Added adafactor implementation that handles stochastic rounding of up…

…date and accumulation

jobs/process/BaseSDTrainProcess.py

@@ -1750,7 +1750,10 @@ def run(self):
 
             with torch.no_grad():
                 # torch.cuda.empty_cache()
-                if self.train_config.optimizer.lower().startswith('dadaptation') or \
+                # if optimizer has get_lrs method, then use it
+                if hasattr(optimizer, 'get_learning_rates'):
+                    learning_rate = optimizer.get_learning_rates()[0]
+                elif self.train_config.optimizer.lower().startswith('dadaptation') or \
                         self.train_config.optimizer.lower().startswith('prodigy'):
                     learning_rate = (
                             optimizer.param_groups[0]["d"] *

toolkit/optimizer.py

@@ -77,18 +77,16 @@ def get_optimizer(
         except ImportError:
             raise ImportError("Please install lion_pytorch to use Lion optimizer -> pip install lion-pytorch")
     elif lower_type == 'adagrad':
-        optimizer = torch.optim.Adagrad(params, lr=float(learning_rate), eps=1e-6, **optimizer_params)
+        optimizer = torch.optim.Adagrad(params, lr=float(learning_rate), **optimizer_params)
     elif lower_type == 'adafactor':
-        # hack in stochastic rounding
+        from toolkit.optimizers.adafactor import Adafactor
         if 'relative_step' not in optimizer_params:
             optimizer_params['relative_step'] = False
         if 'scale_parameter' not in optimizer_params:
             optimizer_params['scale_parameter'] = False
         if 'warmup_init' not in optimizer_params:
             optimizer_params['warmup_init'] = False
         optimizer = Adafactor(params, lr=float(learning_rate), eps=1e-6, **optimizer_params)
-        from toolkit.util.adafactor_stochastic_rounding import step_adafactor
-        optimizer.step = step_adafactor.__get__(optimizer, Adafactor)
     else:
         raise ValueError(f'Unknown optimizer type {optimizer_type}')
     return optimizer

toolkit/optimizers/adafactor.py

@@ -0,0 +1,305 @@
+import math
+from typing import List
+import torch
+from toolkit.optimizers.optimizer_utils import copy_stochastic, stochastic_grad_accummulation
+
+
+class Adafactor(torch.optim.Optimizer):
+    """
+    Adafactor implementation with stochastic rounding accumulation and stochastic rounding on apply.
+    Modified from transformers Adafactor implementation to support stochastic rounding accumulation and apply.
+
+    AdaFactor pytorch implementation can be used as a drop in replacement for Adam original fairseq code:
+    https://github.com/pytorch/fairseq/blob/master/fairseq/optim/adafactor.py
+
+    Paper: *Adafactor: Adaptive Learning Rates with Sublinear Memory Cost* https://arxiv.org/abs/1804.04235 Note that
+    this optimizer internally adjusts the learning rate depending on the `scale_parameter`, `relative_step` and
+    `warmup_init` options. To use a manual (external) learning rate schedule you should set `scale_parameter=False` and
+    `relative_step=False`.
+
+    Arguments:
+        params (`Iterable[nn.parameter.Parameter]`):
+            Iterable of parameters to optimize or dictionaries defining parameter groups.
+        lr (`float`, *optional*):
+            The external learning rate.
+        eps (`Tuple[float, float]`, *optional*, defaults to `(1e-30, 0.001)`):
+            Regularization constants for square gradient and parameter scale respectively
+        clip_threshold (`float`, *optional*, defaults to 1.0):
+            Threshold of root mean square of final gradient update
+        decay_rate (`float`, *optional*, defaults to -0.8):
+            Coefficient used to compute running averages of square
+        beta1 (`float`, *optional*):
+            Coefficient used for computing running averages of gradient
+        weight_decay (`float`, *optional*, defaults to 0.0):
+            Weight decay (L2 penalty)
+        scale_parameter (`bool`, *optional*, defaults to `True`):
+            If True, learning rate is scaled by root mean square
+        relative_step (`bool`, *optional*, defaults to `True`):
+            If True, time-dependent learning rate is computed instead of external learning rate
+        warmup_init (`bool`, *optional*, defaults to `False`):
+            Time-dependent learning rate computation depends on whether warm-up initialization is being used
+
+    This implementation handles low-precision (FP16, bfloat) values, but we have not thoroughly tested.
+
+    Recommended T5 finetuning settings (https://discuss.huggingface.co/t/t5-finetuning-tips/684/3):
+
+        - Training without LR warmup or clip_threshold is not recommended.
+
+           - use scheduled LR warm-up to fixed LR
+           - use clip_threshold=1.0 (https://arxiv.org/abs/1804.04235)
+        - Disable relative updates
+        - Use scale_parameter=False
+        - Additional optimizer operations like gradient clipping should not be used alongside Adafactor
+
+    Example:
+
+    ```python
+    Adafactor(model.parameters(), scale_parameter=False, relative_step=False, warmup_init=False, lr=1e-3)
+    ```
+
+    Others reported the following combination to work well:
+
+    ```python
+    Adafactor(model.parameters(), scale_parameter=True, relative_step=True, warmup_init=True, lr=None)
+    ```
+
+    When using `lr=None` with [`Trainer`] you will most likely need to use [`~optimization.AdafactorSchedule`]
+    scheduler as following:
+
+    ```python
+    from transformers.optimization import Adafactor, AdafactorSchedule
+
+    optimizer = Adafactor(model.parameters(), scale_parameter=True, relative_step=True, warmup_init=True, lr=None)
+    lr_scheduler = AdafactorSchedule(optimizer)
+    trainer = Trainer(..., optimizers=(optimizer, lr_scheduler))
+    ```
+
+    Usage:
+
+    ```python
+    # replace AdamW with Adafactor
+    optimizer = Adafactor(
+        model.parameters(),
+        lr=1e-3,
+        eps=(1e-30, 1e-3),
+        clip_threshold=1.0,
+        decay_rate=-0.8,
+        beta1=None,
+        weight_decay=0.0,
+        relative_step=False,
+        scale_parameter=False,
+        warmup_init=False,
+    )
+    ```"""
+
+    def __init__(
+        self,
+        params,
+        lr=None,
+        eps=(1e-30, 1e-3),
+        clip_threshold=1.0,
+        decay_rate=-0.8,
+        beta1=None,
+        weight_decay=0.0,
+        scale_parameter=True,
+        relative_step=True,
+        warmup_init=False,
+    ):
+        if lr is not None and relative_step:
+            raise ValueError(
+                "Cannot combine manual `lr` and `relative_step=True` options")
+        if warmup_init and not relative_step:
+            raise ValueError(
+                "`warmup_init=True` requires `relative_step=True`")
+
+        defaults = {
+            "lr": lr,
+            "eps": eps,
+            "clip_threshold": clip_threshold,
+            "decay_rate": decay_rate,
+            "beta1": beta1,
+            "weight_decay": weight_decay,
+            "scale_parameter": scale_parameter,
+            "relative_step": relative_step,
+            "warmup_init": warmup_init,
+        }
+        super().__init__(params, defaults)
+
+        self.base_lrs: List[float] = [
+            lr for group in self.param_groups
+        ]
+
+        self.is_stochastic_rounding_accumulation = False
+
+        # setup stochastic grad accum hooks
+        for group in self.param_groups:
+            for param in group['params']:
+                if param.requires_grad and param.dtype != torch.float32:
+                    self.is_stochastic_rounding_accumulation = True
+                    param.register_post_accumulate_grad_hook(
+                        stochastic_grad_accummulation
+                    )
+
+    @staticmethod
+    def _get_lr(param_group, param_state):
+        rel_step_sz = param_group["lr"]
+        if param_group["relative_step"]:
+            min_step = 1e-6 * \
+                param_state["step"] if param_group["warmup_init"] else 1e-2
+            rel_step_sz = min(min_step, 1.0 / math.sqrt(param_state["step"]))
+        param_scale = 1.0
+        if param_group["scale_parameter"]:
+            param_scale = max(param_group["eps"][1], param_state["RMS"])
+        return param_scale * rel_step_sz
+
+    @staticmethod
+    def _get_options(param_group, param_shape):
+        factored = len(param_shape) >= 2
+        use_first_moment = param_group["beta1"] is not None
+        return factored, use_first_moment
+
+    @staticmethod
+    def _rms(tensor):
+        return tensor.norm(2) / (tensor.numel() ** 0.5)
+
+    @staticmethod
+    def _approx_sq_grad(exp_avg_sq_row, exp_avg_sq_col):
+        # copy from fairseq's adafactor implementation:
+        # https://github.com/huggingface/transformers/blob/8395f14de6068012787d83989c3627c3df6a252b/src/transformers/optimization.py#L505
+        r_factor = (exp_avg_sq_row / exp_avg_sq_row.mean(dim=-
+                    1, keepdim=True)).rsqrt_().unsqueeze(-1)
+        c_factor = exp_avg_sq_col.unsqueeze(-2).rsqrt()
+        return torch.mul(r_factor, c_factor)
+
+    def step_hook(self):
+        if not self.is_stochastic_rounding_accumulation:
+            return
+        # copy over stochastically rounded grads
+        for group in self.param_groups:
+            for param in group['params']:
+                if param.requires_grad and hasattr(param, "_accum_grad"):
+                    param.grad = param._accum_grad
+                    del param._accum_grad
+
+    # adafactor manages its own lr
+    def get_learning_rates(self):
+        lrs = [
+            self._get_lr(group, self.state[group["params"][0]])
+            for group in self.param_groups
+            if group["params"][0].grad is not None
+        ]
+        if len(lrs) == 0:
+            lrs = self.base_lrs  # if called before stepping
+        return lrs
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """
+        Performs a single optimization step
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        self.step_hook()
+        loss = None
+        if closure is not None:
+            loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+
+                grad = p.grad
+                if grad.dtype != torch.float32:
+                    grad = grad.to(torch.float32)
+                if grad.is_sparse:
+                    raise RuntimeError(
+                        "Adafactor does not support sparse gradients.")
+
+                state = self.state[p]
+                grad_shape = grad.shape
+
+                factored, use_first_moment = self._get_options(
+                    group, grad_shape)
+                # State Initialization
+                if len(state) == 0:
+                    state["step"] = 0
+
+                    if use_first_moment:
+                        # Exponential moving average of gradient values
+                        state["exp_avg"] = torch.zeros_like(grad)
+                    if factored:
+                        state["exp_avg_sq_row"] = torch.zeros(
+                            grad_shape[:-1]).to(grad)
+                        state["exp_avg_sq_col"] = torch.zeros(
+                            grad_shape[:-2] + grad_shape[-1:]).to(grad)
+                    else:
+                        state["exp_avg_sq"] = torch.zeros_like(grad)
+
+                    state["RMS"] = 0
+                else:
+                    if use_first_moment:
+                        state["exp_avg"] = state["exp_avg"].to(grad)
+                    if factored:
+                        state["exp_avg_sq_row"] = state["exp_avg_sq_row"].to(
+                            grad)
+                        state["exp_avg_sq_col"] = state["exp_avg_sq_col"].to(
+                            grad)
+                    else:
+                        state["exp_avg_sq"] = state["exp_avg_sq"].to(grad)
+
+                p_data_fp32 = p
+                if p.dtype != torch.float32:
+                    p_data_fp32 = p_data_fp32.clone().float()
+
+                state["step"] += 1
+                state["RMS"] = self._rms(p_data_fp32)
+                lr = self._get_lr(group, state)
+
+                beta2t = 1.0 - math.pow(state["step"], group["decay_rate"])
+                eps = group["eps"]
+                if isinstance(eps, tuple) or isinstance(eps, list):
+                    eps = eps[0]
+                update = (grad**2) + eps
+                if factored:
+                    exp_avg_sq_row = state["exp_avg_sq_row"]
+                    exp_avg_sq_col = state["exp_avg_sq_col"]
+
+                    exp_avg_sq_row.mul_(beta2t).add_(
+                        update.mean(dim=-1), alpha=(1.0 - beta2t))
+                    exp_avg_sq_col.mul_(beta2t).add_(
+                        update.mean(dim=-2), alpha=(1.0 - beta2t))
+
+                    # Approximation of exponential moving average of square of gradient
+                    update = self._approx_sq_grad(
+                        exp_avg_sq_row, exp_avg_sq_col)
+                    update.mul_(grad)
+                else:
+                    exp_avg_sq = state["exp_avg_sq"]
+
+                    exp_avg_sq.mul_(beta2t).add_(update, alpha=(1.0 - beta2t))
+                    update = exp_avg_sq.rsqrt().mul_(grad)
+
+                update.div_(
+                    (self._rms(update) / group["clip_threshold"]).clamp_(min=1.0))
+                update.mul_(lr)
+
+                if use_first_moment:
+                    exp_avg = state["exp_avg"]
+                    exp_avg.mul_(group["beta1"]).add_(
+                        update, alpha=(1 - group["beta1"]))
+                    update = exp_avg
+
+                if group["weight_decay"] != 0:
+                    p_data_fp32.add_(
+                        p_data_fp32, alpha=(-group["weight_decay"] * lr))
+
+                p_data_fp32.add_(-update)
+
+                if p.dtype != torch.float32:
+                    # apply stochastic rounding
+                    copy_stochastic(p, p_data_fp32)
+
+        return loss