clip.py

import torch
from torch import nn
from torch.nn import functional as F
from attention import SelfAttention

"""
File Name: clip.py
Description: This script is used to produce pre-trained CLIP weights (v1-5-pruned-emaonly.ckpt from https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main) text encoder and prompt embedding 

How to implement the original CLIP architecture model trained on?
load the model of v1-5-pruned-emaonly.ckpt 
 >>> models = model_loader.preload_models_from_standard_weights(model_file='../checkpoint/v1-5-pruned-emaonly.ckpt ', DEVICE='cpu')
 >>> model['clip']
to get following model specification:
CLIP(
  (embedding): CLIPEmbedding(
    (token_embedding): Embedding(49408, 768)
  )
  (layers): ModuleList(
    (0-11): 12 x CLIPLayer(
      (layernorm_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      (attention): SelfAttention(
        (in_proj): Linear(in_features=768, out_features=2304, bias=True)
        (out_proj): Linear(in_features=768, out_features=768, bias=True)
      )
      (layernorm_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      (linear_1): Linear(in_features=768, out_features=3072, bias=True)
      (linear_2): Linear(in_features=3072, out_features=768, bias=True)
    )
  )
  (layernorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
)
Where the Embedding input dim comes from?
vocab.json ther are 49408 key values pairs mapping from words/special letters to numbers
 downloaded from https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main/tokenizer
 
How to use pre-trained CLIP for inference?
load the weights into Class CLIP
>>> CLIP.load_state_dict(state_dict['clip'], strict=True)

see how to convert state_dict from model_converter.py
"""

class CLIPEmbedding(nn.Module):
    def __init__(self, n_vocab: int, n_embd: int, n_token: int):
        super().__init__()
        
        self.token_embedding = nn.Embedding(n_vocab, n_embd)
        # A learnable weight matrix encodes the position information for each token
        self.position_embedding = nn.Parameter(torch.zeros((n_token, n_embd)))
    
    def forward(self, tokens):
        # (Batch_Size, Seq_Len) -> (Batch_Size, Seq_Len, Dim) 
        x = self.token_embedding(tokens)
        # (Batch_Size, Seq_Len) -> (Batch_Size, Seq_Len, Dim)
        x += self.position_embedding
        
        return x

class CLIPLayer(nn.Module):
    def __init__(self, n_head: int, n_embd: int):
        super().__init__()
        
        # Pre-attention norm
        self.layernorm_1 = nn.LayerNorm(n_embd)
        # Self attention
        self.attention = SelfAttention(n_head, n_embd)
        # Pre-FNN norm
        self.layernorm_2 = nn.LayerNorm(n_embd)
        # Feedforward layer
        self.linear_1 = nn.Linear(n_embd, 4 * n_embd)
        self.linear_2 = nn.Linear(4 * n_embd, n_embd)

    def forward(self, x):
        # (Batch_Size, Seq_Len, Dim)
        residue = x
        
        ### SELF ATTENTION ###

        # (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
        x = self.layernorm_1(x)
        
        # (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
        x = self.attention(x, causal_mask=True)
        
        # (Batch_Size, Seq_Len, Dim) + (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
        x += residue

        ### FEEDFORWARD LAYER ###
        # Apply a feedforward layer where the hidden dimension is 4 times the embedding dimension. 

        residue = x
        # (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
        x = self.layernorm_2(x)
        
        # (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, 4 * Dim)
        x = self.linear_1(x)
        
        # (Batch_Size, Seq_Len, 4 * Dim) -> (Batch_Size, Seq_Len, 4 * Dim)
        x = x * torch.sigmoid(1.702 * x)   # QuickGELU activation function
        
        # (Batch_Size, Seq_Len, 4 * Dim) -> (Batch_Size, Seq_Len, Dim)
        x = self.linear_2(x)
        
        # (Batch_Size, Seq_Len, Dim) + (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
        x += residue

        return x

class CLIP(nn.Module):
    def __init__(self):
        super().__init__()
        self.embedding = CLIPEmbedding(49408, 768, 77)

        self.layers = nn.ModuleList([
            CLIPLayer(12, 768) for i in range(12)
        ])

        self.layernorm = nn.LayerNorm(768)
    
    def forward(self, tokens: torch.LongTensor) -> torch.FloatTensor:
        tokens = tokens.type(torch.long)
        
        # (Batch_Size, Seq_Len) -> (Batch_Size, Seq_Len, Dim)
        state = self.embedding(tokens)

        # Apply encoder layers similar to the Transformer's encoder.
        for layer in self.layers: 
            # (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
            state = layer(state)
        # (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
        output = self.layernorm(state)
        
        return output