Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
|
|
| 1 |
import os
|
| 2 |
import json
|
| 3 |
-
import math
|
| 4 |
import time
|
| 5 |
import random
|
| 6 |
import traceback
|
|
@@ -9,12 +9,21 @@ from collections import defaultdict
|
|
| 9 |
from dataclasses import dataclass
|
| 10 |
|
| 11 |
import numpy as np
|
| 12 |
-
import matplotlib.pyplot as plt
|
| 13 |
import gradio as gr
|
| 14 |
import torch
|
| 15 |
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
|
| 16 |
from sklearn.cluster import KMeans
|
| 17 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
# ---------------------------
|
| 19 |
# Config
|
| 20 |
# ---------------------------
|
|
@@ -116,10 +125,14 @@ _HF_MODEL = None
|
|
| 116 |
def load_flan(model_id: str = MODEL_ID_DEFAULT, device: str = DEVICE):
|
| 117 |
global _HF_TOKENIZER, _HF_MODEL
|
| 118 |
if _HF_MODEL is None:
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 123 |
return _HF_TOKENIZER, _HF_MODEL
|
| 124 |
|
| 125 |
PROMPT_TEMPLATE = """INSTRUCTION: Given the profiles and situation, output compact JSON with keys: mu, k, A_lowrank, sigma.
|
|
@@ -153,6 +166,13 @@ def flan_generate_params(prompt: str, model_id: str = MODEL_ID_DEFAULT, device:
|
|
| 153 |
outs = model.generate(**inputs, max_length=max_length, do_sample=False)
|
| 154 |
text = tokenizer.decode(outs[0], skip_special_tokens=True)
|
| 155 |
parsed = _parse_json_loose(text)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 156 |
return parsed
|
| 157 |
except Exception as e:
|
| 158 |
print('[warn] Flan generation failed:', e)
|
|
@@ -170,7 +190,8 @@ def _parse_json_loose(text: str) -> Dict[str,Any]:
|
|
| 170 |
return json.loads(text[s:e+1])
|
| 171 |
except Exception:
|
| 172 |
pass
|
| 173 |
-
|
|
|
|
| 174 |
|
| 175 |
|
| 176 |
def _fallback_theta() -> Dict[str,Any]:
|
|
@@ -221,13 +242,15 @@ TEMPLATE_UTTERANCES = [
|
|
| 221 |
|
| 222 |
def extract_modes_from_paths(paths: np.ndarray, n_modes: int = 3):
|
| 223 |
n_samps, d, steps = paths.shape
|
|
|
|
|
|
|
| 224 |
feats = np.concatenate([paths.mean(axis=2), paths[:,:, -1]], axis=1)
|
| 225 |
-
kmeans = KMeans(n_clusters=
|
| 226 |
modes = []
|
| 227 |
labels = kmeans.labels_
|
| 228 |
for mode in range(kmeans.n_clusters):
|
| 229 |
idx = np.where(labels == mode)[0]
|
| 230 |
-
weight = len(idx)/n_samps
|
| 231 |
mean_path = paths[idx].mean(axis=0) if len(idx)>0 else np.zeros((d, steps))
|
| 232 |
last_state = mean_path[:, -1]
|
| 233 |
modes.append({'mode': mode, 'weight': float(weight), 'last_state': last_state.tolist()})
|
|
@@ -302,7 +325,7 @@ def get_initial_question(session: SessionState):
|
|
| 302 |
return q
|
| 303 |
|
| 304 |
# ---------------------------
|
| 305 |
-
# Logging (local JSONL)
|
| 306 |
# ---------------------------
|
| 307 |
LOG_DIR = 'runs'
|
| 308 |
os.makedirs(LOG_DIR, exist_ok=True)
|
|
@@ -313,8 +336,16 @@ def log_run(entry: Dict[str,Any]):
|
|
| 313 |
try:
|
| 314 |
with open(LOG_PATH, 'a', encoding='utf-8') as f:
|
| 315 |
f.write(json.dumps(entry, ensure_ascii=False) + '\n')
|
| 316 |
-
|
| 317 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 318 |
|
| 319 |
# ---------------------------
|
| 320 |
# Gradio UI callbacks
|
|
@@ -394,7 +425,7 @@ def plot_paths(paths: np.ndarray, labels: List[str] = None):
|
|
| 394 |
ax.plot(t, mean[i], label=(labels[i] if labels else f'dim{i}'))
|
| 395 |
ax.set_ylim(0,1.2)
|
| 396 |
ax.legend(loc='upper right')
|
| 397 |
-
ax.set_xlabel('Normalized time')
|
| 398 |
plt.tight_layout()
|
| 399 |
return fig
|
| 400 |
|
|
@@ -402,7 +433,7 @@ def plot_paths(paths: np.ndarray, labels: List[str] = None):
|
|
| 402 |
# Gradio app layout
|
| 403 |
# ---------------------------
|
| 404 |
with gr.Blocks(title='Flan-SDE Behavioral Profiler') as demo:
|
| 405 |
-
gr.Markdown("# Flan-SDE Behavioral Profiler
|
| 406 |
with gr.Row():
|
| 407 |
with gr.Column(scale=1):
|
| 408 |
start_btn = gr.Button('Start Session')
|
|
@@ -434,7 +465,8 @@ with gr.Blocks(title='Flan-SDE Behavioral Profiler') as demo:
|
|
| 434 |
modes = json.loads(next_q[6:])['modes']
|
| 435 |
out = predict_from_session(SESSIONS[sid2], use_flan=use_flan, model_id=model_id, n_samps=int(n_samps))
|
| 436 |
fig = plot_paths(out['paths'], labels=["Strategy", "Empathy", "Control", "Masking", "RiskAppetite"])
|
| 437 |
-
except Exception:
|
|
|
|
| 438 |
modes = []
|
| 439 |
fig = None
|
| 440 |
return sid2, profdict, '', 'Prediction complete', json.dumps(modes), fig
|
|
@@ -442,4 +474,4 @@ with gr.Blocks(title='Flan-SDE Behavioral Profiler') as demo:
|
|
| 442 |
return sid2, profdict, next_q, '', None, None
|
| 443 |
|
| 444 |
start_btn.click(on_start, outputs=[sid_out, profile_view, qbox, out_console])
|
| 445 |
-
send_btn.click(on_send, inputs=[sid_out, user_in, profile_view, qbox, use_flan_chk, model_in, n_samps_slider], outputs=[sid_out, profile_view, qbox, out_console, modes_json, plot_out])
|
|
|
|
| 1 |
+
|
| 2 |
import os
|
| 3 |
import json
|
|
|
|
| 4 |
import time
|
| 5 |
import random
|
| 6 |
import traceback
|
|
|
|
| 9 |
from dataclasses import dataclass
|
| 10 |
|
| 11 |
import numpy as np
|
|
|
|
| 12 |
import gradio as gr
|
| 13 |
import torch
|
| 14 |
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
|
| 15 |
from sklearn.cluster import KMeans
|
| 16 |
|
| 17 |
+
import matplotlib
|
| 18 |
+
matplotlib.use('Agg') # Use non-GUI backend to avoid GLX dependencies
|
| 19 |
+
import matplotlib.pyplot as plt
|
| 20 |
+
|
| 21 |
+
try:
|
| 22 |
+
from datasets import Dataset
|
| 23 |
+
HAS_DATASETS = True
|
| 24 |
+
except ImportError:
|
| 25 |
+
HAS_DATASETS = False
|
| 26 |
+
|
| 27 |
# ---------------------------
|
| 28 |
# Config
|
| 29 |
# ---------------------------
|
|
|
|
| 125 |
def load_flan(model_id: str = MODEL_ID_DEFAULT, device: str = DEVICE):
|
| 126 |
global _HF_TOKENIZER, _HF_MODEL
|
| 127 |
if _HF_MODEL is None:
|
| 128 |
+
try:
|
| 129 |
+
print(f"[info] Loading Flan model {model_id} on {device}...")
|
| 130 |
+
_HF_TOKENIZER = AutoTokenizer.from_pretrained(model_id)
|
| 131 |
+
_HF_MODEL = AutoModelForSeq2SeqLM.from_pretrained(model_id)
|
| 132 |
+
_HF_MODEL.to(torch.device(device))
|
| 133 |
+
except Exception as e:
|
| 134 |
+
print(f"[error] Failed to load model {model_id}: {e}")
|
| 135 |
+
raise
|
| 136 |
return _HF_TOKENIZER, _HF_MODEL
|
| 137 |
|
| 138 |
PROMPT_TEMPLATE = """INSTRUCTION: Given the profiles and situation, output compact JSON with keys: mu, k, A_lowrank, sigma.
|
|
|
|
| 166 |
outs = model.generate(**inputs, max_length=max_length, do_sample=False)
|
| 167 |
text = tokenizer.decode(outs[0], skip_special_tokens=True)
|
| 168 |
parsed = _parse_json_loose(text)
|
| 169 |
+
# Validate shapes
|
| 170 |
+
if len(parsed.get('mu', [])) != 5 or len(parsed.get('k', [])) != 5 or len(parsed.get('sigma', [])) != 5:
|
| 171 |
+
raise ValueError("Invalid parameter shapes")
|
| 172 |
+
U = parsed['A_lowrank'].get('U', [])
|
| 173 |
+
V = parsed['A_lowrank'].get('V', [])
|
| 174 |
+
if len(U) != 5 or any(len(row) != len(V[0]) for row in U) or len(V) != 5:
|
| 175 |
+
raise ValueError("Invalid A_lowrank shapes")
|
| 176 |
return parsed
|
| 177 |
except Exception as e:
|
| 178 |
print('[warn] Flan generation failed:', e)
|
|
|
|
| 190 |
return json.loads(text[s:e+1])
|
| 191 |
except Exception:
|
| 192 |
pass
|
| 193 |
+
print('[warn] Failed to parse JSON, using fallback')
|
| 194 |
+
return _fallback_theta()
|
| 195 |
|
| 196 |
|
| 197 |
def _fallback_theta() -> Dict[str,Any]:
|
|
|
|
| 242 |
|
| 243 |
def extract_modes_from_paths(paths: np.ndarray, n_modes: int = 3):
|
| 244 |
n_samps, d, steps = paths.shape
|
| 245 |
+
if n_samps < n_modes:
|
| 246 |
+
n_modes = n_samps
|
| 247 |
feats = np.concatenate([paths.mean(axis=2), paths[:,:, -1]], axis=1)
|
| 248 |
+
kmeans = KMeans(n_clusters=n_modes, random_state=0).fit(feats)
|
| 249 |
modes = []
|
| 250 |
labels = kmeans.labels_
|
| 251 |
for mode in range(kmeans.n_clusters):
|
| 252 |
idx = np.where(labels == mode)[0]
|
| 253 |
+
weight = len(idx)/n_samps if n_samps > 0 else 0
|
| 254 |
mean_path = paths[idx].mean(axis=0) if len(idx)>0 else np.zeros((d, steps))
|
| 255 |
last_state = mean_path[:, -1]
|
| 256 |
modes.append({'mode': mode, 'weight': float(weight), 'last_state': last_state.tolist()})
|
|
|
|
| 325 |
return q
|
| 326 |
|
| 327 |
# ---------------------------
|
| 328 |
+
# Logging (local JSONL + optional HF Hub push)
|
| 329 |
# ---------------------------
|
| 330 |
LOG_DIR = 'runs'
|
| 331 |
os.makedirs(LOG_DIR, exist_ok=True)
|
|
|
|
| 336 |
try:
|
| 337 |
with open(LOG_PATH, 'a', encoding='utf-8') as f:
|
| 338 |
f.write(json.dumps(entry, ensure_ascii=False) + '\n')
|
| 339 |
+
# Optional: Push to HF Hub if configured
|
| 340 |
+
if HAS_DATASETS and 'HF_TOKEN' in os.environ and 'HF_DATASET_REPO' in os.environ:
|
| 341 |
+
try:
|
| 342 |
+
ds = Dataset.from_json(LOG_PATH)
|
| 343 |
+
ds.push_to_hub(os.environ['HF_DATASET_REPO'], token=os.environ['HF_TOKEN'])
|
| 344 |
+
print("[info] Pushed logs to HF Hub dataset")
|
| 345 |
+
except Exception as e:
|
| 346 |
+
print(f"[warn] Failed to push logs to HF Hub: {e}")
|
| 347 |
+
except Exception as e:
|
| 348 |
+
print(f'Warning: failed to write log: {e}')
|
| 349 |
|
| 350 |
# ---------------------------
|
| 351 |
# Gradio UI callbacks
|
|
|
|
| 425 |
ax.plot(t, mean[i], label=(labels[i] if labels else f'dim{i}'))
|
| 426 |
ax.set_ylim(0,1.2)
|
| 427 |
ax.legend(loc='upper right')
|
| 428 |
+
ax.set_xlabel('Normalized time'); ax.set_ylabel('State')
|
| 429 |
plt.tight_layout()
|
| 430 |
return fig
|
| 431 |
|
|
|
|
| 433 |
# Gradio app layout
|
| 434 |
# ---------------------------
|
| 435 |
with gr.Blocks(title='Flan-SDE Behavioral Profiler') as demo:
|
| 436 |
+
gr.Markdown("# Flan-SDE Behavioral Profiler. Ask questions, build a profile, and simulate behavior using a Flan-T5 generated SDE.")
|
| 437 |
with gr.Row():
|
| 438 |
with gr.Column(scale=1):
|
| 439 |
start_btn = gr.Button('Start Session')
|
|
|
|
| 465 |
modes = json.loads(next_q[6:])['modes']
|
| 466 |
out = predict_from_session(SESSIONS[sid2], use_flan=use_flan, model_id=model_id, n_samps=int(n_samps))
|
| 467 |
fig = plot_paths(out['paths'], labels=["Strategy", "Empathy", "Control", "Masking", "RiskAppetite"])
|
| 468 |
+
except Exception as e:
|
| 469 |
+
print(f"[error] Prediction failed: {e}")
|
| 470 |
modes = []
|
| 471 |
fig = None
|
| 472 |
return sid2, profdict, '', 'Prediction complete', json.dumps(modes), fig
|
|
|
|
| 474 |
return sid2, profdict, next_q, '', None, None
|
| 475 |
|
| 476 |
start_btn.click(on_start, outputs=[sid_out, profile_view, qbox, out_console])
|
| 477 |
+
send_btn.click(on_send, inputs=[sid_out, user_in, profile_view, qbox, use_flan_chk, model_in, n_samps_slider], outputs=[sid_out, profile_view, qbox, out_console, modes_json, plot_out])
|