Omartificial-Intelligence-Space
commited on
Commit
•
daf1d9d
1
Parent(s):
e1559b3
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,146 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import random
|
3 |
+
import numpy as np
|
4 |
+
import pandas as pd
|
5 |
+
from datasets import load_dataset
|
6 |
+
from sentence_transformers import CrossEncoder
|
7 |
+
from sklearn.metrics import average_precision_score
|
8 |
+
import matplotlib.pyplot as plt
|
9 |
+
import torch
|
10 |
+
import spaces
|
11 |
+
|
12 |
+
# Check for GPU support and configure appropriately
|
13 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
14 |
+
zero = torch.Tensor([0]).to(device)
|
15 |
+
print(f"Device being used: {zero.device}")
|
16 |
+
|
17 |
+
# Define evaluation metrics
|
18 |
+
def mean_reciprocal_rank(relevance_labels, scores):
|
19 |
+
sorted_indices = np.argsort(scores)[::-1]
|
20 |
+
for rank, idx in enumerate(sorted_indices, start=1):
|
21 |
+
if relevance_labels[idx] == 1:
|
22 |
+
return 1 / rank
|
23 |
+
return 0
|
24 |
+
|
25 |
+
def mean_average_precision(relevance_labels, scores):
|
26 |
+
return average_precision_score(relevance_labels, scores)
|
27 |
+
|
28 |
+
def ndcg_at_k(relevance_labels, scores, k=10):
|
29 |
+
sorted_indices = np.argsort(scores)[::-1]
|
30 |
+
relevance_sorted = np.take(relevance_labels, sorted_indices[:k])
|
31 |
+
dcg = sum(rel / np.log2(rank + 2) for rank, rel in enumerate(relevance_sorted))
|
32 |
+
idcg = sum(1 / np.log2(rank + 2) for rank in range(min(k, sum(relevance_labels))))
|
33 |
+
return dcg / idcg if idcg > 0 else 0
|
34 |
+
|
35 |
+
# Load datasets
|
36 |
+
datasets = {
|
37 |
+
"Relevance_Labels_Dataset": load_dataset("NAMAA-Space/Ar-Reranking-Eval")["train"],
|
38 |
+
"Positive_Negatives_Dataset": load_dataset("NAMAA-Space/Arabic-Reranking-Triplet-5-Eval")["train"]
|
39 |
+
}
|
40 |
+
|
41 |
+
@spaces.GPU
|
42 |
+
def evaluate_model_with_insights(model_name):
|
43 |
+
model = CrossEncoder(model_name, device=device)
|
44 |
+
results = []
|
45 |
+
sample_outputs = []
|
46 |
+
|
47 |
+
for dataset_name, dataset in datasets.items():
|
48 |
+
all_mrr, all_map, all_ndcg = [], [], []
|
49 |
+
dataset_samples = []
|
50 |
+
|
51 |
+
if 'candidate_document' in dataset.column_names:
|
52 |
+
grouped_data = dataset.to_pandas().groupby("query")
|
53 |
+
for query, group in grouped_data:
|
54 |
+
candidate_texts = group['candidate_document'].tolist()
|
55 |
+
relevance_labels = group['relevance_label'].tolist()
|
56 |
+
pairs = [(query, doc) for doc in candidate_texts]
|
57 |
+
scores = model.predict(pairs)
|
58 |
+
|
59 |
+
# Collecting top-5 results for display
|
60 |
+
sorted_indices = np.argsort(scores)[::-1]
|
61 |
+
top_docs = [(candidate_texts[i], scores[i], relevance_labels[i]) for i in sorted_indices[:5]]
|
62 |
+
dataset_samples.append({
|
63 |
+
"Query": query,
|
64 |
+
"Top 5 Candidates": top_docs
|
65 |
+
})
|
66 |
+
|
67 |
+
# Metrics
|
68 |
+
all_mrr.append(mean_reciprocal_rank(relevance_labels, scores))
|
69 |
+
all_map.append(mean_average_precision(relevance_labels, scores))
|
70 |
+
all_ndcg.append(ndcg_at_k(relevance_labels, scores, k=10))
|
71 |
+
else:
|
72 |
+
for entry in dataset:
|
73 |
+
query = entry['query']
|
74 |
+
candidate_texts = [entry['positive'], entry['negative1'], entry['negative2'], entry['negative3'], entry['negative4']]
|
75 |
+
relevance_labels = [1, 0, 0, 0, 0]
|
76 |
+
pairs = [(query, doc) for doc in candidate_texts]
|
77 |
+
scores = model.predict(pairs)
|
78 |
+
|
79 |
+
# Collecting top-5 results for display
|
80 |
+
sorted_indices = np.argsort(scores)[::-1]
|
81 |
+
top_docs = [(candidate_texts[i], scores[i], relevance_labels[i]) for i in sorted_indices[:5]]
|
82 |
+
dataset_samples.append({
|
83 |
+
"Query": query,
|
84 |
+
"Top 5 Candidates": top_docs
|
85 |
+
})
|
86 |
+
|
87 |
+
# Metrics
|
88 |
+
all_mrr.append(mean_reciprocal_rank(relevance_labels, scores))
|
89 |
+
all_map.append(mean_average_precision(relevance_labels, scores))
|
90 |
+
all_ndcg.append(ndcg_at_k(relevance_labels, scores, k=10))
|
91 |
+
|
92 |
+
# Metrics for this dataset
|
93 |
+
results.append({
|
94 |
+
"Dataset": dataset_name,
|
95 |
+
"MRR": np.mean(all_mrr),
|
96 |
+
"MAP": np.mean(all_map),
|
97 |
+
"nDCG@10": np.mean(all_ndcg)
|
98 |
+
})
|
99 |
+
|
100 |
+
# Collect sample outputs for inspection
|
101 |
+
sample_outputs.extend(dataset_samples)
|
102 |
+
|
103 |
+
results_df = pd.DataFrame(results)
|
104 |
+
|
105 |
+
# Plot results as a bar chart
|
106 |
+
fig, ax = plt.subplots(figsize=(8, 6))
|
107 |
+
results_df.plot(kind='bar', x='Dataset', y=['MRR', 'MAP', 'nDCG@10'], ax=ax)
|
108 |
+
ax.set_title(f"Evaluation Results for {model_name}")
|
109 |
+
ax.set_ylabel("Score")
|
110 |
+
plt.xticks(rotation=0)
|
111 |
+
|
112 |
+
return results_df, fig, sample_outputs
|
113 |
+
|
114 |
+
# Gradio app interface
|
115 |
+
def gradio_app_with_insights(model_name):
|
116 |
+
results_df, chart, samples = evaluate_model_with_insights(model_name)
|
117 |
+
sample_display = []
|
118 |
+
for sample in samples:
|
119 |
+
sample_display.append(f"Query: {sample['Query']}")
|
120 |
+
for doc, score, label in sample["Top 5 Candidates"]:
|
121 |
+
sample_display.append(f" Doc: {doc[:50]}... | Score: {score:.2f} | Relevance: {label}")
|
122 |
+
sample_display.append("\n")
|
123 |
+
return results_df, chart, "\n".join(sample_display)
|
124 |
+
|
125 |
+
interface = gr.Interface(
|
126 |
+
fn=gradio_app_with_insights,
|
127 |
+
inputs=gr.Textbox(label="Enter Model Name", placeholder="e.g., NAMAA-Space/GATE-Reranker-V1"),
|
128 |
+
outputs=[
|
129 |
+
gr.Dataframe(label="Evaluation Results"),
|
130 |
+
gr.Plot(label="Evaluation Metrics Chart"),
|
131 |
+
gr.Textbox(label="Sample Reranking Insights", lines=15)
|
132 |
+
],
|
133 |
+
title="Arabic Reranking Model Evaluation and Insights",
|
134 |
+
description=(
|
135 |
+
"This app evaluates Arabic reranking models on two datasets:\n"
|
136 |
+
"1. **Relevance Labels Dataset**\n"
|
137 |
+
"2. **Positive-Negatives Dataset**\n\n"
|
138 |
+
"### Metrics Used:\n"
|
139 |
+
"- **MRR (Mean Reciprocal Rank)**: Measures how quickly the first relevant document appears.\n"
|
140 |
+
"- **MAP (Mean Average Precision)**: Reflects ranking quality across all relevant documents.\n"
|
141 |
+
"- **nDCG@10 (Normalized Discounted Cumulative Gain)**: Focuses on the ranking of relevant documents in the top-10.\n\n"
|
142 |
+
"Input a model name to evaluate its performance, view metrics, and examine sample reranking results."
|
143 |
+
)
|
144 |
+
)
|
145 |
+
|
146 |
+
interface.launch(debug=True)
|