Closed dddzg closed 4 years ago
There is a bug for WeightEMA in the original code. the assignment param = ema_param will not copy the weight from the "emamodel" to "model" due to the variable reference mechanism. So, it should be replaced with param.data.copy(ema_param.data)
There is a bug for WeightEMA in the original code. the assignment param = ema_param will not copy the weight from the "emamodel" to "model" due to the variable reference mechanism. So, it should be replaced with param.data.copy(ema_param.data)