Aranwer's picture
Create app.py
c4ca5b1 verified
raw
history blame
17.4 kB
import streamlit as st
import javalang
import torch
import torch.nn as nn
import torch.nn.functional as F
import re
import numpy as np
import networkx as nx
from transformers import AutoTokenizer, AutoModel
from torch_geometric.data import Data
from torch_geometric.nn import GCNConv
import warnings
import pandas as pd
import zipfile
import os
from collections import defaultdict
# Set up page config
st.set_page_config(
page_title="Advanced Java Code Clone Detector (IJaDataset 2.1)",
page_icon="πŸ”",
layout="wide"
)
# Suppress warnings
warnings.filterwarnings("ignore")
# Constants
MODEL_NAME = "microsoft/codebert-base"
MAX_LENGTH = 512
DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
DATASET_PATH = "archive (1).zip" # Update this path if needed
# Initialize models with caching
@st.cache_resource
def load_models():
try:
# Load CodeBERT for semantic analysis
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
code_model = AutoModel.from_pretrained(MODEL_NAME).to(DEVICE)
# Initialize RNN model
class RNNModel(nn.Module):
def __init__(self, input_size, hidden_size, num_layers):
super(RNNModel, self).__init__()
self.hidden_size = hidden_size
self.num_layers = num_layers
self.rnn = nn.RNN(input_size, hidden_size, num_layers, batch_first=True)
self.fc = nn.Linear(hidden_size, 1)
def forward(self, x):
h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(DEVICE)
out, _ = self.rnn(x, h0)
out = self.fc(out[:, -1, :])
return out
rnn_model = RNNModel(input_size=768, hidden_size=256, num_layers=2).to(DEVICE)
# Initialize GNN model
class GNNModel(nn.Module):
def __init__(self, node_features):
super(GNNModel, self).__init__()
self.conv1 = GCNConv(node_features, 128)
self.conv2 = GCNConv(128, 64)
self.fc = nn.Linear(64, 1)
def forward(self, data):
x, edge_index = data.x, data.edge_index
x = F.relu(self.conv1(x, edge_index))
x = F.dropout(x, training=self.training)
x = self.conv2(x, edge_index)
x = self.fc(x)
return torch.sigmoid(x.mean())
gnn_model = GNNModel(node_features=128).to(DEVICE)
return tokenizer, code_model, rnn_model, gnn_model
except Exception as e:
st.error(f"Failed to load models: {str(e)}")
return None, None, None, None
@st.cache_resource
def load_dataset():
try:
# Extract dataset if needed
if not os.path.exists("Diverse_100K_Dataset"):
with zipfile.ZipFile(DATASET_PATH, 'r') as zip_ref:
zip_ref.extractall(".")
# Load sample pairs (modify this based on your dataset structure)
clone_pairs = []
base_path = "Subject_CloneTypes_Directories"
# Load pairs from all clone types
for clone_type in ["Clone_Type1", "Clone_Type2", "Clone_Type3 - ST", "Clone_Type4"]:
type_path = os.path.join(base_path, clone_type)
if os.path.exists(type_path):
for root, _, files in os.walk(type_path):
if files:
# Take first two files as a pair
if len(files) >= 2:
with open(os.path.join(root, files[0]), 'r', encoding='utf-8') as f1:
code1 = f1.read()
with open(os.path.join(root, files[1]), 'r', encoding='utf-8') as f2:
code2 = f2.read()
clone_pairs.append({
"type": clone_type,
"code1": code1,
"code2": code2
})
break # Just take one pair per type for demo
return clone_pairs[:10] # Return first 10 pairs for demo
except Exception as e:
st.error(f"Error loading dataset: {str(e)}")
return []
tokenizer, code_model, rnn_model, gnn_model = load_models()
dataset_pairs = load_dataset()
# AST Processing Functions
def parse_ast(code):
try:
tokens = javalang.tokenizer.tokenize(code)
parser = javalang.parser.Parser(tokens)
tree = parser.parse()
return tree
except Exception as e:
st.warning(f"AST parsing error: {str(e)}")
return None
def build_ast_graph(ast_tree):
if not ast_tree:
return None
G = nx.DiGraph()
node_id = 0
node_map = {}
def traverse(node, parent_id=None):
nonlocal node_id
current_id = node_id
node_label = str(type(node).__name__)
node_map[current_id] = {'type': node_label, 'node': node}
G.add_node(current_id, type=node_label)
if parent_id is not None:
G.add_edge(parent_id, current_id)
node_id += 1
for child in node.children:
if isinstance(child, javalang.ast.Node):
traverse(child, current_id)
elif isinstance(child, (list, tuple)):
for item in child:
if isinstance(item, javalang.ast.Node):
traverse(item, current_id)
traverse(ast_tree)
return G, node_map
def ast_to_pyg_data(ast_graph):
if not ast_graph:
return None
# Convert AST to PyTorch Geometric Data format
node_features = []
node_types = []
for node in ast_graph.nodes():
node_type = ast_graph.nodes[node]['type']
node_types.append(node_type)
# Simple one-hot encoding of node types (in practice, use better encoding)
feature = [0] * 50 # Assuming max 50 node types
feature[hash(node_type) % 50] = 1
node_features.append(feature)
# Convert networkx graph to edge_index format
edge_index = list(ast_graph.edges())
if not edge_index:
# Add self-loop if no edges
edge_index = [(0, 0)]
edge_index = torch.tensor(edge_index, dtype=torch.long).t().contiguous()
x = torch.tensor(node_features, dtype=torch.float)
return Data(x=x, edge_index=edge_index)
# Normalization function
def normalize_code(code):
try:
code = re.sub(r'//.*', '', code) # Remove single-line comments
code = re.sub(r'/\*.*?\*/', '', code, flags=re.DOTALL) # Multi-line comments
code = re.sub(r'\s+', ' ', code).strip() # Normalize whitespace
return code
except Exception:
return code
# Feature extraction functions
def get_lexical_features(code):
"""Extract lexical features (for Type-1 and Type-2 clones)"""
normalized = normalize_code(code)
tokens = re.findall(r'\b\w+\b', normalized)
return {
'token_count': len(tokens),
'unique_tokens': len(set(tokens)),
'avg_token_length': np.mean([len(t) for t in tokens]) if tokens else 0
}
def get_syntactic_features(ast_tree):
"""Extract syntactic features (for Type-3 clones)"""
if not ast_tree:
return {}
# Count different node types in AST
node_counts = defaultdict(int)
def count_nodes(node):
node_counts[type(node).__name__] += 1
for child in node.children:
if isinstance(child, javalang.ast.Node):
count_nodes(child)
elif isinstance(child, (list, tuple)):
for item in child:
if isinstance(item, javalang.ast.Node):
count_nodes(item)
count_nodes(ast_tree)
return dict(node_counts)
def get_semantic_features(code):
"""Extract semantic features (for Type-4 clones)"""
embedding = get_embedding(code)
return embedding.cpu().numpy().flatten() if embedding is not None else None
# Embedding generation
def get_embedding(code):
try:
code = normalize_code(code)
inputs = tokenizer(
code,
return_tensors="pt",
truncation=True,
max_length=MAX_LENGTH,
padding='max_length'
).to(DEVICE)
with torch.no_grad():
outputs = code_model(**inputs)
return outputs.last_hidden_state.mean(dim=1) # Pooled embedding
except Exception as e:
st.error(f"Error processing code: {str(e)}")
return None
# Clone detection models
def rnn_similarity(emb1, emb2):
"""Calculate similarity using RNN model"""
if emb1 is None or emb2 is None:
return None
# Prepare input for RNN (sequence of embeddings)
combined = torch.cat([emb1.unsqueeze(0), emb2.unsqueeze(0)], dim=0)
with torch.no_grad():
similarity = rnn_model(combined.permute(1, 0, 2))
return torch.sigmoid(similarity).item()
def gnn_similarity(ast1, ast2):
"""Calculate similarity using GNN model"""
if ast1 is None or ast2 is None:
return None
data1 = ast_to_pyg_data(ast1)
data2 = ast_to_pyg_data(ast2)
if data1 is None or data2 is None:
return None
# Move data to device
data1 = data1.to(DEVICE)
data2 = data2.to(DEVICE)
with torch.no_grad():
sim1 = gnn_model(data1)
sim2 = gnn_model(data2)
return F.cosine_similarity(sim1, sim2).item()
def hybrid_similarity(code1, code2):
"""Combined similarity score using all models"""
# Get embeddings
emb1 = get_embedding(code1)
emb2 = get_embedding(code2)
# Parse ASTs
ast_tree1 = parse_ast(code1)
ast_tree2 = parse_ast(code2)
ast_graph1 = build_ast_graph(ast_tree1) if ast_tree1 else None
ast_graph2 = build_ast_graph(ast_tree2) if ast_tree2 else None
# Calculate individual similarities
codebert_sim = F.cosine_similarity(emb1, emb2).item() if emb1 is not None and emb2 is not None else 0
rnn_sim = rnn_similarity(emb1, emb2) if emb1 is not None and emb2 is not None else 0
gnn_sim = gnn_similarity(ast_graph1[0] if ast_graph1 else None,
ast_graph2[0] if ast_graph2 else None) or 0
# Combine with weights (can be tuned)
weights = {
'codebert': 0.4,
'rnn': 0.3,
'gnn': 0.3
}
combined = (weights['codebert'] * codebert_sim +
weights['rnn'] * rnn_sim +
weights['gnn'] * gnn_sim)
return {
'combined': combined,
'codebert': codebert_sim,
'rnn': rnn_sim,
'gnn': gnn_sim
}
# Comparison function
def compare_code(code1, code2):
if not code1 or not code2:
return None
with st.spinner('Analyzing code with multiple techniques...'):
# Get lexical features
lex1 = get_lexical_features(code1)
lex2 = get_lexical_features(code2)
# Get AST trees
ast_tree1 = parse_ast(code1)
ast_tree2 = parse_ast(code2)
# Get syntactic features
syn1 = get_syntactic_features(ast_tree1)
syn2 = get_syntactic_features(ast_tree2)
# Get semantic features
sem1 = get_semantic_features(code1)
sem2 = get_semantic_features(code2)
# Calculate hybrid similarity
similarities = hybrid_similarity(code1, code2)
return {
'similarities': similarities,
'lexical_features': (lex1, lex2),
'syntactic_features': (syn1, syn2),
'ast_trees': (ast_tree1, ast_tree2)
}
# UI Elements
st.title("πŸ” Advanced Java Code Clone Detector (IJaDataset 2.1)")
st.markdown("""
Detect all types of code clones (Type 1-4) using hybrid approach with:
- **CodeBERT** for semantic analysis
- **RNN** for sequence modeling
- **GNN** for AST structural analysis
""")
# Dataset selector
selected_pair = None
if dataset_pairs:
pair_options = {f"{i+1}: {pair['type']}": pair for i, pair in enumerate(dataset_pairs)}
selected_option = st.selectbox("Select a preloaded example pair:", list(pair_options.keys()))
selected_pair = pair_options[selected_option]
# Layout
col1, col2 = st.columns(2)
with col1:
code1 = st.text_area(
"First Java Code",
height=300,
value=selected_pair["code1"] if selected_pair else "",
help="Enter the first Java code snippet"
)
with col2:
code2 = st.text_area(
"Second Java Code",
height=300,
value=selected_pair["code2"] if selected_pair else "",
help="Enter the second Java code snippet"
)
# Threshold sliders
st.subheader("Detection Thresholds")
col1, col2, col3 = st.columns(3)
with col1:
threshold_type12 = st.slider(
"Type 1/2 Threshold",
min_value=0.5,
max_value=1.0,
value=0.9,
step=0.01,
help="Threshold for exact/syntactic clones"
)
with col2:
threshold_type3 = st.slider(
"Type 3 Threshold",
min_value=0.5,
max_value=1.0,
value=0.8,
step=0.01,
help="Threshold for near-miss clones"
)
with col3:
threshold_type4 = st.slider(
"Type 4 Threshold",
min_value=0.5,
max_value=1.0,
value=0.7,
step=0.01,
help="Threshold for semantic clones"
)
# Compare button
if st.button("Compare Code", type="primary"):
if tokenizer is None or code_model is None or rnn_model is None or gnn_model is None:
st.error("Models failed to load. Please check the logs.")
else:
result = compare_code(code1, code2)
if result is not None:
similarities = result['similarities']
lex1, lex2 = result['lexical_features']
syn1, syn2 = result['syntactic_features']
ast_tree1, ast_tree2 = result['ast_trees']
# Display results
st.subheader("Detection Results")
# Determine clone type
combined_sim = similarities['combined']
clone_type = "No Clone"
if combined_sim >= threshold_type12:
clone_type = "Type 1/2 Clone (Exact/Near-Exact)"
elif combined_sim >= threshold_type3:
clone_type = "Type 3 Clone (Near-Miss)"
elif combined_sim >= threshold_type4:
clone_type = "Type 4 Clone (Semantic)"
# Main metrics
col1, col2, col3 = st.columns(3)
with col1:
st.metric("Combined Similarity", f"{combined_sim:.3f}")
with col2:
st.metric("Detected Clone Type", clone_type)
with col3:
st.metric("CodeBERT Similarity", f"{similarities['codebert']:.3f}")
# Detailed metrics
with st.expander("Detailed Similarity Scores"):
cols = st.columns(3)
with cols[0]:
st.metric("RNN Similarity", f"{similarities['rnn']:.3f}")
with cols[1]:
st.metric("GNN Similarity", f"{similarities['gnn']:.3f}")
with cols[2]:
st.metric("Lexical Similarity",
f"{sum(lex1[k] == lex2[k] for k in lex1)/max(len(lex1),1):.2f}")
# Feature comparison
with st.expander("Feature Analysis"):
st.subheader("Lexical Features")
lex_df = pd.DataFrame([lex1, lex2], index=["Code 1", "Code 2"])
st.dataframe(lex_df)
st.subheader("Syntactic Features (AST Node Counts)")
syn_df = pd.DataFrame([syn1, syn2], index=["Code 1", "Code 2"]).fillna(0)
st.dataframe(syn_df)
# AST Visualization
if ast_tree1 and ast_tree2:
with st.expander("AST Visualization (First 20 nodes)"):
st.write("AST visualization would be implemented here with graphviz")
# In a real implementation, you would use graphviz to render the ASTs
# st.graphviz_chart(ast_to_graphviz(ast_tree1))
# st.graphviz_chart(ast_to_graphviz(ast_tree2))
# Normalized code view
with st.expander("Show normalized code"):
tab1, tab2 = st.tabs(["First Code", "Second Code"])
with tab1:
st.code(normalize_code(code1))
with tab2:
st.code(normalize_code(code2))
# Footer
st.markdown("---")
st.markdown("""
*Dataset Information*:
- Using IJaDataset 2.1 from Kaggle
- Contains 100K Java files with clone annotations
- Clone types: Type-1, Type-2, Type-3, and Type-4 clones
*Model Architecture*:
- **CodeBERT**: Pre-trained model for semantic analysis
- **RNN**: Processes token sequences for sequential patterns
- **GNN**: Analyzes AST structure for syntactic patterns
- **Hybrid Approach**: Combines all techniques for comprehensive detection
""")