import streamlit as st import pandas as pd import numpy as np @st.cache def prompt_to_csv(df): # IMPORTANT: Cache the conversion to prevent computation on every rerun df_download = df df_download['Filename']='p'+df_download['ID'].astype('str')+'_1.png' df_download = df[['Prompt','Filename']].drop_duplicates(subset='Filename') return df_download.to_csv().encode('utf-8') # Setup ## Load prompt directory prompt_dir = pd.read_csv('Data/Prompt_dir_221216.csv') #second version of prompt_dir st.session_state['prompt_dir'] = prompt_dir ## Create lists of prompts for manual and automated assessments st.session_state['automated_tasks'] = ['Multiple object types', 'Single object','Negation'] automated_prompts = prompt_dir.loc[ (prompt_dir['Auto_assessment']==True)& (prompt_dir['Task']).isin(st.session_state['automated_tasks'])].ID.tolist() manual_prompts = prompt_dir.ID.tolist() # Page st.title('Generative Image Benchmark') st.write('This is an evaluation platform to assess the performance of image generation algorithms developed by Intel Labs. This is the alpha version of the platform.') st.subheader('User guide') st.write('To assess a generative image algorithm, download a set of prompts using the prompt downloader below. Generate one image per prompt and use the file names provided to name your images. Upload these generated images in the data upload section below. The pages for manual assessment and automated assessment allow you to systematically assess the generated images. The results will be presented and ready for download on the assessment summary page.') st.sidebar.image('Graphics/IL_Logo.png') # Add prompt downloading functions prompt_download_dict = {} ## Count how many prompts are in database to allow for max value in selection prompt_task_count = prompt_dir.Task.value_counts(sort=False) prompt_task_count = prompt_task_count.drop(index='Single object') prompt_task_select = prompt_task_count.copy() ## Hide downloader in box with st.expander("Prompt downloader"): st.write('Select the number of prompts you want to download for each task category. The set of prompts will automatically also include all single objects appearing in the selected prompts.') # Create numerical selector for every task in prompt directory for i_task in prompt_task_select.index: prompt_task_select[i_task] = st.number_input( i_task, value = prompt_task_count[i_task], max_value=prompt_task_count[i_task], min_value=0, step = 1) # Create df with selected number of prompts per task for i_task in prompt_task_select.index: temp_df = prompt_dir.loc[prompt_dir['Task']==i_task][0:prompt_task_select[i_task]] if len(temp_df)>0: prompt_download_dict[i_task]=temp_df # Concat all tasks to dataframe prompt_download = pd.concat(prompt_download_dict.values()) # Add relevant single object prompts single_object_ids = prompt_download.Linked_prompts.str.split(',').explode().unique().astype('int') prompt_download = pd.concat([ prompt_download, prompt_dir.loc[prompt_dir['ID'].isin(single_object_ids)] ]) # Add download button for prompts st.download_button( label="Download prompts", data=prompt_to_csv(prompt_download), file_name='prompt_list.csv', mime='text/csv', ) # Generate empty dataset for results, if it does not exist yet try: num_uploaded_images = st.session_state['eval_df'].shape[0] except KeyError: st.session_state['eval_df'] = pd.DataFrame( columns=['File_name','Prompt_no','automated_eval','manual_eval','manual_eval_completed','manual_eval_task_score']) st.session_state['uploaded_img'] = [] # Create dic for automated asssssment if it does not excist yet try: test_dict = st.session_state['results_dict'] except KeyError: st.session_state['results_dict'] = {} # Data upload setup st.subheader('Data upload') #uploaded_files = st.file_uploader('Upload generated images', accept_multiple_files=True) with st.form("my-form", clear_on_submit=True): uploaded_files = st.file_uploader('Select images for upload', accept_multiple_files=True) man_assessment_share = st.selectbox( 'Select share of uploaded images to be used for manual assessment.', ('100%', '50%')) submitted = st.form_submit_button("Add images") st.session_state['uploaded_img'] = st.session_state['uploaded_img']+uploaded_files # Add new uploaded images to session state ## Try to append it to pre-existing list, else create new list in session state ## Always reset uploaded files to empty list after they have been added to state if len(uploaded_files) != 0: try: # Extract prompts of uploaded files file_names = [x.name for x in uploaded_files] files_prompts = [x.split('_')[0][1:] for x in file_names] # Create manual evaluation df df_dict = {'File_name':file_names, 'Prompt_no':files_prompts} eval_df = pd.DataFrame(df_dict) eval_df['automated_eval'] = eval_df['Prompt_no'].astype('int').isin(automated_prompts) eval_df['manual_eval'] = eval_df['Prompt_no'].astype('int').isin(manual_prompts) eval_df['manual_eval_completed'] = False eval_df['manual_eval_task_score'] = np.nan # Exclude given percentage of uploaded images from manual assessment; with random selection if man_assessment_share == '50%': reassign_number = int(len(eval_df)/2) manual_eval_reassign = eval_df['manual_eval'] random_image_indices = np.random.choice(len(manual_eval_reassign),reassign_number, replace=False) manual_eval_reassign.iloc[random_image_indices]=False eval_df['manual_eval'] = manual_eval_reassign # Join new uploaded df with existing df joint_eval_df = pd.concat([st.session_state['eval_df'], eval_df], ignore_index=True) # Add task name to eval_df Prompt_no_task_dict = dict(zip(prompt_dir.ID.astype('str').to_list(),prompt_dir.Task.to_list())) joint_eval_df['Task'] = joint_eval_df.Prompt_no.map(Prompt_no_task_dict) # Save eval_df to session state st.session_state['eval_df'] = joint_eval_df except KeyError: st.session_state['uploaded_img'] = uploaded_files eval_df = st.session_state['eval_df'] if eval_df.shape[0]!=0: # Print current state of uploaded data st.write("{0} images uploaded. Reload the page to reset the image upload.".format(str(eval_df.shape[0]))) st.write("- Available for manual assessment: ", str(sum(eval_df.manual_eval))) manual_eval_available = sum(eval_df.manual_eval) st.write("- Available for automated assessment: ", str(sum(eval_df.automated_eval))) else: st.write("Upload files to start the assessment.") #st.write(eval_df) #st.write(prompt_dir) #st.session_state['eval_df']