uripper/AVA
0
1import streamlit as st2import requests3 4BAD_WORD = False5my_api = st.secrets["my_api"]6bad_words = st.secrets["bad_words"]7 8def rev_generate(text, max_length=500, temperature=0.5, top_k=5, do_sample=False, use_cache=True):9 API_URL = "https://api-inference.huggingface.co/models/uripper/ReviewTrainingBot"10 headers = {"Authorization": f"Bearer {my_api}"}11 12 if do_sample:13 use_cache = False14 15 def query(payload):16 response = requests.post(API_URL, headers=headers, json=payload)17 return response.json()18 19 output = query({20 "inputs": f"{text}",21 "parameters": {"max_new_tokens": max_length, "temperature": temperature, "top_p": .95, "do_sample": do_sample, "no_repeat_ngram_size":2},22 "options": {"wait_for_model": True, "use_cache": use_cache},23 })24 return output25 26 27if "persona_chat_history" not in st.session_state:28 st.session_state.persona_chat_history = []29 30if "gordon_chat_history" not in st.session_state:31 st.session_state.gordon_chat_history = []32 33 34def main_page():35 36 CHAT = False37 REVIEW = False38 39 40 41 st.title("AVA")42 st.write("This model generates reviews of films and can be accessed on the drop down menu on the left. \n\nThe model is named after Ava from the movie Ex Machina. To use the model you can enter the name of a movie and generate a review for it or have Ava randomly generate a review. This was created by finetuning a GPT-2 model on a dataset of movie reviews. The dataset was created via scraping around 500,000 letterboxd reviews.")43 44 st.title("Limitations and biases")45 st.write("The main limitations of the review feature are that it is unable to find links between the movie title and the review itself, and struggles to determine positive and negative sentiment based on the score that is given. It however gives consistently plausible reviews, if not very plausible. It is unable to determine fact, and cannot give truthful reviews or reliably determine actors/directors for any given movie. Its main, and only, use case is for entertainment.")46 st.write("The review bot also has social biases. Due to its underlying model, it has many of the same biases as GPT-2. These biases can be found here: https://huggingface.co/gpt2. In addition to these biases, it also struggles with some of the unique examples of this training dataset. For a concrete example of this, it is fairly common for a review of a movie with gay or lesbian characters to be described as being 'very gay' on letterboxd.com. This is almost always used as a positive thing, but the bot itself is incapable of determining that this is a positive sentiment, and will describe random films this way in a manner that seems more like a slur. This language can likely be extended to other ways that have not been discovered yet, and the model should be handled with care.") 47 48 49def review():50 BAD_WORD = False51 st.title("Review")52 53 temperature = st.slider("Temperature", 0.1, 1.0, 0.8, 0.01)54 top_k = st.slider("Top K", 1, 100, 15, 1)55 max_length = st.slider("Max Length", 1, 250, 100, 1)56 do_sample = st.checkbox("Do Sample (If unchecked, will use greedy decoding, not recommended for review due to repetition)", True)57 58 st.write("Please enter the name of the movie you would like to review. First generation may take up to a minute or more, as the model is loading. Latter generations should load faster.")59 in_movie = st.text_input("Movie")60 review_button = st.button("Generate Review")61 random_review = st.button("Random Review")62 st.write("Please only press Generate Review or Random Review once, it will take a short amount of time to load during the first generation.")63 if review_button: 64 in_movie = "Movie: " + in_movie + " Score:"65 output = rev_generate(in_movie, max_length=max_length, temperature=temperature, top_k=top_k, do_sample=do_sample)66 67 check_output = output[0]["generated_text"]68 check_output = check_output.split(" ")69 for i in check_output:70 for j in bad_words:71 if i.lower() is j:72 BAD_WORD =True73 74 75 print(output)76 output = output[0]["generated_text"]77 78 if BAD_WORD == True:79 80 st.write("The bot generated a slur, please try again.")81 BAD_WORD = False82 else:83 out_movie =output.split("Score:")[0]84 out_movie = out_movie.replace("Movie: ", "").replace("|","")85 score = output.split("Review:")[0]86 score = score.split("Score:")[1]87 score = score.replace("|","")88 review = output.split("Review:")[1] 89 90 review = review.replace("…", ".")91 review = review.replace("...", ".").replace("|","")92 review = review.replace("<br/>", "/n").replace("br/>","").replace("br","").replace("<","").replace(">","")93 94 95 st.write("Movie:")96 st.write(out_movie)97 st.write("Score:")98 st.write(score)99 st.write("Review:")100 st.write(review)101 102 if random_review:103 output = rev_generate("Movie:", max_length=max_length, temperature=temperature, top_k=top_k, do_sample=do_sample) 104 check_output = output[0]["generated_text"]105 check_output = check_output.split(" ")106 for i in check_output:107 for j in bad_words:108 if i.lower() is j:109 BAD_WORD =True110 print(output)111 output = output[0]["generated_text"]112 if BAD_WORD == True:113 st.write(i)114 st.write("The bot generated a slur, please try again.")115 BAD_WORD = False116 else:117 out_movie =output.split("Score:")[0]118 out_movie = out_movie.replace("Movie: ", "").replace("|","")119 score = output.split("Review:")[0]120 score = score.split("Score:")[1]121 score = score.replace("|","")122 review = output.split("Review:")[1] 123 124 review = review.replace("…", ".")125 review = review.replace("...", ".").replace("|","")126 review = review.replace("<br/>", "/n").replace("br/>","").replace("br","").replace("<","").replace(">","")127 128 st.write("Movie:")129 st.write(out_movie)130 st.write("Score:")131 st.write(score)132 st.write("Review:")133 st.write(review)134 135 136 137page_names_to_funcs = {138 "Main Page": main_page,139 "Ava": review,140}141 142selected_page = st.sidebar.selectbox("Select a page", page_names_to_funcs.keys())143page_names_to_funcs[selected_page]()144 145