Enable operators fusion (Conv + BN + ReLu) in FBNet V2

Summary:
Pull Request resolved: #592

X-link: facebookresearch/mobile-vision#172

The FBNet V2 consists of several basic blocks of pattern conv(2d) + BatchNorm + ReLU. These operators can be fused into one operator, greatly speeding up the computation in GPU.

This diff / PR adds option to build an FBNet V2 backbone in fused manner, i.e. it looks up recursively the blocks and sub-blocks for the known patterns, and fuse them using `mobile_cv.arch.utils.fuse_utils`

Differential Revision: D47395999

fbshipit-source-id: b3da40ade8b36238dc9920c59a638357d861f283

Author: antoine-tran

d2go/modeling/backbone/fbnet_cfg.py

@@ -66,6 +66,9 @@ def add_fbnet_v2_default_configs(_C):
     # https://github.com/rbgirshick/yacs/blob/master/yacs/config.py#L410
     _C.MODEL.FBNET_V2.NORM_ARGS = []
 
+    # Set up operators fusion option in the backbone's blocks
+    _C.MODEL.FBNET_V2.FUSE_OPS = False
+
     _C.MODEL.VT_FPN = CN()
 
     _C.MODEL.VT_FPN.IN_FEATURES = ["res2", "res3", "res4", "res5"]

d2go/modeling/backbone/fbnet_v2.py

@@ -197,12 +197,16 @@ def build_fbnet(cfg, name, in_channels):
     stages = []
     trunk_stride_per_stage = _get_stride_per_stage(arch_def_blocks)
     shape_spec_per_stage = []
+    fuse_ops = getattr(cfg.MODEL.FBNET_V2, "FUSE_OPS", False)
+    is_qat = getattr(cfg, "QAT", False)
     for i, stride_i in enumerate(trunk_stride_per_stage):
         stages.append(
             builder.build_blocks(
                 arch_def_blocks,
                 stage_indices=[i],
                 prefix_name=FBNET_BUILDER_IDENTIFIER + "_",
+                fuse_ops=fuse_ops,
+                is_qat=is_qat,
             )
         )
         shape_spec_per_stage.append(