dmitry-vorobiev / generative-models Goto Github PK

Hi @dmitry-vorobiev

I would like to add your project to https://github.com/pytorch/ignite#others
and optionally even create a working Colab with wandb online logs and generated images by trained model.
How would you estimate the readiness of this repository ?

I took a look on your "Projects" kanban and I'd also be interested in adding AMP training.

What do you think about all this ?

PS: optionally you can also ping me on ODS slack (@vfdev-5) if you would like to discuss more.

Hi @dmitry-vorobiev

I'm trying to play around this code and it would cool to add some info on:

• how to download the dataset
• how to use the dataset from a custom location

What do you think ?

PS: thanks for open-sourcing the code

Maybe like this:

def setup_ema(ema_model, ref_model):
    # assumption: ema_model architecture == ref_model architecture
    ema_model.load_state_dict(ref_model.state_dict())
    for param in ema_model.parameters():
        param.detach_()
    # set EMA model's BN buffers as base model BN buffers:
    for m1, m2 in zip(ref_model.modules(), ema_model.modules()):
        if isinstance(m1, nn.BatchNorm2d) and isinstance(m2, nn.BatchNorm2d):
            m2.running_mean = m1.running_mean
            m2.running_var = m1.running_var

?