refactor: 训练循环改为两重迭代并统一参数命名
- 训练循环从三重(epoch→batched→batch)改为二重(epoch→batch) - batch_size → batch_per_device, accumulation_steps → grad_accum_steps - scheduler 移入 step block 对齐 optimizer 更新步 - GradientClippingCallback 改用 on_step_begin 避免零梯度裁剪 - 移除 _train_impl 误导性的 -> Checkpoint 标注 - total_steps 修除为向下取整并精简为一行 - warmup_steps 改为 warmup_ratio (默认0.05)
This commit is contained in:
@@ -25,9 +25,9 @@ def test_callback_integration(base_test_env, random_dataset):
|
||||
scheduler_fn=scheduler_fn,
|
||||
ckpt_dir=base_test_env["test_dir"],
|
||||
n_epoch=1,
|
||||
batch_size=2,
|
||||
batch_per_device=2,
|
||||
ckpt_interval=3,
|
||||
accumulation_steps=1,
|
||||
grad_accum_steps=1,
|
||||
max_grad_norm=1.0,
|
||||
random_seed=42,
|
||||
device_type=base_test_env["device"],
|
||||
|
||||
Reference in New Issue
Block a user