Update flops to use multi-adds and set var_exclude_expr for finetuning.

efficientdet/det_model_fn.py

@@ -565,6 +565,7 @@ def scaffold_fn():
           ckpt_scope=ckpt_scope,
           var_scope=var_scope,
           var_exclude_expr=params.get('var_exclude_expr', None))
+
       tf.train.init_from_checkpoint(checkpoint, var_map)
 
       return tf.train.Scaffold()

efficientdet/hparams_config.py

@@ -188,7 +188,8 @@ def default_detection_configs():
   h.lr_decay_method = 'cosine'
   h.moving_average_decay = 0.9998
   h.ckpt_var_scope = None  # ckpt variable scope.
-  h.var_exclude_expr = None   # exclude vars when loading pretrained ckpts.
+  # exclude vars when loading pretrained ckpts.
+  h.var_exclude_expr = '.*/class-predict/.*'  # exclude class weights in default
 
   h.backbone_name = 'efficientnet-b1'
   h.backbone_config = None

efficientdet/utils.py

@@ -19,8 +19,8 @@
 # gtype import
 from __future__ import print_function
 
-import re
 import os
+import re
 from absl import logging
 import numpy as np
 import tensorflow.compat.v1 as tf
@@ -304,6 +304,8 @@ def num_params_flops(readable_format=True):
   options['output'] = 'none'
   flops = tf.profiler.profile(
       tf.get_default_graph(), options=options).total_float_ops
+  # We use flops to denote multiply-adds, which is counted as 2 ops in tfprof.
+  flops = flops // 2
   if readable_format:
     nparams = float(nparams) * 1e-6
     flops = float(flops) * 1e-9