search

lychahaha / xnor-net-tf

xnor-net by tensorflow
11 stars 3 forks source link

train on custom data #1

Open ajajar opened 6 years ago

lychahaha commented 6 years ago

Hi @ajajar , Yes, you can train on your custom image after you implement the interface of Dataset, including reset(), get_batch(), is_end(). However, this work does not work stably and it can't reduce the forward time as it doesn't implement conv by XNOR op.

ajajar commented 6 years ago

Hi @ljhandlwt understand. Thank you very much.

Also in your init_model function it is specific to your application , it s bit different from the pytorch xnor implementation (alexnet binary part )

Thanks

lychahaha commented 6 years ago

@ajajar Yes, this is NIN, the same as the implement in cifar-10 of pytorch version, not Alexnet. I don't have enough resource to run the imagenet program, so i give up the implement of Alexnet. However, i think it's easy to update it from NIN to Alexnet, and maybe some additional details should be took care in pytorch version.