from_string:9:20 */;
%5 = concatenate(%4, axis=1) /* ty=Tensor[(1, 12, 1, 1, 1), float32] span=from_string:10:12 */;
%6 = cos(%5) /* ty=Tensor[(1, 12, 1, 1, 1), float32] span=from_string:11:12 */;
%7 = sum(%6, axis=[1, 2, 3], keepdims=True) /* ty=Tensor[(1, 1, 1, 1, 1), float32] span=from_string:12:15 */;
%8 = round(%7) /* ty=Tensor[(1, 1, 1, 1, 1), float32] span=from_string:20:23 */;
%9 = nn.conv3d(%8, %x4, strides=[1, 2, 2], padding=[0, 1, 2, 0, 1, 0], dilation=[1, 1, 2], channels=4, kernel_size=[1, 3, 2]) /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:14:23 */;
%10 = nn.batch_norm(%9, %x5, %x6, %x7, %x8, center=False, scale=False) /* ty=(Tensor[(1, 4, 1, 1, 1), float32], Tensor[(4), float32], Tensor[(4), float32]) span=from_string:15:9 */;
%11 = %10.0 /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:16:23 */;
%12 = nn.batch_norm(%11, %x9, %x10, %x11, %x12, center=False, scale=False) /* ty=(Tensor[(1, 4, 1, 1, 1), float32], Tensor[(4), float32], Tensor[(4), float32]) span=from_string:17:9 */;
%13 = %12.0 /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:18:13 */;
%14 = sum(%1, axis=[2], keepdims=True) /* ty=Tensor[(1, 4, 1, 1, 3), float32] span=from_string:20:18 */;
%15 = exp(%13) /* ty=Tensor[(1, 4, 1, 1, 1), float32] span=from_string:21:13 */;
%16 = maximum(%14, %8) /* ty=Tensor[(1, 4, 1, 1, 3), float32] span=from_string:21:18 */;
%17 = add(%15, %16) /* ty=Tensor[(1, 4, 1, 1, 3), float32] span=from_string:22:28 */;
%18 = nn.instance_norm(%17, %x13, %x14) /* ty=Tensor[(1, 4, 1, 1, 3), float32] span=from_string:23:4 */;
(%18,) /* ty=(Tensor[(1, 4, 1, 1, 3), float32],) span=from_string:4:3 */
}
"""
mod = parser.parse(code)
mod = relay.analysis.extract_intermdeiate_expr(mod, 10)#
print(mod)
inputs = np.load('inputs.npz')
params = np.load('params.npz')
with transform.PassContext(opt_level=0):
lib = relay.build(mod, params=dict(params), target='llvm')
gmod = GraphModule(lib['default'](cpu()))
gmod.run(**inputs)
ref_out = gmod.get_output(0).numpy()
with transform.PassContext(opt_level=3):
mod, params = relay.optimize(mod, params=dict(params), target='llvm')
lib = relay.build(mod, params=params, target='llvm')
gmod = GraphModule(lib['default'](cpu()))
gmod.run(**inputs)
opt_out = gmod.get_output(0).numpy()
# print('Reference:\n', ref_out)
# print('Optimized:\n', opt_out)
print(MRE(ref_out,opt_out))
```
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [I] [Bug] Inconsistent results of a Relay program at optimization level 3 [tvm]
Posted by "wzh99 (via GitHub)" <gi...@apache.org>.
wzh99 closed issue #13359: [Bug] Inconsistent results of a Relay program at optimization level 3
URL: https://github.com/apache/tvm/issues/13359
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org