app.py

import streamlit as st
from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed
from transformers import pipeline
import torch
import json
import pandas as pd

@st.cache(allow_output_mutation=True)
def load_tokenizer(model_ckpt):
    return AutoTokenizer.from_pretrained(model_ckpt)

@st.cache(allow_output_mutation=True)
def load_model(model_ckpt):
    model = AutoModelForCausalLM.from_pretrained(model_ckpt, low_cpu_mem_usage=True)
    return model

@st.cache()
def load_examples():
    with open("examples.json", "r") as f:
        examples = json.load(f)
    return examples

st.set_page_config(page_icon=':laptop:', layout="wide")
           

st.sidebar.header("Models")
models = ["CodeParrot", "OPT", "InCoder"]
selected_models = st.sidebar.multiselect('Select code generation models to compare:', 
                         models,
                         default=["CodeParrot"])
st.sidebar.header("Tasks")
tasks = [" ","Model architecture", "Model evaluation", "Pretraining dataset", "Code generation"]
selected_task = st.sidebar.selectbox("Select a task:", tasks)


tokenizer1 = load_tokenizer("lvwerra/codeparrot")
model1 = load_model("lvwerra/codeparrot")
tokenizer2 = load_tokenizer("facebook/incoder-1B")
model2 = load_model("facebook/incoder-1B")
#tokenizer3 = load_tokenizer("facebook/opt-1.3b")
#model3 = load_model("facebook/opt-1.3b")
pipelines = {}
for element in models:
    if element == "CodeParrot":
        pipelines[element] = pipeline("text-generation", model=model1, tokenizer=tokenizer1)
    elif element == "InCoder":
        tokenizer = load_tokenizer("facebook/incoder-1B")
        model = load_model("facebook/incoder-1B")
        pipelines[element] = pipeline("text-generation", model=model2, tokenizer=tokenizer2)
    #else:
    #    tokenizer = load_tokenizer("facebook/opt-1.3b")
    #    model = load_model("facebook/opt-1.3b")
    #    pipelines[element] = pipeline("text-generation", model=model3, tokenizer=tokenizer3)
        
examples = load_examples()
example_names = [example["name"] for example in examples]
name2id = dict([(name, i) for i, name in enumerate(example_names)])
set_seed(42)
gen_kwargs = {}

if selected_task == " ":
    st.title("Code Generation Models comparison")
    with open("intro.txt", "r") as f:
        intro = f.read()
    st.markdown(intro)
elif selected_task == "Pretraining dataset":
    st.title("Pretraining datasets 📚")
    st.markdown("Preview of some code files from Github repositories")   
    df = pd.read_csv("preview-github-data.csv")
    st.dataframe(df)
    for model in selected_models:
        with open(f"datasets/{model.lower()}.txt", "r") as f:
            text = f.read()
        st.markdown(f"### {model}:")
        st.markdown(text)   
elif selected_task == "Model architecture":
    st.title("Model architecture 🔨")
    for model in selected_models:
        with open(f"architectures/{model.lower()}.txt", "r") as f:
            text = f.read()
        st.markdown(f"## {model}:")
        st.markdown(text) 
elif selected_task == "Code generation":
    st.title("Code generation 💻")
    st.sidebar.header("Examples")
    selected_example = st.sidebar.selectbox("Select one of the following examples:", example_names)
    example_text = examples[name2id[selected_example]]["value"]
    default_length = examples[name2id[selected_example]]["length"]
    st.sidebar.header("Generation settings")
    gen_kwargs["do_sample"] = st.sidebar.radio("Decoding strategy:",  ["Greedy", "Sample"]) == "Sample"
    gen_kwargs["max_new_tokens"] = st.sidebar.slider("Number of tokens to generate:", value=default_length, min_value=8, step=8, max_value=256)
    if gen_kwargs["do_sample"]:
        gen_kwargs["temperature"] = 0.2
        gen_kwargs["top_k"] = 0
        gen_kwargs["top_p"] = 0.95
    gen_prompt = st.text_area("Generate code with prompt:", value=example_text, height=220,).strip()
    if st.button("Generate code!"):
        with st.spinner("Generating code..."):
            for model in selected_models:
                if model != "OPT":
                    pipe = pipelines[model]
                    generated_text = pipe(gen_prompt, **gen_kwargs)[0]['generated_text']
                    st.markdown(f"{model}:")
                    st.code(generated_text)