app.py

import gradio as gr
import numpy as np
import torch
import torch.nn as nn

from utils.data_processing import get_data
from utils.utils import NeighborFinder
from modules.tgn import TGN

import logging
import random
import os 

logging.basicConfig(filename="debug.log", level=logging.INFO)

def download_log():
    return "debug.log" if os.path.exists("debug.log") else None


# Load data
node_features, edge_features, full_data, train_data, val_data, test_data, \
    new_node_val_data, new_node_test_data = get_data('wikipedia')

try:
    logging.info(f"Full data dict: {vars(full_data)}")
except TypeError:
    logging.info("vars(full_data) failed. Trying manual extraction...")
    logging.info(f"Sources: {full_data.sources[:10]}")
    logging.info(f"Timestamps: {full_data.timestamps[:10]}")


# Extract edge information
edge_sources = full_data.sources
edge_destinations = full_data.destinations
edge_timestamps = full_data.timestamps
edge_idxs = full_data.edge_idxs

# Construct adjacency list
all_nodes = list(full_data.sources) + list(full_data.destinations)
n_nodes = max(all_nodes) + 1  # assuming node IDs are 0-indexed

adj_list = [[] for _ in range(n_nodes)]
for src, dst, ts, eidx in zip(edge_sources, edge_destinations, edge_timestamps, edge_idxs):
    adj_list[src].append((dst, eidx, ts))
    adj_list[dst].append((src, eidx, ts))  # If the graph is undirected

# Initialize the neighbor finder
nf = NeighborFinder(adj_list=adj_list, uniform=False)    

# Initialize model (must match training args)
tgn = TGN(
    neighbor_finder=nf,  
    node_features=node_features,
    edge_features=edge_features,
    device='cpu',
    n_layers=1,
    n_heads=2,
    dropout=0.1,
    use_memory=True,
    message_dimension=100,
    memory_dimension=172,
    memory_update_at_start=True,  # Should match training
    embedding_module_type='graph_attention',
    message_function='identity',
    aggregator_type='last',
    memory_updater_type='gru',
    n_neighbors=10,
    use_destination_embedding_in_message=False,
    use_source_embedding_in_message=False,
    dyrep=False
)

# Initialize memory buffers with the correct dimensions
new_num_nodes = 9228  # Current number of nodes
memory_dim = 172      # Your defined memory dimension

# Manually reinitialize memory buffers to the correct size as Parameters
tgn.memory.memory = nn.Parameter(torch.randn(new_num_nodes, memory_dim))
tgn.memory.last_update = nn.Parameter(torch.zeros(new_num_nodes))

tgn.memory_updater.memory.memory = nn.Parameter(torch.randn(new_num_nodes, memory_dim))
tgn.memory_updater.memory.last_update = nn.Parameter(torch.zeros(new_num_nodes))

# Load the state_dict from the pretrained model
state_dict = torch.load("tgn-attn-wikipedia.pth", map_location="cpu")

# Remove any conflicting memory-related parameters
state_dict.pop("memory.memory", None)
state_dict.pop("memory.last_update", None)
state_dict.pop("memory_updater.memory.memory", None)
state_dict.pop("memory_updater.memory.last_update", None)

# Now load the state_dict into the model (strict=False to ignore the missing memory buffers)
tgn.load_state_dict(state_dict, strict=False)

tgn.eval()

def predict_old(u_id, i_id, timestamp):
    u = np.array([int(u_id)])
    i = np.array([int(i_id)])
    ts = np.array([float(timestamp)])
    #prob = tgn.predict_edge_probabilities(u, i, ts)
    prob = tgn.compute_edge_probabilities(u, i, ts)
    return f"Predicted interaction probability: {prob[0]:.4f}"

def get_random_negative_node(source_node, destination_node, total_nodes):
    candidates = list(range(total_nodes))
    candidates.remove(destination_node)
    if source_node in candidates:
        candidates.remove(source_node)  # optional
    return random.choice(candidates)

   
def predict(u_id, i_id, timestamp, mem):
    # Before prediction
    if mem:
        tgn.memory.__init_memory__()  # Re-initialize memory

    # Then run prediction
    u_array = np.array([int(u_id)])  # List of source nodes
    i_array = np.array([int(i_id)])  # List of destination nodes
    ts_array = np.array([int(timestamp)])  # List of timestamps
    

    # Look for matching entry in full_data to get the edge_idx
    edge_idx = None
    for idx in range(len(full_data.sources)):
        if (full_data.sources[idx] == u_id and
            full_data.destinations[idx] == i_id and
            full_data.timestamps[idx] == timestamp):
            edge_idx = idx
            break

    # Edge index for the specific source-destination pair
    edge_idx_array = np.array([0])

    EDGE_FEAT_DIM = 172  # Or detect dynamically if needed
    
    if edge_idx is not None:
        edge_idx_array = np.array([int(edge_idx)])  # List of edge indices
        logging.info(f"Edge features shape: {edge_idx_array}")

    
    # Negative node sampling: choose random negative nodes for this example (should be handled in your dataset)
    total_nodes = len(full_data.unique_nodes)  # or whatever gives your total node count
    random_negative_node = get_random_negative_node(u_id, i_id, total_nodes)
    negative_nodes_array = np.array([random_negative_node])  # List of negative nodes

    # Call compute_edge_probabilities
    # You can pass edge_idxs or dummy edge features here depending on how your TGN expects it
    try:
        positive_probs, negative_probs = tgn.compute_edge_probabilities(
            u_array,
            i_array,
            negative_nodes_array,
            ts_array,
            edge_idx_array
        )
    except RuntimeError as e:
        # fallback in case edge_idxs=None causes internal error
        dummy_edge_features = np.zeros((1, EDGE_FEAT_DIM), dtype=np.float32)
        positive_probs, negative_probs = tgn.compute_edge_probabilities(
            u_array,
            i_array,
            negative_nodes_array,
            ts_array,
            dummy_edge_features  # Your TGN must support this case
        )

    # Now positive_probs is a tensor — use it to get top-k values
    topk = torch.topk(positive_probs, k=min(5, positive_probs.shape[0]))
    top_values = [v.item() for v in topk.values]  # Flatten and convert each tensor to float
    
    # Format the top-k values
    #formatted_result = "\n".join([f"{i+1}. {val:.4f}" for i, val in enumerate(top_values)])
    formatted_result = "\n".join([f"{val:.4f}" for i, val in enumerate(top_values)])
    
    #return f"Top {len(top_values)} predicted interaction probabilities:\n{formatted_result}"
    return f"Predicted interaction probability: {formatted_result}"


with gr.Blocks() as demo:
    gr.Markdown("## 🧠 TGN Playground (Wikipedia)")
    gr.Markdown("Enter node IDs and timestamp to predict future interaction probability using Temporal Graph Networks (TGN).")

    with gr.Row():
        src = gr.Number(label="Source Node ID")
        dst = gr.Number(label="Destination Node ID")
        ts  = gr.Number(label="Timestamp")
        mem = gr.Checkbox(label="Reinitialize memory, don't update on prediction!")  # 👈 your checkbox

    predict_btn = gr.Button("Predict")
    result = gr.Textbox(label="Prediction Output")

    predict_btn.click(fn=predict, inputs=[src, dst, ts, mem], outputs=result)

    gr.Markdown("### Download Log File")
    log_file = gr.File(label="Log File", interactive=False)
    download_btn = gr.Button("Download Logs")

    download_btn.click(fn=download_log, inputs=[], outputs=log_file)


demo.launch()