import streamlit as st st.set_page_config(layout="wide") from uuid import uuid4 import model_comparison as MCOMP import model_loading as MLOAD import model_inferencing as MINFER import user_evaluation_variables from pathlib import Path import tab_manager import yaml from yaml.loader import SafeLoader from PIL import Image import os from huggingface_hub import CommitScheduler, login, snapshot_download login(token=os.environ.get("HF_TOKEN"), write_permission=True) TBYB_LOGO = Image.open('./assets/TBYB_logo_light.png') def setup_page_banner(): global USER_LOGGED_IN # for tab in [tab1, tab2, tab3, tab4, tab5]: c1,c2,c3,c4,c5,c6,c7,c8,c9 = st.columns(9) with c5: st.image(TBYB_LOGO, use_column_width=True) for col in [c1,c2,c3,c4,c5,c6,c7,c8,c9]: col = None st.title('Try Before You Bias (TBYB)') st.write('*A Quantitative T2I Bias Evaluation Tool*') def setup_how_to(): expander = st.expander("How to Use") expander.write("1. Watch our tutorial video series on [YouTube](https://www.youtube.com/channel/UCk-0xyUyT0MSd_hkp4jQt1Q) (Note that the tutorial series was developed for V1.0) \n" "2. Navigate to the '\U0001F527 Setup' tab and input the ID of the HuggingFace \U0001F917 T2I model you want to evaluate.\n" " A list of the most downloaded text-to-image models can be found [here](https://huggingface.co/models?pipeline_tag=text-to-image&sort=downloads)") expander.image(Image.open('./assets/HF_MODEL_ID_EXAMPLE.png')) expander.write("3. Test your chosen model by generating an image using an input prompt e.g.: 'A corgi with some cool sunglasses'\n") expander.image(Image.open('./assets/lykon_corgi.png')) expander.write("4. Navigate to the '\U0001F30E Bias Evaluation (BEval)' or '\U0001F3AF Task-Oriented BEval' tabs " " to evaluate your model once it has been loaded\n" "5. Once you have generated some evaluation images, head over to the '\U0001F4C1 Generated Images' tab to have a look at them\n" "6. To check out your evaluations or all of the TBYB Community evaluations, head over to the '\U0001F4CA Model Comparison' tab\n" "7. For more information about the evaluation process, see our paper on [ArXiv](https://arxiv.org/abs/2312.13053) or navigate to the " " '\U0001F4F0 Additional Information' tab for a TL;DR.\n" "8. For any questions or to report any bugs/issues. Please contact jordan.vice@uwa.edu.au.\n") def setup_additional_information_tab(tab): with tab: st.header("1. Quantifying Bias in Text-to-Image (T2I) Generative Models") st.markdown( """ *Based on the pre-print of the same name available on [ArXiv](https://arxiv.org/abs/2312.13053). Authors: Jordan Vice, Naveed Akhtar, Richard Hartley and Ajmal Mian This web-app was developed by **Jordan Vice** to accompany the article, serving as a practical implementation of how T2I model biases can be quantitatively assessed and compared. Evaluation results from all *base* models discussed in the paper have been incorporated into the TBYB community results and we hope that others share their evaluations as we look to further the discussion on transparency and reliability of T2I models. """) st.header('2. A (very) Brief Summary') st.image(Image.open('./assets/TBYB_flowchart.png')) st.markdown( """ Bias in text-to-image models can propagate unfair social representations and could be exploited to aggressively market ideas or push controversial or sinister agendas. Existing T2I model bias evaluation methods focused on social biases. So, we proposed a bias evaluation methodology that considered general and task-oriented biases, spawning the Try Before You Bias (**TBYB**) application as a result. """ ) st.markdown( """ We proposed three novel metrics to quantify T2I model biases: 1. Distribution Bias - $B_D$ 2. Jaccard Hallucination - $H_J$ 3. Generative Miss Rate - $M_G$ Open the appropriate drop-down menu to understand the logic and inspiration behind metric. """ ) c1,c2,c3 = st.columns(3) with c1: with st.expander("Distribution Bias - $B_D$"): st.markdown( """ Using the Area under the Curve (AuC) as an evaluation metric in machine learning is not novel. However, in the context of T2I models, using AuC allows us to define the distribution of objects that have been detected in generated output image scenes. So, everytime an object is detected in a scene, we update a dictionary (which is available for download after running an evaluation). After evaluating a full set of images, you can use this information to determine what objects appear more frequently than others. After all images are evaluated, we sort the objects in descending order and normalize the data. We then use the normalized values to calculate $B_D$, using the trapezoidal AuC rule i.e.: $B_D = \\Sigma_{i=1}^M\\frac{n_i+n_{i=1}}{2}$ So, if a user conducts a task-oriented study on biases related to **dogs** using a model that was heavily biased using pictures of animals in the wild. You might find that after running evaluations, the most common objects detected were trees and grass - even if these objects weren't specified in the prompt. This would result in a very low $B_D$ in comparison to a model that for example was trained on images of dogs and animals in various different scenarios $\\rightarrow$ which would result in a *higher* $B_D$ in comparison. """ ) with c2: with st.expander("Jaccard Hallucination - $H_J$"): st.markdown( """ Hallucination is a very common phenomena that is discussed in relation to generative AI, particularly in relation to some of the most popular large language models. Depending on where you look, hallucinations can be defined as being positive, negative, or just something to observe $\\rightarrow$ a sentiment that we echo in our bias evaluations. Now, how does hallucination tie into bias? In our work, we use hallucination to define how often a T2I model will *add* objects that weren't specified OR, how often it will *omit* objects that were specified. This indicates that there could be an innate shift in bias in the model, causing it to add or omit certain objects. Initially, we considered using two variables $H^+$ and $H^-$ to define these two dimensions of hallucination. Then, we considered the Jaccard similarity coefficient, which measures the similarity *and* diversity of two sets of objects/samples - defining this as Jaccard Hallucination - $H_J$. Simply put, we define the set of objects detected in the input prompt and then detect the objects in the corresponding output image. Then, we determine the intersect over union. For a model, we calculate the average $H_J$ across generated images using: $H_J = \\frac{\Sigma_{i=0}^{N-1}1-\\frac{\mathcal{X}_i\cap\mathcal{Y}_i}{\mathcal{X}_i\cup\mathcal{Y}_i}}{N}$ """ ) with c3: with st.expander("Generative Miss Rate - $M_G$"): st.markdown( """ Whenever fairness and trust are discussed in the context of machine learning and AI systems, performance is always highlighted as a key metric - regardless of the downstream task. So, in terms of evaluating bias, we thought that it would be important to see if there was a correlation between bias and performance (as we predicted). And while the other metrics do evaluate biases in terms of misalignment, they do not consider the relationship between bias and performance. We use an additional CLIP model to assist in calculating Generative Miss Rate - $M_G$. Logically, as a model becomes more biased, it will begin to diverge away from the intended target and so, the miss rate of the generative model will increase as a result. This was a major consideration when designing this metric. We use the CLIP model as a binary classifier, differentiating between two classes: - the prompt used to generate the image - **NOT** the prompt Through our experiments on intentionally-biased T2I models, we found that there was a clear relationship between $M_G$ and the extent of bias. So, we can use this metric to quantify and infer how badly model performances have been affected by their biases. """ ) st.header('3. TBYB Constraints') st.markdown( """ While we have attempted to design a comprehensive, automated bias evaluation tool. We must acknowledge that in its infancy, TBYB has some constraints: - We have not checked the validity of *every* single T2I model and model type on HuggingFace so we cannot promise that all T2I models will work - if you run into any issues that you think should be possible, feel free to reach out! - Currently, a model_index.json file is required to load models and use them with TBYB, we will look to address other models in the future. - Target models must be public. Gated models are currently not supported. - TBYB only works on T2I models hosted on HuggingFace, other model repositories are not currently supported. - Adaptor models are currently not supported, we will look to add evaluation functionalities of these models in the future. - Download, generation, inference and evaluation times are hardware and model dependent. We are currently deploying the TBYB space on a T4 GPU. With more funding and interest (usage). - BLIP and CLIP models used for evaluations are limited by their own biases and object recognition capabilities. However, manual evaluations of bias could result in subjective labelling biases. Keep in mind that these constraints may be removed or added to any time. """) st.header('4. Disclaimer') st.markdown( """ Given this application is used for the assessment of T2I biases and relies on pre-trained models available on HuggingFace, we are not responsible for any content generated by public-facing models that have been used to generate images using this application. Bias cannot be easily measured and we do not claim that our approach is without any faults. TBYB is proposed as an auxiliary tool to assess model biases and thus, if a chosen model is found to output insensitive, disturbing, distressing or offensive images that propagate harmful stereotypes or representations of marginalised groups, please address your concerns to the model providers. However, given the TBYB tool is designed for bias quantification and is driven by transparency, it would be beneficial to the TBYB community to share evaluations of biased T2I models! Despite only being able to assess HuggingFace \U0001F917 models, we share no association with them outside of hosting TBYB as a HuggingFace space. Given their growth in popularity in the computer science community and their host of T2I model repositories, we have decided to host our web-app here. For further questions/queries or if you want to simply strike a conversation, please reach out to Jordan Vice at: jordan.vice@uwa.edu.au""") st.write("*V2.0 - Released May 17 2024*") setup_page_banner() setup_how_to() tab1, tab2, tab3, tab4, tab5, tab6 = st.tabs(["\U0001F527 Setup", "\U0001F30E Bias Evaluation (BEval)", "\U0001F3AF Task-Oriented BEval.", "\U0001F4CA Model Comparison", "\U0001F4C1 Generated Images", "\U0001F4F0 Additional Information"]) setup_additional_information_tab(tab6) # PLASTER THE LOGO EVERYWHERE tab2.subheader("General Bias Evaluation") tab2.write("Waiting for \U0001F527 Setup to be complete...") tab3.subheader("Task-Oriented Bias Evaluation") tab3.write("Waiting for \U0001F527 Setup to be complete...") tab4.write("Check out other model evaluation results from users across the **TBYB** Community! \U0001F30E ") tab4.write("You can also just compare your own model evaluations by clicking the '*Personal Evaluation*' buttons") MCOMP.initialise_page(tab4) tab5.subheader("Generated Images from General and Task-Oriented Bias Evaluations") tab5.write("Waiting for \U0001F527 Setup to be complete...") with tab1: with st.form("model_definition_form", clear_on_submit=True): modelID = st.text_input('Input the HuggingFace \U0001F917 T2I model_id for the model you ' 'want to analyse e.g.: "runwayml/stable-diffusion-v1-5\n' '[(Shortcut to T2I models - sorted by the most downloads)](https://huggingface.co/models?pipeline_tag=text-to-image&sort=downloads)') submitted1 = st.form_submit_button("Submit") if modelID: with st.spinner('Checking if ' + modelID + ' is valid and downloading it (if required)'): modelLoaded = MLOAD.check_if_model_exists(modelID) if modelLoaded is not None: # st.write("Located " + modelID + " model_index.json file") st.write("Located " + modelID) modelType = MLOAD.get_model_info(modelLoaded) if modelType is not None: st.write("Model is of Type: ", modelType) if submitted1: MINFER.TargetModel = MLOAD.import_model(modelID, modelType) if MINFER.TargetModel is not None: st.write("Text-to-image pipeline looks like this:") st.write(MINFER.TargetModel) user_evaluation_variables.MODEL = modelID user_evaluation_variables.MODEL_TYPE = modelType else: st.error('The Model: ' + modelID + ' does not appear to exist or the model does not contain a model_index.json file.' ' Please check that that HuggingFace repo ID is valid.' ' For more help, please see the "How to Use" Tab above.', icon="🚨") if modelID: with st.form("example_image_gen_form", clear_on_submit=True): testPrompt = st.text_input('Input a random test prompt to test out your ' 'chosen model and see if its generating images:') submitted2 = st.form_submit_button("Submit") if testPrompt and submitted2: with st.spinner("Generating an image with the prompt:\n"+testPrompt+"(This may take some time)"): testImage = MINFER.generate_test_image(MINFER.TargetModel, testPrompt) st.image(testImage, caption='Model: ' + modelID + ' Prompt: ' + testPrompt) st.write('''If you are happy with this model, navigate to the other tabs to evaluate bias! Otherwise, feel free to load up a different model and run it again''') if MINFER.TargetModel is not None: tab_manager.completed_setup([tab2, tab3, tab4, tab5], modelID)