Add new model weights for bilinear/bilinear FPN up/down sample. Add support for training models with mish or other activations with torchscript considerations.

Author: rwightman

README.md

@@ -16,6 +16,12 @@ Aside from the default model configs, there is a lot of flexibility to facilitat
 
 ## Updates
 
+### 2021-02-18
+* Add some new model weights with bilinear interpolation for upsample and downsample in FPN.
+  * 40.9 mAP - `efficientdet_q1`  (replace prev model at 40.6)
+  * 43.2 mAP -`cspresdet50`
+  * 45.2 mAP - `cspdarkdet5m`
+
 ### 2020-12-07
 * Training w/ fully jit scripted model + bench (`--torchscript`) is possible with inclusion of ModelEmaV2 from `timm` and previous torchscript compat additions. Big speed gains for CPU bound training.
 * Add weights for alternate FPN layouts. QuadFPN experiments (`efficientdet_q0/q1/q2`) and CSPResDeXt + PAN (`cspresdext50pan`). See updated table below. Special thanks to [Artus](https://twitter.com/artuskg) for providing resources for training the Q2 model.
@@ -114,11 +120,13 @@ The table below contains models with pretrained weights. There are quite a numbe
 | efficientdet_q0.pth | 35.7 | TBD | N/A | N/A | 4.13 |
 | efficientdet_d1.pth | 39.4 | 39.5 | 39.1 | 39.6 | 6.62 |
 | tf_efficientdet_d1.pth | 40.1 | TBD | 40.2 | 40.5 | 6.63 |
-| efficientdet_q1.pth | 40.6 | TBD | N/A | N/A | 6.98 |
+| efficientdet_q1.pth | 40.9 | TBD | N/A | N/A | 6.98 |
 | cspresdext50pan | 41.2 | TBD | N/A | N/A | 22.2 |
 | resdet50 | 41.6 | TBD | N/A | N/A | 27.6 |
 | efficientdet_q2.pth | 43.1 | TBD | N/A | N/A | 8.81 |
+| cspresdet50 | 43.2 | TBD | N/A | N/A | 24.3 |
 | tf_efficientdet_d2.pth | 43.4 | TBD | 42.5 | 43 | 8.10 |
+| cspdarkdet53m | 45.2 | TBD | N/A | N/A | 35.6 |
 | tf_efficientdet_d3.pth | 47.1 | TBD | 47.2 | 47.5 | 12.0 |
 | tf_efficientdet_d4.pth | 49.2 | TBD | 49.3 | 49.7 | 20.7 |
 | tf_efficientdet_d5.pth | 51.2 | TBD | 51.2 | 51.5 | 33.7 |

effdet/config/model_config.py

@@ -18,6 +18,7 @@ def default_detection_model_configs():
 
     h.backbone_name = 'tf_efficientnet_b1'
     h.backbone_args = None  # FIXME sort out kwargs vs config for backbone creation
+    h.backbone_indices = None
 
     # model specific, input preprocessing parameters
     h.image_size = (640, 640)
@@ -167,21 +168,21 @@ def default_detection_model_configs():
     cspresdet50=dict(
         name='cspresdet50',
         backbone_name='cspresnet50',
-        image_size=(640, 640),
+        image_size=(768, 768),
         aspect_ratios=[1.0, 2.0, 0.5],
         fpn_channels=88,
         fpn_cell_repeats=4,
         box_class_repeats=3,
         pad_type='',
         act_type='leaky_relu',
         head_act_type='silu',
-        downsample_type='max',
+        downsample_type='bilinear',
         upsample_type='bilinear',
         redundant_bias=False,
         separable_conv=False,
         head_bn_level_first=True,
         backbone_args=dict(drop_path_rate=0.2),
-        url='',
+        url='https://github.com/rwightman/efficientdet-pytorch/releases/download/v0.1/cspresdet50b-386da277.pth',
     ),
     cspresdext50=dict(
         name='cspresdext50',
@@ -230,8 +231,30 @@ def default_detection_model_configs():
         separable_conv=False,
         head_bn_level_first=True,
         backbone_args=dict(drop_path_rate=0.2),
+        backbone_indices=(3, 4, 5),
         url='',
     ),
+    cspdarkdet53m=dict(
+        name='cspdarkdet53m',
+        backbone_name='cspdarknet53',
+        image_size=(768, 768),
+        aspect_ratios=[1.0, 2.0, 0.5],
+        fpn_channels=96,
+        fpn_cell_repeats=4,
+        box_class_repeats=3,
+        pad_type='',
+        fpn_name='qufpn_fa',
+        act_type='leaky_relu',
+        head_act_type='mish',
+        downsample_type='bilinear',
+        upsample_type='bilinear',
+        redundant_bias=False,
+        separable_conv=False,
+        head_bn_level_first=True,
+        backbone_args=dict(drop_path_rate=0.2),
+        backbone_indices=(3, 4, 5),
+        url='https://github.com/rwightman/efficientdet-pytorch/releases/download/v0.1/cspdarkdet53m-79062b2d.pth',
+    ),
     mixdet_m=dict(
         name='mixdet_m',
         backbone_name='mixnet_m',
@@ -328,10 +351,12 @@ def default_detection_model_configs():
         box_class_repeats=3,
         pad_type='',
         fpn_name='qufpn_fa',  # quad-fpn + fast attn experiment
+        downsample_type='bilinear',
+        upsample_type='bilinear',
         redundant_bias=False,
         head_bn_level_first=True,
         backbone_args=dict(drop_path_rate=0.2),
-        url='https://github.com/rwightman/efficientdet-pytorch/releases/download/v0.1/efficientdet_q1-b238aba5.pth',
+        url='https://github.com/rwightman/efficientdet-pytorch/releases/download/v0.1/efficientdet_q1b-d0612140.pth',
     ),
     efficientdet_q2=dict(
         name='efficientdet_q2',

effdet/efficientdet.py

@@ -557,7 +557,8 @@ def __init__(self, config, pretrained_backbone=True, alternate_init=False):
         self.config = config
         set_config_readonly(self.config)
         self.backbone = create_model(
-            config.backbone_name, features_only=True, out_indices=(2, 3, 4),
+            config.backbone_name, features_only=True,
+            out_indices=self.config.backbone_indices or (2, 3, 4),
             pretrained=pretrained_backbone, **config.backbone_args)
         feature_info = get_feature_info(self.backbone)
         self.fpn = BiFpn(self.config, feature_info)

train.py

@@ -40,6 +40,7 @@
 from effdet.data import resolve_input_config, SkipSubset
 from effdet.anchors import Anchors, AnchorLabeler
 from timm.models import resume_checkpoint, load_checkpoint
+from timm.models.layers import set_layer_config
 from timm.utils import *
 from timm.optim import create_optimizer
 from timm.scheduler import create_scheduler
@@ -267,20 +268,21 @@ def main():
 
     torch.manual_seed(args.seed + args.rank)
 
-    model = create_model(
-        args.model,
-        bench_task='train',
-        num_classes=args.num_classes,
-        pretrained=args.pretrained,
-        pretrained_backbone=args.pretrained_backbone,
-        redundant_bias=args.redundant_bias,
-        label_smoothing=args.smoothing,
-        legacy_focal=args.legacy_focal,
-        jit_loss=args.jit_loss,
-        soft_nms=args.soft_nms,
-        bench_labeler=args.bench_labeler,
-        checkpoint_path=args.initial_checkpoint,
-    )
+    with set_layer_config(scriptable=args.torchscript):
+        model = create_model(
+            args.model,
+            bench_task='train',
+            num_classes=args.num_classes,
+            pretrained=args.pretrained,
+            pretrained_backbone=args.pretrained_backbone,
+            redundant_bias=args.redundant_bias,
+            label_smoothing=args.smoothing,
+            legacy_focal=args.legacy_focal,
+            jit_loss=args.jit_loss,
+            soft_nms=args.soft_nms,
+            bench_labeler=args.bench_labeler,
+            checkpoint_path=args.initial_checkpoint,
+        )
     model_config = model.config  # grab before we obscure with DP/DDP wrappers
 
     if args.local_rank == 0:

validate.py

@@ -11,9 +11,8 @@
 
 from effdet import create_model, create_evaluator, create_dataset, create_loader
 from effdet.data import resolve_input_config
-from effdet.evaluator import CocoEvaluator, PascalEvaluator
 from timm.utils import AverageMeter, setup_default_logging
-
+from timm.models.layers import set_layer_config
 
 has_apex = False
 try:
@@ -107,16 +106,17 @@ def validate(args):
     args.prefetcher = not args.no_prefetcher
 
     # create model
-    bench = create_model(
-        args.model,
-        bench_task='predict',
-        num_classes=args.num_classes,
-        pretrained=args.pretrained,
-        redundant_bias=args.redundant_bias,
-        soft_nms=args.soft_nms,
-        checkpoint_path=args.checkpoint,
-        checkpoint_ema=args.use_ema,
-    )
+    with set_layer_config(scriptable=args.torchscript):
+        bench = create_model(
+            args.model,
+            bench_task='predict',
+            num_classes=args.num_classes,
+            pretrained=args.pretrained,
+            redundant_bias=args.redundant_bias,
+            soft_nms=args.soft_nms,
+            checkpoint_path=args.checkpoint,
+            checkpoint_ema=args.use_ema,
+        )
     model_config = bench.config
 
     param_count = sum([m.numel() for m in bench.parameters()])