optimization package pr comments

official/vision/beta/projects/yolo/common/registry_imports.py

@@ -31,6 +31,5 @@ from official.vision.beta.projects.yolo.tasks import yolo
 
 # import optimization packages
 from official.vision.beta.projects.yolo.optimization import optimizer_factory
-from official.vision.beta.projects.yolo.optimization.configs import learning_rate_config
 from official.vision.beta.projects.yolo.optimization.configs import optimization_config
 from official.vision.beta.projects.yolo.optimization.configs import optimizer_config
\ No newline at end of file

official/vision/beta/projects/yolo/configs/experiments/yolov4/inference/512.yaml

+runtime:
+  distribution_strategy: 'mirrored'
+  mixed_precision_dtype: 'float16'
+  num_gpus: 1
+task:
+  smart_bias_lr: 0.1
+  model:
+    darknet_based_model: True
+    input_size: [512, 512, 3]
+    backbone:
+      type: 'darknet'
+      darknet:
+        model_id: 'cspdarknet53'
+        max_level: 5
+        min_level: 3
+    decoder:
+      type: yolo_decoder
+      yolo_decoder:
+        version: v4
+        type: regular
+        activation: leaky
+    head:
+      smart_bias: true
+    detection_generator:
+      box_type:
+        'all': original
+      scale_xy:
+        '5': 1.05
+        '4': 1.1
+        '3': 1.2
+      max_boxes: 200
+      nms_type: greedy
+      iou_thresh: 0.25
+      nms_thresh: 0.45
+      pre_nms_points: 500
+    loss:
+      use_scaled_loss: False
+      update_on_repeat: False
+      box_loss_type:  
+        'all': ciou
+      ignore_thresh:
+        'all': 0.7
+      iou_normalizer: 
+        'all': 0.07
+      cls_normalizer: 
+        'all': 1.0
+      object_normalizer: 
+        'all': 1.0
+      objectness_smooth: 
+        'all': 0.0
+      max_delta:
+        'all': 5.0
+    norm_activation:
+      activation: mish
+      norm_epsilon: 0.0001
+      norm_momentum: 0.99
+      use_sync_bn: true
+    num_classes: 80
+    anchor_boxes:
+      anchors_per_scale: 3
+      boxes: [box: [12, 16],   box: [19, 36],   box: [40, 28], 
+              box: [36, 75],   box: [76, 55],   box: [72, 146], 
+              box: [142, 110], box: [192, 243], box: [459, 401]]
+  train_data:
+    global_batch_size: 1
+    # dtype: float32
+    input_path: '/media/vbanna/DATA_SHARE/CV/datasets/COCO_raw/records/train*'
+    # is_training: true
+    # drop_remainder: true
+    # seed: 1000
+    parser:
+      mosaic: 
+        mosaic_frequency: 1.0
+        mixup_frequency: 0.0
+        mosaic_crop_mode: 'scale'
+        mosaic_center: 0.25
+        aug_scale_min: 0.1
+        aug_scale_max: 1.9
+      max_num_instances: 200
+      letter_box: True
+      random_flip: True
+      aug_rand_translate: 0.1
+      random_pad: False
+  validation_data:
+    # global_batch_size: 1
+    # dtype: float32
+    input_path: '/media/vbanna/DATA_SHARE/CV/datasets/COCO_raw/records/val*'
+#     is_training: false
+#     drop_remainder: true
+#     parser:
+#       max_num_instances: 200
+#       letter_box: True
+#       use_tie_breaker: True
+#       anchor_thresh: 0.213
+#   weight_decay: 0.000
+#   init_checkpoint: '../checkpoints/512-wd-baseline-e1'
+#   init_checkpoint_modules: 'all'
+#   annotation_file: null
+trainer:
+  optimizer_config:
+    ema: null
+#   train_steps: 500500 # 160 epochs at 64 batchsize -> 500500 * 64/2
+#   validation_steps: 625
+#   steps_per_loop: 1850
+#   summary_interval: 1850
+#   validation_interval: 9250 
+#   checkpoint_interval: 1850
+#   optimizer_config:
+#     ema:
+#       average_decay: 0.9998
+#       trainable_weights_only: False
+#       dynamic_decay: True
+#     learning_rate:
+#       type: stepwise
+#       stepwise:
+#         boundaries: [400000, 450000]
+#         name: PiecewiseConstantDecay
+#         values: [0.00131, 0.000131, 0.0000131] 
+#     optimizer:
+#       type: sgd_torch
+#       sgd_torch:
+#         momentum: 0.949
+#         momentum_start: 0.949
+#         nesterov: True
+#         warmup_steps: 1000
+#         weight_decay: 0.0005
+#         sim_torch: true
+#         name: SGD
+#     warmup:
+#       type: 'linear'
+#       linear:
+#         warmup_steps: 1000 #learning rate rises from 0 to 0.0013 over 1000 steps

official/vision/beta/projects/yolo/configs/experiments/yolov4/tpu/512.yaml

@@ -41,7 +41,7 @@ task:
         'all': 0.07
       cls_normalizer: 
         'all': 1.0
-      object_normalizer: 
+      obj_normalizer: 
         'all': 1.0
       objectness_smooth: 
         'all': 0.0

official/vision/beta/projects/yolo/optimization/configs/learning_rate_config.py

-# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Dataclasses for learning rate schedule config."""
-from typing import List, Optional
-
-import dataclasses
-from official.modeling.hyperparams import base_config
-# zfrom official.modeling.optimization.configs import learning_rate_config

official/vision/beta/projects/yolo/optimization/configs/optimization_config.py

@@ -21,9 +21,6 @@ from typing import Optional
 
 import dataclasses
 
-from official.modeling.hyperparams import base_config
-from official.modeling.hyperparams import oneof
-from official.vision.beta.projects.yolo.optimization.configs import learning_rate_config as lr_cfg
 from official.vision.beta.projects.yolo.optimization.configs import optimizer_config as opt_cfg
 from official.modeling.optimization.configs import optimization_config as optimization_cfg
 

official/vision/beta/projects/yolo/optimization/optimizer_factory.py

@@ -62,11 +62,28 @@ class OptimizerFactory(optimizer_factory.OptimizerFactory):
   """
 
   def get_bias_lr_schedule(self, bias_lr):
-    """Used to build additional learning rate schedules."""
-    temp = self._warmup_config.warmup_learning_rate
-    self._warmup_config.warmup_learning_rate = bias_lr
-    lr = self.build_learning_rate()
-    self._warmup_config.warmup_learning_rate = temp
+    """Build learning rate.
+
+    Builds learning rate from config. Learning rate schedule is built according
+    to the learning rate config. If learning rate type is consant,
+    lr_config.learning_rate is returned.
+
+    Returns:
+      tf.keras.optimizers.schedules.LearningRateSchedule instance. If
+      learning rate type is consant, lr_config.learning_rate is returned.
+    """
+    if self._lr_type == 'constant':
+      lr = self._lr_config.learning_rate
+    else:
+      lr = LR_CLS[self._lr_type](**self._lr_config.as_dict())
+
+    if self._warmup_config:
+      if self._warmup_type != 'linear':
+        raise ValueError("Smart Bias is only supported currently with a"
+                         "linear warm up.")
+      warm_up_cfg = self._warmup_config.as_dict()
+      warm_up_cfg["warmup_learning_rate"] = bias_lr
+      lr = WARMUP_CLS['linear'](lr, **warm_up_cfg)
     return lr
 
   @gin.configurable