How to run this model？ about rml-cnn

[bbenligiray]

I've updated the readme for instructions about the datasets and the pretrained model.

The code allows you to train from scratch, for example python main.py nus_wide random robust_warp 100 0 would train a randomly initialized model on NUS-WIDE with fully supervised rml-cnn loss. See here for how these parameters work. You can add an --optimize parameter to do hyperparameter optimization.

[mymuli]

@bbenligiray Thank you very much first!
But in the params.py file, I am confused about these parameters，[No_lr_steps、Lr_patience、No_random_start、No_opt_iters] ，Can you explain them?
At the same time, what's the difference between parameter imagenet and parameter imagenet2?
Thank you again！

[bbenligiray]

no_lr_steps = 2 -> there will be two learning rate steps. So if we start training with a learning rate of 0.1, after val_loss stops improving, continue training with 0.1*0.1=0.01 learning rate (so the first step with 0.1 and the second step with 0.01 learning rate)
lr_patience=10 -> stop training if val_loss doesn't improve for 10 epochs

no_random_starts=8
no_opt_iters=15 this means first try 8 random hyperparameter sets between the given intervals, then try to converge to the best result with 15-8=7 additional tries. See this.

So let's you started optimizing your hyperparameters with --optimize and ended up on the border (for example you were searching between 0.1 and 0.001 learning rate and ended up on 0.001). This means that the optimal hyperparameter was likely outside your initial interval (the first interval is named imagenet). Then you would set the second interval (imagenet2) between 0.001 and 0.00001 and run main.py with --optimize --cont. This parses the earlier training result between 0.1 and 0.001 to make a more educated estimation rather than starting from the scratch.

Basically it's a bit of over-engineering on my part. If you are not going to use the --cont parameter, you can ignore imagenet2.

[mymuli]

@bbenligiray I was inspired，Thank you again！