Stable_Diffusion/diffusion.py

from torch import nn
from torch.nn import functional as F
from .unet import UNET, UNET_OutputLayer


class TimeEmbedding(nn.Module):
    def __init__(self, n_embd):
        super().__init__()
        self.linear_1 = nn.Linear(n_embd, 4 * n_embd)
        self.linear_2 = nn.Linear(4 * n_embd, 4 * n_embd)

    def forward(self, x):
        # x: (1, 320)

        # (1, 320) -> (1, 1280)
        x = self.linear_1(x)
        
        # (1, 1280) -> (1, 1280)
        x = F.silu(x) 
        
        # (1, 1280) -> (1, 1280)
        x = self.linear_2(x)

        return x


class Diffusion(nn.Module):
    def __init__(self):
        super().__init__()
        self.time_embedding = TimeEmbedding(320)
        self.unet = UNET()
        self.final = UNET_OutputLayer(320, 4)
    
    def forward(self, latent, context, time):
        # latent: (Batch_Size, 4, Height / 8, Width / 8)
        # context: (Batch_Size, Seq_Len, Dim)
        # time: (1, 320)

        # (1, 320) -> (1, 1280)
        time = self.time_embedding(time)
        
        # (Batch, 4, Height / 8, Width / 8) -> (Batch, 320, Height / 8, Width / 8)
        output = self.unet(latent, context, time)
        
        # (Batch, 320, Height / 8, Width / 8) -> (Batch, 4, Height / 8, Width / 8)
        output = self.final(output)
        
        # (Batch, 4, Height / 8, Width / 8)
        return output