Added all app files for gradio demo for Binary classification on structured data using GRN-VSN model

app.py

@@ -0,0 +1,59 @@
+import gradio as gr
+from .constants import CSV_HEADER, NUMERIC_FEATURE_NAMES, CATEGORICAL_FEATURES_WITH_VOCABULARY, NUMBER_INPUT_COLS
+from .preprocess import create_max_values_map, create_dropdown_default_values_map, create_sample_test_data
+from .predict import batch_predict, user_input_predict
+
+inputs_list = []
+max_values_map = create_max_values_map()
+dropdown_default_values_map = create_dropdown_default_values_map()
+sample_input_df_val = create_sample_test_data()
+
+demo = gr.Blocks()
+
+with demo:
+    
+    gr.Markdown("# **Binary Classification using Gated Residual and Variable Selection Networks** \n")
+    gr.Markdown("This demo demonstrates the use of Gated Residual Networks (GRN) and Variable Selection Networks (VSN), proposed by Bryan Lim et al. in <a href=\"https://arxiv.org/abs/1912.09363/\">Temporal Fusion Transformers (TFT) for Interpretable Multi-horizon Time Series Forecasting</a> for structured data classification")
+    gr.Markdown("Play around and see yourself 🤗 ")
+    
+    with gr.Tabs():
+        
+        with gr.TabItem("Predict using batch of inputs"):
+            gr.Markdown("**Input DataFrame** \n")
+            input_df = gr.Dataframe(headers=CSV_HEADER,value=samp,)
+            gr.Markdown("**Output DataFrame** \n")
+            output_df = gr.Dataframe()
+            gr.Markdown("**Make Predictions**")
+            with gr.Row():
+                compute_button = gr.Button("Predict")
+        
+        with gr.TabItem("Tweak inputs Yourself"):
+            with gr.Tabs():
+                
+                with gr.TabItem("Numerical Inputs"):
+                    gr.Markdown("Set values for numerical inputs here.")
+                    for num_variable in NUMERIC_FEATURE_NAMES:
+                        with gr.Column():
+                            if num_variable in NUMBER_INPUT_COLS:
+                                numeric_input = gr.Number(label=num_variable)
+                            else:
+                                curr_max_val = max_values_map["max_"+num_variable]
+                                numeric_input = gr.Slider(0,curr_max_val, label=num_variable,step=1)
+                                inputs_list.append(numeric_input)
+                
+                with gr.TabItem("Categorical Inputs"):
+                    gr.Markdown("Choose values for categorical inputs here.")
+                    for cat_variable in CATEGORICAL_FEATURES_WITH_VOCABULARY.keys():
+                        with gr.Column():
+                            categorical_input = gr.Dropdown(CATEGORICAL_FEATURES_WITH_VOCABULARY[cat_variable], label=cat_variable, value=str(dropdown_default_values_map["max_"+cat_variable]))
+                            inputs_list.append(categorical_input)
+                            
+                    predict_button = gr.Button("Predict")
+                    final_output = gr.Label()
+    
+    predict_button.click(user_input_predict, inputs=inputs_list, outputs=final_output)
+    compute_button.click(batch_predict, inputs=input_df, outputs=output_df)
+    gr.Markdown('\n Author: <a href=\"https://huggingface.co/shivi\">Shivalika Singh</a> <br> Based on this <a href=\"https://keras.io/examples/structured_data/classification_with_grn_and_vsn/\">Keras example</a> by <a href=\"https://www.linkedin.com/in/khalid-salama-24403144/\">Khalid Salama</a> <br> Demo Powered by this <a href=\"https://huggingface.co/shivi/classification-grn-vsn\">GRN-VSN model</a>')
+    
+    
+demo.launch()

constants.py

@@ -0,0 +1,88 @@
+import pandas as pd
+from .preprocess import load_test_data
+
+# Column names.
+CSV_HEADER = [
+    "age",
+    "class_of_worker",
+    "detailed_industry_recode",
+    "detailed_occupation_recode",
+    "education",
+    "wage_per_hour",
+    "enroll_in_edu_inst_last_wk",
+    "marital_stat",
+    "major_industry_code",
+    "major_occupation_code",
+    "race",
+    "hispanic_origin",
+    "sex",
+    "member_of_a_labor_union",
+    "reason_for_unemployment",
+    "full_or_part_time_employment_stat",
+    "capital_gains",
+    "capital_losses",
+    "dividends_from_stocks",
+    "tax_filer_stat",
+    "region_of_previous_residence",
+    "state_of_previous_residence",
+    "detailed_household_and_family_stat",
+    "detailed_household_summary_in_household",
+    "instance_weight",
+    "migration_code-change_in_msa",
+    "migration_code-change_in_reg",
+    "migration_code-move_within_reg",
+    "live_in_this_house_1_year_ago",
+    "migration_prev_res_in_sunbelt",
+    "num_persons_worked_for_employer",
+    "family_members_under_18",
+    "country_of_birth_father",
+    "country_of_birth_mother",
+    "country_of_birth_self",
+    "citizenship",
+    "own_business_or_self_employed",
+    "fill_inc_questionnaire_for_veterans_admin",
+    "veterans_benefits",
+    "weeks_worked_in_year",
+    "year",
+    "income_level",
+]
+
+# Target feature name.
+TARGET_FEATURE_NAME = "income_level"
+
+# Weight column name.
+WEIGHT_COLUMN_NAME = "instance_weight"
+
+# Numeric feature names.
+NUMERIC_FEATURE_NAMES = [
+    "age",
+    "wage_per_hour",
+    "capital_gains",
+    "capital_losses",
+    "dividends_from_stocks",
+    "num_persons_worked_for_employer",
+    "weeks_worked_in_year",
+]
+
+##Cols which will use "Number" component of gradio for taking user input
+NUMBER_INPUT_COLS = ['age', 'num_persons_worked_for_employer','weeks_worked_in_year']
+
+test_data = load_test_data()
+
+CATEGORICAL_FEATURES_WITH_VOCABULARY = {
+    feature_name: sorted([str(value) for value in list(test_data[feature_name].unique())])
+    for feature_name in CSV_HEADER
+    if feature_name
+    not in list(NUMERIC_FEATURE_NAMES + [WEIGHT_COLUMN_NAME, TARGET_FEATURE_NAME])
+}
+# All features names.
+FEATURE_NAMES = NUMERIC_FEATURE_NAMES + list(
+    CATEGORICAL_FEATURES_WITH_VOCABULARY.keys()
+)
+# Feature default values.
+COLUMN_DEFAULTS = [
+    [0.0]
+    if feature_name in NUMERIC_FEATURE_NAMES + [TARGET_FEATURE_NAME, WEIGHT_COLUMN_NAME]
+    else ["NA"]
+    for feature_name in CSV_HEADER
+]

predict.py

@@ -0,0 +1,107 @@
+import pandas as pd
+from .preprocess import get_dataset_from_csv
+from huggingface_hub import from_pretrained_keras
+
+##Load Model
+model = from_pretrained_keras("shivi/classification-grn-vsn")
+
+def batch_predict(input_data):
+    """
+    This function is used for fetching predictions corresponding to input_dataframe.
+    It outputs another dataframe containing: 
+        1. prediction probability for each class
+        2. actual expected outcome for each entry in the input dataframe
+    """
+    input_data_file = "prod_data.csv"
+    labels = ['Probability of Income greater than 50000',"Probability of Income less than 50000","Actual Income"]
+    
+    predictions_df = pd.DataFrame(columns=labels)
+
+    input_data.to_csv(input_data_file, index=None, header=None)
+
+    prod_dataset = get_dataset_from_csv(input_data_file, shuffle=True)
+    
+    pred = model.predict(prod_dataset)
+    
+    for prediction, actual_gt in zip(pred, input_data['income_level'].values.tolist()):
+        y_pred_prob = round(prediction.flatten()[0] * 100, 2)
+        y_not_prob = round((1-prediction.flatten()[0]) * 100, 2)
+        y_pred = ">50000" if prediction.flatten()[0] > 0.5 else "<50000"
+        prob_scores = {labels[0]: str(y_pred_prob)+"%" , labels[1]: str(y_not_prob)+"%", labels[2]: y_pred}
+        predictions_df = predictions_df.append(prob_scores,ignore_index=True)
+        
+    return predictions_df
+
+
+def user_input_predict(age, wage, cap_gains, cap_losses, dividends, num_persons, weeks_worked_in_year,
+            class_of_worker, detailed_industry_recode,detailed_occupation_recode,education,
+            enroll_in_edu_inst_last_wk, marital_stat, major_industry_code,major_occupation_code,
+            race, hispanic_origin, sex, member_of_a_labor_union,reason_for_unemployment,
+            full_or_part_time_employment_stat, tax_filer_stat,region_of_previous_residence,
+            state_of_previous_residence,detailed_household_and_family_stat,detailed_household_summary_in_household,
+            migration_codechange_in_msa,migration_codechange_in_reg, migration_codemove_within_reg,
+            live_in_this_house_1_year_ago,migration_prev_res_in_sunbelt,family_members_under_18,
+            country_of_birth_father,country_of_birth_mother,country_of_birth_self,
+            citizenship,own_business_or_self_employed,fill_inc_questionnaire_for_veterans_admin,
+            veterans_benefits, year):
+    
+    """
+    This function is used for fetching model predictions based on inputs given by user on demo app
+    """
+    
+    input_dict = {"age": [age],
+    "class_of_worker": [class_of_worker],
+    "detailed_industry_recode": [detailed_industry_recode],
+    "detailed_occupation_recode": [detailed_occupation_recode],
+    "education":[education],
+    "wage_per_hour": [wage],
+    "enroll_in_edu_inst_last_wk": [enroll_in_edu_inst_last_wk],
+    "marital_stat": [marital_stat],
+    "major_industry_code": [major_industry_code],
+    "major_occupation_code": [major_occupation_code],
+    "race": [race],
+    "hispanic_origin": [hispanic_origin],
+    "sex": [sex],
+    "member_of_a_labor_union": [member_of_a_labor_union],
+    "reason_for_unemployment": [reason_for_unemployment],
+    "full_or_part_time_employment_stat": [full_or_part_time_employment_stat],
+    "capital_gains": [cap_gains],
+    "capital_losses": [cap_losses],
+    "dividends_from_stocks": [dividends],
+    "tax_filer_stat": [tax_filer_stat],
+    "region_of_previous_residence": [region_of_previous_residence],
+    "state_of_previous_residence": [state_of_previous_residence],
+    "detailed_household_and_family_stat": [detailed_household_and_family_stat],
+    "detailed_household_summary_in_household": [detailed_household_summary_in_household],
+    "instance_weight": [0.0],
+    "migration_code-change_in_msa": [migration_codechange_in_msa],
+    "migration_code-change_in_reg": [migration_codechange_in_reg],
+    "migration_code-move_within_reg": [migration_codemove_within_reg],
+    "live_in_this_house_1_year_ago": [live_in_this_house_1_year_ago],
+    "migration_prev_res_in_sunbelt": [migration_prev_res_in_sunbelt],
+    "num_persons_worked_for_employer": [num_persons],
+    "family_members_under_18": [family_members_under_18],
+    "country_of_birth_father": [country_of_birth_father],
+    "country_of_birth_mother": [country_of_birth_mother],
+    "country_of_birth_self": [country_of_birth_self],
+    "citizenship": [citizenship],
+    "own_business_or_self_employed": [own_business_or_self_employed],
+    "fill_inc_questionnaire_for_veterans_admin": [fill_inc_questionnaire_for_veterans_admin],
+    "veterans_benefits": [veterans_benefits],
+    "weeks_worked_in_year": [weeks_worked_in_year],
+    "year": [year],
+    "income_level": [0],
+  }
+    input_df = pd.DataFrame.from_dict(input_dict)
+    input_data_file = "input_data.csv"
+    
+    input_df.to_csv(input_data_file, index=None, header=None)
+    prod_dataset = get_dataset_from_csv(input_data_file, shuffle=True)
+    
+    labels = ['Income greater than 50000',"Income less than 50000"]
+    prediction = model.predict(prod_dataset)
+    y_pred_prob = round(prediction[0].flatten()[0],5)
+    y_not_prob = round(1-prediction[0].flatten()[0],3)
+    
+    confidences = {labels[0]: float(y_pred_prob), labels[1]: float(y_not_prob)}
+    return confidences

preprocess.py

@@ -0,0 +1,65 @@
+import tensorflow as tf
+import pandas as pd
+from .constants import CSV_HEADER, TARGET_FEATURE_NAME, WEIGHT_COLUMN_NAME, NUMERIC_FEATURE_NAMES, COLUMN_DEFAULTS, CATEGORICAL_FEATURES_WITH_VOCABULARY
+
+
+##Helper functions for preprocessing of data:
+
+def process(features, target):
+    for feature_name in features:
+        if feature_name in CATEGORICAL_FEATURES_WITH_VOCABULARY:
+            # Cast categorical feature values to string.
+            features[feature_name] = tf.cast(features[feature_name], tf.dtypes.string)
+    # Get the instance weight.
+    weight = features.pop(WEIGHT_COLUMN_NAME)
+    return features, target, weight
+
+
+def get_dataset_from_csv(csv_file_path, shuffle=False, batch_size=128):
+
+    dataset = tf.data.experimental.make_csv_dataset(
+        csv_file_path,
+        batch_size=batch_size,
+        column_names=CSV_HEADER,
+        column_defaults=COLUMN_DEFAULTS,
+        label_name=TARGET_FEATURE_NAME,
+        num_epochs=1,
+        header=False,
+        shuffle=shuffle,
+    ).map(process)
+
+    return dataset
+
+def create_max_values_map():
+    max_values_map = {}
+    for col in NUMERIC_FEATURE_NAMES:
+        max_val = max(test_data[col])
+        max_values_map["max_"+col] = max_val
+    return max_values_map
+
+def create_dropdown_default_values_map():
+    dropdown_default_values_map = {}
+    for col in CATEGORICAL_FEATURES_WITH_VOCABULARY.keys():
+        max_val = test_data[col].max()
+        dropdown_default_values_map["max_"+col] = max_val
+    return dropdown_default_values_map
+
+def load_test_data():
+    
+    test_data_url = "https://archive.ics.uci.edu/ml/machine-learning-databases/census-income-mld/census-income.test.gz"
+    test_data = pd.read_csv(test_data_url, header=None, names=CSV_HEADER)
+    
+    return test_data
+
+def create_sample_test_data():
+    
+    test_data = load_test_data()
+    
+    test_data["income_level"] = test_data["income_level"].apply(
+    lambda x: 0 if x == " - 50000." else 1)
+    
+    sample_df = test_data.loc[:20,:]
+    sample_df_values = samp.values.tolist()
+    
+    return sample_df_values
+    

requirements.txt

@@ -0,0 +1,2 @@
+tensorflow
+gradio