make validation more comparable across runs

data/every_dream_validation.py

@@ -1,4 +1,3 @@
-import copy
 import json
 import logging
 import math
@@ -64,6 +63,9 @@ class EveryDreamValidator:
             with open(val_config_path, 'rt') as f:
                 self.config.update(json.load(f))
 
+        self.train_overlapping_dataloader_loss_offset = None
+        self.val_loss_offset = None
+
         self.loss_val_history = []
         self.val_loss_window_size = 4 # todo: arg for this?
 
@@ -100,28 +102,42 @@ class EveryDreamValidator:
                                          [Any, Any], tuple[torch.Tensor, torch.Tensor]]):
         if (epoch % self.every_n_epochs) == 0:
             if self.train_overlapping_dataloader is not None:
-                self._do_validation('stabilize-train', global_step, self.train_overlapping_dataloader,
+                mean_loss = self._calculate_validation_loss('stabilize-train',
+                                                            self.train_overlapping_dataloader,
                                                             get_model_prediction_and_target_callable)
+                if self.train_overlapping_dataloader_loss_offset is None:
+                    self.train_overlapping_dataloader_loss_offset = -mean_loss
+                self.log_writer.add_scalar(tag=f"loss/stabilize-train",
+                                           scalar_value=self.train_overlapping_dataloader_loss_offset + mean_loss,
+                                           global_step=global_step)
             if self.val_dataloader is not None:
-                val_loss = self._do_validation('val', global_step, self.val_dataloader, get_model_prediction_and_target_callable)
-            
-                self.loss_val_history.append(val_loss)
+                mean_loss = self._calculate_validation_loss('val',
+                                                            self.val_dataloader,
+                                                            get_model_prediction_and_target_callable)
+                if self.val_loss_offset is None:
+                    self.val_loss_offset = -mean_loss
+                self.log_writer.add_scalar(tag=f"loss/val",
+                                           scalar_value=self.val_loss_offset + mean_loss,
+                                           global_step=global_step)
+                self.loss_val_history.append(mean_loss)
                 if len(self.loss_val_history) > (self.val_loss_window_size * 2 + 1):
                     dy = np.diff(self.loss_val_history[-self.val_loss_window_size:])
                     if np.average(dy) > 0:
                         logging.warning(f"Validation loss shows diverging")
                         # todo: signal stop?
 
-    def _do_validation(self, tag, global_step, dataloader, get_model_prediction_and_target: Callable[
-        [Any, Any], tuple[torch.Tensor, torch.Tensor]]):
+    def _calculate_validation_loss(self, tag, dataloader, get_model_prediction_and_target: Callable[
+        [Any, Any], tuple[torch.Tensor, torch.Tensor]]) -> float:
         with torch.no_grad(), isolate_rng():
+            # ok to override seed here because we are in a `with isolate_rng():` block
+            random.seed(self.seed)
+            torch.manual_seed(self.seed)
+
             loss_validation_epoch = []
             steps_pbar = tqdm(range(len(dataloader)), position=1, leave=False)
             steps_pbar.set_description(f"{Fore.LIGHTCYAN_EX}Validate ({tag}){Style.RESET_ALL}")
 
             for step, batch in enumerate(dataloader):
-                # ok to override seed here because we are in a `with isolate_rng():` block
-                torch.manual_seed(self.seed + step)
                 model_pred, target = get_model_prediction_and_target(batch["image"], batch["tokens"])
 
                 loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
@@ -136,8 +152,6 @@ class EveryDreamValidator:
             steps_pbar.close()
 
         loss_validation_local = sum(loss_validation_epoch) / len(loss_validation_epoch)
-        self.log_writer.add_scalar(tag=f"loss/{tag}", scalar_value=loss_validation_local, global_step=global_step)
-
         return loss_validation_local
 
 
@@ -153,13 +167,10 @@ class EveryDreamValidator:
             val_items = list(disable_multiplier_and_flip(val_items))
             logging.info(f" * Removed {len(val_items)} images from the training set to use for validation")
         elif val_split_mode == 'manual':
-            args = Namespace(
-                aspects=aspects.get_aspect_buckets(self.resolution),
-                flip_p=0.0,
-                seed=self.seed,
-            )
-            val_data_root = self.config['val_data_root']
-            val_items = resolver.resolve_root(val_data_root, args)
+            val_data_root = self.config.get('val_data_root', None)
+            if val_data_root is None:
+                raise ValueError("Manual validation split requested but `val_data_root` is not defined in validation config")
+            val_items = self._load_manual_val_split(val_data_root)
             logging.info(f" * Loaded {len(val_items)} validation images from {val_data_root}")
         else:
             raise ValueError(f"Unrecognized validation split mode '{val_split_mode}'")
@@ -181,6 +192,17 @@ class EveryDreamValidator:
         stabilize_dataloader = build_torch_dataloader(stabilize_ed_batch, batch_size=self.batch_size)
         return stabilize_dataloader
 
+    def _load_manual_val_split(self, val_data_root: str):
+        args = Namespace(
+            aspects=aspects.get_aspect_buckets(self.resolution),
+            flip_p=0.0,
+            seed=self.seed,
+        )
+        val_items = resolver.resolve_root(val_data_root, args)
+        val_items.sort(key=lambda i: i.pathname)
+        random.shuffle(val_items)
+        return val_items
+
     def _build_ed_batch(self, items: list[ImageTrainItem], batch_size: int, tokenizer, name='val'):
         batch_size = self.batch_size
         seed = self.seed
@@ -196,5 +218,6 @@ class EveryDreamValidator:
             tokenizer=tokenizer,
             seed=seed,
             name=name,
+            crop_jitter=0
         )
         return ed_batch