chore: 更新项目名称

This commit is contained in:
2026-03-31 09:34:11 +08:00
parent 780b9e1855
commit 2e009cf59a
51 changed files with 118 additions and 118 deletions
+30
View File
@@ -0,0 +1,30 @@
from astrai.trainer.trainer import Trainer
from astrai.trainer.strategy import StrategyFactory, BaseStrategy
from astrai.trainer.schedule import SchedulerFactory, BaseScheduler
from astrai.trainer.train_callback import (
TrainCallback,
GradientClippingCallback,
SchedulerCallback,
CheckpointCallback,
ProgressBarCallback,
MetricLoggerCallback,
)
__all__ = [
# Main trainer
"Trainer",
# Strategy factory
"StrategyFactory",
"BaseStrategy",
# Scheduler factory
"SchedulerFactory",
"BaseScheduler",
# Callbacks
"TrainCallback",
"GradientClippingCallback",
"SchedulerCallback",
"CheckpointCallback",
"ProgressBarCallback",
"MetricLoggerCallback",
]
+103
View File
@@ -0,0 +1,103 @@
import torch.nn as nn
from typing import Dict
def grad_norm(model: nn.Module, norm_type: int = 2) -> Dict[str, float]:
"""Compute gradient norm for each parameter in the model."""
norms = {}
for name, param in model.named_parameters():
norms[name] = 0.0
if param.grad:
norm = param.grad.data.norm(norm_type).item()
norms[name] = norm
return norms
def grad_std(model: nn.Module) -> Dict[str, float]:
"""Compute standard deviation of gradients for each parameter."""
stds = {}
for name, param in model.named_parameters():
stds[name] = 0.0
if param.grad:
std = param.grad.data.std().item()
stds[name] = std
return stds
def grad_max(model: nn.Module) -> Dict[str, float]:
"""Find the maximum absolute gradient value for each parameter."""
max_vals = {}
for name, param in model.named_parameters():
max_vals[name] = -float("inf")
if param.grad:
max_val = param.grad.data.max().item()
max_vals[name] = max_val
return max_vals
def grad_min(model: nn.Module) -> Dict[str, float]:
"""Find the minimum absolute gradient value for each parameter."""
min_vals = {}
for name, param in model.named_parameters():
min_vals[name] = float("inf")
if param.grad:
min_val = param.grad.data.min().item()
min_vals[name] = min_val
return min_vals
def grad_mean(model: nn.Module) -> Dict[str, float]:
"""Compute mean of gradients for each parameter."""
means = {}
for name, param in model.named_parameters():
means[name] = 0.0
if param.grad:
mean = param.grad.data.mean().item()
means[name] = mean
return means
def grad_nan_num(model: nn.Module) -> Dict[str, int]:
"""Count the number of NaNs in gradients for each parameter."""
nan_nums = {}
for name, param in model.named_parameters():
nan_nums[name] = 0
if param.grad:
nan_num = param.grad.isnan().sum().item()
nan_nums[name] = nan_num
return nan_nums
def ctx_get_loss(ctx):
return ctx.loss
def ctx_get_lr(ctx):
return ctx.optimizer.param_groups[-1]["lr"]
def ctx_get_grad_norm(ctx):
return grad_norm(ctx.model)
def ctx_get_grad_std(ctx):
return grad_std(ctx.model)
def ctx_get_grad_max(ctx):
return grad_max(ctx.model)
def ctx_get_grad_min(ctx):
return grad_min(ctx.model)
def ctx_get_grad_mean(ctx):
return grad_mean(ctx.model)
def ctx_get_grad_nan_num(ctx):
return grad_nan_num(ctx.model)
+237
View File
@@ -0,0 +1,237 @@
"""Learning rate scheduler implementations with factory pattern."""
import math
from abc import abstractmethod, ABC
from typing import Any, Dict, List, Type
from torch.optim.lr_scheduler import LRScheduler
from astrai.config.schedule_config import ScheduleConfig
class BaseScheduler(LRScheduler, ABC):
"""Base scheduler class for all other schedulers."""
def __init__(self, optimizer, last_epoch: int = -1):
super().__init__(optimizer, last_epoch)
@abstractmethod
def get_lr(self) -> List[float]:
"""Calculate the current learning rate."""
raise NotImplementedError
def state_dict(self) -> Dict[str, Any]:
return super().state_dict()
def load_state_dict(self, state_dict: Dict[str, Any]):
super().load_state_dict(state_dict)
class SchedulerFactory:
"""Factory class for creating learning rate schedulers.
Supports decorator-based registration for extensible scheduler types.
Also supports creation from ScheduleConfig objects.
Example usage:
@SchedulerFactory.register("custom")
class CustomScheduler(BaseScheduler):
...
scheduler = SchedulerFactory.create(optimizer, "custom", **kwargs)
# Or from config
config = CosineScheduleConfig(total_steps=10000)
scheduler = SchedulerFactory.load(optimizer, config)
"""
SCHEDULER_MAP: Dict[str, Type[BaseScheduler]] = {}
@classmethod
def register(cls, name: str):
"""Decorator to register a new scheduler class.
Args:
name: Registration name for the scheduler
Returns:
Decorator function that registers the scheduler class
"""
def decorator(scheduler_cls: Type[BaseScheduler]) -> Type[BaseScheduler]:
if not issubclass(scheduler_cls, BaseScheduler):
raise TypeError(
f"{scheduler_cls.__name__} must inherit from BaseScheduler"
)
cls.SCHEDULER_MAP[name] = scheduler_cls
return scheduler_cls
return decorator
@classmethod
def create(cls, optimizer, schedule_type: str, **kwargs) -> BaseScheduler:
"""Create a scheduler instance by type name.
Args:
optimizer: PyTorch optimizer
schedule_type: Type of scheduler ("cosine", "sgdr")
**kwargs: Arguments passed to the scheduler constructor
Returns:
Scheduler instance
Raises:
ValueError: If schedule_type is not supported
"""
if schedule_type not in cls.SCHEDULER_MAP:
raise ValueError(
f"Unknown schedule type: '{schedule_type}'. "
f"Supported types: {sorted(cls.SCHEDULER_MAP.keys())}"
)
scheduler_cls = cls.SCHEDULER_MAP[schedule_type]
return scheduler_cls(optimizer, **kwargs)
@staticmethod
def load(optimizer, schedule_config: ScheduleConfig) -> BaseScheduler:
"""Create a scheduler from a ScheduleConfig object.
Args:
optimizer: PyTorch optimizer
schedule_config: ScheduleConfig instance
Returns:
Scheduler instance
"""
kwargs = schedule_config.get_kwargs()
schedule_type = kwargs.pop("schedule_type")
return SchedulerFactory.create(optimizer, schedule_type, **kwargs)
@classmethod
def available_types(cls) -> list:
"""Return list of registered scheduler type names."""
return list(cls.SCHEDULER_MAP.keys())
# ============== Scheduler Classes ==============
# All scheduler classes are registered at class definition time using the decorator
@SchedulerFactory.register("cosine")
class CosineScheduler(BaseScheduler):
"""Cosine decay scheduler with warmup, implemented as PyTorch LRScheduler."""
def __init__(
self,
optimizer,
warmup_steps: int,
lr_decay_steps: int,
min_rate: float = 0.05,
last_epoch: int = -1,
):
self.warmup_steps = warmup_steps
self.lr_decay_steps = lr_decay_steps
self.min_rate = min_rate
self.total_steps = warmup_steps + lr_decay_steps
super().__init__(optimizer, last_epoch)
def get_lr(self) -> List[float]:
# warmup
if self.last_epoch < self.warmup_steps:
warmup_factor = max(self.min_rate, self.last_epoch / self.warmup_steps)
return [base_lr * warmup_factor for base_lr in self.base_lrs]
# cosine decay
decay_progress = (self.last_epoch - self.warmup_steps) / self.lr_decay_steps
decay_progress = min(decay_progress, 1.0)
cosine_decay = 0.5 * (1.0 + math.cos(math.pi * decay_progress))
decay_factor = max(self.min_rate, cosine_decay)
return [base_lr * decay_factor for base_lr in self.base_lrs]
def state_dict(self):
state = super().state_dict()
state.update(
{
"warmup_steps": self.warmup_steps,
"lr_decay_steps": self.lr_decay_steps,
"min_rate": self.min_rate,
"total_steps": self.total_steps,
}
)
return state
def load_state_dict(self, state_dict):
self.warmup_steps = state_dict.pop("warmup_steps")
self.lr_decay_steps = state_dict.pop("lr_decay_steps")
self.min_rate = state_dict.pop("min_rate")
self.total_steps = state_dict.pop("total_steps")
super().load_state_dict(state_dict)
@SchedulerFactory.register("sgdr")
class SGDRScheduler(BaseScheduler):
"""SGDR (Stochastic Gradient Descent with Warm Restarts) scheduler."""
def __init__(
self,
optimizer,
warmup_steps: int,
cycle_length: int,
min_rate: float = 0.05,
t_mult: int = 2,
last_epoch: int = -1,
):
self.warmup_steps = warmup_steps
self.cycle_length = cycle_length
self.min_rate = min_rate
self.t_mult = t_mult
super().__init__(optimizer, last_epoch)
def get_lr(self):
# warmup
if self.last_epoch < self.warmup_steps:
warmup_factor = max(self.min_rate, self.last_epoch / self.warmup_steps)
return [base_lr * warmup_factor for base_lr in self.base_lrs]
# SGDR
steps_since_warmup = self.last_epoch - self.warmup_steps
# 1. Calculate current cycle and position within cycle
current_cycle_length = self.cycle_length
total_cycles_length = 0
cycle_num = 0
while total_cycles_length + current_cycle_length <= steps_since_warmup:
total_cycles_length += current_cycle_length
current_cycle_length *= self.t_mult
cycle_num += 1
steps_in_cycle = steps_since_warmup - total_cycles_length
# 2. Cosine annealing within the current cycle
cosine_factor = 0.5 * (
1 + math.cos(math.pi * steps_in_cycle / current_cycle_length)
)
learning_rate_factor = self.min_rate + (1 - self.min_rate) * cosine_factor
return [base_lr * learning_rate_factor for base_lr in self.base_lrs]
def state_dict(self):
"""Returns the state of the scheduler as a dict."""
state = super().state_dict()
state.update(
{
"warmup_steps": self.warmup_steps,
"cycle_length": self.cycle_length,
"min_rate": self.min_rate,
"t_mult": self.t_mult,
}
)
return state
def load_state_dict(self, state_dict):
"""Loads the scheduler's state."""
self.warmup_steps = state_dict.pop("warmup_steps")
self.cycle_length = state_dict.pop("cycle_length")
self.min_rate = state_dict.pop("min_rate")
self.t_mult = state_dict.pop("t_mult")
super().load_state_dict(state_dict)
+362
View File
@@ -0,0 +1,362 @@
"""Training strategy implementations with factory pattern."""
import copy
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.parallel import DistributedDataParallel as DDP
from torch import Tensor
from typing import Any, Callable, Dict, Union
from abc import ABC, abstractmethod
def unwrap_model(model: nn.Module) -> nn.Module:
"""Unwrap DDP wrapper if present to get the original model."""
if isinstance(model, DDP):
return model.module
return model
def create_ref_model(model: nn.Module) -> nn.Module:
"""Create a reference model for DPO/GRPO training.
Handles DDP-wrapped models safely by unwrapping first,
then creating a deep copy with frozen gradients.
"""
original_model = unwrap_model(model)
ref_model = copy.deepcopy(original_model)
ref_model.requires_grad_(False)
ref_model.eval()
return ref_model
def move_to_device(batch: Dict[str, Tensor], device: str) -> Any:
"""Move batch tensors to specified device with non-blocking transfer."""
return {key: value.to(device, non_blocking=True) for key, value in batch.items()}
def get_logprobs(
model: Union[nn.Module, Callable[..., Dict[str, Tensor]]],
input_ids: Tensor,
mask: Tensor,
reduction: str,
):
"""Compute token-wise log probabilities from model outputs.
Args:
model: The language model
input_ids: Input token IDs of shape [batch_size, seq_len]
mask: Attention mask of shape [batch_size, seq_len]
reduction: How to reduce over sequence dimension ("mean", "sum", "none")
Returns:
Log probabilities with reduction applied over sequence dimension
"""
allowed_reductions = ["mean", "sum", "none"]
if reduction not in allowed_reductions:
raise ValueError(
f"reduction must be one of {allowed_reductions}, got '{reduction}'"
)
shifted_input_ids = input_ids[:, 1:]
shifted_mask = mask[:, 1:]
logits = model(input_ids[:, :-1], mask[:, :-1])["logits"]
log_probs = torch.log_softmax(logits.float(), dim=-1)
token_logprobs = torch.gather(
log_probs, dim=-1, index=shifted_input_ids.unsqueeze(-1)
).squeeze(-1)
if reduction == "mean":
return (token_logprobs * shifted_mask).sum(dim=-1) / shifted_mask.sum(
dim=-1
).clamp(min=1.0)
elif reduction == "sum":
return (token_logprobs * shifted_mask).sum(dim=-1)
else:
return token_logprobs * shifted_mask
class BaseStrategy(ABC):
"""Abstract base class for training strategies."""
def __init__(
self, model: Union[Callable[..., Dict[str, Tensor]]], device: str, **kwargs
):
self.model = model
self.device = device
self.extra_kwargs = kwargs
@abstractmethod
def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
"""Compute loss for the given batch.
Args:
batch: Dictionary containing batch tensors
Returns:
Computed loss tensor
"""
raise NotImplementedError
def __call__(self, batch: Dict[str, Tensor]) -> Tensor:
"""Allow calling strategy directly as a callable."""
return self.compute_loss(batch)
class StrategyFactory:
"""Factory class for creating training strategy instances.
Supports decorator-based registration for extensible strategy types.
All default strategies (seq, sft, dpo, grpo) are automatically registered.
Example usage:
@StrategyFactory.register("custom")
class CustomStrategy(BaseStrategy):
...
strategy = StrategyFactory.create(model, "custom", device)
"""
SUPPORTED_STRATEGIES = frozenset({"seq", "sft", "dpo", "grpo"})
STRATEGY_MAP: Dict[str, type] = {}
@classmethod
def register(cls, name: str):
"""Decorator to register a new strategy class.
Args:
name: Registration name for the strategy
Returns:
Decorator function that registers the strategy class
"""
def decorator(strategy_cls: type) -> type:
if not issubclass(strategy_cls, BaseStrategy):
raise TypeError(
f"{strategy_cls.__name__} must inherit from BaseStrategy"
)
cls.STRATEGY_MAP[name] = strategy_cls
return strategy_cls
return decorator
@classmethod
def create(cls, model, train_type: str, device: str, **kwargs) -> BaseStrategy:
"""Create a strategy instance based on training type.
Args:
model: Model instance for the strategy
train_type: Type of training ("seq", "sft", "dpo", "grpo")
device: Device to run the strategy on
**kwargs: Additional arguments passed to strategy constructor
Returns:
Strategy instance
Raises:
ValueError: If train_type is not supported
NotImplementedError: If train_type is in supported list but not implemented
"""
if train_type not in cls.SUPPORTED_STRATEGIES:
raise ValueError(
f"Unknown training strategy: '{train_type}'. "
f"Supported strategies: {sorted(cls.SUPPORTED_STRATEGIES)}"
)
if train_type not in cls.STRATEGY_MAP:
raise NotImplementedError(
f"Strategy '{train_type}' is supported but not yet implemented."
)
strategy_cls = cls.STRATEGY_MAP[train_type]
return strategy_cls(model, device, **kwargs)
@classmethod
def available_strategies(cls) -> list:
"""Return list of registered strategy names."""
return list(cls.STRATEGY_MAP.keys())
# ============== Strategy Classes ==============
# All strategies are registered at class definition time using the decorator
@StrategyFactory.register("seq")
class SEQStrategy(BaseStrategy):
"""Standard next-token prediction training strategy.
Computes cross-entropy loss for next token prediction.
"""
def __init__(self, model, device, label_smoothing: float = 0.0, **kwargs):
super().__init__(model, device, **kwargs)
self.label_smoothing = label_smoothing
def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
batch = move_to_device(batch, self.device)
input_ids, target_ids = batch["input_ids"], batch["target_ids"]
logits = self.model(input_ids=input_ids)["logits"]
loss = F.cross_entropy(
input=logits.flatten(0, 1).float(),
target=target_ids.flatten(),
label_smoothing=self.label_smoothing,
)
return loss
@StrategyFactory.register("sft")
class SFTStrategy(BaseStrategy):
"""Supervised Fine-tuning strategy with loss masking.
Applies cross-entropy loss only to tokens where loss_mask is True.
"""
def __init__(self, model, device, label_smoothing: float = 0.0, **kwargs):
super().__init__(model, device, **kwargs)
self.label_smoothing = label_smoothing
def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
batch = move_to_device(batch, self.device)
input_ids, target_ids, loss_mask = (
batch["input_ids"],
batch["target_ids"],
batch["loss_mask"],
)
ignore_index = -100
logits = self.model(input_ids=input_ids)["logits"]
target_ids = target_ids.masked_fill(loss_mask == 0, ignore_index)
loss = F.cross_entropy(
input=logits.flatten(0, 1).float(),
target=target_ids.flatten(),
ignore_index=ignore_index,
label_smoothing=self.label_smoothing,
)
return loss
@StrategyFactory.register("dpo")
class DPOStrategy(BaseStrategy):
"""Direct Preference Optimization strategy.
Implements the DPO loss from the paper "Direct Preference Optimization".
Uses a reference model to compute KL divergence penalty.
"""
def __init__(
self,
model: nn.Module,
device: str,
beta: float = 0.1,
reduction: str = "mean",
**kwargs,
):
super().__init__(model, device, **kwargs)
self.ref_model = create_ref_model(model)
self.beta = beta
self.reduction = reduction
def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
batch = move_to_device(batch, self.device)
chosen_ids, rejected_ids = batch["chosen"], batch["rejected"]
chosen_mask, rejected_mask = batch["chosen_mask"], batch["rejected_mask"]
contact_ids = torch.cat([chosen_ids, rejected_ids], dim=0)
contact_mask = torch.cat([chosen_mask, rejected_mask], dim=0)
log_pi = get_logprobs(self.model, contact_ids, contact_mask, self.reduction)
with torch.no_grad():
log_ref = get_logprobs(
self.ref_model, contact_ids, contact_mask, self.reduction
)
log_pi_chosen = log_pi[: chosen_ids.shape[0]]
log_pi_rejected = log_pi[chosen_ids.shape[0] :]
log_ref_chosen = log_ref[: chosen_ids.shape[0]]
log_ref_rejected = log_ref[chosen_ids.shape[0] :]
pi_log_ratio = log_pi_chosen - log_pi_rejected
ref_log_ratio = log_ref_chosen - log_ref_rejected
ratio_diff = pi_log_ratio - ref_log_ratio
dpo_loss = -F.logsigmoid(self.beta * ratio_diff).mean()
return dpo_loss
@StrategyFactory.register("grpo")
class GRPOStrategy(BaseStrategy):
"""Group Relative Policy Optimization strategy.
Implements GRPO with clipping and KL penalty.
"""
def __init__(
self,
model: nn.Module,
device: str,
clip_eps: float = 0.2,
kl_coef: float = 0.01,
group_size: int = 4,
reduction: str = "mean",
**kwargs,
):
super().__init__(model, device, **kwargs)
self.ref_model = create_ref_model(model)
self.clip_eps = clip_eps
self.kl_coef = kl_coef
self.group_size = group_size
self.reduction = reduction
def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
batch = move_to_device(batch, self.device)
prompts = batch["prompts"]
responses = batch["responses"]
masks = batch["masks"]
rewards = batch["rewards"]
batch_size, group_size, response_len = responses.shape
responses_flat = responses.view(-1, response_len)
masks_flat = masks.view(-1, response_len)
prompt_expanded = prompts.unsqueeze(1).repeat(1, group_size, 1).flatten(0, 1)
# Shape: (batch_size * group_size, seq_len + response_len)
full_sequences = torch.cat([prompt_expanded, responses_flat], dim=-1)
full_masks = torch.cat([torch.ones_like(prompt_expanded), masks_flat], dim=-1)
log_probs_policy = get_logprobs(
self.model, full_sequences, full_masks, self.reduction
)
log_probs_policy = log_probs_policy.view(batch_size, group_size)
with torch.no_grad():
log_probs_ref = get_logprobs(
self.ref_model, full_sequences, full_masks, self.reduction
)
log_probs_ref = log_probs_ref.view(batch_size, group_size)
# Compute advantages from rewards with normalization
eps = torch.finfo(log_probs_policy.dtype).eps
mean = rewards.mean(dim=-1, keepdim=True)
std = rewards.std(dim=-1, keepdim=True)
advantages = (rewards - mean) / (std + eps)
# PPO-style clipped surrogate objective
ratio = torch.exp(0) # Off-policy: policy_model = old_model
surr1 = ratio * advantages
surr2 = torch.clamp(ratio, 1 - self.clip_eps, 1 + self.clip_eps) * advantages
policy_loss = -torch.min(surr1, surr2).mean()
kl_penalty = self.kl_coef * (log_probs_policy - log_probs_ref).square().mean()
total_loss = policy_loss + kl_penalty
return total_loss
+235
View File
@@ -0,0 +1,235 @@
import os
import json
import time
import torch.nn as nn
from pathlib import Path
from tqdm import tqdm
from torch.nn.utils import clip_grad_norm_
from typing import Callable, List, Optional, Protocol
from astrai.parallel import only_on_rank
from astrai.trainer.metric_util import (
ctx_get_loss,
ctx_get_lr,
ctx_get_grad_max,
ctx_get_grad_min,
ctx_get_grad_norm,
ctx_get_grad_mean,
ctx_get_grad_std,
ctx_get_grad_nan_num,
)
from astrai.data.serialization import Checkpoint
from astrai.trainer.train_context import TrainContext
class TrainCallback(Protocol):
"""
Callback interface for trainer.
"""
def on_train_begin(self, context: TrainContext):
"""Called at the beginning of training."""
def on_train_end(self, context: TrainContext):
"""Called at the end of training."""
def on_epoch_begin(self, context: TrainContext):
"""Called at the beginning of each epoch."""
def on_epoch_end(self, context: TrainContext):
"""Called at the end of each epoch."""
def on_step_begin(self, context: TrainContext):
"""Called at the beginning of each step."""
def on_step_end(self, context: TrainContext):
"""Called at the end of each step."""
def on_batch_begin(self, context: TrainContext):
"""Called at the beginning of each batch."""
def on_batch_end(self, context: TrainContext):
"""Called at the end of each batch."""
def on_error(self, context: TrainContext):
"""Called when an error occurs during training."""
class GradientClippingCallback(TrainCallback):
"""
Gradient clipping callback for trainer.
"""
def __init__(self, max_grad_norm: float):
self.max_grad_norm = max_grad_norm
def on_step_begin(self, context: TrainContext):
_ = context
clip_grad_norm_(context.model.parameters(), self.max_grad_norm)
class SchedulerCallback(TrainCallback):
"""
Scheduler callback for trainer.
"""
def __init__(self):
pass
def on_train_begin(self, context: TrainContext):
for group in context.optimizer.param_groups:
if "initial_lr" not in group:
group["initial_lr"] = group["lr"]
def on_batch_end(self, context: TrainContext):
if context.scheduler:
context.scheduler.step()
class CheckpointCallback(TrainCallback):
"""
Checkpoint callback for trainer.
"""
def __init__(
self,
save_dir: str,
interval: int,
weight_only: bool = False,
state_dict_fn: Optional[Callable[[nn.Module], dict]] = None,
):
self.save_dir = save_dir
self.interval = interval
self.weight_only = weight_only
self.state_dict_fn = state_dict_fn
self.last_ckpt_iter = 0
@only_on_rank(0)
def _save_checkpoint(self, context: TrainContext):
save_path = os.path.join(
self.save_dir, f"epoch_{context.epoch}_iter_{context.iteration}"
)
state_dict = (
self.state_dict_fn(context.model)
if self.state_dict_fn
else context.model.state_dict()
)
context.checkpoint = Checkpoint(
state_dict=state_dict, epoch=context.epoch, iteration=context.iteration
)
context.checkpoint.save(save_path)
self.last_ckpt_iter = context.iteration
def on_batch_end(self, context: TrainContext):
if context.iteration - self.last_ckpt_iter >= self.interval:
self._save_checkpoint(context)
def on_train_end(self, context: TrainContext):
if context.iteration != self.last_ckpt_iter:
self._save_checkpoint(context)
def on_error(self, context: TrainContext):
self._save_checkpoint(context)
class ProgressBarCallback(TrainCallback):
"""
Progress bar callback for trainer.
"""
def __init__(self, num_epoch: int):
self.num_epoch = num_epoch
self.progress_bar: tqdm = None
@only_on_rank(0)
def on_epoch_begin(self, context: TrainContext):
self.progress_bar = tqdm(
context.dataloader,
desc=f"Epoch {context.epoch + 1}/{self.num_epoch}",
dynamic_ncols=True,
)
@only_on_rank(0)
def on_batch_end(self, context: TrainContext):
self.progress_bar.set_postfix(
{
"loss": f"{context.loss:.4f}",
"lr": f"{context.optimizer.param_groups[-1]['lr']:.2e}",
}
)
self.progress_bar.update(1)
@only_on_rank(0)
def on_epoch_end(self, context: TrainContext):
_ = context
if self.progress_bar:
self.progress_bar.close()
class MetricLoggerCallback(TrainCallback):
def __init__(
self,
log_dir: str,
save_interval: int,
log_interval: int = 10,
metrics: List[str] = None,
):
self.last_log_iter = 0
self.save_interval = save_interval
self.log_interval = log_interval
self.metrics = metrics or ["loss", "lr"]
self.log_dir = Path(log_dir) if log_dir else Path.cwd() / "logs"
self.log_dir.mkdir(parents=True, exist_ok=True)
self.log_cache = []
self._metric_funcs = {
"loss": ctx_get_loss,
"lr": ctx_get_lr,
"grad_norm": ctx_get_grad_norm,
"grad_std": ctx_get_grad_std,
"grad_max": ctx_get_grad_max,
"grad_min": ctx_get_grad_min,
"grad_mean": ctx_get_grad_mean,
"grad_nan_num": ctx_get_grad_nan_num,
}
def _get_log_data(self, context: TrainContext):
return {
"timestamp": time.strftime("%Y-%m-%d %H:%M:%S"),
"epoch": context.epoch,
"iter": context.iteration,
**{m: self._metric_funcs[m](context) for m in self.metrics},
}
@only_on_rank(0)
def _add_log(self, log_data):
self.log_cache.append(log_data)
@only_on_rank(0)
def _save_log(self, epoch, iter):
log_file = self.log_dir / f"epoch_{epoch}_iter_{iter}_metric.jsonl"
with open(log_file, "w") as f:
for log in self.log_cache:
f.write(json.dumps(log) + "\n")
def on_batch_end(self, context):
if context.iteration % self.log_interval == 0:
log_data = self._get_log_data(context)
self._add_log(log_data)
if context.iteration - self.last_log_iter >= self.save_interval:
self._save_log(context.epoch, context.iteration)
self.last_log_iter = context.iteration
def on_train_end(self, context):
if context.iteration != self.last_log_iter:
self._save_log(context.epoch, context.iteration)
def on_error(self, context):
self._save_log(context.epoch, context.iteration)
+99
View File
@@ -0,0 +1,99 @@
import torch.nn as nn
from torch.optim import Optimizer
from torch.optim.lr_scheduler import LRScheduler
from torch.utils.data import DataLoader
from astrai.data import ResumableDistributedSampler
from astrai.data.serialization import Checkpoint
from astrai.trainer.strategy import StrategyFactory, BaseStrategy
from astrai.config.train_config import TrainConfig
from astrai.parallel.setup import get_current_device, get_world_size, get_rank
from dataclasses import dataclass, field
from typing import Optional, Self
@dataclass
class TrainContext:
model: nn.Module = field(default=None)
strategy: BaseStrategy = field(default=None)
dataloader: DataLoader = field(default=None)
optimizer: Optimizer = field(default=None)
scheduler: LRScheduler = field(default=None)
checkpoint: Checkpoint = field(default=None)
epoch: int = field(default=0)
iteration: int = field(default=0)
loss: float = field(default=0.0)
world_size: int = field(default=1)
rank: int = field(default=0)
kwargs: dict = field(default_factory=dict)
class TrainContextBuilder:
def __init__(self, config: TrainConfig):
self.config = config
self._context = TrainContext(
model=config.model,
world_size=get_world_size(),
rank=get_rank(),
)
device = get_current_device()
self._context.model = self._context.model.to(device=device)
if self.config.nprocs > 1:
fn = self.config.parallel_wrapper
self._context.model = fn(self._context.model)
self._context.optimizer = self.config.optimizer_fn(self._context.model)
self._context.scheduler = self.config.scheduler_fn(self._context.optimizer)
def with_checkpoint(self, checkpoint: Optional[Checkpoint]) -> Self:
if checkpoint is None:
checkpoint = Checkpoint(
state_dict=self._context.model.state_dict(),
)
else:
# resume from the assigned checkpoint or assigned iteration
self._context.epoch = max(checkpoint.epoch, self.config.start_epoch)
self._context.iteration = max(checkpoint.iteration, self.config.start_batch)
self._context.model.load_state_dict(checkpoint.state_dict)
self._context.checkpoint = checkpoint
return self
def with_dataloader(self) -> Self:
# fix: change batch level iteration to sample level offset
config = self.config
sampler_offset = self._context.iteration * config.batch_size
resumeable_sampler = ResumableDistributedSampler(
data_source=config.dataset,
start_epoch=self._context.epoch,
start_iter=sampler_offset,
seed=config.random_seed,
)
dataloader = DataLoader(
config.dataset,
batch_size=config.batch_size,
sampler=resumeable_sampler,
num_workers=config.num_workers,
pin_memory=config.pin_memory,
prefetch_factor=config.prefetch_factor,
)
self._context.dataloader = dataloader
return self
def with_strategy(self) -> Self:
self._context.strategy = StrategyFactory.create(
model=self._context.model,
train_type=self.config.strategy,
device=get_current_device(),
**self.config.extra_kwargs,
)
return self
def build(self) -> TrainContext:
return self._context
+105
View File
@@ -0,0 +1,105 @@
import logging
from typing import Optional, List
from astrai.config import TrainConfig
from astrai.trainer.train_callback import (
TrainCallback,
ProgressBarCallback,
CheckpointCallback,
MetricLoggerCallback,
GradientClippingCallback,
SchedulerCallback,
)
from astrai.trainer.train_context import TrainContext, TrainContextBuilder
from astrai.data.serialization import Checkpoint
from astrai.parallel.setup import spawn_parallel_fn
logger = logging.getLogger(__name__)
class Trainer:
def __init__(
self, train_config: TrainConfig, callbacks: Optional[List[TrainCallback]] = None
):
self.train_config = train_config
default_callbacks = self._get_default_callbacks()
self.callbacks = (
default_callbacks + callbacks if callbacks else default_callbacks
)
def _get_default_callbacks(self) -> List[TrainCallback]:
train_config = self.train_config
return [
ProgressBarCallback(train_config.n_epoch),
CheckpointCallback(train_config.ckpt_dir, train_config.ckpt_interval),
MetricLoggerCallback(train_config.ckpt_dir, train_config.ckpt_interval),
GradientClippingCallback(train_config.max_grad_norm),
SchedulerCallback(),
]
def _build_context(self, checkpoint: Optional[Checkpoint]) -> TrainContext:
return (
TrainContextBuilder(self.train_config)
.with_checkpoint(checkpoint)
.with_dataloader()
.with_strategy()
.build()
)
def _call_callbacks(self, method_name: str, context: TrainContext):
for callback in self.callbacks:
method = getattr(callback, method_name, None)
if method:
method(context)
def train(self, checkpoint: Optional[Checkpoint] = None):
config = self.train_config
spawn_parallel_fn(
self._train_impl,
backend=config.backend,
world_size=config.nprocs,
master_addr=config.master_addr,
master_port=config.master_port,
device_type=config.device_type,
device_ids=config.device_ids,
checkpoint=checkpoint,
)
def _train_impl(self, checkpoint: Optional[Checkpoint] = None) -> Checkpoint:
context = self._build_context(checkpoint)
self._call_callbacks("on_train_begin", context)
try:
context.model.train()
# 1.epoch
for epoch in range(context.epoch, self.train_config.n_epoch):
context.epoch = epoch
self._call_callbacks("on_epoch_begin", context)
for batch in context.dataloader:
if context.iteration % self.train_config.accumulation_steps == 0:
# 2. step
self._call_callbacks("on_step_begin", context)
context.optimizer.step()
context.optimizer.zero_grad()
self._call_callbacks("on_step_end", context)
# 3. batch
self._call_callbacks("on_batch_begin", context)
loss = context.strategy(batch)
context.loss = loss.item()
context.iteration += 1
# to make the loss normalized by accumulation steps
stand_loss = loss / self.train_config.accumulation_steps
stand_loss.backward()
self._call_callbacks("on_batch_end", context)
self._call_callbacks("on_epoch_end", context)
except Exception as e:
logger.error(f"Training failed: {str(e)}", exc_info=True)
self._call_callbacks("on_error", context)
raise
finally:
self._call_callbacks("on_train_end", context)