wkcn edited a comment on pull request #18707:
URL: https://github.com/apache/incubator-mxnet/pull/18707#issuecomment-659304799
Performance Benchmark:
transpose operator on CPU, **axes is generated randomly**
ndim | max use (kb) | avg time (ms)
---|---|--
1|12582.9121|1.0786
2|12582.9121|1.0851
3|12582.9121|0.6763
4|12582.9121|1.2172
5|12582.9121|6.4305
6|12582.9121|11.7841
7|12583.3604|65.7184
8|12583.4238|65.2171
9|12583.4883|82.4930
The increase of memory footprint is slight, but the time is intolerable when
`axes.ndim() > 6`.
If **axes is monotonically increasing** (namely [0, 1, 2, 3, ..., ndim - 1])
(comment the line 21st `random.shuffle(axes)`),
ndim | max use (kb) | avg time (ms)
---|---|--
1|12582.9121|1.1492
2|12582.9121|1.1732
3|12582.9121|1.3264
4|12582.9121|1.3896
5|12582.9121|0.9107
6|12582.9121|0.8965
7|12583.3604|0.9028
8|12583.4238|0.9105
9|12583.4883|0.8981
If **axes is monotonically decreasing** (namely [ndim - 1, ndim -2, ..., 2,
1, 0])
ndim | max use (kb) | avg time (ms)
---|---|--
1|12582.9121|1.1290
2|12582.9121|1.1204
3|12582.9121|1.1874
4|12582.9121|1.4240
5|12582.9121|7.7080
6|12582.9121|24.0448
7|12583.3604|115.1126
8|12583.4238|105.9091
9|12583.4883|106.3913
Compare with NumPy Transpose:
ndim | numpy time (s) | mxnet time (s)
--|--|--
1 | 0.1621077060699463 | 0.31803297996520996
2 | 0.2637207508087158 | 0.33347415924072266
3 | 0.4311816692352295 | 0.47667574882507324
4 | 0.5303101539611816 | 0.49021244049072266
5 | 0.5940566062927246 | 1.48443603515625
6 | 0.8220541477203369 | 2.03752064704895
7 | 0.8727006912231445 | 9.488046169281006
8 | 1.0004301071166992 | 9.947605848312378
9 | 1.2341070175170898 | 12.262272119522095
Test Code:
```python
import mxnet as mx
from mxnet import profiler
print(mx)
import numpy as np
from numpy.testing import assert_allclose
import time
import random
seed = 42
np.random.seed(seed)
mx.random.seed(seed)
#configure the profiler
profiler.set_config(profile_all=True, aggregate_stats=True,
filename='trace_profile.json')
#start the profiler collecting data
def test_transpose(ndim):
for t in range(20):
dims = [4 for _ in range(ndim)]
dims[-1] *= 4 ** (10 - ndim)
axes = list(range(ndim))
random.shuffle(axes)
axes = tuple(axes)
x = mx.nd.array(np.random.normal(size=dims))
y = mx.nd.transpose(x, axes=axes)
assert_allclose(np.transpose(x.asnumpy(), axes=axes), y.asnumpy())
for ndim in range(1, 10):
profiler.set_state('run')
tic = time.time()
test_transpose(ndim)
print(ndim, "====", time.time() - tic)
#stop the profiler
profiler.set_state('stop')
#dump the profiling data as a string
print(profiler.dumps(reset=True))
print("Over")
```
Test Code, compare with NumPy
```python
import mxnet as mx
from mxnet import profiler
print(mx)
import numpy as np
from numpy.testing import assert_allclose
import time
import random
seed = 42
np.random.seed(seed)
mx.random.seed(seed)
def test_transpose(ndim):
np_time = 0
mx_time = 0
for t in range(20):
dims = [5 for _ in range(ndim)]
dims[-1] *= 5 ** (10 - ndim)
axes = list(range(ndim))
random.shuffle(axes)
axes = tuple(axes)
x_np = np.array(np.random.normal(size=dims), dtype=np.float32)
x_mx = mx.nd.array(x_np, dtype=np.float32)
for _ in range(2):
y_np = np.transpose(x_np, axes=axes).copy()
y_mx = mx.nd.transpose(x_mx, axes=axes)
y_mx.asnumpy()
tic_np = time.time()
for _ in range(1):
y_np = np.transpose(x_np, axes=axes).copy()
np_time += time.time() - tic_np
tic_mx = time.time()
for _ in range(1):
y_mx = mx.nd.transpose(x_mx, axes=axes)
y_mx.asnumpy()
mx_time += time.time() - tic_mx
print(f"{ndim} | {np_time} | {mx_time}")
for ndim in range(1, 10):
test_transpose(ndim)
print("Over")
```
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
