source from discord SD Training Labs.

how_to_v2.pt.txt

modules\prompt_parser.py file.

v2.pt can be loaded by putting it in the main folder of the repo and adding
---------------------------------------------------------------------------

import torch
from torch import nn
from modules import devices
 
class VectorAdjustPrior(nn.Module):
    def __init__(self, hidden_size, inter_dim=64):
        super().__init__()
        self.vector_proj = nn.Linear(hidden_size*2, inter_dim, bias=True)
        self.out_proj = nn.Linear(hidden_size+inter_dim, hidden_size, bias=True)
 
    def forward(self, z):
        b, s = z.shape[0:2]
        x1 = torch.mean(z, dim=1).repeat(s, 1)
        x2 = z.reshape(b*s, -1)
        x = torch.cat((x1, x2), dim=1)
        x = self.vector_proj(x)
        x = torch.cat((x2, x), dim=1)
        x = self.out_proj(x)
        x = x.reshape(b, s, -1)
        return x
 
    @classmethod
    def load_model(cls, model_path, hidden_size=768, inter_dim=64):
        model = cls(hidden_size=hidden_size, inter_dim=inter_dim)
        model.load_state_dict(torch.load(model_path)["state_dict"])
        model.to(devices.device)
        
        return model
 
 
vap = VectorAdjustPrior.load_model('v2.pt').cuda()

------------------------------------------------------------------


after the

----------------------------
import lark
---------------------------


line and adding
---------------------
conds = vap(conds)
---------------------

between
---------------------------------------------
conds = model.get_learned_conditioning(texts)
---------------------------------------------

and

---------------------------------------------
cond_schedule = [] lines.
---------------------------------------------