Closed betterhalfwzm closed 4 years ago
- Use larger lambda
- Start from thinner supernet
Thank you for your answer.And where can i set the lambda or thinner supernet in the code or yml? How many hours with 8 cards to search and retrain? Looking forward to your reply. Thank you very much. @meijieru
- Check a YAML config of the training example.
- We didn't run the code for 8 cards and could not guarantee the performance.
Thank you for your answer. Start from thinner supernet that means to change the channels and numbers of block in the inverted_residual_setting? If i only use larger lambda( like rho: 2.4e-4), can it work? @meijieru Thank you very much. 'inverted_residual_setting': [[1, 16, 1, 1, [3]], [6, 24, 4, 2, [3, 5, 7]], [6, 40, 4, 2, [3, 5, 7]], [6, 80, 4, 2, [3, 5, 7]], [6, 96, 4, 1, [3, 5, 7]], [6, 192, 4, 2, [3, 5, 7]], [6, 320, 1, 1, [3, 5, 7]]] prune_params: { method: network_slimming, bn_prune_filter: expansion_only_skip_expand1, rho: 1.8e-4, epoch_free: 0, epoch_warmup: 25, scheduler: linear, stepwise: True, logging_verbose: False }
To start from a thinner supernet, you could try the followings:
- modify the output channels of the blocks
- use fewer blocks
- start from smaller expand ratio
- If the model you want is at the same level computation cost, I recommend a larger lambda. But if you want a much lighter model, I suggest to use a thinner supernet and then adjust the lambda for better performance.
Thank you for your advice. I'll try it out
Thx!!