Spaces:
Build error
Build error
kyleledbetter
commited on
Commit
·
0ec25a0
1
Parent(s):
7bd4255
feat(app): support more models and datasets
Browse files
app.py
CHANGED
@@ -1,104 +1,321 @@
|
|
1 |
import gradio as gr
|
2 |
import requests
|
3 |
import json
|
4 |
-
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
|
|
5 |
from datasets import load_dataset
|
|
|
6 |
import plotly.io as pio
|
7 |
import plotly.graph_objects as go
|
8 |
import plotly.express as px
|
|
|
9 |
import pandas as pd
|
10 |
from sklearn.metrics import confusion_matrix
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
|
12 |
-
|
13 |
-
tokenizer = AutoTokenizer.from_pretrained(endpoint)
|
14 |
-
model = AutoModelForSequenceClassification.from_pretrained(endpoint)
|
15 |
-
return tokenizer, model
|
16 |
|
17 |
|
18 |
def test_model(tokenizer, model, test_data: list, label_map: dict):
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
|
|
27 |
|
28 |
def generate_label_map(dataset):
|
29 |
-
|
30 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
return label_map
|
32 |
|
|
|
|
|
|
|
33 |
|
34 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
true_labels = [r[1] for r in results]
|
36 |
pred_labels = [r[2] for r in results]
|
37 |
|
38 |
-
|
39 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
|
41 |
-
fig = go.Figure(
|
42 |
-
data=go.Heatmap(
|
43 |
-
z=cm,
|
44 |
-
x=list(label_map.values()),
|
45 |
-
y=list(label_map.values()),
|
46 |
-
colorscale='Viridis',
|
47 |
-
colorbar=dict(title='Number of Samples')
|
48 |
-
),
|
49 |
-
layout=go.Layout(
|
50 |
-
title='Confusion Matrix',
|
51 |
-
xaxis=dict(title='Predicted Labels'),
|
52 |
-
yaxis=dict(title='True Labels', autorange='reversed')
|
53 |
-
)
|
54 |
-
)
|
55 |
|
56 |
-
fig.update_layout(height=600, width=800)
|
57 |
|
58 |
-
|
59 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
return fig
|
66 |
|
67 |
-
def app(model_endpoint: str, dataset_name: str, config_name: str, dataset_split: str, num_samples: int):
|
68 |
-
tokenizer, model = load_model(model_endpoint)
|
69 |
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
78 |
|
79 |
-
|
80 |
-
|
|
|
|
|
|
|
81 |
|
82 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
83 |
|
84 |
interface = gr.Interface(
|
85 |
fn=app,
|
86 |
inputs=[
|
87 |
-
gr.inputs.
|
88 |
-
|
|
|
|
|
89 |
gr.inputs.Textbox(lines=1, label="Dataset Name",
|
90 |
-
placeholder="ex: glue"),
|
91 |
gr.inputs.Textbox(lines=1, label="Config Name",
|
92 |
-
placeholder="ex: sst2"),
|
93 |
gr.inputs.Dropdown(
|
94 |
-
choices=["train", "validation", "test"], label="Dataset Split"),
|
95 |
gr.inputs.Number(default=100, label="Number of Samples"),
|
|
|
|
|
|
|
96 |
],
|
97 |
-
# outputs=gr.
|
98 |
# outputs=gr.outputs.HTML(),
|
99 |
-
outputs=gr.Plot(),
|
|
|
|
|
|
|
|
|
100 |
title="Fairness and Bias Testing",
|
101 |
-
description="Enter a model
|
102 |
)
|
103 |
|
104 |
# Define the label map globally
|
|
|
1 |
import gradio as gr
|
2 |
import requests
|
3 |
import json
|
4 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoModelForTokenClassification, AutoModelForQuestionAnswering
|
5 |
+
|
6 |
from datasets import load_dataset
|
7 |
+
import datasets
|
8 |
import plotly.io as pio
|
9 |
import plotly.graph_objects as go
|
10 |
import plotly.express as px
|
11 |
+
from plotly.subplots import make_subplots
|
12 |
import pandas as pd
|
13 |
from sklearn.metrics import confusion_matrix
|
14 |
+
import importlib
|
15 |
+
import torch
|
16 |
+
from dash import Dash, html, dcc
|
17 |
+
import numpy as np
|
18 |
+
from sklearn.metrics import accuracy_score
|
19 |
+
from sklearn.metrics import f1_score
|
20 |
+
|
21 |
+
|
22 |
+
def load_model(model_type: str, model_name_or_path: str, dataset_name: str, config_name: str):
|
23 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
|
24 |
+
|
25 |
+
if model_type == "text_classification":
|
26 |
+
dataset = load_dataset(dataset_name, config_name)
|
27 |
+
num_labels = len(dataset["train"].features["label"].names)
|
28 |
+
|
29 |
+
if "roberta" in model_name_or_path.lower():
|
30 |
+
from transformers import RobertaForSequenceClassification
|
31 |
+
model = RobertaForSequenceClassification.from_pretrained(
|
32 |
+
model_name_or_path, num_labels=num_labels)
|
33 |
+
else:
|
34 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
35 |
+
model_name_or_path, num_labels=num_labels)
|
36 |
+
elif model_type == "token_classification":
|
37 |
+
dataset = load_dataset(dataset_name, config_name)
|
38 |
+
num_labels = len(
|
39 |
+
dataset["train"].features["ner_tags"].feature.names)
|
40 |
+
model = AutoModelForTokenClassification.from_pretrained(
|
41 |
+
model_name_or_path, num_labels=num_labels)
|
42 |
+
elif model_type == "question_answering":
|
43 |
+
model = AutoModelForQuestionAnswering.from_pretrained(model_name_or_path)
|
44 |
+
else:
|
45 |
+
raise ValueError(f"Invalid model type: {model_type}")
|
46 |
|
47 |
+
return tokenizer, model
|
|
|
|
|
|
|
48 |
|
49 |
|
50 |
def test_model(tokenizer, model, test_data: list, label_map: dict):
|
51 |
+
results = []
|
52 |
+
for text, _, true_label in test_data:
|
53 |
+
inputs = tokenizer(text, return_tensors="pt",
|
54 |
+
truncation=True, padding=True)
|
55 |
+
outputs = model(**inputs)
|
56 |
+
pred_label = label_map[int(outputs.logits.argmax(dim=-1))]
|
57 |
+
results.append((text, true_label, pred_label))
|
58 |
+
return results
|
59 |
+
|
60 |
|
61 |
def generate_label_map(dataset):
|
62 |
+
if "label" not in dataset.features or dataset.features["label"] is None:
|
63 |
+
return {}
|
64 |
+
|
65 |
+
if isinstance(dataset.features["label"], datasets.ClassLabel):
|
66 |
+
num_labels = dataset.features["label"].num_classes
|
67 |
+
label_map = {i: label for i, label in enumerate(dataset.features["label"].names)}
|
68 |
+
else:
|
69 |
+
num_labels = len(set(dataset["label"]))
|
70 |
+
label_map = {i: label for i, label in enumerate(set(dataset["label"]))}
|
71 |
return label_map
|
72 |
|
73 |
+
def calculate_fairness_score(results, label_map):
|
74 |
+
true_labels = [r[1] for r in results]
|
75 |
+
pred_labels = [r[2] for r in results]
|
76 |
|
77 |
+
# Overall accuracy
|
78 |
+
# accuracy = (true_labels == pred_labels).mean()
|
79 |
+
accuracy = accuracy_score(true_labels, pred_labels)
|
80 |
+
# Calculate confusion matrix for each group
|
81 |
+
group_names = label_map.values()
|
82 |
+
group_cms = {}
|
83 |
+
for group in group_names:
|
84 |
+
true_group_indices = [i for i, label in enumerate(true_labels) if label == group]
|
85 |
+
pred_group_labels = [pred_labels[i] for i in true_group_indices]
|
86 |
+
true_group_labels = [true_labels[i] for i in true_group_indices]
|
87 |
+
|
88 |
+
cm = confusion_matrix(true_group_labels, pred_group_labels, labels=list(group_names))
|
89 |
+
group_cms[group] = cm
|
90 |
+
|
91 |
+
# Calculate fairness score
|
92 |
+
score = 0
|
93 |
+
for i, group1 in enumerate(group_names):
|
94 |
+
for j, group2 in enumerate(group_names):
|
95 |
+
if i < j:
|
96 |
+
cm1 = group_cms[group1]
|
97 |
+
cm2 = group_cms[group2]
|
98 |
+
diff = np.abs(cm1 - cm2)
|
99 |
+
score += (diff.sum() / 2) / cm1.sum()
|
100 |
+
|
101 |
+
return accuracy, score
|
102 |
+
|
103 |
+
def calculate_per_class_metrics(true_labels, pred_labels, label_map, metric='accuracy'):
|
104 |
+
unique_labels = sorted(label_map.values())
|
105 |
+
metrics = []
|
106 |
+
|
107 |
+
if metric == 'accuracy':
|
108 |
+
for label in unique_labels:
|
109 |
+
label_indices = [i for i, true_label in enumerate(true_labels) if true_label == label]
|
110 |
+
true_label_subset = [true_labels[i] for i in label_indices]
|
111 |
+
pred_label_subset = [pred_labels[i] for i in label_indices]
|
112 |
+
accuracy = accuracy_score(true_label_subset, pred_label_subset)
|
113 |
+
metrics.append(accuracy)
|
114 |
+
elif metric == 'f1':
|
115 |
+
f1_scores = f1_score(true_labels, pred_labels, labels=unique_labels, average=None)
|
116 |
+
metrics = f1_scores.tolist()
|
117 |
+
else:
|
118 |
+
raise ValueError(f"Invalid metric: {metric}")
|
119 |
+
|
120 |
+
return metrics
|
121 |
+
|
122 |
+
def generate_visualization(visualization_type, results, label_map):
|
123 |
true_labels = [r[1] for r in results]
|
124 |
pred_labels = [r[2] for r in results]
|
125 |
|
126 |
+
if visualization_type == "confusion_matrix":
|
127 |
+
return generate_report_card(results, label_map)["fig"]
|
128 |
+
elif visualization_type == "per_class_accuracy":
|
129 |
+
per_class_accuracy = calculate_per_class_metrics(
|
130 |
+
true_labels, pred_labels, label_map, metric='accuracy')
|
131 |
+
|
132 |
+
colors = px.colors.qualitative.Plotly
|
133 |
+
fig = go.Figure()
|
134 |
+
for i, label in enumerate(label_map.values()):
|
135 |
+
fig.add_trace(go.Bar(
|
136 |
+
x=[label],
|
137 |
+
y=[per_class_accuracy[i]],
|
138 |
+
name=label,
|
139 |
+
marker_color=colors[i % len(colors)]
|
140 |
+
))
|
141 |
+
|
142 |
+
fig.update_layout(title='Per-Class Accuracy',
|
143 |
+
xaxis_title='Class', yaxis_title='Accuracy')
|
144 |
+
return fig
|
145 |
+
elif visualization_type == "per_class_f1":
|
146 |
+
per_class_f1 = calculate_per_class_metrics(
|
147 |
+
true_labels, pred_labels, label_map, metric='f1')
|
148 |
+
|
149 |
+
colors = px.colors.qualitative.Plotly
|
150 |
+
fig = go.Figure()
|
151 |
+
for i, label in enumerate(label_map.values()):
|
152 |
+
fig.add_trace(go.Bar(
|
153 |
+
x=[label],
|
154 |
+
y=[per_class_f1[i]],
|
155 |
+
name=label,
|
156 |
+
marker_color=colors[i % len(colors)]
|
157 |
+
))
|
158 |
+
|
159 |
+
fig.update_layout(title='Per-Class F1-Score',
|
160 |
+
xaxis_title='Class', yaxis_title='F1-Score')
|
161 |
+
return fig
|
162 |
+
else:
|
163 |
+
raise ValueError(f"Invalid visualization type: {visualization_type}")
|
164 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
165 |
|
|
|
166 |
|
167 |
+
def generate_report_card(results, label_map):
|
168 |
+
true_labels = [r[1] for r in results]
|
169 |
+
pred_labels = [r[2] for r in results]
|
170 |
+
|
171 |
+
cm = confusion_matrix(true_labels, pred_labels,
|
172 |
+
labels=list(label_map.values()))
|
173 |
+
|
174 |
+
# Create the plotly figure
|
175 |
+
fig = make_subplots(rows=1, cols=1)
|
176 |
+
fig.add_trace(go.Heatmap(
|
177 |
+
z=cm,
|
178 |
+
x=list(label_map.values()),
|
179 |
+
y=list(label_map.values()),
|
180 |
+
colorscale='RdYlGn',
|
181 |
+
colorbar=dict(title='# of Samples')
|
182 |
+
))
|
183 |
+
fig.update_layout(
|
184 |
+
height=500, width=600,
|
185 |
+
title='Confusion Matrix',
|
186 |
+
xaxis=dict(title='Predicted Labels'),
|
187 |
+
yaxis=dict(title='True Labels', autorange='reversed')
|
188 |
+
)
|
189 |
+
|
190 |
+
# Create the text output
|
191 |
+
# accuracy = pd.Series(true_labels) == pd.Series(pred_labels)
|
192 |
+
accuracy = accuracy_score(true_labels, pred_labels, normalize=False)
|
193 |
+
fairness_score = calculate_fairness_score(results, label_map)
|
194 |
|
195 |
+
per_class_accuracy = calculate_per_class_metrics(
|
196 |
+
true_labels, pred_labels, label_map, metric='accuracy')
|
197 |
+
per_class_f1 = calculate_per_class_metrics(
|
198 |
+
true_labels, pred_labels, label_map, metric='f1')
|
|
|
199 |
|
|
|
|
|
200 |
|
201 |
+
text_output = html.Div(children=[
|
202 |
+
html.H2('Performance Metrics'),
|
203 |
+
html.Div(children=[
|
204 |
+
html.Div(children=[
|
205 |
+
html.H3('Accuracy'),
|
206 |
+
html.H4(f'{accuracy}')
|
207 |
+
], className='metric'),
|
208 |
+
html.Div(children=[
|
209 |
+
html.H3('Fairness Score'),
|
210 |
+
# html.H4(f'{fairness_score}')
|
211 |
+
html.H4(
|
212 |
+
f'Accuracy: {fairness_score[0]:.2f}, Score: {fairness_score[1]:.2f}')
|
213 |
+
], className='metric'),
|
214 |
+
], className='metric-container'),
|
215 |
+
], className='text-output')
|
216 |
|
217 |
+
# Combine the plot and text output into a Dash container
|
218 |
+
# report_card = html.Div([
|
219 |
+
# dcc.Graph(figure=fig),
|
220 |
+
# text_output,
|
221 |
+
# ])
|
222 |
|
223 |
+
# return report_card
|
224 |
+
|
225 |
+
report_card = {
|
226 |
+
"fig": fig,
|
227 |
+
"accuracy": accuracy,
|
228 |
+
"fairness_score": fairness_score,
|
229 |
+
"per_class_accuracy": per_class_accuracy,
|
230 |
+
"per_class_f1": per_class_f1
|
231 |
+
}
|
232 |
+
return report_card
|
233 |
+
|
234 |
+
# return fig, text_output
|
235 |
+
|
236 |
+
|
237 |
+
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):
|
238 |
+
tokenizer, model = load_model(
|
239 |
+
model_type, model_name_or_path, dataset_name, config_name)
|
240 |
+
|
241 |
+
# Load the dataset
|
242 |
+
# Add this line to cast num_samples to an integer
|
243 |
+
num_samples = int(num_samples)
|
244 |
+
dataset = load_dataset(
|
245 |
+
dataset_name, config_name, split=f"{dataset_split}[:{num_samples}]")
|
246 |
+
test_data = []
|
247 |
+
|
248 |
+
if dataset_name == "glue":
|
249 |
+
test_data = [(item["sentence"], None,
|
250 |
+
dataset.features["label"].names[item["label"]]) for item in dataset]
|
251 |
+
elif dataset_name == "tweet_eval":
|
252 |
+
test_data = [(item["text"], None, dataset.features["label"].names[item["label"]])
|
253 |
+
for item in dataset]
|
254 |
+
else:
|
255 |
+
test_data = [(item["sentence"], None,
|
256 |
+
dataset.features["label"].names[item["label"]]) for item in dataset]
|
257 |
+
|
258 |
+
# if model_type == "text_classification":
|
259 |
+
# for item in dataset:
|
260 |
+
# text = item["sentence"]
|
261 |
+
# context = None
|
262 |
+
# true_label = item["label"]
|
263 |
+
# test_data.append((text, context, true_label))
|
264 |
+
# elif model_type == "question_answering":
|
265 |
+
# for item in dataset:
|
266 |
+
# text = item["question"]
|
267 |
+
# context = item["context"]
|
268 |
+
# true_label = None
|
269 |
+
# test_data.append((text, context, true_label))
|
270 |
+
# else:
|
271 |
+
# raise ValueError(f"Invalid model type: {model_type}")
|
272 |
+
|
273 |
+
label_map = generate_label_map(dataset)
|
274 |
+
|
275 |
+
results = test_model(tokenizer, model, test_data, label_map)
|
276 |
+
# fig, text_output = generate_report_card(results, label_map)
|
277 |
+
|
278 |
+
# return fig, text_output
|
279 |
+
|
280 |
+
report_card = generate_report_card(results, label_map)
|
281 |
+
visualization = generate_visualization(visualization_type, results, label_map)
|
282 |
+
|
283 |
+
per_class_metrics_str = "\n".join([f"{label}: Acc {acc:.2f}, F1 {f1:.2f}" for label, acc, f1 in zip(
|
284 |
+
label_map.values(), report_card['per_class_accuracy'], report_card['per_class_f1'])])
|
285 |
+
|
286 |
+
|
287 |
+
# return report_card["fig"], f"Accuracy: {report_card['accuracy']}, Fairness Score: {report_card['fairness_score'][1]:.2f}"
|
288 |
+
# return f"Accuracy: {report_card['accuracy']}, Fairness Score: {report_card['fairness_score'][1]:.2f}", report_card["fig"]
|
289 |
+
return (f"Accuracy: {report_card['accuracy']}, Fairness Score: {report_card['fairness_score'][1]:.2f}\n\n"
|
290 |
+
f"Per-Class Metrics:\n{per_class_metrics_str}"), visualization
|
291 |
|
292 |
interface = gr.Interface(
|
293 |
fn=app,
|
294 |
inputs=[
|
295 |
+
gr.inputs.Radio(["text_classification", "token_classification",
|
296 |
+
"question_answering"], label="Model Type", default="text_classification"),
|
297 |
+
gr.inputs.Textbox(lines=1, label="Model Name or Path",
|
298 |
+
placeholder="ex: distilbert-base-uncased-finetuned-sst-2-english", default="distilbert-base-uncased-finetuned-sst-2-english"),
|
299 |
gr.inputs.Textbox(lines=1, label="Dataset Name",
|
300 |
+
placeholder="ex: glue", default="glue"),
|
301 |
gr.inputs.Textbox(lines=1, label="Config Name",
|
302 |
+
placeholder="ex: sst2", default="cola"),
|
303 |
gr.inputs.Dropdown(
|
304 |
+
choices=["train", "validation", "test"], label="Dataset Split", default="validation"),
|
305 |
gr.inputs.Number(default=100, label="Number of Samples"),
|
306 |
+
gr.inputs.Dropdown(
|
307 |
+
choices=["confusion_matrix", "per_class_accuracy", "per_class_f1"], label="Visualization Type", default="confusion_matrix"
|
308 |
+
),
|
309 |
],
|
310 |
+
# outputs=gr.Plot(),
|
311 |
# outputs=gr.outputs.HTML(),
|
312 |
+
# outputs=[gr.outputs.HTML(), gr.Plot()],
|
313 |
+
outputs=[
|
314 |
+
gr.outputs.Textbox(label="Fairness and Bias Metrics"),
|
315 |
+
gr.Plot(label="Graph")
|
316 |
+
],
|
317 |
title="Fairness and Bias Testing",
|
318 |
+
description="Enter a model and dataset to test for fairness and bias.",
|
319 |
)
|
320 |
|
321 |
# Define the label map globally
|