Spaces:
Running
Running
| # type: ignore -- ignores linting import issues when using multiple virtual environments | |
| import streamlit.components.v1 as components | |
| import streamlit as st | |
| import pandas as pd | |
| import logging | |
| from deeploy import Client | |
| from constants import ( | |
| relationship_dict, | |
| occupation_dict, | |
| education_dict, | |
| type_of_work_dict, | |
| countries_dict, | |
| marital_status_dict, | |
| ) | |
| # reset Plotly theme after streamlit import | |
| import plotly.io as pio | |
| pio.templates.default = "plotly" | |
| logging.basicConfig(level=logging.INFO) | |
| st.set_page_config(layout="wide") | |
| st.title("Credit-Scoring Model Explainability") | |
| st.markdown( | |
| """ | |
| <style> | |
| section[data-testid="stSidebar"] { | |
| width: 300px !important; # Set the width to your desired value | |
| } | |
| </style> | |
| """, | |
| unsafe_allow_html=True, | |
| ) | |
| def get_model_url(): | |
| model_url = st.text_area( | |
| "Model URL (without the /explain endpoint, default is the demo deployment)", | |
| "https://api.app.deeploy.ml/workspaces/708b5808-27af-461a-8ee5-80add68384c7/deployments/dc8c359d-5f61-4107-8b0f-de97ec120289/", | |
| height=125, | |
| ) | |
| elems = model_url.split("/") | |
| try: | |
| workspace_id = elems[4] | |
| deployment_id = elems[6] | |
| except IndexError: | |
| workspace_id = "" | |
| deployment_id = "" | |
| return model_url, workspace_id, deployment_id | |
| def ChangeButtonColour(widget_label, font_color, background_color="transparent"): | |
| # func to change button colors | |
| htmlstr = f""" | |
| <script> | |
| var elements = window.parent.document.querySelectorAll('button'); | |
| for (var i = 0; i < elements.length; ++i) {{ | |
| if (elements[i].innerText == '{widget_label}') {{ | |
| elements[i].style.color ='{font_color}'; | |
| elements[i].style.background = '{background_color}' | |
| }} | |
| }} | |
| </script> | |
| """ | |
| components.html(f"{htmlstr}", height=0, width=0) | |
| with st.sidebar: | |
| st.image("deeploy_logo_wide.png", width=250) | |
| # Ask for model URL and token | |
| host = st.text_input("Host (Changing is optional)", "app.deeploy.ml") | |
| model_url, workspace_id, deployment_id = get_model_url() | |
| deployment_token = st.text_input("Deeploy Model Token", "my-secret-token") | |
| if deployment_token == "my-secret-token": | |
| st.warning("Please enter Deeploy API token.") | |
| # Split model URL into workspace and deployment ID | |
| # st.write("Values below are for debug only:") | |
| # st.write("Workspace ID: ", workspace_id) | |
| # st.write("Deployment ID: ", deployment_id) | |
| client_options = { | |
| "host": host, | |
| "deployment_token": deployment_token, | |
| "workspace_id": workspace_id, | |
| } | |
| client = Client(**client_options) | |
| if "expander_toggle" not in st.session_state: | |
| st.session_state.expander_toggle = True | |
| def submit_and_clear(): | |
| try: | |
| # Call the explain endpoint as it also includes the prediction | |
| client.evaluate( | |
| deployment_id, request_log_id, prediction_log_id, st.session_state.evaluation_input | |
| ) | |
| st.toast(":green[Feedback submitted successfully.]", icon="✅") | |
| st.session_state.eval_selected = False | |
| st.session_state.predict_button_clicked = False | |
| st.session_state.eval_selected = False | |
| show_expander() | |
| except Exception as e: | |
| logging.error(e) | |
| st.error( | |
| "Failed to submit feedback." | |
| + "Check whether you are using the right model URL and token for evaluations. " | |
| + "Contact Deeploy if the problem persists." | |
| ) | |
| st.toast(f"Failed to submit feedback: {e}") | |
| def hide_expander(): | |
| st.session_state.expander_toggle = False | |
| def show_expander(): | |
| st.session_state.expander_toggle = True | |
| # Attributes | |
| st.subheader("Loan Application") | |
| with st.expander("Application form", expanded=st.session_state.expander_toggle): | |
| # Split view in 2 columns | |
| col1, col2 = st.columns(2) | |
| with col1: | |
| # Create input fields for attributes from constant dicts | |
| age = st.number_input("Age", min_value=0, max_value=100, value=30) | |
| marital_status = st.selectbox("Marital Status", marital_status_dict.keys()) | |
| marital_status_id = marital_status_dict[marital_status] | |
| native_country = st.selectbox( | |
| "Native Country", countries_dict.keys(), index=len(countries_dict) - 1 | |
| ) | |
| native_country_id = countries_dict[native_country] | |
| relationship = st.selectbox("Family situation", relationship_dict.keys()) | |
| relationship_id = relationship_dict[relationship] | |
| occupation = st.selectbox("Occupation", occupation_dict.keys(), index=1) | |
| occupation_id = occupation_dict[occupation] | |
| with col2: | |
| education = st.selectbox("Highest education level", education_dict.keys(), index=4) | |
| education_id = education_dict[education] | |
| type_of_work = st.selectbox("Type of work", type_of_work_dict.keys()) | |
| type_of_work_id = type_of_work_dict[type_of_work] | |
| hours_per_week = st.number_input( | |
| "Working hours per week", min_value=0, max_value=100, value=40 | |
| ) | |
| capital_gain = st.number_input( | |
| "Yearly income [€]", min_value=0, max_value=1000000, value=70000 | |
| ) | |
| capital_loss = st.number_input( | |
| "Yearly expenditures [€]", min_value=0, max_value=1000000, value=60000 | |
| ) | |
| data_df = pd.DataFrame( | |
| [ | |
| [ | |
| age, | |
| type_of_work, | |
| education, | |
| marital_status, | |
| occupation, | |
| relationship, | |
| capital_gain, | |
| capital_loss, | |
| hours_per_week, | |
| native_country, | |
| ] | |
| ], | |
| columns=[ | |
| "Age", | |
| "Type of work", | |
| "Highest education level", | |
| "Marital Status", | |
| "Occupation", | |
| "Family situation", | |
| "Yearly Income [€]", | |
| "Yearly expenditures [€]", | |
| "Working hours per week", | |
| "Native Country", | |
| ], | |
| ) | |
| data_df_t = data_df.T | |
| request_body = { | |
| "instances": [ | |
| [ | |
| age, | |
| type_of_work_id, | |
| education_id, | |
| marital_status_id, | |
| occupation_id, | |
| relationship_id, | |
| capital_gain, | |
| capital_loss, | |
| hours_per_week, | |
| native_country_id, | |
| ] | |
| ] | |
| } | |
| if "predict_button_clicked" not in st.session_state: | |
| st.session_state.predict_button_clicked = False | |
| if deployment_token != "my-secret-token": | |
| predict_button = st.button( | |
| "Predict", key="predict_button", help="Click to get the AI prediction.", on_click=hide_expander | |
| ) | |
| if predict_button: | |
| st.session_state.predict_button_clicked = True | |
| # st.session_state.selected = "Loan Decision" | |
| with st.spinner("Loading prediction and explanation..."): | |
| # Call the explain endpoint as it also includes the prediction | |
| exp = client.explain( | |
| request_body=request_body, deployment_id=deployment_id | |
| ) | |
| st.session_state.exp = exp | |
| if st.session_state.predict_button_clicked: | |
| try: | |
| exp = st.session_state.exp | |
| # Read explanation to dataframe from json | |
| predictions = exp["predictions"] | |
| request_log_id = exp["requestLogId"] | |
| prediction_log_id = exp["predictionLogIds"][0] | |
| exp_df = pd.DataFrame( | |
| [exp["explanations"][0]["shap_values"]], columns=exp["featureLabels"] | |
| ) | |
| exp_df.columns = data_df.columns | |
| exp_df_t = exp_df.T | |
| # Merge data and explanation | |
| exp_df_t = data_df_t.merge(exp_df_t, left_index=True, right_index=True) | |
| weight_feat = "Weight" | |
| exp_df_t.columns = ["Feature value", weight_feat] | |
| exp_df_t["Feature"] = exp_df_t.index | |
| exp_df_t = exp_df_t[["Feature", "Feature value", weight_feat]] | |
| exp_df_t["Feature value"] = exp_df_t["Feature value"].astype(str) | |
| # Filter values below 0.01 | |
| exp_df_t = exp_df_t[ | |
| (exp_df_t[weight_feat] > 0.01) | (exp_df_t[weight_feat] < -0.01) | |
| ] | |
| exp_df_t[weight_feat] = exp_df_t[weight_feat].astype(float).round(2) | |
| pos_exp_df_t = exp_df_t[exp_df_t[weight_feat] > 0] | |
| pos_exp_df_t = pos_exp_df_t.sort_values(by=weight_feat, ascending=False) | |
| neg_exp_df_t = exp_df_t[exp_df_t[weight_feat] < 0] | |
| neg_exp_df_t = neg_exp_df_t.sort_values(by=weight_feat, ascending=True) | |
| neg_exp_df_t[weight_feat] = neg_exp_df_t[weight_feat].abs() | |
| # Get 3 features with highest positive relevance score | |
| pos_feats = pos_exp_df_t[weight_feat].nlargest(3).index.tolist() | |
| # For feature, get feature value and concatenate into a single string | |
| pos_feats = [ | |
| f"{feat}: {pos_exp_df_t.loc[feat, 'Feature value']}" | |
| for feat in pos_feats | |
| ] | |
| # Get 3 features with highest negative relevance score | |
| neg_feats = neg_exp_df_t[weight_feat].nlargest(3).index.tolist() | |
| # For feature, get feature value and concatenate into a single string | |
| neg_feats = [ | |
| f"{feat}: {neg_exp_df_t.loc[feat, 'Feature value']}" | |
| for feat in neg_feats | |
| ] | |
| if predictions[0]: | |
| # Show prediction | |
| st.subheader("Loan Decision: :green[POSITIVE]", divider="green") | |
| # Format subheader to green | |
| st.markdown( | |
| "<style>.css-1v3fvcr{color: green;}</style>", unsafe_allow_html=True | |
| ) | |
| # If prediction is positive, first show positive features, then negative features | |
| st.success( | |
| "**Positive creditworthiness**. This is primarily attributed to: \n - " | |
| + " \n- ".join(pos_feats) | |
| ) | |
| st.warning( | |
| "However, the following features weight ***against*** the loan application: \n - " | |
| + " \n- ".join(neg_feats) | |
| + " \n For more details, see full explanation of the credit assessment below.", | |
| ) | |
| else: | |
| st.subheader("Loan Decision: :red[NEGATIVE]", divider="red") | |
| # If prediction is negative, first show negative features, then positive features | |
| st.error( | |
| "**Negative creditworthiness**. This is primarily attributed to: \n - " | |
| + " \n - ".join(neg_feats) | |
| ) | |
| st.warning( | |
| "However, the following factors weigh ***in favor*** of the loan applicant: \n - " | |
| + " \n - ".join(pos_feats) | |
| + " \n For more details, see full explanation of the credit assessment below.", | |
| ) | |
| explanation_expander = st.expander("Show explanation") | |
| with explanation_expander: | |
| # Show explanation | |
| col_pos, col_neg = st.columns(2) | |
| with col_pos: | |
| st.subheader("Factors :green[in favor] of loan approval") | |
| # st.success("**Factors in favor of loan approval**") | |
| st.dataframe( | |
| pos_exp_df_t, | |
| hide_index=True, | |
| width=600, | |
| column_config={ | |
| "Weight": st.column_config.ProgressColumn( | |
| "Weight", | |
| width="small", | |
| format=" ", | |
| min_value=0, | |
| max_value=1, | |
| ) | |
| }, | |
| ) | |
| with col_neg: | |
| st.subheader("Factors :red[against] loan approval") | |
| # st.error("**Factors against loan approval**") | |
| st.dataframe( | |
| neg_exp_df_t, | |
| hide_index=True, | |
| width=600, | |
| column_config={ | |
| "Weight": st.column_config.ProgressColumn( | |
| "Weight", | |
| width="small", | |
| format=" ", | |
| min_value=0, | |
| max_value=1, | |
| ) | |
| }, | |
| ) | |
| st.divider() | |
| # Add prediction evaluation | |
| st.subheader("Evaluation: Do you agree with the predicted creditworthiness?") | |
| st.info( | |
| "AI model predictions always come with a certain level of uncertainty. \nEvaluate the correctness of the prediction based on your expertise and experience." | |
| ) | |
| cols = st.columns(4) | |
| col_yes, col_no = cols[:2] | |
| with col_yes: | |
| yes_button = st.button( | |
| "Yes, I agree", | |
| key="yes_button", | |
| use_container_width=True, | |
| help="Click if you agree with the prediction", | |
| ) | |
| ChangeButtonColour("Yes, I agree", "white", "green") | |
| with col_no: | |
| no_button = st.button( | |
| "No, I disagree", | |
| key="no_button", | |
| use_container_width=True, | |
| help="Click if you disagree with the prediction", | |
| type="primary", | |
| ) | |
| ChangeButtonColour("No, I disagree", "white", "#DD360C") | |
| # ChangeButtonColour("No, I disagree", "#DD360C", "#F0F0F0") | |
| if "eval_selected" not in st.session_state: | |
| st.session_state["eval_selected"] = False | |
| if yes_button: | |
| st.session_state.eval_selected = True | |
| st.session_state.evaluation_input = { | |
| "result": 0 # Agree with the prediction | |
| } | |
| st.session_state.placeholder = "Income is sufficient, given applicant's background" | |
| if no_button: | |
| st.session_state.eval_selected = True | |
| desired_output = not predictions[0] | |
| st.session_state.evaluation_input = { | |
| "result": 1, # Disagree with the prediction | |
| "value": {"predictions": [desired_output]}, | |
| } | |
| st.session_state.placeholder = "Income is too low, given applicant's background" | |
| if st.session_state.eval_selected: | |
| comment = st.text_input("Would you like to add a comment?", placeholder=st.session_state.placeholder) | |
| if comment: | |
| st.session_state.evaluation_input["explanation"] = comment | |
| st.button("Submit", key="submit_button", on_click=submit_and_clear) | |
| except Exception as e: | |
| logging.error(e) | |
| st.error( | |
| "Failed to retrieve the prediction or explanation." | |
| + "Check whether you are using the right model URL and Token. " | |
| + "Contact Deeploy if the problem persists." | |
| ) | |