Balanced Trainer

• Source: /trainers/balanced_trainer.py

fake, real = self.gan.forward_discriminator()
if d_real < (fake+config.imbalance):
  self.train_d()
else:
  self.train_g()

examples

• Configurations: /trainers/balanced_trainer/

{
  "class": "class:hypergan.trainers.balanced_trainer.BalancedTrainer",
  "imbalance": 0.06,
  "pretrain_d": 1000,
  "d_optimizer": {
    "class": "class:torch.optim.Adam",
    "lr": 1e-4,
    "betas":[0.0,0.999]
  },
  "g_optimizer": {
    "class": "class:torch.optim.Adam",
    "lr": 1e-4,
    "betas":[0.0,0.999]
  },
  "hooks": [
    {
      "class": "function:hypergan.train_hooks.adversarial_norm_train_hook.AdversarialNormTrainHook",
      "gamma": 2e4,
      "loss": ["d"]
    },
    {
      "class": "function:hypergan.train_hooks.initialize_as_autoencoder.InitializeAsAutoencoder",
      "steps": 10000,
      "optimizer": {
        "class": "class:torch.optim.Adam",
        "lr": 1e-4,
        "betas":[0.9,0.999]
      },
      "encoder": {
        "class": "class:hypergan.discriminators.configurable_discriminator.ConfigurableDiscriminator",
        "layers":[
          "conv 32 stride=1", "adaptive_avg_pool", "relu",
          "conv 64 stride=1", "adaptive_avg_pool", "relu",
          "conv 128 stride=1", "adaptive_avg_pool", "relu",
          "conv 256 stride=1", "adaptive_avg_pool", "relu",
          "conv 512 stride=1", "adaptive_avg_pool", "relu",
          "conv 512 stride=1", "adaptive_avg_pool", "relu",
          "flatten",
          "linear 256 bias=false", "tanh"
        ]
      }
    }
  ]
}

options

attribute	description	type
g_optimizer	Optimizer configuration for G	Config (required)
d_optimizer	Optimizer configuration for D	Config (required)
hooks	Train Hooks	Array of configs (optional)
pretrain_d	First N steps only trains D	Integer (optional)
imbalance	Threshold distance for G training. Defaults to 0.1	Float (optional)
d_fake_balance	Changes conditional to d_fake(t) > d_fake(t-1)	Boolean (optional)