det,path: add more dropout

torchdrive/models/det_deform.py

@@ -43,7 +43,10 @@ def __init__(
         self.num_queries = num_queries
 
         self.query_embed = nn.Embedding(num_queries, dim)
-        self.reference_points_project = nn.Linear(dim, 2)
+        self.reference_points_project = nn.Sequential(
+            nn.Dropout(dropout),
+            nn.Linear(dim, 2),
+        )
 
         decoder_layer = DeformableTransformerDecoderLayer(
             d_model=dim,
@@ -58,8 +61,11 @@ def __init__(
             num_layers=num_layers,
         )
 
-        self.bbox_decoder = MLP(dim, dim, 9, num_layers=4)
-        self.class_decoder = nn.Conv1d(dim, num_classes + 1, 1)
+        self.bbox_decoder = MLP(dim, dim, 9, num_layers=4, dropout=dropout)
+        self.class_decoder = nn.Sequential(
+            nn.Dropout(dropout),
+            nn.Conv1d(dim, num_classes + 1, 1),
+        )
 
         bev_encoders = []
         for i in range(num_levels - 1, -1, -1):
@@ -78,6 +84,7 @@ def __init__(
                         group_width=dim,
                         bottleneck_multiplier=1.0,
                     ),
+                    nn.Dropout(dropout),
                     nn.Conv2d(dim, dim, 1),
                     LearnedPositionalEncoding2d((h * 2**i, w * 2**i), dim),
                 )

torchdrive/models/mlp.py

@@ -8,7 +8,12 @@ class MLP(nn.Module):
     """
 
     def __init__(
-        self, input_dim: int, hidden_dim: int, output_dim: int, num_layers: int
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        output_dim: int,
+        num_layers: int,
+        dropout: float = 0.0,
     ) -> None:
         super().__init__()
 
@@ -21,13 +26,15 @@ def __init__(
 
         for i in range(num_layers - 2):
             layers += [
+                nn.Dropout(dropout),
                 nn.Linear(hidden_dim, hidden_dim),
                 nn.ReLU(inplace=True),
             ]
 
-        layers.append(
+        layers += [
+            nn.Dropout(dropout),
             nn.Linear(hidden_dim, output_dim),
-        )
+        ]
 
         self.decoder = nn.Sequential(*layers)
 

torchdrive/models/path.py

@@ -136,6 +136,7 @@ def __init__(
                 group_width=bev_dim,
                 bottleneck_multiplier=1.0,
             ),
+            nn.Dropout(dropout),
             nn.Conv2d(bev_dim, dim, 1),
             LearnedPositionalEncoding2d(bev_shape, dim),
         )
@@ -147,22 +148,27 @@ def __init__(
             nn.Sequential(
                 nn.Linear(pos_dim, dim),
                 nn.ReLU(inplace=True),
+                nn.Dropout(dropout),
                 nn.Linear(dim, dim),
                 LearnedPositionalEncodingSeq(max_seq_len, dim),
             )
         )
         # pyre-fixme[4]: Attribute must be annotated.
         self.pos_decoder = compile_fn(
             nn.Sequential(
+                nn.Dropout(dropout),
                 nn.Linear(dim, dim),
                 nn.ReLU(inplace=True),
+                nn.Dropout(dropout),
                 nn.Linear(dim, pos_dim),
             )
         )
 
         static_features = 2 * pos_dim
         # pyre-fixme[4]: Attribute must be annotated.
-        self.static_encoder = compile_fn(MLP(static_features, dim, dim, num_layers=3))
+        self.static_encoder = compile_fn(
+            MLP(static_features, dim, dim, num_layers=3, dropout=dropout)
+        )
 
         self.transformer = StockTransformerDecoder(
             dim=dim,
@@ -267,6 +273,7 @@ def __init__(
         self.query_embed = nn.Embedding(num_queries, dim)
 
         self.bev_encoder = nn.Sequential(
+            nn.Dropout(dropout),
             models.regnet.AnyStage(
                 bev_dim,
                 bev_dim,
@@ -278,6 +285,7 @@ def __init__(
                 group_width=bev_dim,
                 bottleneck_multiplier=1.0,
             ),
+            nn.Dropout(dropout),
             nn.Conv2d(bev_dim, dim, 1),
             LearnedPositionalEncoding2d(bev_shape, dim),
         )
@@ -288,10 +296,13 @@ def __init__(
         # pyre-fixme[4]: Attribute must be annotated.
         self.pos_decoder = compile_fn(
             nn.Sequential(
+                nn.Dropout(dropout),
                 nn.Linear(decoder_dim, inter_dim),
                 nn.ReLU(inplace=True),
+                nn.Dropout(dropout),
                 nn.Linear(inter_dim, inter_dim),
                 nn.ReLU(inplace=True),
+                nn.Dropout(dropout),
                 nn.Linear(inter_dim, out_dim),
             )
         )
@@ -301,11 +312,14 @@ def __init__(
             dim_feedforward=dim * 4,
             nhead=num_heads,
             batch_first=True,
+            dropout=dropout,
         )
         self.transformer = nn.TransformerDecoder(decoder_layer, num_layers=num_layers)
 
         # pyre-fixme[4]: Attribute must be annotated.
-        self.static_encoder = compile_fn(MLP(static_features, dim, dim, num_layers=3))
+        self.static_encoder = compile_fn(
+            MLP(static_features, dim, dim, num_layers=3, dropout=dropout)
+        )
 
         transformer_init(self)
 

torchdrive/tasks/det.py

@@ -351,7 +351,7 @@ def forward(
 
         if len(unmatched_classes) > 0:
             losses["unmatched"] = (
-                F.cross_entropy(unmatched_classes, target_classes) * 40
+                F.cross_entropy(unmatched_classes, target_classes) * 20
             )
 
         if ctx.log_img:
@@ -366,7 +366,7 @@ def forward(
             ctx.add_scalar("classes/accuracy", self.accuracy.compute())
             self.accuracy.reset()
 
-        losses = {k: v for k, v in losses.items()}
+        losses = {k: v * 5 for k, v in losses.items()}
 
         return losses