asi/igpt-fr-cased-base

Model description

iGPT-fr 🇫🇷 is a GPT model for French pre-trained incremental language model developped by the Laboratoire de Linguistique Formelle (LLF). We adapted GPT-fr 🇫🇷 model to generate images conditionned by text inputs.

Intended uses & limitations

The model can be leveraged for image generation tasks. The model is currently under a developpment phase.

How to use

The model might be used through the 🤗 Transformers librairie. You will also need to install the Taming Transformers library for high-resolution image synthesis:

pip install git+https://github.com/CompVis/taming-transformers.git

from transformers import GPT2Tokenizer, GPT2LMHeadModel
from huggingface_hub import hf_hub_download
from omegaconf import OmegaConf
from taming.models import vqgan
import torch
from PIL import Image
import numpy as np

# Load VQGAN model
vqgan_ckpt = hf_hub_download(repo_id="boris/vqgan_f16_16384", filename="model.ckpt", force_download=False)
vqgan_config = hf_hub_download(repo_id="boris/vqgan_f16_16384", filename="config.yaml", force_download=False)

config = OmegaConf.load(vqgan_config)
vqgan_model = vqgan.VQModel(**config.model.params)
vqgan_model.eval().requires_grad_(False)
vqgan_model.init_from_ckpt(vqgan_ckpt)

# Load pretrained model
model = GPT2LMHeadModel.from_pretrained("asi/igpt-fr-cased-base")
model.eval()
tokenizer = GPT2Tokenizer.from_pretrained("asi/igpt-fr-cased-base")

# Generate a sample of text
input_sentence = "Une carte de l'europe"
input_ids = tokenizer.encode(input_sentence, return_tensors='pt')
input_ids = torch.cat((input_ids, torch.tensor([[50000]])), 1)  # Add image generation token

greedy_output = model.generate(
  input_ids.to(device), 
  max_length=256+input_ids.shape[1],
  do_sample=True, 
  top_p=0.92, 
  top_k=0)

def custom_to_pil(x):
  x = x.detach().cpu()
  x = torch.clamp(x, -1., 1.)
  x = (x + 1.)/2.
  x = x.permute(1,2,0).numpy()
  x = (255*x).astype(np.uint8)
  x = Image.fromarray(x)
  if not x.mode == "RGB":
    x = x.convert("RGB")
  return x

z_idx = greedy_output[0, input_ids.shape[1]:] - 50001
z_quant = vqgan_model.quantize.get_codebook_entry(z_idx, shape=(1, 16, 16, 256))
x_rec = vqgan_model.decode(z_quant).to('cpu')[0]
display(custom_to_pil(x_rec))

You may also filter results based on CLIP:

from tqdm import tqdm

def hallucinate(prompt, num_images=64):
    input_ids = tokenizer.encode(prompt, return_tensors='pt')
    input_ids = torch.cat((input_ids, torch.tensor([[50000]])), 1).to(device)  # Add image generation token

    all_images = []
    for i in tqdm(range(num_images)):
        greedy_output = model.generate(
          input_ids.to(device), 
          max_length=256+input_ids.shape[1],
          do_sample=True,
          top_p=0.92, 
          top_k=0)

        z_idx = greedy_output[0, input_ids.shape[1]:] - 50001
        z_quant = vqgan_model.quantize.get_codebook_entry(z_idx, shape=(1, 16, 16, 256))
        x_rec = vqgan_model.decode(z_quant).to('cpu')[0]
        all_images.append(custom_to_pil(x_rec))
    return all_images

input_sentence = "Une carte de l'europe"
all_images = hallucinate(input_sentence)

from transformers import pipeline

opus_model = "Helsinki-NLP/opus-mt-fr-en"
opus_translator = pipeline("translation", model=opus_model)

opus_translator(input_sentence)

from transformers import CLIPProcessor, CLIPModel

clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
clip_processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")

def clip_top_k(prompt, images, k=8):
  prompt_fr = opus_translator(input_sentence)[0]['translation_text']
  inputs = clip_processor(text=prompt_fr, images=images, return_tensors="pt", padding=True)
  outputs = clip_model(**inputs)
  logits = outputs.logits_per_text # this is the image-text similarity score
  scores = np.array(logits[0].detach()).argsort()[-k:][::-1]
  return [images[score] for score in scores]

filtered_images = clip_top_k(input_sentence, all_images)

for fi in filtered_images:
  display(fi)

Training data

We created a dedicated corpus to train our generative model. The training corpus consists in text-image pairs. We aggregated portions from existing corpora: Laion-5B and WIT. The final dataset includes 10,807,534 samples.

Training procedure

We pre-trained the model on the new CNRS (French National Centre for Scientific Research) Jean Zay supercomputer. We perform the training within a total of 140 hours of computation on Tesla V-100 hardware (TDP of 300W). The training was distributed on 8 compute nodes of 8 GPUs. We used data parallelization in order to divide each micro-batch on the computing units. We estimated the total emissions at 1161.22 kgCO2eq, using the Machine Learning Impact calculator presented in Lacoste et al., (2019).