Hi,
I’m running AutoTVM for tuning a network written using the relay interface. My tuner is based on the tutorial, tune_relay_x86.py.
Here’s my ‘tuning option’
tuning_option = {
'log_filename': log_file,
'tuner': 'random',
'early_stopping': 500,
'measure_option': autotvm.measure_option(
builder=autotvm.LocalBuilder(),
runner=autotvm.LocalRunner(number=10, repeat=1,
min_repeat_ms=1000),
),
}
And here’s my output.
Extract tasks...
Tuning...
[Task 1/ 5] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/1152) | 0.00 s
[Task 1/ 5] Current/Best: 1414.35/6037.21 GFLOPS | Progress: (560/1152) | 1654.08 s Done.
[Task 2/ 5] Current/Best: 551.32/6469.98 GFLOPS | Progress: (336/2048) | 1691.42 s
[Task 2/ 5] Current/Best: 458.95/6469.98 GFLOPS | Progress: (672/2048) | 3531.38 s
[Task 2/ 5] Current/Best: 1044.44/6469.98 GFLOPS | Progress: (784/2048) | 4109.08 s Done.
[Task 3/ 5] Current/Best: 0.00/ 0.00 GFLOPS | Progress: (0/1792) | 0.00 s
[Task 3/ 5] Current/Best: 762.93/6319.01 GFLOPS | Progress: (1680/1792) | 5208.71 s Done.
[Task 4/ 5] Current/Best: 1997.02/6840.22 GFLOPS | Progress: (784/784) | 3528.16 s Done.
[Task 5/ 5] Current/Best: 1310.34/6094.27 GFLOPS | Progress: (112/112) | 101.15 s Done.
Cannot find config for target=llvm -device=tracing, workload=('conv2d', (128, 3, 227, 227, 'float32'), (64, 3, 11, 11, 'float32'), (4, 4), (0, 0), (1, 1), 'NCHW', 'float32'). A fallback configuration is used, which may bring great performance regression.
Cannot find config for target=llvm -device=tracing, workload=('conv2d', (128, 64, 55, 55, 'float32'), (192, 64, 5, 5, 'float32'), (1, 1), (2, 2), (1, 1), 'NCHW', 'float32'). A fallback configuration is used, which may bring great performance regression.
Cannot find config for target=llvm -device=tracing, workload=('conv2d', (128, 192, 27, 27, 'float32'), (384, 192, 3, 3, 'float32'), (1, 1), (1, 1), (1, 1), 'NCHW', 'float32'). A fallback configuration is used, which may bring great performance regression.
Cannot find config for target=llvm -device=tracing, workload=('conv2d', (128, 384, 27, 27, 'float32'), (384, 384, 3, 3, 'float32'), (1, 1), (1, 1), (1, 1), 'NCHW', 'float32'). A fallback configuration is used, which may bring great performance regression.
Cannot find config for target=llvm -device=tracing, workload=('conv2d', (128, 384, 27, 27, 'float32'), (256, 384, 3, 3, 'float32'), (1, 1), (1, 1), (1, 1), 'NCHW', 'float32'). A fallback configuration is used, which may bring great performance regression.
2019-08-19 16:24:52,311 INFO Start to benchmark layout transformation...
2019-08-19 16:27:42,939 INFO Benchmarking layout transformation successful.
2019-08-19 16:27:42,941 INFO Start to run PBQP algorithm...
2019-08-19 16:27:42,942 INFO Finished PBQPExecutor run. Got optimal solution.
2019-08-19 16:27:42,944 INFO Writing optimal schedules to alexnet_graph_opt.log successfully.
Compile...
Config for target=llvm -mcpu=skylake-avx512, workload=('dense', (128, 4096, 'float32'), (1008, 4096, 'float32'), 0, 'float32') is missing in ApplyGraphBest context. A fallback configuration is used, which may bring great performance regression.
Config for target=llvm -mcpu=skylake-avx512, workload=('dense', (128, 4096, 'float32'), (4096, 4096, 'float32'), 0, 'float32') is missing in ApplyGraphBest context. A fallback configuration is used, which may bring great performance regression.
Config for target=llvm -mcpu=skylake-avx512, workload=('dense', (128, 43264, 'float32'), (4096, 43264, 'float32'), 0, 'float32') is missing in ApplyGraphBest context. A fallback configuration is used, which may bring great performance regression.
Evaluate inference time cost...
Mean inference time (std dev): 484.82 ms (6.10 ms)
-
As you can see, all the operators (5 convs, and 3 dense layers) fall back to default configurations. Can someone please explain why this happens?
-
Also, what does `-device=tracing’ mean?
Thanks,