code_srt_sgwi_v1 / src /sgwi_trainer.py
natmin322's picture
Upload folder using huggingface_hub
e841b45 verified
"""
SGWI + Dual Fisher Trainer (Contribution 2)
============================================
Inherits from SRT_Trainer. Adds:
- SGWI (SRT-Guided Warm Initialization): warm-init LoRA from similar past tasks
- Dual Fisher: SRT-weighted embedding Fisher penalty during training
Usage in run_t5.py:
from sgwi_trainer import SGWI_DualFisher_Trainer
trainer = SGWI_DualFisher_Trainer(
...,
sgwi_mode='sgwi', # 'sgwi', 'inflora', 'random', 'sgwi+inflora'
lambda_emb=0.01, # Dual Fisher strength (0 = disabled)
)
"""
import os
import math
import logging
import numpy as np
import torch
import torch.nn as nn
from typing import Optional, Dict, List, Tuple
from cl_trainer_srt import SRT_Trainer
logger = logging.getLogger(__name__)
class SGWI_DualFisher_Trainer(SRT_Trainer):
"""SRT_Trainer + SGWI initialization + Dual Fisher regularization."""
def __init__(
self,
model,
args,
train_dataset,
cur_task_id,
task_order,
# SGWI params
sgwi_mode: str = 'sgwi', # 'sgwi', 'inflora', 'random', 'sgwi+inflora'
lambda_emb: float = 0.0, # Dual Fisher λ (0 = disabled)
# SRT params (passed to SRT_Trainer)
srt_metric_mode='hard',
srt_shrink=True,
srt_shrink_factor=0.1,
srt_max_emb_samples=500,
srt_load_path=None,
srt_skip_forward=False,
# Standard trainer params
data_collator_replay=None,
replay_dataset_dict=None,
replay_label_dict=None,
eval_dataset=None,
tokenizer=None,
data_collator=None,
compute_metrics=None,
callbacks=None,
**kwargs,
):
super().__init__(
model=model,
args=args,
train_dataset=train_dataset,
cur_task_id=cur_task_id,
task_order=task_order,
srt_metric_mode=srt_metric_mode,
srt_shrink=srt_shrink,
srt_shrink_factor=srt_shrink_factor,
srt_max_emb_samples=srt_max_emb_samples,
srt_load_path=srt_load_path,
srt_skip_forward=srt_skip_forward,
data_collator_replay=data_collator_replay,
replay_dataset_dict=replay_dataset_dict,
replay_label_dict=replay_label_dict,
eval_dataset=eval_dataset,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics,
callbacks=callbacks,
**kwargs,
)
self.sgwi_mode = sgwi_mode
self.lambda_emb = lambda_emb
self.theta_stars: Dict[int, Dict[str, torch.Tensor]] = {}
self._current_task_mu = None # cached for Dual Fisher
logger.info(f"[SGWI] mode={sgwi_mode}, lambda_emb={lambda_emb}, task_id={cur_task_id}")
# =========================================================================
# SGWI: Override get_reg_matrix
# =========================================================================
def get_reg_matrix(self):
"""Override InfLoRA initialization with SGWI or other modes.
6 Ablation Configs:
Config 1 'inflora': InfLoRA baseline (null-space + GPM)
Config 2 'random': No GPM, freeze A(kaiming), B=0
Config 3 'full_lora': No GPM, train A(kaiming)+B(0)
Config 5 'sgwi_freeze_a': No GPM, SGWI→A(frozen), B=0
Config 6 'sgwi_train_a': No GPM, SGWI→A(trainable), B=0
Config 4 'sgwi_full': No GPM, SGWI→A(trainable)+B(warm)
"""
mode = self.sgwi_mode
logger.info(f"[SGWI] get_reg_matrix: mode={mode}, task_id={self.cur_task_id}")
# ── Config 1: InfLoRA baseline ──────────────────────────────
if mode == 'inflora':
super().get_reg_matrix()
return
# ── All other configs: NO GPM, NO null-space ────────────────
self._init_gpm_attrs_skip()
# Task 0: no prior tasks → nothing to warm-init
if self.cur_task_id == 0:
logger.info(f"[SGWI] Task 0, mode={mode} → standard init (no prior tasks)")
return
# ── Config 2: random (freeze A kaiming, B=0) ────────────────
if mode == 'random':
logger.info("[SGWI] Config 2: random — no warm init")
return
# ── Config 3: full_lora (train A kaiming, B=0) ──────────────
if mode == 'full_lora':
logger.info("[SGWI] Config 3: full_lora — no warm init, A will be unfrozen by run_t5.py")
return
# ── Configs 4,5,6: need SGWI weights ────────────────────────
srt_weights = self._compute_sgwi_weights()
if not srt_weights:
logger.warning("[SGWI] No SRT weights. Falling back to random init.")
return
logger.info(f"[SGWI] SRT weights: {srt_weights}")
# ── Config 5: sgwi_freeze_a (SGWI→A frozen, B=0) ────────────
if mode == 'sgwi_freeze_a':
logger.info("[SGWI] Config 5: warm-init A only (frozen)")
self._sgwi_init_a(srt_weights)
return
# ── Config 6: sgwi_train_a (SGWI→A trainable, B=0) ──────────
if mode == 'sgwi_train_a':
logger.info("[SGWI] Config 6: warm-init A only (will be unfrozen by run_t5.py)")
self._sgwi_init_a(srt_weights)
return
# ── Config 4: sgwi_full (SGWI→A+B, both trainable) ──────────
if mode == 'sgwi_full':
logger.info("[SGWI] Config 4: warm-init both A and B")
self._sgwi_init_a(srt_weights)
self._fuse_past_lora_adapters(srt_weights) # warm-init B
return
# Legacy modes
if mode == 'sgwi':
self._sgwi_init()
return
if mode == 'sgwi+inflora':
self._sgwi_init()
super().get_reg_matrix()
return
raise ValueError(f"Unknown sgwi_mode: {mode}")
def _init_gpm_attrs_skip(self):
"""Initialize GPM attributes to safe defaults so _inner_training_loop doesn't crash.
Sets _cur_task=0 (falsy) so GPM projection is skipped during training.
Use for sgwi-only and random modes."""
self._cur_task = 0
self.feature_list = []
self.feature_trans_list = []
self.feature_mat = []
self.feature_trans_mat = []
logger.info("[SGWI] GPM attrs initialized (skip mode): _cur_task=0, feature_list=[]")
def _sgwi_init_a(self, srt_weights: Dict[int, float]):
"""SGWI warm-init for lora_A only, via SVD of weighted past ΔW.
delta_W = Σ w_s * (B_s @ A_s) [out_dim, in_dim]
SVD: delta_W = U @ S @ V^T
A_new = sqrt(S[:r]) * V^T[:r, :] — captures the important input directions
"""
model = self.model
device = next(model.parameters()).device
lora_r = self.args.lora_r if hasattr(self.args, 'lora_r') else 8
fused_count = 0
skipped_count = 0
for name, module in model.named_modules():
if not (hasattr(module, 'lora_q') and hasattr(module, 'lora_v')):
continue
if not hasattr(module, 'previous_lora_weights_q'):
continue
prev_q = getattr(module, 'previous_lora_weights_q', None)
prev_v = getattr(module, 'previous_lora_weights_v', None)
if prev_q is None or len(prev_q) == 0:
continue
for lora_tag, lora_cur, prev_list in [
('lora_q', module.lora_q, prev_q),
('lora_v', module.lora_v, prev_v),
]:
if prev_list is None or len(prev_list) == 0:
continue
# Compute delta_W from past tasks
delta_W = None
for past_task_id, w_s in srt_weights.items():
if isinstance(past_task_id, str):
try:
pos = self.task_order.index(past_task_id)
idx = (self.cur_task_id - 1) - pos
except ValueError:
skipped_count += 1
continue
else:
idx = int(past_task_id)
if idx < 0 or idx >= len(prev_list):
skipped_count += 1
continue
past_lora = prev_list[idx]
BA = past_lora.lora_B.data.float() @ past_lora.lora_A.data.float()
delta_W = w_s * BA if delta_W is None else delta_W + w_s * BA
if delta_W is None or delta_W.norm().item() < 1e-10:
skipped_count += 1
continue
# SVD decomposition → extract top-r input directions for A
try:
U, S, Vt = torch.linalg.svd(delta_W.to(device), full_matrices=False)
current_r = lora_cur.lora_A.data.shape[0]
r = min(current_r, len(S))
S_sqrt = torch.sqrt(S[:r] + 1e-12)
A_new = S_sqrt.unsqueeze(1) * Vt[:r, :] # [r, in_dim]
old_norm = lora_cur.lora_A.data.norm().item()
lora_cur.lora_A.data[:r, :].copy_(A_new.to(lora_cur.lora_A.data.device))
new_norm = lora_cur.lora_A.data.norm().item()
if fused_count < 3:
logger.info(f"[SGWI-A] {name}.{lora_tag}: A_new norm={new_norm:.4f} (was {old_norm:.4f})")
fused_count += 1
except Exception as e:
logger.warning(f"[SGWI-A] SVD failed for {name}.{lora_tag}: {e}")
skipped_count += 1
logger.info(f"[SGWI-A] Warm-init A for {fused_count} modules, skipped {skipped_count}")
def _sgwi_init(self):
"""SRT-Guided Warm Initialization for current task's LoRA."""
logger.info(f"[SGWI] Initializing task {self.cur_task_id} from past LoRA adapters...")
# Step 1: Get SRT weights from router signatures
srt_weights = self._compute_sgwi_weights()
if not srt_weights:
logger.warning("[SGWI] No past task signatures found. Falling back to standard init.")
super().get_reg_matrix()
return
logger.info(f"[SGWI] SRT weights: {srt_weights}")
# Step 2: Load past LoRA weights and compute weighted fusion
self._fuse_past_lora_adapters(srt_weights)
def _compute_sgwi_weights(self) -> Dict[int, float]:
"""Compute softmax weights from SRT distances to past tasks."""
if self.srt_router is None or len(self.srt_router.signatures) == 0:
return {}
# Extract current task embeddings to compute signature
logger.info("[SGWI] Extracting current task embeddings for distance computation...")
result = self._extract_task_embeddings(max_samples=self.srt_max_emb_samples)
# _extract_task_embeddings returns (h_train, task_ids) tuple
if isinstance(result, tuple):
h_train, _ = result
else:
h_train = result
if h_train is None or (hasattr(h_train, '__len__') and len(h_train) == 0):
return {}
current_mu = h_train.mean(dim=0).cpu().numpy()
self._current_task_mu = current_mu
# Compute distances to all past tasks
distances = {}
for task_id, sig in self.srt_router.signatures.items():
diff = current_mu - sig.mu
# Use pooled covariance if available, else L2
if hasattr(self.srt_router, 'pooled_cov') and self.srt_router.pooled_cov is not None:
try:
cov_inv = np.linalg.inv(self.srt_router.pooled_cov + 1e-6 * np.eye(len(diff)))
d = float(diff @ cov_inv @ diff)
except np.linalg.LinAlgError:
d = float(np.sum(diff ** 2))
else:
d = float(np.sum(diff ** 2))
distances[task_id] = d
if len(distances) == 0:
return {}
# Trivial case: only 1 past task
if len(distances) == 1:
return {k: 1.0 for k in distances}
# Softmax with τ = median heuristic
d_values = list(distances.values())
tau = float(np.median(d_values)) + 1e-8
weights = {}
for k, d in distances.items():
weights[k] = math.exp(-d / tau)
Z = sum(weights.values()) + 1e-12
weights = {k: w / Z for k, w in weights.items()}
return weights
def _fuse_past_lora_adapters(self, srt_weights: Dict[int, float]):
"""Weighted fusion of past LoRA adapters → warm-init current task's lora_B.
Architecture note (GainLoRA-InfLoRA):
- lora_A / lora_B: SINGLE nn.Parameter per module = CURRENT task's adapter
- previous_lora_weights_q[i] / previous_lora_weights_v[i]: past task i's LoRALayer
with .lora_A (nn.Parameter) and .lora_B (nn.Parameter)
Strategy:
1. For each past task s, get B_s and A_s from previous_lora_weights
2. Compute delta_W = Σ w_s * (B_s @ A_s)
3. Warm-init current lora_B via least-squares: B_new = delta_W @ A_cur^+
"""
model = self.model
device = next(model.parameters()).device
fused_count = 0
skipped_count = 0
for name, module in model.named_modules():
# Look for T5Attention modules with previous_lora_weights
if not (hasattr(module, 'lora_q') and hasattr(module, 'lora_v')):
continue
if not hasattr(module, 'previous_lora_weights_q'):
continue
prev_q = getattr(module, 'previous_lora_weights_q', None)
prev_v = getattr(module, 'previous_lora_weights_v', None)
if prev_q is None or len(prev_q) == 0:
continue
for lora_tag, lora_cur, prev_list in [
('lora_q', module.lora_q, prev_q),
('lora_v', module.lora_v, prev_v),
]:
if prev_list is None or len(prev_list) == 0:
continue
# Compute weighted ΔW = Σ w_s * (B_s @ A_s)
delta_W = None
for past_task_id, w_s in srt_weights.items():
# past_task_id may be a string task name (from srt_router.signatures)
# or an int. Convert to previous_lora_weights index.
# In run_t5.py: previous_lora_path is "t1,t2,...,t_{n-1}" in task_order,
# then reversed → prev_list[0] = most recent = task_order[cur_task_id-1]
if isinstance(past_task_id, str):
try:
pos = self.task_order.index(past_task_id)
idx = (self.cur_task_id - 1) - pos
except ValueError:
logger.debug(f"[SGWI] task name '{past_task_id}' not in task_order, skipping")
skipped_count += 1
continue
else:
idx = int(past_task_id)
if idx < 0 or idx >= len(prev_list):
logger.debug(f"[SGWI] idx={idx} out of range for prev_list len={len(prev_list)}, skipping")
skipped_count += 1
continue
past_lora = prev_list[idx]
A_s = past_lora.lora_A.data.float() # [r, in_features]
B_s = past_lora.lora_B.data.float() # [out_features, r]
BA = B_s @ A_s # [out_features, in_features]
if delta_W is None:
delta_W = w_s * BA
else:
delta_W = delta_W + w_s * BA
if delta_W is None:
continue
delta_W_norm = delta_W.norm().item()
if delta_W_norm < 1e-10:
logger.debug(f"[SGWI] delta_W ≈ 0 for {name}.{lora_tag} (norm={delta_W_norm:.2e}), skipping")
skipped_count += 1
continue
# Warm-init lora_B (trainable) via least-squares given frozen lora_A:
# B_warm = delta_W @ A_cur^T @ (A_cur @ A_cur^T + εI)^{-1}
try:
A_cur = lora_cur.lora_A.data.float().to(device) # [r, in_features]
AtA = A_cur @ A_cur.T # [r, r]
eps_mat = 1e-4 * torch.eye(A_cur.shape[0], device=device)
B_warm = delta_W.to(device) @ A_cur.T @ torch.linalg.inv(AtA + eps_mat) # [out_features, r]
old_norm = lora_cur.lora_B.data.norm().item()
lora_cur.lora_B.data.copy_(B_warm.to(lora_cur.lora_B.data.device))
new_norm = lora_cur.lora_B.data.norm().item()
if fused_count < 3:
logger.info(f"[SGWI] {name}.{lora_tag}: delta_W={delta_W_norm:.4f}, "
f"B_warm={new_norm:.4f} (was {old_norm:.4f})")
fused_count += 1
except Exception as e:
logger.warning(f"[SGWI] lora_B warm init failed for {name}.{lora_tag}: {e}")
continue
logger.info(f"[SGWI] Fused {fused_count} LoRA modules, skipped {skipped_count}")
def _get_lora_weight(self, lora_module, attr_name, task_id):
"""Safely get LoRA weight tensor for a given task_id."""
attr = getattr(lora_module, attr_name, None)
if attr is None:
return None
# Case 1: nn.ModuleDict or dict keyed by task_id
if isinstance(attr, (nn.ModuleDict, dict)):
key = str(task_id) if isinstance(attr, nn.ModuleDict) else task_id
if key in attr:
m = attr[key]
return m.weight.data if hasattr(m, 'weight') else m.data
# Try integer key
if task_id in attr:
m = attr[task_id]
return m.weight.data if hasattr(m, 'weight') else m.data
# Case 2: nn.ModuleList indexed by task_id
if isinstance(attr, nn.ModuleList):
if task_id < len(attr):
m = attr[task_id]
return m.weight.data if hasattr(m, 'weight') else m.data
# Case 3: Single nn.Linear (only for current task)
if isinstance(attr, nn.Linear):
return attr.weight.data
# Case 4: raw tensor
if isinstance(attr, torch.Tensor):
return attr
return None
def _set_lora_weight(self, lora_module, attr_name, task_id, new_weight):
"""Safely set LoRA weight tensor for current task_id."""
attr = getattr(lora_module, attr_name, None)
if attr is None:
return
try:
if isinstance(attr, (nn.ModuleDict, dict)):
key = str(task_id) if isinstance(attr, nn.ModuleDict) else task_id
if key in attr:
m = attr[key]
if hasattr(m, 'weight'):
m.weight.data.copy_(new_weight.to(m.weight.device))
else:
attr[key] = new_weight.to(next(lora_module.parameters()).device)
elif isinstance(attr, nn.ModuleList):
if task_id < len(attr):
m = attr[task_id]
if hasattr(m, 'weight'):
m.weight.data.copy_(new_weight.to(m.weight.device))
elif isinstance(attr, nn.Linear):
attr.weight.data.copy_(new_weight.to(attr.weight.device))
except Exception as e:
logger.debug(f"[SGWI] Could not set weight for {attr_name}[{task_id}]: {e}")
# =========================================================================
# Dual Fisher: Override compute_loss
# =========================================================================
def compute_loss(self, model, inputs, return_outputs=False):
"""Standard loss + Dual Fisher penalty."""
outputs = model(**inputs)
loss = outputs["loss"] if isinstance(outputs, dict) else outputs[0]
if self.lambda_emb > 0 and self.cur_task_id > 0 and len(self.theta_stars) > 0:
fisher_loss = self._dual_fisher_penalty()
loss = loss + fisher_loss
return (loss, outputs) if return_outputs else loss
def _dual_fisher_penalty(self) -> torch.Tensor:
"""SRT-weighted L2 penalty around past task parameters."""
total = torch.tensor(0.0, device=next(self.model.parameters()).device)
# Get SRT weights (cached or recompute)
srt_weights = {}
if self.srt_router and len(self.srt_router.signatures) > 0:
for task_id in self.theta_stars:
if task_id in self.srt_router.signatures:
srt_weights[task_id] = 1.0 # uniform for now
# Normalize
if srt_weights:
Z = sum(srt_weights.values())
srt_weights = {k: v / Z for k, v in srt_weights.items()}
if not srt_weights:
return total
for task_id, w_s in srt_weights.items():
if task_id not in self.theta_stars:
continue
for name, param in self.model.named_parameters():
if not param.requires_grad:
continue
if 'lora_' not in name:
continue
if name not in self.theta_stars[task_id]:
continue
theta_star = self.theta_stars[task_id][name].to(param.device)
delta = param - theta_star
total = total + w_s * (delta ** 2).sum()
return self.lambda_emb * total
# =========================================================================
# Override on_task_end to save θ* for Dual Fisher
# =========================================================================
def on_task_end(self, task_id):
"""Save θ* for Dual Fisher, then do standard SRT task end."""
# Save current task's final LoRA weights for future Dual Fisher
logger.info(f"[SGWI] Saving θ* for task {task_id} (Dual Fisher)")
self.theta_stars[task_id] = {}
for name, param in self.model.named_parameters():
if 'lora_' in name and param.requires_grad:
self.theta_stars[task_id][name] = param.detach().clone().cpu()
saved_count = len(self.theta_stars[task_id])
logger.info(f"[SGWI] Saved {saved_count} parameters for task {task_id}")
# Save theta_stars to disk for persistence
theta_stars_path = os.path.join(self.args.output_dir, 'saved_weights', 'theta_stars.pt')
os.makedirs(os.path.dirname(theta_stars_path), exist_ok=True)
torch.save(self.theta_stars, theta_stars_path)
# Standard SRT: compute signature, wire router
super().on_task_end(task_id)
def _load_theta_stars(self, srt_load_path: str):
"""Load previously saved θ* from checkpoint."""
theta_path = os.path.join(srt_load_path, 'theta_stars.pt')
if os.path.exists(theta_path):
self.theta_stars = torch.load(theta_path, map_location='cpu')
logger.info(f"[SGWI] Loaded θ* for {len(self.theta_stars)} past tasks from {theta_path}")
else:
logger.info(f"[SGWI] No θ* found at {theta_path}")