import streamlit as st import model_inferencing as MINFER import general_bias_measurement as GBM import model_comparison as MCOMP import user_evaluation_variables import pandas as pd import numpy as np import json import csv from itertools import cycle import random import time import datetime import zipfile from io import BytesIO, StringIO def completed_setup(tabs, modelID): with tabs[0]: st.write("\U0001F917 ", modelID, " has been loaded!") st.write("Ready for General Bias Evaluation") # general_bias_eval_setup(tabs[0]) with tabs[1]: st.write("\U0001F917 ", modelID, " has been loaded!") st.write("Ready for Task-Oriented Bias Evaluation") with tabs[3]: if not all([user_evaluation_variables.OBJECT_IMAGES_IN_UI, user_evaluation_variables.OCCUPATION_IMAGES_IN_UI, user_evaluation_variables.TASK_IMAGES_IN_UI]): st.write("\U0001F917 ", modelID, " has been loaded!") st.write("Waiting for Images to be generated.") # if any([user_evaluation_variables.OBJECT_IMAGES_IN_UI, user_evaluation_variables.OCCUPATION_IMAGES_IN_UI, # user_evaluation_variables.TASK_IMAGES_IN_UI]): update_images_tab(tabs[3]) with tabs[0]: general_bias_eval_setup(tabs[0], modelID, tabs[3]) with tabs[1]: task_oriented_bias_eval_setup(tabs[1],modelID, tabs[3]) def general_bias_eval_setup(tab, modelID, imagesTab): generalBiasSetupDF_EVAL = pd.DataFrame( { "GEN Eval. Variable": ["No. Images to Generate per prompt", "No. Inference Steps", "Image Size (N x N)"], "GEN Values": ["10", "100", "512"], } ) generalBiasSetupDF_TYPE = pd.DataFrame( { "Image Types": ["Objects", "Person in Frame", "Occupations / Label"], "Check": [True, True, True], } ) tableColumn1, tableColumn2 = st.columns(2) with tab: with tableColumn1: GENValTable = st.data_editor( generalBiasSetupDF_EVAL, column_config={ "GEN Eval. Variable": st.column_config.Column( "Variable", help="General Bias Evaluation variable to control extent of evaluations", width=None, required=None, disabled=True, ), "GEN Values": st.column_config.Column( "Values", help="Input values in this column", width=None, required=True, disabled=False, ), }, hide_index=True, num_rows="fixed", ) with tableColumn2: GENCheckTable = st.data_editor( generalBiasSetupDF_TYPE, column_config={ "Check": st.column_config.CheckboxColumn( "Select", help="Select the types of images you want to generate", default=False, ) }, disabled=["Image Types"], hide_index=True, num_rows="fixed", ) if st.button('Evaluate!', key="EVAL_BUTTON_GEN"): initiate_general_bias_evaluation(tab, modelID, [GENValTable, GENCheckTable], imagesTab) st.rerun() if user_evaluation_variables.RUN_TIME and user_evaluation_variables.CURRENT_EVAL_TYPE == 'general': GBM.output_eval_results(user_evaluation_variables.EVAL_METRICS, 21, 'general') st.write("\U0001F553 Time Taken: ", user_evaluation_variables.RUN_TIME) saveEvalsButton = st.button("Save + Upload Evaluations", key='SAVE_EVAL_GEN') saveDistButton = st.button("Download Object Distribution", key='SAVE_TOP_GEN') if saveEvalsButton: st.write("Saving and uploading evaluations") user_evaluation_variables.update_evaluation_table('general',False) user_evaluation_variables.reset_variables('general') if saveDistButton: download_word_distribution_csv(user_evaluation_variables.EVAL_METRICS, user_evaluation_variables.EVAL_ID, 'general') def task_oriented_bias_eval_setup(tab,modelID,imagesTab): biasSetupDF_EVAL = pd.DataFrame( { "TO Eval. Variable": ["No. Images to Generate per prompt", "No. Inference Steps", "Image Size (N x N)"], "TO Values": ["10", "100", "512"], } ) with tab: TOValTable = st.data_editor( biasSetupDF_EVAL, column_config={ "TO Eval. Variable": st.column_config.Column( "Variable", help="General Bias Evaluation variable to control extent of evaluations", width=None, required=None, disabled=True, ), "TO Values": st.column_config.Column( "Values", help="Input values in this column", width=None, required=True, disabled=False, ), }, hide_index=True, num_rows="fixed", ) target = st.text_input('What is the single-token target of your task-oriented evaluation study ' 'e.g.: "burger", "coffee", "men", "women"') if st.button('Evaluate!', key="EVAL_BUTTON_TO"): if len(target) > 0: initiate_task_oriented_bias_evaluation(tab, modelID, TOValTable, target, imagesTab) st.rerun() else: st.error('Please input a target for your task-oriented analysis', icon="🚨") # update_images_tab(imagesTab) if user_evaluation_variables.RUN_TIME and user_evaluation_variables.CURRENT_EVAL_TYPE == 'task-oriented': GBM.output_eval_results(user_evaluation_variables.EVAL_METRICS, 21, 'task-oriented') st.write("\U0001F553 Time Taken: ", user_evaluation_variables.RUN_TIME) saveEvalsButton = st.button("Save + Upload Evaluations", key='SAVE_EVAL_TASK') saveDistButton = st.button("Download Object Distribution", key='SAVE_TOP_TASK') if saveEvalsButton: st.write("Saving and uploading evaluations") user_evaluation_variables.update_evaluation_table('task-oriented',False) user_evaluation_variables.reset_variables('task-oriented') if saveDistButton: download_word_distribution_csv(user_evaluation_variables.EVAL_METRICS, user_evaluation_variables.EVAL_ID, user_evaluation_variables.TASK_TARGET) # update_images_tab(imagesTab) def download_word_distribution_csv(data, evalID, evalType): filePath = './'+evalID+'_'+evalType+'_word_distribution.csv' listOfObjects = list(data[0].items()) with open(filePath, 'w', newline='') as fp: csvwriter = csv.writer(fp) csvwriter.writerows([["Evaluation ID", evalID], ["Distribution Bias", data[2]], ["Jaccard hallucination", np.mean(data[3])], ["Generative Miss Rate", np.mean(data[4])]]) csvwriter.writerow(['Position', 'Object', 'No. Occurences', 'Normalized']) for obj, val, norm, ii in zip(listOfObjects, data[0].values(), data[1], range(len(listOfObjects))): csvwriter.writerow([ii, obj[0], val, norm]) st.success('Successfully downloaded word distribution data!', icon="✅") def initiate_general_bias_evaluation(tab, modelID, specs, imagesTab): startTime = time.time() objectData = None occupationData = None objects = [] actions = [] occupations = [] occupationDescriptors = [] objectPrompts = None occupationPrompts = None objectImages = [] objectCaptions = [] occupationImages = [] occupationCaptions = [] evaluationImages = [] evaluationCaptions = [] with tab: st.write("Initiating General Bias Evaluation Experiments with the following setup:") st.write(" ***Model*** = ", modelID) infoColumn1, infoColumn2 = st.columns(2) with infoColumn1: st.write(" ***No. Images per prompt*** = ", specs[0]["GEN Values"][0]) st.write(" ***No. Steps*** = ", specs[0]["GEN Values"][1]) st.write(" ***Image Size*** = ", specs[0]["GEN Values"][2], "$\\times$", specs[0]["GEN Values"][2]) with infoColumn2: st.write(" ***Objects*** = ", specs[1]["Check"][0]) st.write(" ***Objects and Actions*** = ", specs[1]["Check"][1]) st.write(" ***Occupations*** = ", specs[1]["Check"][2]) st.markdown("___") if specs[1]["Check"][0]: objectData = read_csv_to_list("./data/list_of_objects.csv") if specs[1]["Check"][2]: occupationData = read_csv_to_list("./data/list_of_occupations.csv") if objectData == None and occupationData == None: st.error('Make sure that at least one of the "Objects" or "Occupations" rows are checked', icon="🚨") else: if specs[1]["Check"][0]: for row in objectData[1:]: objects.append(row[0]) if specs[1]["Check"][1]: for row in objectData[1:]: actions.append(row[1:]) if specs[1]["Check"][2]: for row in occupationData[1:]: occupations.append(row[0]) occupationDescriptors.append(row[1:]) with infoColumn1: st.write("***No. Objects*** = ", len(objects)) st.write("***No. Actions*** = ", len(actions)*3) with infoColumn2: st.write("***No. Occupations*** = ", len(occupations)) st.write("***No. Occupation Descriptors*** = ", len(occupationDescriptors)*3) if len(objects) > 0: objectPrompts = MINFER.construct_general_bias_evaluation_prompts(objects, actions) if len(occupations) > 0: occupationPrompts = MINFER.construct_general_bias_evaluation_prompts(occupations, occupationDescriptors) if objectPrompts is not None: OBJECTprogressBar = st.progress(0, text="Generating Object-related images. Please wait.") objectImages, objectCaptions = MINFER.generate_test_images(OBJECTprogressBar, "Generating Object-related images. Please wait.", objectPrompts, int(specs[0]["GEN Values"][0]), int(specs[0]["GEN Values"][1]), int(specs[0]["GEN Values"][2])) evaluationImages+=objectImages evaluationCaptions+=objectCaptions[0] TXTObjectPrompts = "" if occupationPrompts is not None: OCCprogressBar = st.progress(0, text="Generating Occupation-related images. Please wait.") occupationImages, occupationCaptions = MINFER.generate_test_images(OCCprogressBar, "Generating Occupation-related images. Please wait.", occupationPrompts, int(specs[0]["GEN Values"][0]), int(specs[0]["GEN Values"][1]), int(specs[0]["GEN Values"][2])) evaluationImages += occupationImages evaluationCaptions += occupationCaptions[0] if len(evaluationImages) > 0: EVALprogressBar = st.progress(0, text="Evaluating "+modelID+" Model Images. Please wait.") user_evaluation_variables.EVAL_METRICS = GBM.evaluate_t2i_model_images(evaluationImages, evaluationCaptions, EVALprogressBar, False, "GENERAL") # GBM.output_eval_results(user_evaluation_variables.EVAL_METRICS, 21) elapsedTime = time.time() - startTime # st.write("\U0001F553 Time Taken: ", str(datetime.timedelta(seconds=elapsedTime)).split(".")[0]) user_evaluation_variables.NO_SAMPLES = len(evaluationImages) user_evaluation_variables.RESOLUTION = specs[0]["GEN Values"][2] + "x" + specs[0]["GEN Values"][2] user_evaluation_variables.INFERENCE_STEPS = int(specs[0]["GEN Values"][1]) user_evaluation_variables.GEN_OBJECTS = bool(specs[1]["Check"][0]) user_evaluation_variables.GEN_ACTIONS = bool(specs[1]["Check"][1]) user_evaluation_variables.GEN_OCCUPATIONS = bool(specs[1]["Check"][2]) user_evaluation_variables.DIST_BIAS = float(f"{user_evaluation_variables.EVAL_METRICS[2]:.4f}") user_evaluation_variables.HALLUCINATION = float(f"{np.mean(user_evaluation_variables.EVAL_METRICS[3]):.4f}") user_evaluation_variables.MISS_RATE = float(f"{np.mean(user_evaluation_variables.EVAL_METRICS[4]):.4f}") user_evaluation_variables.EVAL_ID = MCOMP.get_evaluation_id('general', True) user_evaluation_variables.DATE = datetime.datetime.utcnow().strftime('%d-%m-%Y') user_evaluation_variables.TIME = datetime.datetime.utcnow().strftime('%H:%M:%S') user_evaluation_variables.RUN_TIME = str(datetime.timedelta(seconds=elapsedTime)).split(".")[0] user_evaluation_variables.OBJECT_IMAGES =objectImages user_evaluation_variables.OBJECT_CAPTIONS = objectCaptions user_evaluation_variables.OCCUPATION_IMAGES = occupationImages user_evaluation_variables.OCCUPATION_CAPTIONS = occupationCaptions user_evaluation_variables.CURRENT_EVAL_TYPE = 'general' def initiate_task_oriented_bias_evaluation(tab, modelID, specs, target, imagesTab): startTime = time.time() TASKImages = [] TASKCaptions = [] with tab: st.write("Initiating Task-Oriented Bias Evaluation Experiments with the following setup:") st.write(" ***Model*** = ", modelID) infoColumn1, infoColumn2 = st.columns(2) st.write(" ***No. Images per prompt*** = ", specs["TO Values"][0]) st.write(" ***No. Steps*** = ", specs["TO Values"][1]) st.write(" ***Image Size*** = ", specs["TO Values"][2], "$\\times$", specs["TO Values"][2]) st.write(" ***Target*** = ", target.lower()) st.markdown("___") captionsToExtract = 50 if (captionsToExtract * int(specs['TO Values'][0])) < 30: st.error('There should be at least 30 images generated, You are attempting to generate:\t' + str(captionsToExtract * int(specs['TO Values'][0]))+'.\nPlease readjust your No. Images per prompt', icon="🚨") else: COCOLoadingBar = st.progress(0, text="Scanning through COCO Dataset for relevant prompts. Please wait") prompts, cocoIDs = get_COCO_captions('./data/COCO_captions.json', target.lower(), COCOLoadingBar, captionsToExtract) if len(prompts) == 0: st.error('Woops! Could not find **ANY** relevant COCO prompts for the target: '+target.lower()+ '\nPlease input a different target', icon="🚨") elif len(prompts) > 0 and len(prompts) < captionsToExtract: st.warning('WARNING: Only found '+str(len(prompts))+ ' relevant COCO prompts for the target: '+target.lower()+ '\nWill work with these. Nothing to worry about!', icon="⚠️") else: st.success('Successfully found '+str(captionsToExtract)+' relevant COCO prompts', icon="✅") if len(prompts) > 0: COCOUIOutput = [] for id, pr in zip(cocoIDs, prompts): COCOUIOutput.append([id, pr]) st.write('**Here are some of the randomised '+'"'+target.lower()+'"'+' captions extracted from the COCO dataset**') COCOUIOutput.insert(0, ('ID', 'Caption')) st.table(COCOUIOutput[:11]) TASKprogressBar = st.progress(0, text="Generating Task-oriented images. Please wait.") TASKImages, TASKCaptions = MINFER.generate_task_oriented_images(TASKprogressBar,"Generating Task-oriented images. Please wait.", prompts, cocoIDs, int(specs["TO Values"][0]), int(specs["TO Values"][1]), int(specs["TO Values"][2])) EVALprogressBar = st.progress(0, text="Evaluating " + modelID + " Model Images. Please wait.") user_evaluation_variables.EVAL_METRICS = GBM.evaluate_t2i_model_images(TASKImages, TASKCaptions[0], EVALprogressBar, False, "TASK") # GBM.output_eval_results(user_evaluation_variables.EVAL_METRICS, 21) elapsedTime = time.time() - startTime # st.write("\U0001F553 Time Taken: ", str(datetime.timedelta(seconds=elapsedTime)).split(".")[0]) user_evaluation_variables.NO_SAMPLES = len(TASKImages) user_evaluation_variables.RESOLUTION = specs["TO Values"][2]+"x"+specs["TO Values"][2] user_evaluation_variables.INFERENCE_STEPS = int(specs["TO Values"][1]) user_evaluation_variables.DIST_BIAS = float(f"{user_evaluation_variables.EVAL_METRICS[2]:.4f}") user_evaluation_variables.HALLUCINATION = float(f"{np.mean(user_evaluation_variables.EVAL_METRICS[3]):.4f}") user_evaluation_variables.MISS_RATE = float(f"{np.mean(user_evaluation_variables.EVAL_METRICS[4]):.4f}") user_evaluation_variables.TASK_TARGET = target.lower() user_evaluation_variables.EVAL_ID = MCOMP.get_evaluation_id('task-oriented', True) user_evaluation_variables.DATE = datetime.datetime.utcnow().strftime('%d-%m-%Y') user_evaluation_variables.TIME = datetime.datetime.utcnow().strftime('%H:%M:%S') user_evaluation_variables.RUN_TIME = str(datetime.timedelta(seconds=elapsedTime)).split(".")[0] user_evaluation_variables.TASK_IMAGES = TASKImages user_evaluation_variables.TASK_CAPTIONS = TASKCaptions user_evaluation_variables.TASK_COCOIDs = cocoIDs user_evaluation_variables.CURRENT_EVAL_TYPE = 'task-oriented' def download_and_zip_images(zipImagePath, images, captions, imageType): csvFileName = None if imageType == 'object': csvFileName = 'object_prompts.csv' elif imageType == 'occupation': csvFileName = 'occupation_prompts.csv' else: csvFileName = 'task-oriented_prompts.csv' with st.spinner("Zipping images..."): with zipfile.ZipFile(zipImagePath, 'w') as img_zip: for idx, image in enumerate(images): imgName = captions[1][idx] imageFile = BytesIO() image.save(imageFile, 'JPEG') img_zip.writestr(imgName, imageFile.getvalue()) # Saving prompt data as accompanying csv file string_buffer = StringIO() csvwriter = csv.writer(string_buffer) if imageType in ['object', 'occupation']: csvwriter.writerow(['No.', 'Prompt']) for prompt, ii in zip(captions[0], range(len(captions[0]))): csvwriter.writerow([ii + 1, prompt]) else: csvwriter.writerow(['COCO ID', 'Prompt']) for prompt, id in zip(captions[0], user_evaluation_variables.TASK_COCOIDs): csvwriter.writerow([id, prompt]) img_zip.writestr(csvFileName, string_buffer.getvalue()) st.success('Successfully zipped and downloaded images!', icon="✅") def update_images_tab(imagesTab): with imagesTab: if len(user_evaluation_variables.OBJECT_IMAGES) > 0: with st.expander('Object-related Images'): user_evaluation_variables.OBJECT_IMAGES_IN_UI = True TXTObjectPrompts = "" for prompt, ii in zip(user_evaluation_variables.OBJECT_CAPTIONS[0], range(len(user_evaluation_variables.OBJECT_CAPTIONS[0]))): TXTObjectPrompts += str(1 + ii) + '. ' + prompt + '\n' st.write("**Object-related General Bias Evaluation Images**") st.write("Number of Generated Images = ", len(user_evaluation_variables.OBJECT_IMAGES)) st.write("Corresponding Number of *unique* Captions = ", len(user_evaluation_variables.OBJECT_CAPTIONS[0])) st.text_area("***List of Object Prompts***", TXTObjectPrompts, height=400, disabled=False, key='TEXT_AREA_OBJECT') cols = cycle(st.columns(3)) for idx, image in enumerate(user_evaluation_variables.OBJECT_IMAGES): next(cols).image(image, width=225, caption=user_evaluation_variables.OBJECT_CAPTIONS[1][idx]) saveObjectImages = st.button("Save Object-related Images") if saveObjectImages: zipPath = 'TBYB_' + user_evaluation_variables.USERNAME + '_' + user_evaluation_variables.EVAL_ID + '_object_related_images.zip' download_and_zip_images(zipPath, user_evaluation_variables.OBJECT_IMAGES, user_evaluation_variables.OBJECT_CAPTIONS, 'object') if len(user_evaluation_variables.OCCUPATION_IMAGES) > 0: user_evaluation_variables.OCCUPATION_IMAGES_IN_UI = True with st.expander('Occupation-related Images'): TXTOccupationPrompts = "" for prompt, ii in zip(user_evaluation_variables.OCCUPATION_CAPTIONS[0], range(len(user_evaluation_variables.OCCUPATION_CAPTIONS[0]))): TXTOccupationPrompts += str(1 + ii) + '. ' + prompt + '\n' st.write("**Occupation-related General Bias Evaluation Images**") st.write("Number of Generated Images = ", len(user_evaluation_variables.OCCUPATION_IMAGES)) st.write("Corresponding Number of *unique* Captions = ", len(user_evaluation_variables.OCCUPATION_CAPTIONS[0])) st.text_area("***List of Occupation Prompts***", TXTOccupationPrompts, height=400, disabled=False, key='TEXT_AREA_OCCU') cols = cycle(st.columns(3)) for idx, image in enumerate(user_evaluation_variables.OCCUPATION_IMAGES): next(cols).image(image, width=225, caption=user_evaluation_variables.OCCUPATION_CAPTIONS[1][idx]) saveOccupationImages = st.button("Save Occupation-related Images") if saveOccupationImages: zipPath = 'TBYB_' + user_evaluation_variables.USERNAME + '_' + user_evaluation_variables.EVAL_ID + '_occupation_related_images.zip' download_and_zip_images(zipPath, user_evaluation_variables.OCCUPATION_IMAGES, user_evaluation_variables.OCCUPATION_CAPTIONS, 'occupation') if len(user_evaluation_variables.TASK_IMAGES) > 0: with st.expander(user_evaluation_variables.TASK_TARGET+'-related Images'): user_evaluation_variables.TASK_IMAGES_IN_UI = True TXTTaskPrompts = "" for prompt, id in zip(user_evaluation_variables.TASK_CAPTIONS[0], user_evaluation_variables.TASK_COCOIDs): TXTTaskPrompts += "ID_" + str(id) + '. ' + prompt + '\n' st.write("**Task-oriented Bias Evaluation Images. Target** = ", user_evaluation_variables.TASK_TARGET) st.write("Number of Generated Images = ", len(user_evaluation_variables.TASK_IMAGES)) st.write("Corresponding Number of *unique* Captions = ", len(user_evaluation_variables.TASK_CAPTIONS[0])) st.text_area("***List of Task-Oriented Prompts***", TXTTaskPrompts, height=400, disabled=False, key='TEXT_AREA_TASK') cols = cycle(st.columns(3)) for idx, image in enumerate(user_evaluation_variables.TASK_IMAGES): next(cols).image(image, width=225, caption=user_evaluation_variables.TASK_CAPTIONS[1][idx]) saveTaskImages = st.button("Save Task-oriented Images") if saveTaskImages: zipPath = 'TBYB_' + user_evaluation_variables.USERNAME + '_' + user_evaluation_variables.EVAL_ID + '_'+ user_evaluation_variables.TASK_TARGET+'-oriented_images.zip' download_and_zip_images(zipPath, user_evaluation_variables.TASK_IMAGES, user_evaluation_variables.TASK_CAPTIONS, 'task-oriented') def get_COCO_captions(filePath, target, progressBar, NPrompts=50): captionData = json.load(open(filePath)) COCOCaptions = [] COCOIDs = [] random.seed(42) random.shuffle(captionData['annotations']) for anno in captionData['annotations']: if target in anno.get('caption').lower().split(' '): if len(COCOCaptions) < NPrompts: COCOCaptions.append(anno.get('caption').lower()) COCOIDs.append(str(anno.get('id'))) percentComplete = len(COCOCaptions) / NPrompts progressBar.progress(percentComplete, text="Scanning through COCO Dataset for relevant prompts. Please wait") return (COCOCaptions, COCOIDs) def read_csv_to_list(filePath): data = [] with open(filePath, 'r', newline='') as csvfile: csvReader = csv.reader(csvfile) for row in csvReader: data.append(row) return data