app.py

import gradio as gr
import torch
import pandas as pd
import numpy as np
from torch_geometric.data import Data
from torch_geometric.nn import GATConv
from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity

# Define the GATConv model architecture
class ModeratelySimplifiedGATConvModel(torch.nn.Module):
    def __init__(self, in_channels, hidden_channels, out_channels):
        super().__init__()
        self.conv1 = GATConv(in_channels, hidden_channels, heads=2)
        self.dropout1 = torch.nn.Dropout(0.45)
        self.conv2 = GATConv(hidden_channels * 2, out_channels, heads=1)

    def forward(self, x, edge_index, edge_attr=None):
        x = self.conv1(x, edge_index, edge_attr)
        x = torch.relu(x)
        x = self.dropout1(x)
        x = self.conv2(x, edge_index, edge_attr)
        return x

# Load the dataset and the GATConv model
data = torch.load("graph_data.pt", map_location=torch.device("cpu"))

# Correct the state dictionary's key names
original_state_dict = torch.load("graph_model.pth", map_location=torch.device("cpu"))
corrected_state_dict = {}
for key, value in original_state_dict.items():
    if "lin.weight" in key:
        corrected_state_dict[key.replace("lin.weight", "lin_src.weight")] = value
        corrected_state_dict[key.replace("lin.weight", "lin_dst.weight")] = value
    else:
        corrected_state_dict[key] = value

# Initialize the GATConv model with the corrected state dictionary
gatconv_model = ModeratelySimplifiedGATConvModel(
    in_channels=data.x.shape[1], hidden_channels=32, out_channels=768
)
gatconv_model.load_state_dict(corrected_state_dict)

# Load the BERT-based sentence transformer model
model_bert = SentenceTransformer("all-mpnet-base-v2")

# Ensure the DataFrame is loaded properly
df = pd.read_feather("EmbeddedCombined.feather")

# Generate GNN-based embeddings
with torch.no_grad():
    all_video_embeddings = gatconv_model(data.x, data.edge_index, data.edge_attr).cpu()

# Function to find the most similar video and recommend the top 10 based on GNN embeddings
def get_similar_and_recommend(input_text):
    # Find the most similar video based on input text
    embeddings_matrix = np.array(df["embeddings"].tolist())
    input_embedding = model_bert.encode([input_text])[0]
    similarities = cosine_similarity([input_embedding], embeddings_matrix)[0]
    most_similar_index = np.argmax(similarities)

    most_similar_video = {
        "title": df["title"].iloc[most_similar_index],
        "description": df["description"].iloc[most_similar_index],
        "similarity_score": similarities[most_similar_index],
    }

    # Recommend the top 10 videos based on GNN embeddings and dot product
    def recommend_next_10_videos(given_video_index, all_video_embeddings):
        dot_products = [
            torch.dot(all_video_embeddings[given_video_index].cpu(), all_video_embeddings[i].cpu())
            for i in range(all_video_embeddings.shape[0])
        ]
        dot_products[given_video_index] = -float("inf")

        top_10_indices = np.argsort(dot_products)[::-1][:10]
        recommendations = [df["title"].iloc[idx] for idx in top_10_indices]
        return recommendations

    top_10_recommendations = recommend_next_10_videos(
        most_similar_index, all_video_embeddings
    )

    return (
        most_similar_video["title"],
        most_similar_video["description"],
        most_similar_video["similarity_score"],
        top_10_recommendations,
    )

# Update the Gradio interface to fix the output type
interface = gr.Interface(
    fn=get_similar_and_recommend,
    inputs=gr.components.Textbox(label="Enter Text to Find Most Similar Video"),
    outputs=[
        gr.components.Textbox(label="Video Title"),
        gr.components.Textbox(label="Video Description"),
        gr.components.Textbox(label="Similarity Score"),
        gr.components.Textbox(label="Top 10 Recommended Videos", lines=10),  # Handle a list
    ],
    title="Video Recommendation System with GNN-based Recommendations",
    description="Enter text to find the most similar video and get the top 10 recommended videos based on dot product and GNN embeddings.",
)

# Launch the Gradio interface
interface.launch()