Align with main

micromind/core.py

@@ -459,7 +459,7 @@ def train(
         epochs: int = 1,
         datasets: Dict = {},
         metrics: List[Metric] = [],
-        checkpointer: Optional[Checkpointer] = None,  # fix type hints
+        checkpointer: Optional[Checkpointer] = None,
         debug: Optional[bool] = False,
     ) -> None:
         """
@@ -500,6 +500,7 @@ def train(
                 f"Starting from epoch {self.start_epoch + 1}."
                 + f" Training is scheduled for {epochs} epochs."
             )
+
         for e in range(self.start_epoch + 1, epochs + 1):
             self.current_epoch = e
             pbar = tqdm(
@@ -521,6 +522,7 @@ def train(
                 with self.accelerator.autocast():
                     model_out = self(batch)
                     loss = self.compute_loss(model_out, batch)
+                    loss_epoch += loss.item()
 
                 self.accelerator.backward(loss)
                 self.opt.step()

micromind/utils/checkpointer.py

@@ -96,6 +96,7 @@ def __init__(
         if hparams is not None and os.path.exists(self.root_dir):
             with open(os.path.join(self.root_dir, "args.yaml"), "w") as args_f:
                 args_f.write(yaml.dump(vars(hparams)))
+
         else:
             warnings.warn(
                 "You did not specify the configuration to the checkpointer, \

recipes/image_classification/README.md

@@ -1,5 +1,6 @@
 ## Image classification
 
+
 **Disclaimer**: we will shortly release HuggingFace checkpoints for ImageNet, CIFAR-100, and CIFAR-10 for both PhiNet and XiNet.
 
 This image classification recipe uses the PyTorch image models library (`timm`) to augment the data. It supports most data augmentation strategies, and datasets of the original implementation. However, it is implemented using `micromind` and thus, it exploits all the exportability and functionalities of the library.

recipes/object_detection/README.md

@@ -1,6 +1,6 @@
 ## Object Detection using YOLO
 
-**[1 Dec 2023] Fix DDP handling and computational graph**
+**[1 Dec 2023]** Fix DDP handling and computational graph
 
 **Disclaimer**: we will shortly release HuggingFace checkpoints for COCO and VOC for both PhiNet and XiNet.
 

recipes/object_detection/train.py

@@ -172,6 +172,7 @@ def mAP(self, pred, batch):
                         batch["ori_shape"][i],
                     )
                 )
+
         batch_bboxes = torch.stack(batch_bboxes).to(self.device)
         mmAP = mean_average_precision(post_predictions, batch, batch_bboxes)