import gradio as gr import requests import json from transformers import pipeline, AutoTokenizer, AutoModelForSequenceClassification, AutoModelForTokenClassification, AutoModelForQuestionAnswering from datasets import load_dataset import datasets import plotly.io as pio import plotly.graph_objects as go import plotly.express as px from plotly.subplots import make_subplots import pandas as pd from sklearn.metrics import confusion_matrix import importlib import torch from dash import Dash, html, dcc import numpy as np from sklearn.metrics import accuracy_score from sklearn.metrics import f1_score def load_model(model_type: str, model_name_or_path: str, dataset_name: str, config_name: str): tokenizer = AutoTokenizer.from_pretrained(model_name_or_path) if model_type == "text_classification": dataset = load_dataset(dataset_name, config_name) num_labels = len(dataset["train"].features["label"].names) if "roberta" in model_name_or_path.lower(): from transformers import RobertaForSequenceClassification model = RobertaForSequenceClassification.from_pretrained( model_name_or_path, num_labels=num_labels) else: model = AutoModelForSequenceClassification.from_pretrained( model_name_or_path, num_labels=num_labels) elif model_type == "token_classification": dataset = load_dataset(dataset_name, config_name) num_labels = len( dataset["train"].features["ner_tags"].feature.names) model = AutoModelForTokenClassification.from_pretrained( model_name_or_path, num_labels=num_labels) elif model_type == "question_answering": model = AutoModelForQuestionAnswering.from_pretrained(model_name_or_path) else: raise ValueError(f"Invalid model type: {model_type}") return tokenizer, model def test_model(tokenizer, model, test_data: list, label_map: dict): results = [] for text, _, true_label in test_data: inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True) outputs = model(**inputs) pred_label = label_map[int(outputs.logits.argmax(dim=-1))] results.append((text, true_label, pred_label)) return results def generate_label_map(dataset): if "label" not in dataset.features or dataset.features["label"] is None: return {} if isinstance(dataset.features["label"], datasets.ClassLabel): num_labels = dataset.features["label"].num_classes label_map = {i: label for i, label in enumerate(dataset.features["label"].names)} else: num_labels = len(set(dataset["label"])) label_map = {i: label for i, label in enumerate(set(dataset["label"]))} return label_map # Explain fairness score: https://arxiv.org/pdf/1908.09635.pdf def calculate_fairness_score(results, label_map): true_labels = [r[1] for r in results] pred_labels = [r[2] for r in results] # Overall accuracy # accuracy = (true_labels == pred_labels).mean() accuracy = accuracy_score(true_labels, pred_labels) # Calculate confusion matrix for each group group_names = label_map.values() group_cms = {} for group in group_names: true_group_indices = [i for i, label in enumerate(true_labels) if label == group] pred_group_labels = [pred_labels[i] for i in true_group_indices] true_group_labels = [true_labels[i] for i in true_group_indices] cm = confusion_matrix(true_group_labels, pred_group_labels, labels=list(group_names)) group_cms[group] = cm # Calculate fairness score which means the average difference between confusion matrices score = 0 for i, group1 in enumerate(group_names): for j, group2 in enumerate(group_names): if i < j: cm1 = group_cms[group1] cm2 = group_cms[group2] diff = np.abs(cm1 - cm2) score += (diff.sum() / 2) / cm1.sum() return accuracy, score # Per-class metrics means the metrics for each class, and the class is defined by the label_map def calculate_per_class_metrics(true_labels, pred_labels, label_map, metric='accuracy'): unique_labels = sorted(label_map.values()) metrics = [] if metric == 'accuracy': for label in unique_labels: label_indices = [i for i, true_label in enumerate(true_labels) if true_label == label] true_label_subset = [true_labels[i] for i in label_indices] pred_label_subset = [pred_labels[i] for i in label_indices] accuracy = accuracy_score(true_label_subset, pred_label_subset) metrics.append(accuracy) elif metric == 'f1': f1_scores = f1_score(true_labels, pred_labels, labels=unique_labels, average=None) metrics = f1_scores.tolist() else: raise ValueError(f"Invalid metric: {metric}") return metrics def generate_fairness_statement(accuracy, fairness_score): accuracy_level = "high" if accuracy >= 0.85 else "moderate" if accuracy >= 0.7 else "low" fairness_level = "low" if fairness_score <= 0.15 else "moderate" if fairness_score <= 0.3 else "high" # statement = f"The model has a {accuracy_level} overall accuracy of {accuracy * 100:.2f}% and a {fairness_level} fairness score of {fairness_score:.2f}. " statement = f"Assessment: " if fairness_level == "low": statement += f"The low fairness score ({fairness_score:.2f}) and accuracy ({accuracy * 100:.2f}%) indicate that the model is relatively fair and does not exhibit significant bias across different groups." elif fairness_level == "moderate": statement += f"The moderate fairness score ({fairness_score:.2f}) and accuracy ({accuracy * 100:.2f}%) suggest that the model may have some bias across different groups, and further investigation is needed to ensure it does not disproportionately affect certain groups." else: statement += f"The high fairness score ({fairness_score:.2f}) and accuracy ({accuracy * 100:.2f}%) indicate that the model exhibits significant bias across different groups, and it's recommended to address this issue to ensure fair predictions for all groups." return statement def generate_visualization(visualization_type, results, label_map, chart_mode): true_labels = [r[1] for r in results] pred_labels = [r[2] for r in results] background_color = "white" if chart_mode == "Light" else "black" text_color = "black" if chart_mode == "Light" else "white" if visualization_type == "confusion_matrix": return generate_report_card(results, label_map, chart_mode)["fig"] elif visualization_type == "per_class_accuracy": per_class_accuracy = calculate_per_class_metrics( true_labels, pred_labels, label_map, metric='accuracy') colors = px.colors.qualitative.Plotly fig = go.Figure() for i, label in enumerate(label_map.values()): fig.add_trace(go.Bar( x=[label], y=[per_class_accuracy[i]], name=label, marker_color=colors[i % len(colors)] )) fig.update_xaxes(showgrid=True, gridwidth=1, gridcolor='LightGray', linecolor='black', linewidth=1) fig.update_yaxes(showgrid=True, gridwidth=1, gridcolor='LightGray', linecolor='black', linewidth=1) fig.update_layout(plot_bgcolor=background_color, paper_bgcolor=background_color, font=dict(color=text_color), title='Per-Class Accuracy', xaxis_title='Class', yaxis_title='Accuracy' ) return fig elif visualization_type == "per_class_f1": per_class_f1 = calculate_per_class_metrics( true_labels, pred_labels, label_map, metric='f1') colors = px.colors.qualitative.Plotly fig = go.Figure() for i, label in enumerate(label_map.values()): fig.add_trace(go.Bar( x=[label], y=[per_class_f1[i]], name=label, marker_color=colors[i % len(colors)] )) fig.update_xaxes(showgrid=True, gridwidth=1, gridcolor='LightGray', linecolor='black', linewidth=1) fig.update_yaxes(showgrid=True, gridwidth=1, gridcolor='LightGray', linecolor='black', linewidth=1) fig.update_layout(plot_bgcolor=background_color, paper_bgcolor=background_color, font=dict(color=text_color), title='Per-Class F1-Score', xaxis_title='Class', yaxis_title='F1-Score' ) return fig elif visualization_type == "interactive_dashboard": return generate_interactive_dashboard(results, label_map, chart_mode) else: raise ValueError(f"Invalid visualization type: {visualization_type}") def generate_interactive_dashboard(results, label_map, chart_mode): true_labels = [r[1] for r in results] pred_labels = [r[2] for r in results] colors = ['#EF553B', '#00CC96', '#636EFA', '#AB63FA', '#FFA15A', '#19D3F3', '#FF6692', '#B6E880', '#FF97FF', '#FECB52'] background_color = "white" if chart_mode == "Light" else "black" text_color = "black" if chart_mode == "Light" else "white" # Create confusion matrix cm_fig = generate_report_card(results, label_map, chart_mode)["fig"] # Create per-class accuracy bar chart pca_data = calculate_per_class_metrics(true_labels, pred_labels, label_map, metric='accuracy') pca_fig = go.Bar(x=list(label_map.values()), y=pca_data, marker=dict(color=colors)) # Create per-class F1-score bar chart pcf_data = calculate_per_class_metrics(true_labels, pred_labels, label_map, metric='f1') pcf_fig = go.Bar(x=list(label_map.values()), y=pcf_data, marker=dict(color=colors)) # Combine all charts into a mixed subplot fig = make_subplots(rows=2, cols=2, shared_xaxes=True, specs=[[{"colspan": 2}, None], [{}, {}]], print_grid=True,subplot_titles=( "Confusion Matrix", "Per-Class Accuracy", "Per-Class F1-Score")) fig.add_trace(cm_fig['data'][0], row=1, col=1) fig.add_trace(pca_fig, row=2, col=1) fig.add_trace(pcf_fig, row=2, col=2) fig.update_xaxes(showgrid=True, gridwidth=1, gridcolor='LightGray', linecolor='black', linewidth=1) fig.update_yaxes(showgrid=True, gridwidth=1, gridcolor='LightGray', linecolor='black', linewidth=1) # Update layout fig.update_layout(height=700, width=650, plot_bgcolor=background_color, paper_bgcolor=background_color, font=dict(color=text_color), title="Fairness Report", showlegend=False ) return fig def generate_report_card(results, label_map, chart_mode): true_labels = [r[1] for r in results] pred_labels = [r[2] for r in results] background_color = "white" if chart_mode == "Light" else "black" text_color = "black" if chart_mode == "Light" else "white" cm = confusion_matrix(true_labels, pred_labels) # Normalize the confusion matrix cm_normalized = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis] # Create a custom color scale custom_color_scale = np.zeros(cm_normalized.shape, dtype='str') for i in range(cm_normalized.shape[0]): for j in range(cm_normalized.shape[1]): custom_color_scale[i, j] = '#EF553B' if i == j else '#00CC96' fig = go.Figure(go.Heatmap(z=cm_normalized, x=list(label_map.values()), y=list(label_map.values()), text=cm, hovertemplate='%{text}', colorscale=[[0, '#EF553B'], [ 1, '#00CC96']], showscale=False, zmin=0, zmax=1, customdata=custom_color_scale)) fig.update_xaxes(showgrid=True, gridwidth=1, gridcolor='LightGray', linecolor='black', linewidth=1) fig.update_yaxes(showgrid=True, gridwidth=1, gridcolor='LightGray', linecolor='black', linewidth=1) fig.update_layout( plot_bgcolor=background_color, paper_bgcolor=background_color, font=dict(color=text_color), height=500, width=600, title='Confusion Matrix', xaxis=dict(title='Predicted Labels'), yaxis=dict(title='True Labels') ) # Create the text output # accuracy = pd.Series(true_labels) == pd.Series(pred_labels) accuracy = accuracy_score(true_labels, pred_labels, normalize=False) fairness_score = calculate_fairness_score(results, label_map) per_class_accuracy = calculate_per_class_metrics( true_labels, pred_labels, label_map, metric='accuracy') per_class_f1 = calculate_per_class_metrics( true_labels, pred_labels, label_map, metric='f1') report_card = { "fig": fig, "accuracy": accuracy, "fairness_score": fairness_score, "per_class_accuracy": per_class_accuracy, "per_class_f1": per_class_f1 } return report_card # return fig, text_output def generate_insights(custom_text, model_name, dataset_name, accuracy, fairness_score, report_card, generator): per_class_metrics = { 'accuracy': report_card.get('per_class_accuracy', []), 'f1': report_card.get('per_class_f1', []) } if not per_class_metrics['accuracy'] or not per_class_metrics['f1']: input_text = f"{custom_text} The model {model_name} has been evaluated on the {dataset_name} dataset. It has an overall accuracy of {accuracy * 100:.2f}%. The fairness score is {fairness_score:.2f}. Per-class metrics could not be calculated. Please provide some interesting insights about the fairness and bias of the model." else: input_text = f"{custom_text} The model {model_name} has been evaluated on the {dataset_name} dataset. It has an overall accuracy of {accuracy * 100:.2f}%. The fairness score is {fairness_score:.2f}. The per-class metrics are: {per_class_metrics}. Please provide some interesting insights about the fairness, bias, and per-class performance." insights = generator(input_text, max_length=600, do_sample=True, temperature=0.7) return insights[0]['generated_text'] def app(model_type: str, model_name_or_path: str, dataset_name: str, config_name: str, dataset_split: str, num_samples: int, visualization_type: str, chart_mode: str): tokenizer, model = load_model( model_type, model_name_or_path, dataset_name, config_name) # Load the dataset # Add this line to cast num_samples to an integer num_samples = int(num_samples) dataset = load_dataset( dataset_name, config_name, split=f"{dataset_split}[:{num_samples}]") test_data = [] if dataset_name == "glue": test_data = [(item["sentence"], None, dataset.features["label"].names[item["label"]]) for item in dataset] elif dataset_name == "tweet_eval": test_data = [(item["text"], None, dataset.features["label"].names[item["label"]]) for item in dataset] else: test_data = [(item["sentence"], None, dataset.features["label"].names[item["label"]]) for item in dataset] # if model_type == "text_classification": # for item in dataset: # text = item["sentence"] # context = None # true_label = item["label"] # test_data.append((text, context, true_label)) # elif model_type == "question_answering": # for item in dataset: # text = item["question"] # context = item["context"] # true_label = None # test_data.append((text, context, true_label)) # else: # raise ValueError(f"Invalid model type: {model_type}") label_map = generate_label_map(dataset) results = test_model(tokenizer, model, test_data, label_map) # fig, text_output = generate_report_card(results, label_map) # return fig, text_output report_card = generate_report_card(results, label_map, chart_mode) visualization = generate_visualization(visualization_type, results, label_map, chart_mode) per_class_metrics_str = "\n".join([f"{label}: Acc {acc:.2f}, F1 {f1:.2f}" for label, acc, f1 in zip( label_map.values(), report_card['per_class_accuracy'], report_card['per_class_f1'])]) accuracy, fairness_score = calculate_fairness_score(results, label_map) fairness_statement = generate_fairness_statement(accuracy, fairness_score) # Use a GPU if available, otherwise use -1 for CPU. generator = pipeline( 'text-generation', model='gpt2', device=-1) # Use EleutherAI/gpt-neo-1.3B or EleutherAI/GPT-J-6B for GPT3 for distilgpt2 for GPT2 per_class_metrics = { 'accuracy': report_card['per_class_accuracy'], 'f1': report_card['per_class_f1'] } custom_text = fairness_statement insights = generate_insights(custom_text, model_name_or_path, dataset_name, accuracy, fairness_score, report_card, generator) # return report_card["fig"], f"Accuracy: {report_card['accuracy']}, Fairness Score: {report_card['fairness_score'][1]:.2f}" # return f"Accuracy: {report_card['accuracy']}, Fairness Score: {report_card['fairness_score'][1]:.2f}", report_card["fig"] return (f"{insights}\n\n" f"Accuracy: {report_card['accuracy']}, Fairness Score: {report_card['fairness_score'][1]: .2f}\n\n" f"Per-Class Metrics:\n{per_class_metrics_str}"), visualization interface = gr.Interface( fn=app, inputs=[ gr.inputs.Radio(["text_classification", "token_classification", "question_answering"], label="Model Type", default="text_classification"), gr.inputs.Textbox(lines=1, label="Model Name or Path", placeholder="ex: distilbert-base-uncased-finetuned-sst-2-english", default="distilbert-base-uncased-finetuned-sst-2-english"), gr.inputs.Textbox(lines=1, label="Dataset Name", placeholder="ex: glue", default="glue"), gr.inputs.Textbox(lines=1, label="Config Name", placeholder="ex: sst2", default="cola"), gr.inputs.Dropdown( choices=["train", "validation", "test"], label="Dataset Split", default="validation"), gr.inputs.Number(default=100, label="Number of Samples"), gr.inputs.Dropdown( choices=["interactive_dashboard", "confusion_matrix", "per_class_accuracy", "per_class_f1"], label="Visualization Type", default="interactive_dashboard" ), gr.inputs.Radio(["Light", "Dark"], label="Chart Mode", default="Light"), ], # outputs=gr.Plot(), # outputs=gr.outputs.HTML(), # outputs=[gr.outputs.HTML(), gr.Plot()], outputs=[ gr.outputs.Textbox(label="Fairness and Bias Metrics"), gr.Plot(label="Graph") ], title="Fairness and Bias Testing", description="Enter a model and dataset to test for fairness and bias.", ) # Define the label map globally label_map = {0: "negative", 1: "positive"} if __name__ == "__main__": interface.launch()