Closed vict0rsch closed 3 years ago
For a multi-batch like:
{"rf: batch0, "rn": batch1, "sf": batch2 "sn": batch3}
Updates will be:
- real D["a"]["r"] (batch0)
- fake D["a"]["s"] (batch0)
- real D["t"]["f"] (batch0)
- fake D["t"]["n"] (batch0)
- real D["a"]["r"] (batch1)
- fake D["a"]["s"] (batch1)
- real D["t"]["n"] (batch1)
- fake D["t"]["f"] (batch1)
- real D["a"]["s"] (batch2)
- fake D["a"]["r"] (batch2)
- real D["t"]["f"] (batch2)
- fake D["t"]["n"] (batch2)
- real D["a"]["s"] (batch3)
- fake D["a"]["r"] (batch3)
- real D["t"]["n"] (batch3)
- fake D["t"]["f"] (batch3)
Questions:
As is, the model updates twice each of the 4 discriminator (in each direction)
@51N84D @melisandeteng @alexrey88 @tianyu-z that was a pb with translation. Do we agree it's not the case anymore and we can close this?
As per #18