sacred config for faster rcnn

dmlc · Jerryzcn · Jul 8, 2020 · May 6, 2020 · May 6, 2020 · May 6, 2020
commit 3bc987fbf8fce11d72b16b8b0b41fcdbdd4edada
diff --git a/gluoncv/pipelines/estimators/base_estimator.py b/gluoncv/pipelines/estimators/base_estimator.py
@@ -95,7 +95,7 @@ def __getitem__(self, key):
 
 class BaseEstimator:
     def __init__(self, config, logger=None):
-        self._log = logger if logger is not None else logging.getLogger(__name__)
+        self._logger = logger if logger is not None else logging.getLogger(__name__)
 
         # finalize the config
         r = self._ex.run('_get_config', config_updates=config, options={'--loglevel': 50})
@@ -105,7 +105,7 @@ def __init__(self, config, logger=None):
         logdir = r.config.get('logdir', None)
         self._logdir = os.path.abspath(logdir) if logdir else os.getcwd()
         config_file = os.path.join(self._logdir, config_fn)
-        save_config(r.config, self._log, config_file)
+        save_config(r.config, self._logger, config_file)
         self._cfg = DotDict(r.config)
         self._cfg.freeze()
         _random.seed(self._cfg.seed)

diff --git a/gluoncv/pipelines/estimators/rcnn/__init__.py b/gluoncv/pipelines/estimators/rcnn/__init__.py
@@ -1,3 +1,4 @@
 """R-CNN Estimator implementations"""
 
 from .faster_rcnn import FasterRCNNEstimator
+from .default import ex
diff --git a/gluoncv/pipelines/estimators/rcnn/default.py b/gluoncv/pipelines/estimators/rcnn/default.py
@@ -0,0 +1,92 @@
+from sacred import Experiment, Ingredient
+
+faster_rcnn = Ingredient('faster_rcnn')
+train_hp = Ingredient('train_hp')
+valid_hp = Ingredient('valid_hp')
+
+
+@faster_rcnn.config
+def faster_rcnn_default():
+    network = 'resnet50_v1b'  # base feature network
+    dataset = 'voc'  # dataset
+    nms_thresh = 0.5
+    nms_topk = -1
+    post_nms = -1
+    roi_mode = 'align'
+    roi_size = (7, 7)
+    strides = (4, 8, 16, 32, 64)
+    clip = 4.14
+    rpn_channel = 256
+    anchor_base_size = 16
+    anchor_aspect_ratio = (0.5, 1, 2)
+    anchor_scales = (2, 4, 8, 16, 32)
+    anchor_alloc_size = (384, 384)
+    rpn_nms_thresh = 0.7
+    max_num_gt = 100
+    gpus = (0, 1, 2, 3, 4, 5, 6, 7)
+    norm_layer = None
+    use_fpn = True
+    custom_model = True
+    num_fpn_filters = 256
+    num_box_head_conv = 4
+    num_box_head_conv_filters = 256
+    num_box_head_dense_filters = 1024
+    image_short = 800
+    image_max_size = 1333
+    amp = False
+    static_alloc = False
+
+
+@train_hp.config
+def train_cfg():
+    pretrained_base = True  # whether load the imagenet pre-trained base
+    batch_size = 16
+    start_epoch = 0
+    epochs = 26
+    lr = 0.01  # learning rate
+    lr_decay = 0.1  # decay rate of learning rate.
+    lr_decay_epoch = (20, 24)  # epochs at which learning rate decays
+    lr_mode = 'step'  # learning rate scheduler mode. options are step, poly and cosine
+    lr_warmup = 500  # number of iterations for warmup.
+    lr_warmup_factor = 1. / 3.  # starging lr warmup factor.
+    momentum = 0.9  # momentum
+    wd = 1e-4  # weight decay
+    log_interval = 100  # log interval
+    seed = 233
+    verbose = False
+    mixup = False
+    no_mixup_epochs = 20
+    rpn_smoothl1_rho = 0.001
+    rcnn_smoothl1_rho = 0.001
+    horovod = False
+    no_pretrained_base = False
+    rpn_train_pre_nms = 12000
+    rpn_train_post_nms = 2000
+    rpn_min_size = 1
+    rcnn_num_samples = 512
+    rcnn_pos_iou_thresh = 0.5
+    rcnn_pos_ratio = 0.25
+    executor_threads = 4
+
+
+@valid_hp.config
+def valid_cfg():
+    rpn_test_pre_nms = 6000
+    rpn_test_post_nms = 1000
+    val_interval = 1  # Epoch interval for validation
+
+
+ex = Experiment('faster_rcnn_default', ingredients=[train_hp, valid_hp, faster_rcnn])
+
+
+@ex.config
+def default_configs():
+    dataset = 'coco'
+    resume = ''
+    save_prefix = ''
+    save_interval = 1  # save interval in epoch
+    horovod = False
+    num_workers = 16
+    kv_store = 'nccl'
+    disable_hybridization = False
+