search

dmlc / minpy

NumPy interface with mixed backend execution
https://minpy.readthedocs.io/en/latest/
Other
1.11k stars 112 forks source link

Symbolic Reuse fix #144

Closed lryta closed 7 years ago

lryta commented 7 years ago

Fix symbol cannot be reused. @HrWangChengdu @jermainewang

Taco-W commented 7 years ago

@lryta The travis is done now. But I guess the failure is caused by perform test, as the extra CNN layer might impact the network performance. What's the accuracy you would get after using extra another CNN layer?

lryta commented 7 years ago

@HrWangChengdu It's due to a weird pip error: pip cannot install minpy. I restarted the job on Travis. Regarding my modification on testcase, I just added a dumb layer which won't change the output result (it's not sent to the final result anyway). I just want to address symbols can be reused in the testcases.

jermainewang commented 7 years ago

Could you test the performance loss due to this change?

lryta commented 7 years ago

I roughly tested in cnn_sym.py and I saw no performance difference. My scant modification should not change performance logically. (No computational change in fact)

Taco-W commented 7 years ago

Could you please use trash as the output and test the network gain? I just want to make sure the gradient could work.