Spaces:
Paused
Paused
File size: 10,592 Bytes
a52f96d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 |
"""
DistilBERT-based student agent with online learning and memory decay.
Uses DistilBERT for Multiple Choice to answer reading comprehension tasks.
Implements online learning (fine-tune on 1 example at a time).
"""
import torch
from torch.optim import AdamW
from transformers import (
DistilBertForMultipleChoice,
DistilBertTokenizer,
)
from typing import List, Dict
import numpy as np
from collections import defaultdict
from interfaces import StudentAgentInterface, StudentState, Task
from memory_decay import MemoryDecayModel
class StudentAgent(StudentAgentInterface):
"""
DistilBERT-based student that learns reading comprehension.
Features:
- Online learning (1 example at a time)
- Memory decay (Ebbinghaus forgetting)
- Per-topic skill tracking
- Gradient accumulation for stability
"""
def __init__(
self,
learning_rate: float = 5e-5,
retention_constant: float = 80.0,
device: str = 'cpu',
max_length: int = 256,
gradient_accumulation_steps: int = 4
):
"""
Args:
learning_rate: LM fine-tuning learning rate
retention_constant: Forgetting speed (higher = slower forgetting)
device: 'cpu' or 'cuda'
max_length: Max tokens for passage + question + choices
gradient_accumulation_steps: Accumulate gradients for stability
"""
self.device = device
self.max_length = max_length
self.gradient_accumulation_steps = gradient_accumulation_steps
# Load DistilBERT for multiple choice
# Allow silent mode for testing
verbose = True # Can be overridden
try:
if verbose:
print("Loading DistilBERT model...", end=" ", flush=True)
self.model = DistilBertForMultipleChoice.from_pretrained(
"distilbert-base-uncased"
).to(self.device)
self.tokenizer = DistilBertTokenizer.from_pretrained(
"distilbert-base-uncased"
)
if verbose:
print("✅")
except Exception as e:
if verbose:
print(f"⚠️ (Model unavailable, using dummy mode)")
self.model = None
self.tokenizer = None
# Optimizer
if self.model:
self.optimizer = AdamW(self.model.parameters(), lr=learning_rate)
else:
self.optimizer = None
# Memory decay model
self.memory = MemoryDecayModel(retention_constant=retention_constant)
# Track per-topic base skills (before forgetting)
self.topic_base_skills: Dict[str, float] = {}
# Track learning history
self.topic_attempts: Dict[str, int] = defaultdict(int)
self.topic_correct: Dict[str, int] = defaultdict(int)
# Gradient accumulation counter
self.grad_step = 0
# Training mode flag
if self.model:
self.model.train()
def answer(self, task: Task) -> int:
"""
Predict answer without updating weights.
Prediction accuracy is modulated by memory decay.
"""
if not self.model:
# Dummy model: random guessing
return np.random.randint(0, 4)
self.model.eval()
# Prepare inputs
inputs = self._prepare_inputs(task)
with torch.no_grad():
outputs = self.model(**inputs)
logits = outputs.logits
predicted_idx = torch.argmax(logits, dim=-1).item()
# Apply memory decay to prediction
# If student has forgotten, prediction becomes more random
effective_skill = self.memory.get_effective_skill(task.topic)
# Probability of using learned answer vs random guess
# MCQ baseline = 0.25 (random guessing)
use_learned_prob = 0.25 + 0.75 * effective_skill
if np.random.random() < use_learned_prob:
return predicted_idx
else:
# Random guess
return np.random.randint(0, 4)
def learn(self, task: Task) -> bool:
"""
Fine-tune on a single task (online learning).
Returns:
True if prediction was correct, False otherwise
"""
if not self.model:
# Dummy learning: track statistics only
predicted = np.random.randint(0, 4)
was_correct = (predicted == task.answer)
self._update_stats(task, was_correct)
return was_correct
self.model.train()
# Get prediction before learning
predicted = self.answer(task)
was_correct = (predicted == task.answer)
# Prepare inputs with correct answer
inputs = self._prepare_inputs(task)
inputs['labels'] = torch.tensor([task.answer], device=self.device)
# Forward pass
outputs = self.model(**inputs)
loss = outputs.loss
# Backward pass with gradient accumulation
loss = loss / self.gradient_accumulation_steps
loss.backward()
self.grad_step += 1
# Update weights every N steps
if self.grad_step % self.gradient_accumulation_steps == 0:
self.optimizer.step()
self.optimizer.zero_grad()
# Update statistics
self._update_stats(task, was_correct)
return was_correct
def _update_stats(self, task: Task, was_correct: bool):
"""Update topic statistics and memory."""
self.topic_attempts[task.topic] += 1
if was_correct:
self.topic_correct[task.topic] += 1
# Compute base skill (accuracy without forgetting)
base_skill = self.topic_correct[task.topic] / self.topic_attempts[task.topic]
self.topic_base_skills[task.topic] = base_skill
# Update memory (record practice)
self.memory.update_practice(task.topic, base_skill)
def evaluate(self, eval_tasks: List[Task]) -> float:
"""
Evaluate on held-out tasks without updating weights.
Returns:
Accuracy (0.0-1.0)
"""
if not eval_tasks:
return 0.0
if not self.model:
# Dummy evaluation: return random
return 0.25
self.model.eval()
correct = 0
for task in eval_tasks:
predicted = self.answer(task)
if predicted == task.answer:
correct += 1
return correct / len(eval_tasks)
def get_state(self) -> StudentState:
"""
Get current state for teacher observation.
Returns per-topic accuracies accounting for forgetting.
"""
topic_accuracies = {}
time_since_practice = {}
for topic in self.topic_base_skills:
# Get effective skill (with forgetting)
effective_skill = self.memory.get_effective_skill(topic)
# Convert to expected accuracy on MCQ
topic_accuracies[topic] = 0.25 + 0.75 * effective_skill
# Time since last practice
time_since_practice[topic] = self.memory.get_time_since_practice(topic)
return StudentState(
topic_accuracies=topic_accuracies,
topic_attempts=dict(self.topic_attempts),
time_since_practice=time_since_practice,
total_timesteps=sum(self.topic_attempts.values()),
current_time=self.memory.current_time
)
def _prepare_inputs(self, task: Task) -> Dict[str, torch.Tensor]:
"""
Prepare inputs for DistilBERT multiple choice model.
Format: [CLS] passage [SEP] question [SEP] choice [SEP]
Repeated for each of 4 choices.
"""
if not self.tokenizer:
return {}
# Create 4 input sequences (one per choice)
input_texts = []
for choice in task.choices:
# Format: passage + question + choice
text = f"{task.passage} {task.question} {choice}"
input_texts.append(text)
# Tokenize
encoded = self.tokenizer(
input_texts,
padding=True,
truncation=True,
max_length=self.max_length,
return_tensors='pt'
)
# Reshape for multiple choice format
# (batch_size=1, num_choices=4, seq_length)
input_ids = encoded['input_ids'].unsqueeze(0).to(self.device)
attention_mask = encoded['attention_mask'].unsqueeze(0).to(self.device)
return {
'input_ids': input_ids,
'attention_mask': attention_mask
}
def advance_time(self, delta: float = 1.0):
"""Advance time for memory decay."""
self.memory.advance_time(delta)
def save(self, path: str):
"""Save model checkpoint."""
if not self.model:
print("⚠️ No model to save (using dummy model)")
return
torch.save({
'model_state_dict': self.model.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict() if self.optimizer else None,
'topic_base_skills': self.topic_base_skills,
'topic_attempts': dict(self.topic_attempts),
'topic_correct': dict(self.topic_correct),
'memory': self.memory,
'grad_step': self.grad_step
}, path)
print(f"💾 Saved checkpoint to {path}")
def load(self, path: str):
"""Load model checkpoint."""
checkpoint = torch.load(path, map_location=self.device)
if self.model:
self.model.load_state_dict(checkpoint['model_state_dict'])
if self.optimizer and checkpoint.get('optimizer_state_dict'):
self.optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
self.topic_base_skills = checkpoint['topic_base_skills']
self.topic_attempts = defaultdict(int, checkpoint['topic_attempts'])
self.topic_correct = defaultdict(int, checkpoint['topic_correct'])
self.memory = checkpoint['memory']
self.grad_step = checkpoint.get('grad_step', 0)
print(f"✅ Loaded checkpoint from {path}")
|