Downstream task weights are potentially missing layers about emoclip

[NickyFot]

The default VClip() constructor initializes the network with 4 temporal layers, however, in the paper we specify that we use 2.

[NickyFot]

Hey, this is a DDP/pytorch error; please see the relevant StackOverflow explanation here.

[SethGA]

Hi again,

Thank you for your quick reply! The issue I am raising is different from the module. prefix being present; in the gist I sent, after I load the weights, I do strip the module. prefix from the state dict. I realize that I should have cleared the output of the first cell to make this more clear. I am specifically referring to certain temporal transformer layers (e.g. temporal.temporal_transformer.layers.2.0.fn.norm.weight) that don't appear to be present in the state dict.