T5/generate-dict-embeddingsT5.py

#!/bin/env python

"""
    (T5 counterpart of "generate-dict-embeddingsXL.py".
"""

outputfile="embeddingsT5.temp.safetensors"

import sys
import torch
from safetensors.torch import save_file
from transformers import T5Tokenizer,T5EncoderModel

processor=None
tmodel=None

device=torch.device("cuda")

def initT5model():
    global processor,tmodel
    T="mcmonkey/google_t5-v1_1-xxl_encoderonly"
    processor = T5Tokenizer.from_pretrained(T)
    tmodel = T5EncoderModel.from_pretrained(T).to(device)


def embed_from_text(text):
    global processor,tmodel
    #print("Word:"+text)
    tokens = processor(text, return_tensors="pt")
    tokens.to(device)

    if len(tokens.input_ids) >2:
        print("ERROR: expected single token per word")
        print(text)
        exit(1)
        # We can only accept single-token words, because we want our output 
        # to be a single embedding per word, and we dont have an official
        # way to merge multiple T5 embeddings into one, like CLIP does

    with torch.no_grad():
        outputs = tmodel(tokens.input_ids)

    embedding = outputs.last_hidden_state[0][0]
    #print(encoding.shape)
    # Shape of this is (1,2,4096)
    
    return embedding

initT5model()

print("Reading in 'dictionary'")
with open("dictionary","r") as f:
    tokendict = f.readlines()
    tokendict = [token.strip() for token in tokendict]  # Remove trailing newlines


count=1
all_embeddings = []

for word in tokendict:
    emb = embed_from_text(word)
    emb=emb.unsqueeze(0) # stupid matrix magic to make torch.cat work
    all_embeddings.append(emb)
    count+=1
    if (count %100) ==0:
        print(count)


embs = torch.cat(all_embeddings,dim=0)
print("Shape of result = ",embs.shape)

if len(embs.shape) != 2:
    print("Sanity check: result is wrong shape: it wont work")

print(f"Saving the calculatiuons to {outputfile}...")
save_file({"embeddings": embs}, outputfile)