Add Adam optimizer (#390)

* add adam optimizer

* fix bugs

* simple code

* fix bug

Author: fsx950223

efficientdet/anchors.py

@@ -335,6 +335,7 @@ def _generate_detections_tf(cls_outputs,
     scores = detections[:, 4]
     boxes = detections[:, :4]
 
+  image_size = utils.parse_image_size(image_size)
   detections = tf.stack([
       tf.cast(tf.tile(image_id, tf.shape(top_detection_idx)), tf.float32),
       tf.clip_by_value(boxes[:, 0], 0, image_size[0]) * image_scale,
@@ -455,10 +456,7 @@ def __init__(self, min_level, max_level, num_scales, aspect_ratios,
     self.num_scales = num_scales
     self.aspect_ratios = aspect_ratios
     self.anchor_scale = anchor_scale
-    if isinstance(image_size, int):
-      self.image_size = (image_size, image_size)
-    else:
-      self.image_size = image_size
+    self.image_size = utils.parse_image_size(image_size)
     self.feat_sizes = utils.get_feat_sizes(image_size, max_level)
     self.config = self._generate_configs()
     self.boxes = self._generate_boxes()

efficientdet/det_model_fn.py

@@ -502,8 +502,14 @@ def _model_outputs(inputs):
     ema_vars = utils.get_ema_vars()
 
   if mode == tf.estimator.ModeKeys.TRAIN:
-    optimizer = tf.train.MomentumOptimizer(
-        learning_rate, momentum=params['momentum'])
+    if params['optimizer'].lower() == 'sgd':
+      optimizer = tf.train.MomentumOptimizer(
+          learning_rate, momentum=params['momentum'])
+    elif params['optimizer'].lower() == 'adam':
+      optimizer = tf.train.AdamOptimizer(
+          learning_rate)
+    else:
+      raise ValueError('optimizers should be adam or sgd')
     if params['use_tpu']:
       optimizer = tf.tpu.CrossShardOptimizer(optimizer)
 

efficientdet/hparams_config.py

@@ -196,7 +196,11 @@ def default_detection_configs():
   h.is_training_bn = True
   # optimization
   h.momentum = 0.9
+  # 'adam', 'sgd'
+  h.optimizer = 'sgd'
+  # 0.008 for adam
   h.learning_rate = 0.08
+  # 0.0008 for adam
   h.lr_warmup_init = 0.008
   h.lr_warmup_epoch = 1.0
   h.first_lr_drop_epoch = 200.0

efficientdet/inference.py

@@ -249,6 +249,7 @@ def det_post_process_combined(params, cls_outputs, box_outputs, scales,
           tf.expand_dims(tf.range(batch_size), axis=1), [1, max_boxes_to_draw]),
       dtype=tf.float32)
   image_size = params['image_size']
+  image_size = utils.parse_image_size(image_size)
   ymin = tf.clip_by_value(nmsed_boxes[..., 0], 0, image_size[0]) * scales
   xmin = tf.clip_by_value(nmsed_boxes[..., 1], 0, image_size[1]) * scales
   ymax = tf.clip_by_value(nmsed_boxes[..., 2], 0, image_size[0]) * scales