search

enyac-group / single-path-nas

Single-Path NAS: Designing Hardware-Efficient ConvNets in less than 4 Hours
https://arxiv.org/abs/1904.02877
Apache License 2.0
396 stars 60 forks source link

multi-gpus, reproduce loss #15

Open MissyDu opened 5 years ago

MissyDu commented 5 years ago

I have implmented a multi-gpus version. 1>But I am not sure how to evaluate reproduce to be successful for the search process.What's the correct evaluation accuracy in the training process? 2>And I find the loss is large in the search training process, like this issue mentioned. Is that a normal value for loss? I found it's related with "runtime_lambda_val", and 0.02 is the correct setting for it? 3>And "base_learning_rate" is 0.016, it seem to be small for training from scratch, though it's enlarged by " FLAGS.base_learning_rate * (FLAGS.train_batch_size / 256.0)", and is it the correct setting?

SUNziwei0527 commented 3 years ago

can you give me your code ? please!