MemeWorld / app.py
EnigmaOfTheWorld's picture
Update app.py
bbc856f
import os
import re
import gradio as gr
from transformers import AutoTokenizer, ViTFeatureExtractor, VisionEncoderDecoderModel
import openai
openai.api_key = os.environ['OPENAI_KEY']
## Training models
device='cpu'
encoder_checkpoint = "nlpconnect/vit-gpt2-image-captioning"
decoder_checkpoint = "nlpconnect/vit-gpt2-image-captioning"
model_checkpoint = "nlpconnect/vit-gpt2-image-captioning"
feature_extractor = ViTFeatureExtractor.from_pretrained(encoder_checkpoint)
tokenizer = AutoTokenizer.from_pretrained(decoder_checkpoint)
model = VisionEncoderDecoderModel.from_pretrained(model_checkpoint).to(device)
## READING THE IMAGE
## Extracting features from image
## then create a context for the image like
## Then input the department and context extracted and send it to LLM to get captio meme
def predict(department,image,max_length=64, num_beams=4):
image = image.convert('RGB')
image = feature_extractor(image, return_tensors="pt").pixel_values.to(device)
print(image)
clean_text = lambda x: x.replace('<|endoftext|>','').split('\n')[0]
print(clean_text)
caption_ids = model.generate(image, max_length = max_length)[0]
print(caption_ids)
caption_text = clean_text(tokenizer.decode(caption_ids))
print(caption_text)
dept=department
context= caption_text
response = openai.Completion.create(
model="text-davinci-003",
prompt=f'create non offensive one line meme for given department and context\n\ndepartment- data science\ncontext-a man sitting on a bench with a laptop\nmeme- \"I\'m not a data scientist, but I play one on my laptop.\"\n\ndepartment-startup\ncontext-a young boy is smiling while using a laptop\nmeme-\"When your startup gets funded and you can finally afford a new laptop\"\n\ndepartment- {dept}\ncontext-{context}\nmeme-',
max_tokens=20,
temperature=0.8)
reponse = response.choices[0].text
reponse = reponse.replace("department", "")
Feedback_SQL="DEPT"+dept+"CAPT"+caption_text+"MAMAY"+reponse
return reponse
output = gr.outputs.Textbox(type="text",label="Meme")
examples = [f"example{i}.png" for i in range(1,7)]
## GRADIO INTERFACE
description= " Looking for a fun and easy way to generate memes? Look no further than Meme world! Leveraging large language models like GPT-3PT-3 / Ai21 / Cohere, you can create memes that are sure to be a hit with your friends or network. Created with ♥️ by Arsalan @[Xaheen](https://www.linkedin.com/in/sallu-mandya/). kindly share your thoughts in discussion session and use the app responsibly #NO_Offense \n \n built with ❤️ @[Xhaheen](https://www.linkedin.com/in/sallu-mandya/)"
title = "Meme world 🖼️"
dropdown=["data science", "product management","marketing","startup" ,"agile","crypto" , "SEO" ]
article = "Created By : Xaheen "
theme = gr.themes.Glass(
primary_hue="cyan",
neutral_hue="gray",font_mono=['IBM Plex Mono', 'ui-monospace', 'Consolas', 'monospace'],
)
interface = gr.Interface(
fn=predict,
inputs = [gr.inputs.Dropdown(dropdown),gr.inputs.Image(label="Upload your Image", type = 'pil', optional=True)],
theme=theme,
outputs=output,
examples =[['data science', 'example5.png'],
['product management', 'example2.png'],
['startup', 'example3.png'],
['marketing', 'example4.png'],
['agile', 'example1.png'],
['crypto', 'example6.png']],
title=title,
description=description,
article = article
)
interface.launch(debug=True)