Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,124 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
import torch
|
3 |
+
import pandas as pd
|
4 |
+
import numpy as np
|
5 |
+
from torch_geometric.data import Data
|
6 |
+
from torch_geometric.nn import GATConv
|
7 |
+
from sentence_transformers import SentenceTransformer
|
8 |
+
from sklearn.metrics.pairwise import cosine_similarity
|
9 |
+
|
10 |
+
# Define the GATConv model architecture
|
11 |
+
class ModeratelySimplifiedGATConvModel(torch.nn.Module):
|
12 |
+
def __init__(self, in_channels, hidden_channels, out_channels):
|
13 |
+
super().__init__()
|
14 |
+
self.conv1 = GATConv(in_channels, hidden_channels, heads=2)
|
15 |
+
self.dropout1 = torch.nn.Dropout(0.45)
|
16 |
+
self.conv2 = GATConv(hidden_channels * 2, out_channels, heads=1)
|
17 |
+
|
18 |
+
def forward(self, x, edge_index, edge_attr=None):
|
19 |
+
x = self.conv1(x, edge_index, edge_attr)
|
20 |
+
x = torch.relu(x)
|
21 |
+
x = self.dropout1(x)
|
22 |
+
x = self.conv2(x, edge_index, edge_attr)
|
23 |
+
return x
|
24 |
+
|
25 |
+
# Load the dataset and the GATConv model
|
26 |
+
data = torch.load("graph_data.pt", map_location=torch.device("cpu"))
|
27 |
+
|
28 |
+
# Correct the state dictionary's key names
|
29 |
+
original_state_dict = torch.load("graph_model.pth", map_location=torch.device("cpu"))
|
30 |
+
corrected_state_dict = {}
|
31 |
+
for key, value in original_state_dict.items():
|
32 |
+
if "lin.weight" in key:
|
33 |
+
corrected_state_dict[key.replace("lin.weight", "lin_src.weight")] = value
|
34 |
+
corrected_state_dict[key.replace("lin.weight", "lin_dst.weight")] = value
|
35 |
+
else:
|
36 |
+
corrected_state_dict[key] = value
|
37 |
+
|
38 |
+
# Initialize the GATConv model with the corrected state dictionary
|
39 |
+
gatconv_model = ModeratelySimplifiedGATConvModel(
|
40 |
+
in_channels=data.x.shape[1], hidden_channels=32, out_channels=768
|
41 |
+
)
|
42 |
+
gatconv_model.load_state_dict(corrected_state_dict)
|
43 |
+
|
44 |
+
# Load the BERT-based sentence transformer model
|
45 |
+
model_bert = SentenceTransformer("all-mpnet-base-v2")
|
46 |
+
|
47 |
+
# Ensure the DataFrame is loaded properly
|
48 |
+
try:
|
49 |
+
df = pd.read_json("combined_data.json.gz", orient='records', lines=True, compression='gzip')
|
50 |
+
except Exception as e:
|
51 |
+
st.error(f"Error reading JSON file: {e}")
|
52 |
+
|
53 |
+
# Generate GNN-based embeddings
|
54 |
+
with torch.no_grad():
|
55 |
+
all_video_embeddings = gatconv_model(data.x, data.edge_index, data.edge_attr).cpu()
|
56 |
+
|
57 |
+
# Function to find the most similar video and recommend the top 10 based on GNN embeddings
|
58 |
+
def get_similar_and_recommend(input_text):
|
59 |
+
# Find the most similar video based on cosine similarity
|
60 |
+
embeddings_matrix = np.array(df["embeddings"].tolist())
|
61 |
+
input_embedding = model_bert.encode([input_text])[0]
|
62 |
+
similarities = cosine_similarity([input_embedding], embeddings_matrix)[0]
|
63 |
+
|
64 |
+
most_similar_index = np.argmax(similarities) # Find the most similar video
|
65 |
+
|
66 |
+
# Get all features of the most similar video
|
67 |
+
most_similar_video_features = df.iloc[most_similar_index].to_dict()
|
68 |
+
|
69 |
+
# Clean up certain fields
|
70 |
+
if "text_for_embedding" in most_similar_video_features:
|
71 |
+
del most_similar_video_features["text_for_embedding"]
|
72 |
+
if "embeddings" in most_similar_video_features:
|
73 |
+
del most_similar_video_features["embeddings"]
|
74 |
+
|
75 |
+
# Recommend the top 10 videos based on GNN embeddings
|
76 |
+
def recommend_top_10(given_video_index, all_video_embeddings):
|
77 |
+
dot_products = [
|
78 |
+
torch.dot(all_video_embeddings[given_video_index], all_video_embeddings[i])
|
79 |
+
for i in range(all_video_embeddings.shape[0])
|
80 |
+
]
|
81 |
+
dot_products[given_video_index] = -float("inf") # Exclude the most similar video
|
82 |
+
|
83 |
+
top_10_indices = np.argsort(dot_products)[::-1][:10]
|
84 |
+
return [df.iloc[idx].to_dict() for idx in top_10_indices]
|
85 |
+
|
86 |
+
top_10_recommended_videos_features = recommend_top_10(most_similar_index, all_video_embeddings)
|
87 |
+
|
88 |
+
# Apply search context to determine weights for GNN results
|
89 |
+
user_keywords = input_text.split() # Create a list of keywords from user input
|
90 |
+
video_weights = []
|
91 |
+
weight = 1.0 # Initial weight factor
|
92 |
+
|
93 |
+
for keyword in user_keywords:
|
94 |
+
if keyword.lower() in df["title"].str.lower().tolist(): # Check for matching keywords
|
95 |
+
weight += 0.1 # Increase weight for matching keyword
|
96 |
+
|
97 |
+
# Calculate the weight for each GNN output
|
98 |
+
video_weights = [weight] * len(top_10_recommended_videos_features)
|
99 |
+
|
100 |
+
# Clean up certain fields in recommendations
|
101 |
+
for recommended_video in top_10_recommended_videos_features:
|
102 |
+
if "text_for_embedding" in recommended_video:
|
103 |
+
del recommended_video["text_for_embedding"]
|
104 |
+
if "embeddings" in recommended_video:
|
105 |
+
del recommended_video["embeddings"]
|
106 |
+
|
107 |
+
# Create the output JSON with the most similar video, final recommendations, and weights
|
108 |
+
output = {
|
109 |
+
"search_context": {
|
110 |
+
"input_text": input_text, # What the user provided
|
111 |
+
"weights": video_weights, # Weights for each GNN-based recommendation
|
112 |
+
},
|
113 |
+
"most_similar_video": most_similar_video_features,
|
114 |
+
"final_recommendations": top 10 recommended videos with individual weights for each recommendation
|
115 |
+
}
|
116 |
+
|
117 |
+
return output
|
118 |
+
|
119 |
+
# Create a Streamlit text input widget for entering text and retrieve the most similar video and top 10 recommended videos
|
120 |
+
user_input = st.text_input("Enter text to find the most similar video")
|
121 |
+
|
122 |
+
if user_input:
|
123 |
+
recommendations = get_similar_and_recommend(user_input)
|
124 |
+
st.json(recommendations)
|