rm change to orbit in final, testing backbone training

coco_retinanet_gpu.yaml

+runtime:
+  distribution_strategy: 'mirrored'
+  mixed_precision_dtype: 'float16'
+  loss_scale: 'dynamic'
+  num_gpus: 2
+task:
+  annotation_file: Null 
+  init_checkpoint: Null
+  model:
+    num_classes: 80
+    input_size: [640, 640, 3]
+    min_level: 3
+    max_level: 7
+  losses:
+    l2_weight_decay: 0.0001
+  train_data:
+    input_path: Null
+    tfds_name: 'coco/2017'
+    tfds_split: 'train'
+    tfds_download: True
+    is_training: True
+    global_batch_size: 16
+    dtype: 'float16'
+    cycle_length: 5
+    decoder:
+      type: tfds_decoder
+    shuffle_buffer_size: 2
+  validation_data:
+    input_path: Null
+    tfds_name: 'coco/2017'
+    tfds_split: 'validation'
+    tfds_download: True
+    # tfds_skip_decoding_feature: source_id,image,height,width,groundtruth_classes,groundtruth_is_crowd,groundtruth_area,groundtruth_boxes
+    is_training: False
+    global_batch_size: 16
+    dtype: 'float16'
+    cycle_length: 10
+    decoder:
+      type: tfds_decoder
+    shuffle_buffer_size: 2
+trainer:
+  train_steps: 532224
+  validation_steps: 1564
+  validation_interval: 2000
+  steps_per_loop: 200 #59136
+  summary_interval: 200 #59136
+  checkpoint_interval: 10000
+  optimizer_config:
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+        #    learning_rate:
+        #      type: 'cosine'
+        #      cosine:
+        #        initial_learning_rate: 0.0021875
+        #        decay_steps: 4257792
+        #        alpha: 0.01
+# Stepwise version
+    learning_rate:
+      type: 'stepwise'
+      stepwise:
+        # boundaries: [26334, 30954]
+        boundaries: [421344, 495264]
+        # values: [0.28, 0.028, 0.0028]
+        values: [0.0175, 0.00175, 0.000175]
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 20480
+        warmup_learning_rate: 0.0001634375

official/vision/beta/projects/yolo/configs/experiments/csp_darknet53_tfds.yaml

@@ -16,14 +16,16 @@ task:
   train_data:
     tfds_name: 'imagenet2012'
     tfds_split: 'train'
-    tfds_data_dir: '~/tensorflow_datasets'
+    tfds_data_dir: '~/tensorflow_datasets/'
+    tfds_download: true
     is_training: true
     global_batch_size: 128
     dtype: 'float16'
   validation_data:
     tfds_name: 'imagenet2012'
     tfds_split: 'validation'
-    tfds_data_dir: '~/tensorflow_datasets'
+    tfds_data_dir: '~/tensorflow_datasets/'
+    tfds_download: true
     is_training: true
     global_batch_size: 128
     dtype: 'float16'

official/vision/beta/projects/yolo/configs/experiments/darknet53_tfds.yaml

@@ -15,14 +15,16 @@ task:
   train_data:
     tfds_name: 'imagenet2012'
     tfds_split: 'train'
-    tfds_data_dir: '~/tensorflow_datasets'
+    tfds_data_dir: '~/tensorflow_datasets/'
+    tfds_download: true
     is_training: true
     global_batch_size: 128
     dtype: 'float16'
   validation_data:
     tfds_name: 'imagenet2012'
     tfds_split: 'validation'
-    tfds_data_dir: '~/tensorflow_datasets'
+    tfds_data_dir: '~/tensorflow_datasets/'
+    tfds_download: true
     is_training: true
     global_batch_size: 128
     dtype: 'float16'

official/vision/beta/train.py

@@ -35,6 +35,8 @@ FLAGS = flags.FLAGS
 def main(_):
   gin.parse_config_files_and_bindings(FLAGS.gin_file, FLAGS.gin_params)
   params = train_utils.parse_configuration(FLAGS)
+  import pprint
+  pprint.pprint(params.as_dict())
   model_dir = FLAGS.model_dir
   if 'train' in FLAGS.mode:
     # Pure eval modes do not output yaml files. Otherwise continuous eval job

orbit/utils/common.py

@@ -82,7 +82,8 @@ def make_distributed_dataset(strategy, dataset_or_fn, *args, **kwargs):
       kwargs["input_context"] = input_context
     return dataset_or_fn(*args, **kwargs)
   
-  return strategy.distribute_datasets_from_function(dataset_fn)
+  return strategy.experimental_distribute_datasets_from_function(dataset_fn)
+  #return strategy.distribute_datasets_from_function(dataset_fn)
 
 
 def get_value(x):

test.yaml

+runtime:
+  distribution_strategy: 'mirrored'
+  mixed_precision_dtype: 'float32'
+  loss_scale: 'dynamic'
+  num_gpus: 1
+task:
+  init_checkpoint: Null
+  model:
+    num_classes: 80
+    input_size: [640, 640, 3]
+    min_level: 3
+    max_level: 7
+  losses:
+    l2_weight_decay: 0.0001
+  train_data:
+    input_path: Null
+    tfds_name: 'coco/2017'
+    tfds_split: 'train'
+    tfds_download: True
+    is_training: True
+    global_batch_size: 2
+    dtype: 'float16'
+    cycle_length: 5
+    decoder:
+      type: tfds_decoder
+    shuffle_buffer_size: 2
+  validation_data:
+    input_path: Null
+    tfds_name: 'coco/2017'
+    tfds_split: 'validation'
+    tfds_download: True
+    # tfds_skip_decoding_feature: source_id,image,height,width,groundtruth_classes,groundtruth_is_crowd,groundtruth_area,groundtruth_boxes
+    is_training: False
+    global_batch_size: 2
+    dtype: 'float16'
+    cycle_length: 10
+    decoder:
+      type: tfds_decoder
+    shuffle_buffer_size: 2
+trainer:
+  train_steps: 4257792
+  validation_steps: 2500
+  validation_interval: 5000
+  steps_per_loop: 100 #59136
+  summary_interval: 100 #59136
+  checkpoint_interval: 59136
+  optimizer_config:
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+        #    learning_rate:
+        #      type: 'cosine'
+        #      cosine:
+        #        initial_learning_rate: 0.0021875
+        #        decay_steps: 4257792
+        #        alpha: 0.01
+# Stepwise version
+    learning_rate:
+      type: 'stepwise'
+      stepwise:
+        # boundaries: [26334, 30954]
+        boundaries: [3370752, 3962112]
+        # values: [0.28, 0.028, 0.0028]
+        values: [0.0021875, 0.00021875, 0.000021875]
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 64000
+        warmup_learning_rate: 0.0000523

tfds_coco_example_decoder.py_old

+import tensorflow_datasets as tfds 
+import tensorflow as tf
+from official.vision.beta.dataloaders import decoder
+
+import matplotlib.pyplot as plt
+import cv2
+
+
+class TfdsExampleDecoder(decoder.Decoder):
+  """Tensorflow Dataset Example proto decoder."""
+  def __init__(self,
+               include_mask=False,
+               regenerate_source_id=False):
+    self._include_mask = include_mask
+    self._regenerate_source_id = regenerate_source_id
+
+  def decode(self, serialized_example):
+    """Decode the serialized example.
+    Args:
+      serialized_example: a single serialized tf.Example string.
+    Returns:
+      decoded_tensors: a dictionary of tensors with the following fields:
+        - source_id: a string scalar tensor.
+        - image: a uint8 tensor of shape [None, None, 3].
+        - height: an integer scalar tensor.
+        - width: an integer scalar tensor.
+        - groundtruth_classes: a int64 tensor of shape [None].
+        - groundtruth_is_crowd: a bool tensor of shape [None].
+        - groundtruth_area: a float32 tensor of shape [None].
+        - groundtruth_boxes: a float32 tensor of shape [None, 4].
+        - groundtruth_instance_masks: a float32 tensor of shape
+            [None, None, None].
+        - groundtruth_instance_masks_png: a string tensor of shape [None].
+    """
+    decoded_tensors = {
+        'source_id': serialized_example['image/id'],
+        'image': serialized_example['image'],
+        'height': tf.shape(serialized_example['image'])[0],
+        'width':  tf.shape(serialized_example['image'])[1],
+        'groundtruth_classes': serialized_example['objects']['label'],
+        'groundtruth_is_crowd': serialized_example['objects']['is_crowd'],
+        'groundtruth_area': serialized_example['objects']['area'],
+        'groundtruth_boxes': serialized_example['objects']['bbox'],
+    }
+    return decoded_tensors
+