Eval Callback for Vizard Detector

Summary:
Added an eval callback for the vizard detector to replace evaluation done natively in vizard. Evaluation is now called at the end of cfg.TEST.EVAL_PERIOD train steps and is a more direct migration of the d2go evaluator code.
- EMA not yet enabled, EMA eval code is commented out for now
- Modified COCO evaluation to accept cfg of shape DictConfig
- Removed eval_step override in unit, eval_step will not be used anymore

Reviewed By: luyich

Differential Revision: D65966876

fbshipit-source-id: 1e46d2f59919a883bb4a8b07d5c2417033df151d

Author: Leon Dong

detectron2/evaluation/coco_evaluation.py

@@ -11,6 +11,7 @@
 from collections import OrderedDict
 import pycocotools.mask as mask_util
 import torch
+from omegaconf import DictConfig
 from pycocotools.coco import COCO
 from pycocotools.cocoeval import COCOeval
 from tabulate import tabulate
@@ -114,7 +115,7 @@ def __init__(
             max_dets_per_image = [1, 10, max_dets_per_image]
         self._max_dets_per_image = max_dets_per_image
 
-        if tasks is not None and isinstance(tasks, CfgNode):
+        if tasks is not None and (isinstance(tasks, CfgNode) or isinstance(tasks, DictConfig)):
             kpt_oks_sigmas = (
                 tasks.TEST.KEYPOINT_OKS_SIGMAS if not kpt_oks_sigmas else kpt_oks_sigmas
             )