Weights and Biases Callback for Elegy

elegy/callbacks/__init__.py

@@ -9,6 +9,7 @@
 from .sigint import SigInt
 from .tensorboard import TensorBoard
 from .terminate_nan import TerminateOnNaN
+from .wandb_callback import WandbCallback
 
 __all__ = [
     "CallbackList",
@@ -21,4 +22,5 @@
     "RemoteMonitor",
     "CSVLogger",
     "TensorBoard",
+    "WandbCallback"
 ]

elegy/callbacks/wandb_callback.py

@@ -0,0 +1,211 @@
+# Implementation based on tf.keras.callbacks.py and elegy.callbacks.TensorBoard
+# https://github.com/tensorflow/tensorflow/blob/v2.2.0/tensorflow/python/keras/callbacks.py
+# https://github.com/poets-ai/elegy/blob/master/elegy/callbacks/tensorboard.py
+
+
+import math
+import wandb
+from typing import Union, Optional, Dict
+
+from .callback import Callback
+
+
+class WandbCallback(Callback):
+    """
+    Callback that streams epoch results to a [Weights & Biases](https://wandb.ai/) run.
+
+    ```python
+    wandb.login()
+    wandb_logger = WandbCallback(project="sample-wandb-project", job_type="train")
+    model.fit(X_train, Y_train, callbacks=[wandb_logger])
+    ```
+    """
+
+    def __init__(
+        self,
+        project: Optional[str] = None,
+        name: Optional[str] = None,
+        entity: Optional[str] = None,
+        job_type: Optional[str] = None,
+        config: Union[Dict, str, None] = None,
+        update_freq: Union[str, int] = "epoch",
+        save_model: bool = False,
+        monitor: str = "val_loss",
+        mode: str = "min",
+        **kwargs,
+    ):
+        """
+        Arguments:
+            project: (str, optional) The name of the project where you're sending the new run.
+                If the project is not specified, the run is put in an "Uncategorized" project.
+            name: (str, optional) A short display name for this run, which is how you'll
+                identify this run in the UI. By default we generate a random two-word name that
+                lets you easily cross-reference runs from the table to charts. Keeping these run
+                names short makes the chart legends and tables easier to read. If you're looking
+                for a place to save your hyperparameters, we recommend saving those in config.
+            entity: (str, optional) An entity is a username or team name where you're sending runs.
+                This entity must exist before you can send runs there, so make sure to create your
+                account or team in the UI before starting to log runs. If you don't specify an entity,
+                the run will be sent to your default entity, which is usually your username.
+            job_type: (str, optional) Specify the type of run, which is useful when you're grouping
+                runs together into larger experiments using group. For example, you might have multiple
+                jobs in a group, with job types like train and eval. Setting this makes it easy to
+                filter and group similar runs together in the UI so you can compare apples to apples.
+            config: (dict, argparse, absl.flags, str, optional) This sets `wandb.config`, a dictionary-like
+                object for saving inputs to your job, like hyperparameters for a model or settings for a
+                data preprocessing job. The config will show up in a table in the UI that you can use to
+                group, filter, and sort runs. Keys should not contain . in their names, and values should
+                be under 10 MB. If dict, argparse or `absl.flags`: will load the key value pairs into the
+                wandb.config object. If str: will look for a yaml file by that name, and load config from
+                that file into the `wandb.config` object.
+            update_freq: (str, int)`'batch'` or `'epoch'` or integer. When using `'batch'`, writes the
+                losses and metrics to TensorBoard after each batch. The same applies for `'epoch'`. If
+                using an integer, let's say `1000`, the callback will write the metrics and losses to
+                TensorBoard every 1000 batches. Note that writing too frequently to TensorBoard can slow
+                down your training.
+            save_model: (bool) Save a model when monitor beats all previous epochs if set to `True` otherwise
+                don't save models.
+            monitor: (str) name of metric to monitor. Defaults to `'val_loss'`.
+            mode: (str) one of {`'min'`, `'max'`, `'every'`}. `'min'` - save model when monitor is minimized.
+                `'max'` - save model when monitor is maximized. `'every'` - save model after every epoch.
+        """
+        super().__init__()
+        self.run = (
+            wandb.init(
+                project=project,
+                name=name,
+                entity=entity,
+                job_type=job_type,
+                config=config,
+                **kwargs,
+            )
+            if wandb.run is None
+            else wandb.run
+        )
+        self.keys = None
+        self.write_per_batch = True
+        self._constant_fields = ["size"]
+        self._constants = {}
+        self._save_model = save_model
+        self._monitor = monitor
+        self._mode = mode
+        self._monitor_metric_val = math.inf if mode == "min" else -math.inf
+        self._best_epoch = 0
+        self._model_path = f"model-best-0"
+        self._model_checkpoint = self.model
+        try:
+            self.update_freq = int(update_freq)
+        except ValueError as e:
+            self.update_freq = 1
+            if update_freq == "batch":
+                self.write_per_batch = True
+            elif update_freq == "epoch":
+                self.write_per_batch = False
+            else:
+                raise e
+
+    def _gather_configs(self):
+        module_attributes = vars(vars(self.model)["module"])
+        for _var in module_attributes:
+            if (
+                type(module_attributes[_var]) == str
+                or type(module_attributes[_var]) == int
+            ):
+                wandb.run.config[_var] = module_attributes[_var]
+
+    def _add_model_as_artifact(self, model_path: str, epoch: int):
+        artifact = wandb.Artifact(
+            f"model-{self.run.name}",
+            type="model",
+            metadata={
+                "epoch": epoch,
+                "model_path": model_path
+            }
+        )
+        artifact.add_dir(model_path)
+        self.run.log_artifact(artifact)
+
+    def on_train_begin(self, logs=None):
+        self.steps = self.params["steps"]
+        self.global_step = 0
+        self._gather_configs()
+
+    def on_train_batch_end(self, batch: int, logs=None):
+        if not self.write_per_batch:
+            return
+        logs = logs or {}
+        for key in self._constant_fields:
+            self._constants[key] = logs[key]
+            logs.pop(key, None)
+            logs.pop("val_" + key, None)
+        self.global_step = batch + self.current_epoch * (self.steps)
+        if self.global_step % self.update_freq == 0:
+            if self.keys is None:
+                self.keys = logs.keys()
+            for key in self.keys:
+                log_key = key
+                if log_key[:4] != "val_":
+                    log_key = "train_" + log_key
+                self.run.log({log_key: logs[key]}, step=self.global_step)
+
+    def on_epoch_begin(self, epoch: int, logs=None):
+        self.current_epoch = epoch
+
+    def on_epoch_end(self, epoch: int, logs=None):
+        logs = logs or {}
+        for key in self._constant_fields:
+            self._constants[key] = logs[key]
+            logs.pop(key, None)
+            logs.pop("val_" + key, None)
+
+        if self.keys is None:
+            self.keys = logs.keys()
+
+        if self.write_per_batch:
+            for key in logs:
+                log_key = key
+                if log_key[:4] != "val_":
+                    log_key = "train_" + log_key
+                self.run.log({log_key: logs[key]}, step=self.global_step)
+
+        elif epoch % self.update_freq == 0:
+            for key in logs:
+                log_key = key
+                if log_key[:4] != "val_":
+                    log_key = "train_" + log_key
+                self.run.log({log_key: logs[key]}, step=epoch)
+
+        if self._save_model:
+            if self._mode == "every":
+                self._best_epoch = epoch
+                self._model_path = f"model-{epoch + 1}-{self.run.name}"
+                print(f"Saving Model at {self._model_path}")
+                self._model_checkpoint = self.model
+                self._model_checkpoint.save(self._model_path)
+
+            elif self._mode == "min" and logs[self._monitor] < self._monitor_metric_val:
+                self._best_epoch = epoch
+                self._model_path = f"model-best-{epoch + 1}-{self.run.name}"
+                print(
+                    f"{self._monitor} decreased at epoch {epoch}. Saving Model at {self._model_path}"
+                )
+                self._model_checkpoint = self.model
+                self._model_checkpoint.save(self._model_path)
+                self._monitor_metric_val = logs[self._monitor]
+
+            elif self._mode == "max" and logs[self._monitor] > self._monitor_metric_val:
+                self._best_epoch = epoch
+                self._model_path = f"model-best-{epoch + 1}-{self.run.name}"
+                print(
+                    f"{self._monitor} increased at epoch {epoch}. Saving Model at {self._model_path}"
+                )
+                self._model_checkpoint = self.model
+                self._model_checkpoint.save(self._model_path)
+                self._monitor_metric_val = logs[self._monitor]
+
+            self._add_model_as_artifact(self._model_path, epoch=self._best_epoch)
+
+    def on_train_end(self, logs=None):
+        for key in self._constant_fields:
+            wandb.run.summary[key] = self._constants[key]
+        self.run.finish()