echom / src /streamlit_app.py
bluesketch's picture
Update src/streamlit_app.py
d98928c verified
Raw History Blame Contribute Delete
6.95 kB
import streamlit as st
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.nn import GCNConv
from torch_geometric.utils import add_self_loops
import pandas as pd
import numpy as np
from sklearn.preprocessing import StandardScaler
import time
# ==========================================
# 0. PAGE CONFIG
# ==========================================
st.set_page_config(page_title="GHOST_VIBE", page_icon="👻", layout="centered")
st.markdown("""
<style>
.stApp { background-color: #050505; color: #00ff41; font-family: 'Courier New', monospace; }
.stTextInput > div > div > input { background-color: #111; color: #00ff41; border: 1px solid #333; }
.stButton > button { background-color: #000; border: 1px solid #00ff41; color: #00ff41; }
.stButton > button:hover { background-color: #00ff41; color: #000; }
.user-msg { color: #fff; border-left: 2px solid #00ff41; padding-left: 10px; margin-bottom: 5px; }
.sys-msg { color: #555; font-size: 0.8em; font-style: italic; }
.match-anim { border: 2px solid #ff00ff; padding: 20px; text-align: center; animation: pulse 2s infinite; margin-top: 20px; }
@keyframes pulse { 0% { box-shadow: 0 0 0 0 rgba(255, 0, 255, 0.4); } 70% { box-shadow: 0 0 0 10px rgba(255, 0, 255, 0); } 100% { box-shadow: 0 0 0 0 rgba(255, 0, 255, 0); } }
</style>
""", unsafe_allow_html=True)
# ==========================================
# 1. MODEL DEFINITION
# ==========================================
class MotifConv(nn.Module):
def __init__(self, in_channels, out_channels, num_motifs=2, act=True):
super().__init__()
self.convs = nn.ModuleList([GCNConv(in_channels, out_channels) for _ in range(num_motifs)])
self.bn = nn.BatchNorm1d(out_channels)
self.act = act
def forward(self, x, motif_graphs):
outs = []
for i, conv in enumerate(self.convs):
if i < len(motif_graphs) and motif_graphs[i].size(1) > 0:
out = conv(x, motif_graphs[i])
outs.append(out)
else:
outs.append(torch.zeros(x.size(0), conv.out_channels).to(x.device))
out = self.bn(torch.stack(outs).sum(dim=0))
return F.relu(out) if self.act else out
class VibeMatchMGNN(nn.Module):
def __init__(self, in_dim, hidden_dim, out_dim):
super().__init__()
self.mgnn1 = MotifConv(in_dim, hidden_dim, act=True)
self.mgnn2 = MotifConv(hidden_dim, out_dim, act=False)
def forward(self, x, motif_graphs):
h = self.mgnn1(x, motif_graphs)
h = self.mgnn2(h, motif_graphs)
return h
# ==========================================
# 2. GLOBAL BRAIN (SHARED MEMORY)
# ==========================================
@st.cache_resource
class GlobalBrain:
def __init__(self):
self.users = []
self.vectors = []
self.model = None
self.scaler = None
self.projector = None
self.load_system()
def load_system(self):
# 1. Load CSVs
try:
df = pd.read_csv('user_features.csv')
raw_x = df.drop(columns=['user_id', 'mapped_id'], errors='ignore').fillna(0)
self.scaler = StandardScaler()
self.scaler.fit(raw_x.values)
except:
self.scaler = StandardScaler()
self.scaler.fit(np.random.rand(10, 387))
# 2. Load Model
self.model = VibeMatchMGNN(64, 64, 32)
try:
self.model.load_state_dict(torch.load('vibematch_model.pth', map_location='cpu'))
except:
print("⚠️ .pth not found. Using random weights.")
# --- THE FIX IS HERE: FORCE EVAL MODE ---
self.model.eval()
# 3. Projector
torch.manual_seed(42)
self.projector = nn.Linear(387, 64)
def add_user(self, user_id, vector_387):
self.users.append(user_id)
self.vectors.append(vector_387)
def find_match(self, target_vec_32, current_user_id):
if not self.users: return None, 0
best_score = -999
best_user = None
for i, u_id in enumerate(self.users):
if u_id == current_user_id: continue
vec_raw = self.vectors[i]
# Project & Inference
with torch.no_grad():
self.model.eval() # FORCE EVAL AGAIN
vec_scaled = self.scaler.transform([vec_raw])
vec_t = torch.tensor(vec_scaled, dtype=torch.float32)
vec_64 = self.projector(vec_t)
dummy = [torch.empty((2,0), dtype=torch.long)]*2
cand_vec_32 = self.model(vec_64, dummy)[0]
score = torch.dot(target_vec_32, cand_vec_32).item()
if score > best_score:
best_score = score
best_user = u_id
return best_user, best_score
# Initialize the shared brain
brain = GlobalBrain()
# ==========================================
# 3. TEXT PROCESSING
# ==========================================
def process_text(text):
vec = np.random.rand(387)
vec[0] = len(text.split())
vec[2] = sum(1 for c in text if c.isupper())
# Run Inference
with torch.no_grad():
brain.model.eval() # FORCE EVAL AGAIN
vec_scaled = brain.scaler.transform([vec])
vec_t = torch.tensor(vec_scaled, dtype=torch.float32)
vec_64 = brain.projector(vec_t)
dummy = [torch.empty((2,0), dtype=torch.long)]*2
vec_32 = brain.model(vec_64, dummy)[0]
return vec, vec_32
# ==========================================
# 4. UI LOGIC
# ==========================================
st.title("GHOST_VIBE // LIVE NODE")
if "uid" not in st.session_state:
st.session_state.uid = f"Guest_{int(time.time())}"
st.caption(f"CONNECTED AS: {st.session_state.uid}")
if prompt := st.chat_input("Broadcast to the network..."):
st.markdown(f"<div class='user-msg'>{prompt}</div>", unsafe_allow_html=True)
# 1. Process
raw_vec, vibe_vec = process_text(prompt)
# 2. Save
brain.add_user(st.session_state.uid, raw_vec)
st.markdown(f"<div class='sys-msg'>Your vibe vector has been added to the live network.</div>", unsafe_allow_html=True)
# 3. Match
match_id, score = brain.find_match(vibe_vec, st.session_state.uid)
if match_id:
st.markdown(f"""
<div class='match-anim'>
<h2 style='color:#fff'>MATCH FOUND</h2>
<h1 style='color:#00ff41'>{match_id}</h1>
<p>COMPATIBILITY: {int(score*100)}%</p>
</div>
""", unsafe_allow_html=True)
else:
st.markdown(f"<div class='sys-msg' style='color:orange'>No matches yet. You are the first node in this cluster. Waiting for others...</div>", unsafe_allow_html=True)
# Show Stats
st.sidebar.title("NETWORK STATS")
st.sidebar.metric("Live Users", len(brain.users))