graph-byclip.py

#!/bin/env python

""" Work in progress
Plan:
   Modded version of graph-embeddings.py
   Just to see if using different CLIP module changes values significantly
   (It does not)

   It does have the small bonus feature of being able to accept a purely
   numerical tokenid in liu of a number, if you use the syntax,
   "#345".
   You can input a text string, or a single numeric code, per input

   This code requires
       pip install git+https://github.com/openai/CLIP.git
"""


import sys
import json
import torch
import clip

import PyQt5
import matplotlib
matplotlib.use('QT5Agg')  # Set the backend to TkAgg

import matplotlib.pyplot as plt

### The stablediffusion standard model is Vit-L/14
#CLIPname= "ViT-B/16"
CLIPname= "ViT-L/14"
#CLIPname= "ViT-L/14@336px"
# Available models:
#  'RN50', 'RN101', 'RN50x4', 'RN50x16', 'RN50x64', 'ViT-B/32', 'ViT-B/16', 'ViT-L/14', 'ViT-L/14@336px'

device=torch.device("cuda")
print("loading CLIP model",CLIPname)
model, processor = clip.load(CLIPname,device=device)
model.cuda().eval()
print("done")

def embed_from_tokenid(num):
    # A bit sleazy, but, eh.
    tokens = clip.tokenize("dummy").to(device)
    tokens[0][1]=num

    with torch.no_grad():
        embed = model.encode_text(tokens)
        return embed



def embed_from_text(text):
    if text[0]=="#":
        print("Converting string to number")
        return embed_from_tokenid(int(text[1:]))

    tokens = clip.tokenize(text).to(device)
    print("Tokens for",text,"=",tokens)

    with torch.no_grad():
        embed = model.encode_text(tokens)
        return embed


fig, ax = plt.subplots()


text1 = input("First prompt or #tokenid: ")
text2 = input("Second prompt(or leave blank): ")


print("generating embeddings for each now")
emb1 = embed_from_text(text1)
print("shape of emb1:",emb1.shape)

graph1=emb1[0].tolist()
ax.plot(graph1, label=text1[:20])

if len(text2) >0:
    emb2 = embed_from_text(text2)
    graph2=emb2[0].tolist()
    ax.plot(graph2, label=text2[:20])

# Add labels, title, and legend
#ax.set_xlabel('Index')
ax.set_ylabel('Values')
ax.set_title(f"Graph of Embeddings in {CLIPname}")
ax.legend()

# Display the graph
print("Pulling up the graph")
plt.show()