Some features that I needed when working on deformableDETR

alonet/deformable_detr/deformable_detr.py

@@ -90,7 +90,7 @@ def __init__(
         super().__init__()
         self.device = device
         self.num_feature_levels = num_feature_levels
-        self.transformer = transformer
+        self.backbone = backbone
         self.num_queries = num_queries
         self.return_intermediate_dec = return_intermediate_dec
         self.hidden_dim = transformer.d_model
@@ -105,9 +105,6 @@ def __init__(
         self.background_class = num_classes if self.activation_fn == "softmax" else None
         num_classes += 1 if self.activation_fn == "softmax" else 0  # Add bg class
 
-        self.class_embed = nn.Linear(self.hidden_dim, num_classes)
-        self.bbox_embed = MLP(self.hidden_dim, self.hidden_dim, 4, 3)
-        self.query_embed = nn.Embedding(num_queries, self.hidden_dim * 2)
         # Projection for Multi-scale features
         if num_feature_levels > 1:
             num_backbone_outs = len(backbone.strides) - 1  # Ignore backbone.layer1
@@ -138,8 +135,11 @@ def __init__(
                     )
                 ]
             )
+        self.query_embed = nn.Embedding(num_queries, self.hidden_dim * 2)
+        self.transformer = transformer
+        self.class_embed = nn.Linear(self.hidden_dim, num_classes)
+        self.bbox_embed = MLP(self.hidden_dim, self.hidden_dim, 4, 3)
 
-        self.backbone = backbone
         self.aux_loss = aux_loss
         self.with_box_refine = with_box_refine
         self.tracing = tracing
@@ -617,11 +617,15 @@ def build_decoder_layer(
             n_points=dec_n_points,
         )
 
-    def build_decoder(self, dec_layers: int = 6, return_intermediate_dec: bool = True):
+    def build_decoder(self, dec_layers: int = 6, return_intermediate_dec: bool = True, hidden_dim: int = 256, num_feature_levels: int = 4):
         """Build decoder layer
 
         Parameters
         ----------
+        hidden_dim : int, optional
+            Hidden dimension size, by default 256
+        num_feature_levels : int, optional
+            Number of feature levels, by default 4
         dec_layers : int, optional
             Number of decoder layers, by default 6
         return_intermediate_dec : bool, optional
@@ -632,7 +636,7 @@ def build_decoder(self, dec_layers: int = 6, return_intermediate_dec: bool = Tru
         :class:`~alonet.deformable.deformable_transformer.DeformableTransformerDecoder`
             Transformer decoder
         """
-        decoder_layer = self.build_decoder_layer()
+        decoder_layer = self.build_decoder_layer(hidden_dim=hidden_dim, num_feature_levels=num_feature_levels)
 
         return DeformableTransformerDecoder(decoder_layer, dec_layers, return_intermediate_dec)
 
@@ -679,7 +683,7 @@ def build_transformer(
         :mod:`Transformer <alonet.detr.transformer>`
             Transformer module
         """
-        decoder = self.build_decoder()
+        decoder = self.build_decoder(hidden_dim=hidden_dim, num_feature_levels=num_feature_levels)
 
         return DeformableTransformer(
             decoder=decoder,

alonet/detr/data_modules/coco_detection2detr.py

@@ -88,20 +88,18 @@ def setup(self, stage: Optional[str] = None):
 if __name__ == "__main__":
     # setup data
     loader_kwargs = dict(
-        name="rabbits",
-        train_folder="train",
-        train_ann="train/_annotations.coco.json",
-        val_folder="valid",
-        val_ann="valid/_annotations.coco.json",
+        name="coco",
+        train_folder="train2017",
+        train_ann="annotations/instances_train2017.json",
+        val_folder="val2017",
+        val_ann="annotations/instances_val2017.json",
     )
 
     args = CocoDetection2Detr.add_argparse_args(ArgumentParser()).parse_args()  # Help provider
     coco = CocoDetection2Detr(args, **loader_kwargs)
     coco.prepare_data()
     coco.setup()
-
-    samples = next(iter(coco.train_dataloader()))
-    samples[0].get_view().render()
-
-    samples = next(iter(coco.val_dataloader()))
-    samples[0].get_view().render()
+    iterator = iter(coco.train_dataloader())
+    for i in range(2):
+        samples = next(iterator)
+        samples[0].get_view().render()

alonet/detr/data_modules/data2detr.py

@@ -16,6 +16,7 @@
 
 import alonet
 import aloscene
+from torch.utils.data.sampler import RandomSampler, SequentialSampler
 
 
 class Data2Detr(pl.LightningDataModule):
@@ -122,6 +123,9 @@ def add_argparse_args(parent_parser: ArgumentParser, parser: _ArgumentGroup = No
         parser.add_argument(
             "--sample", action="store_true", help="Download a sample for train/val process (Default: %(default)s)"
         )
+        parser.add_argument(
+            "--sequential", action="store_true", help="Use sequential loading for train (Default: %(default)s)"
+        )
         return parent_parser
 
     def train_transform(self, frame: aloscene.Frame, same_on_sequence: bool = True, same_on_frames: bool = False):
@@ -217,7 +221,10 @@ def train_dataloader(self):
         torch.utils.data.DataLoader
             Dataloader for training process
         """
-        return self.train_dataset.train_loader(batch_size=self.batch_size, num_workers=self.num_workers)
+        return self.train_dataset.train_loader(
+            batch_size=self.batch_size,
+            num_workers=self.num_workers,
+            sampler=SequentialSampler if self.sequential else RandomSampler)
 
     def val_dataloader(self, sampler: torch.utils.data = None):
         """Get val dataloader

alonet/detr/detr.py

@@ -74,7 +74,12 @@ def __init__(
         tracing: bool = False,
     ):
         super().__init__()
+        self.backbone = backbone
         self.num_queries = num_queries
+        hidden_dim = transformer.d_model
+        self.hidden_dim = hidden_dim
+        self.query_embed = nn.Embedding(num_queries, hidden_dim)
+        self.input_proj = nn.Conv2d(backbone.num_channels, hidden_dim, kernel_size=1)
         self.transformer = transformer
         self.num_decoder_layers = transformer.decoder.num_layers
         self.num_classes = num_classes
@@ -87,15 +92,9 @@ def __init__(
         self.background_class = self.num_classes if background_class is None else background_class
         self.num_classes += 1
 
-        hidden_dim = transformer.d_model
-        self.hidden_dim = hidden_dim
-
         self.class_embed = self.build_class_embed()
         self.bbox_embed = self.build_bbox_embed()
 
-        self.query_embed = nn.Embedding(num_queries, hidden_dim)
-        self.input_proj = nn.Conv2d(backbone.num_channels, hidden_dim, kernel_size=1)
-        self.backbone = backbone
         self.aux_loss = aux_loss
         self.tracing = tracing
 

aloscene/mask.py

@@ -234,3 +234,22 @@ def _get_set_children(self, labels_set: Union[str, None] = None):
         else:
             labels = [None] * len(self)
         return labels
+
+
+    def _spatial_shift(self, shift_y: float, shift_x: float, **kwargs):
+            """
+            Spatially shift the Mask.
+            This function is empty for now because the 1st implementation created issues.
+            The only purpose of this is to stop triggering the warning message.
+            Parameters
+            ----------
+            shift_y: float
+                Shift percentage on the y axis. Could be negative or positive
+            shift_x: float
+                Shift percentage on the x axis. Could ne negative or positive.
+            Returns
+            -------
+            shifted_tensor: aloscene.AugmentedTensor
+                shifted tensor
+            """
+            return self