Rolling Memory

(no paper)

Rolling memory is a type of experience replay. Each training step, a memory is replaced with the top scoring batch item.

Each types pairing becomes a discriminator that is added to the loss.

examples

{
    "class": "function:hypergan.train_hooks.experimental.rolling_memory_2_train_hook.RollingMemoryTrainHook",
    "types": ["mx-/g(mz-)"]
}

mx- is a memory of x that gets updated each training step. g(mz-) is a memory of z that gets run through a generator and updated each trainng step.

A discriminator d(mx-, g(mz-)) is created and added to the gan loss.

options

attribute	description	type
types	What memories and how they are paired. See memory types below	array of strings
top_k	How many memory items to replace per frame. Defaults to 1	integer
only	Overrides all other losses when this is set. Defaults to false	boolean

memory types

memory	description
mx-	x reverse sorted by d_real
mx+	x sorted by d_real
mg-	memory of g reverse sorted by d_fake
mg+	memory of g sorted by d_fake
g(mz-)	generator of memory of z reverse sorted by d_fake
g(mz+)	generator of memory of z sorted by d_fake
x	gan.inputs.x
g	gan.generator.sample