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: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
