Sweeps pytorch conv layers

Hi all,

I want to use sweeps to find best architecutre for image classification. My model look like this :

      self.cb1 = conv_block(3,wandb.config.cb1,kernel_size=3,stride=2)
      self.cb2 = conv_block(wandb.config.cb1,wandb.config.cb2,kernel_size=3,stride=2)
      self.cb3 = conv_block(wandb.config.cb2,wandb.config.cb3,kernel_size=3,stride=2)
      self.cb4 = conv_block(wandb.config.cb3,wandb.config.cb4,kernel_size=3,stride=2)
      self.cb5 = conv_block(wandb.config.cb4,wandb.config.cb5,kernel_size=3,stride=2)
      self.cb6 = conv_block(wandb.config.cb5,wandb.config.cb6,kernel_size=3,stride=2)
      self.lin1 = nn.Linear(12800,256)

values: [16,32,64]
values: [32,64,96]
values: [64,96,160]
values: [96,160,256]
values: [160,256,416]
values: [256,384,512]

How can I found value for linear layer after convolutions layers ? It will be all time different ? Is there some trick for this? Is this how you are looking for best architecture ?


Hi @mirokery ,

We can suggest you to try on two things.

  1. Parameter Importance.
    -to see what is the importance of those values to your goal/experiment and look for
  2. Parallel Coordinates Chart
    -to see which of the values best corresponds to the best metric you are tracking.

Hi @mirokery , since we have not heard back from you we are going to close this request. If you would like to re-open the conversation, please let us know!