✨ Make log_steps and save_steps accept float value as a fraction …

…in `TrainerConfig`

docs/tutorial/training/trainer.md

@@ -111,9 +111,10 @@ Let's explore all the available parameters:
 - **metric_for_best_model** (str):Reference metric key to watch for the best model.
   Recommended to have a {`train.` | `evaluation.`} prefix (e.g, evaluation.f1, train.accuracy, etc.) but if not, defaults
   to `evaluation.{metric_for_best_model}`.
-- **save_steps** (int): Save the trainer outputs every `save_steps` steps. Set to `0` or `None` to ignore saving between
-  training steps.
-- **log_steps** (int): Save training metrics every `log_steps` steps.
+- **save_steps** (int): Save the trainer outputs every `save_steps` steps. Leave as None to ignore saving in-between training steps.
+If set to a float value between 0 and 1, it will be interpreted as a fraction of the total steps.
+- **log_steps** (int): Save training metrics every `log_steps` steps. If set to a float value between 0 and 1, it will be
+interpreted as a fraction of the total steps.
 - **checkpoints_dir** (str): Path to the checkpoints' folder. The actual files will be saved under `{output_dir}/{checkpoints_dir}`.
 - **logs_dir** (str): Path to the logs' folder. The actual log files will be saved under `{output_dir}/{logs_dir}`.
 

hezar/configs.py

@@ -453,10 +453,14 @@ class TrainerConfig(Config):
             `evaluation.{metric_for_best_model}`.
         save_freq (int) (DEPRECATED):
             Deprecated and renamed to `save_steps`.
+        save_enabled (bool):
+            Whether to save checkpoints at all. `False` disables even the saves in-between the epochs.
         save_steps (int):
-            Save the trainer outputs every `save_steps` steps. Leave as `0` to ignore saving between training steps.
+            Save the trainer outputs every `save_steps` steps. Leave as None to ignore saving in-between training steps.
+            If set to a float value between 0 and 1, it will be interpreted as a fraction of the total steps.
         log_steps (int):
-            Save training metrics every `log_steps` steps.
+            Save training metrics every `log_steps` steps. If set to a float value between 0 and 1, it will be
+            interpreted as a fraction of the total steps.
         checkpoints_dir (str):
             Path to the checkpoints' folder. The actual files will be saved under `{output_dir}/{checkpoints_dir}`.
         logs_dir (str):
@@ -492,7 +496,7 @@ class TrainerConfig(Config):
     metrics: List[str | MetricConfig] = None
     metric_for_best_model: str = "loss"
     save_enabled: bool = True
-    save_freq: int = None
+    save_freq: int = "deprecated"
     save_steps: int = None
     log_steps: int = None
     checkpoints_dir: str = "checkpoints"
@@ -519,15 +523,15 @@ def __post_init__(self):
                 self.metric_for_best_model = f"train.{self.metric_for_best_model}"
 
         # Validate steps
-        if self.save_steps is not None and self.save_steps % self.gradient_accumulation_steps != 0:
+        if isinstance(self.save_steps, int) and self.save_steps % self.gradient_accumulation_steps != 0:
             logger.warning(
                 f"It's recommended to set a `save_steps` dividable by `gradient_accumulation_steps`, "
                 f"otherwise, the saved model will have non-updated weights!\n"
                 f"`save_steps={self.save_steps}`, `gradient_accumulation_steps={self.gradient_accumulation_steps}`"
             )
 
         # Validate deprecated fields
-        if self.save_freq is not None:
+        if self.save_freq != "deprecated":
             logger.warning(
                 "Trainer argument `save_freq` is deprecated! Use `save_steps` (number of training steps per save)."
                 "Note that saving is also done at the end of each epoch unless you set `save_enabled` to `False` !"

hezar/trainer/trainer.py

@@ -159,11 +159,30 @@ def __init__(
         self.eval_dataset = eval_dataset
         self.data_collator = data_collator or getattr(self.train_dataset, "data_collator", None)
 
+        # Configure steps
         self.num_batches = math.ceil(len(self.train_dataset) / self.config.batch_size)
+        self.config.max_steps = (
+            math.ceil(self.config.max_steps * self.num_batches) * self.config.num_epochs
+            if isinstance(self.config.max_steps, float)
+            and 0 < self.config.max_steps <= 1
+            else self.config.max_steps
+        )
         self.total_steps = min(
             self.config.max_steps or self.num_batches * self.config.num_epochs,
             self.num_batches * self.config.num_epochs
         )
+        self.config.save_steps = (
+            math.ceil(self.config.save_steps * self.total_steps)
+            if isinstance(self.config.save_steps, float)
+            and 0 < self.config.save_steps <= 1
+            else self.config.save_steps
+        )
+        self.config.log_steps = (
+            math.ceil(self.config.log_steps * self.total_steps)
+            if isinstance(self.config.log_steps, float)
+            and 0 < self.config.log_steps <= 1
+            else self.config.log_steps
+        )
         self.steps_in_epoch = min(self.num_batches, self.total_steps)
         self.config.save_steps = self.steps_in_epoch if not self.config.save_steps else self.config.save_steps
         self.saves_in_epoch = math.floor(self.steps_in_epoch / self.config.save_steps)