[tutorial][benchmark] nnvm -> relay #4368
Conversation
|
Shouldn't we update get_network() too? |
|
sorry I need to verify rpc once I get a device tomorrow. |
apps/benchmark/util.py
Outdated
| elif name == 'mobilenet_v2': | ||
| net, params = nnvm.testing.mobilenet_v2.get_workload(batch_size=batch_size, dtype=dtype) | ||
| net, params = testing.mobilenet_v2.get_workload(batch_size=batch_size, dtype=dtype) |
There was a problem hiding this comment.
It seems testing.mobilenet_v2.get_workload doesn't exist, so delete it
apps/benchmark/util.py
Outdated
| elif "densenet" in name: | ||
| n_layer = int(name.split('-')[1]) | ||
| net, params = nnvm.testing.densenet.get_workload(num_layers=n_layer, batch_size=batch_size, dtype=dtype) | ||
| net, params = testing.densenet.get_workload(num_layers=n_layer, batch_size=batch_size, dtype=dtype) |
There was a problem hiding this comment.
num_layers -> densenet_size
Weird, but see https://github.com/apache/incubator-tvm/blob/master/python/tvm/relay/testing/densenet.py#L84
|
@petrex could you also help to verify the arm results? |
|
@yzhliu I can help verify the result on AMD GPU, is that what you are looking for? thx |
|
I already tested it on AMD (also to verify recent AMDGPU codegen fix), @yzhliu maybe you mistook it for somebody else? |
* [tutorial] nnvm -> relay * use relay workload * delete movbilenetv2 option
* [tutorial] nnvm -> relay * use relay workload * delete movbilenetv2 option
|
@yzhliu Hi, I am having hard time to reproduce the results of mobile GPU. I am wondering did you test this PR w/ mobile gpu? Thanks!!! |
|
not yet. did you run auto tvm tuning? |
@tqchen @masahi