update configs

README.md

@@ -55,6 +55,8 @@ Please note that:
 
 For your convenience, we provide the following trained models (more models are coming soon).
 
+**ResNe(x)ts:**
+
 Model | Total training mem (GB) | Multi-scale training | Testing time / im | AP (minival) | AP (test-dev) | Link
 --- |:---:|:---:|:---:|:---:|:--:|:---:
 FCOS_R_50_FPN_1x | 29.3 | No | 71ms | 37.1 | 37.4 | [download](https://cloudstor.aarnet.edu.au/plus/s/dDeDPBLEAt19Xrl/download)
@@ -62,6 +64,13 @@ FCOS_R_101_FPN_2x | 44.1 | Yes | 74ms | 41.4 | 41.5 | [download](https://cloudst
 FCOS_X_101_32x8d_FPN_2x | 72.9 | Yes | 122ms | 42.5 | 42.7 | [download](https://cloudstor.aarnet.edu.au/plus/s/U5myBfGF7MviZ97/download)
 FCOS_X_101_64x4d_FPN_2x | 77.7 | Yes | 140ms | 43.0 | 43.2 | [download](https://cloudstor.aarnet.edu.au/plus/s/wpwoCi4S8iajFi9/download)
 
+*All ResNe(x)t models are trained with 16 images in a mini-batch.*
+
+**MobileNets:**
+Model | Training batch size | Multi-scale training | Testing time / im | AP (minival) | Link
+--- |:---:|:---:|:---:|:---:|:--:|:---:
+FCOS_R_50_FPN_1x | 29.3 | No | 71ms | 37.1 | 37.4 | [download](https://cloudstor.aarnet.edu.au/plus/s/dDeDPBLEAt19Xrl/download)
+
 [1] *1x and 2x mean the model is trained for 90K and 180K iterations, respectively.* \
 [2] *We report total training memory footprint on all GPUs instead of the memory footprint per GPU as in maskrcnn-benchmark*. \
 [3] *All results are obtained with a single model and without any test time data augmentation such as multi-scale, flipping and etc..* \

configs/fcos/fcos_MNV2_FPN_1x.yaml → configs/fcos/fcos_bn_bs16_MNV2_FPN_1x.yaml

@@ -5,10 +5,12 @@ MODEL:
   FCOS_ON: True
   BACKBONE:
     CONV_BODY: "MNV2-FPN-RETINANET"
+    FREEZE_CONV_BODY_AT: 0
   RESNETS:
     BACKBONE_OUT_CHANNELS: 256
   RETINANET:
     USE_C5: False # FCOS uses P5 instead of C5
+  USE_SYNCBN: False
 DATASETS:
   TRAIN: ("coco_2014_train", "coco_2014_valminusminival")
   TEST: ("coco_2014_minival",)

configs/fcos/fcos_syncbn_bs32_MNV2_FPN_1x.yaml

+MODEL:
+  META_ARCHITECTURE: "GeneralizedRCNN"
+  WEIGHT: "https://cloudstor.aarnet.edu.au/plus/s/xtixKaxLWmbcyf7/download#mobilenet_v2-ecbe2b5.pth"
+  RPN_ONLY: True
+  FCOS_ON: True
+  BACKBONE:
+    CONV_BODY: "MNV2-FPN-RETINANET"
+    FREEZE_CONV_BODY_AT: 0
+  RESNETS:
+    BACKBONE_OUT_CHANNELS: 256
+  RETINANET:
+    USE_C5: False # FCOS uses P5 instead of C5
+  USE_SYNCBN: True
+DATASETS:
+  TRAIN: ("coco_2014_train", "coco_2014_valminusminival")
+  TEST: ("coco_2014_minival",)
+INPUT:
+  MIN_SIZE_TRAIN: (800,)
+  MAX_SIZE_TRAIN: 1333
+  MIN_SIZE_TEST: 800
+  MAX_SIZE_TEST: 1333
+DATALOADER:
+  SIZE_DIVISIBILITY: 32
+SOLVER:
+  BASE_LR: 0.01
+  WEIGHT_DECAY: 0.0001
+  STEPS: (60000, 80000)
+  MAX_ITER: 90000
+  IMS_PER_BATCH: 32
+  WARMUP_METHOD: "constant"

maskrcnn_benchmark/config/defaults.py

@@ -34,6 +34,7 @@ _C.MODEL.CLS_AGNOSTIC_BBOX_REG = False
 # the path in paths_catalog. Else, it will use it as the specified absolute
 # path
 _C.MODEL.WEIGHT = ""
+_C.USE_SYNCBN = False
 
 
 # -----------------------------------------------------------------------------

tools/remove_solver_states.py

+# Set up custom environment before nearly anything else is imported
+# NOTE: this should be the first import (no not reorder)
+from maskrcnn_benchmark.utils.env import setup_environment  # noqa F401 isort:skip
+import argparse
+import os
+import torch
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Remove the solver states stored in a trained model")
+    parser.add_argument(
+        "model",
+        default="models/FCOS_R_50_FPN_1x.pth",
+        help="path to the input model file",
+    )
+
+    args = parser.parse_args()
+
+    model = torch.load(args.model)
+    del model["optimizer"]
+    del model["scheduler"]
+
+    filename_wo_ext, ext = os.path.splitext(args.model)
+    output_file = filename_wo_ext + "_wo_solver_states" + ext
+    torch.save(model, output_file)
+    print("Done. The model without solver states is saved to {}".format(output_file))
+
+if __name__ == "__main__":
+    main()
+

tools/train_net.py

@@ -31,7 +31,8 @@ def train(cfg, local_rank, distributed):
     device = torch.device(cfg.MODEL.DEVICE)
     model.to(device)
 
-    model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
+    if cfg.MODEL.USE_SYNCBN:
+        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
 
     optimizer = make_optimizer(cfg, model)
     scheduler = make_lr_scheduler(cfg, optimizer)