Merge branch 'modify-mpa-templates-for-hpo' into develop

Author: eunwoosh

external/deep-object-reid/torchreid_tasks/utils.py

@@ -441,7 +441,10 @@ def on_epoch_end(self, epoch, logs=None):
             print(f'score = {score} at epoch {self.current_epoch} / {self._num_iters}')
             # as a trick, score (at least if it's accuracy not the loss) and iteration number
             # could be assembled just using summation and then disassembeled.
-            score = score + int(self._num_iters)
+            if 1.0 > score:
+                score = score + int(self._num_iters)
+            else:
+                score = -(score + int(self._num_iters))
         self.update_progress_callback(self.get_progress(), score=score)
 
 

external/mmdetection/detection_tasks/apis/detection/ote_utils.py

@@ -126,7 +126,10 @@ def on_epoch_end(self, epoch, logs=None):
                     print(f'score = {score} at epoch {epoch} / {int(iter_num)}')
                     # as a trick, score (at least if it's accuracy not the loss) and iteration number
                     # could be assembled just using summation and then disassembeled.
-                    score = score + int(iter_num)
+                    if 1.0 > score:
+                        score = score + int(iter_num)
+                    else:
+                        score = -(score + int(iter_num))
         self.update_progress_callback(self.get_progress(), score=score)
 
 

external/mmdetection/detection_tasks/extension/utils/runner.py

@@ -57,7 +57,8 @@ def train(self, data_loader: DataLoader, **kwargs):
         self.data_loader = data_loader
         self._max_iters = self._max_epochs * len(self.data_loader)
         self.call_hook('before_train_epoch')
-        time.sleep(2)  # Prevent possible deadlock during epoch transition
+        # TODO: uncomment below line or resolve root cause of deadlock issue if multi-GPUs need to be supported.
+        # time.sleep(2)  # Prevent possible multi-gpu deadlock during epoch transition
         for i, data_batch in enumerate(self.data_loader):
             self._inner_iter = i
             self.call_hook('before_train_iter')

external/mmsegmentation/segmentation_tasks/apis/segmentation/ote_utils.py

@@ -75,7 +75,10 @@ def on_epoch_end(self, epoch, logs=None):
                     print(f'score = {score} at epoch {epoch} / {int(iter_num)}')
                     # as a trick, score (at least if it's accuracy not the loss) and iteration number
                     # could be assembled just using summation and then disassembeled.
-                    score = score + int(iter_num)
+                    if 1.0 > score:
+                        score = score + int(iter_num)
+                    else:
+                        score = -(score + int(iter_num))
         self.update_progress_callback(self.get_progress(), score=score)
 
 

external/mmsegmentation/segmentation_tasks/extension/utils/runner.py

@@ -61,7 +61,8 @@ def train(self, data_loader: DataLoader, **kwargs):
         self.data_loader = data_loader
         self._max_iters = self._max_epochs * len(self.data_loader)
         self.call_hook('before_train_epoch')
-        time.sleep(2)  # Prevent possible deadlock during epoch transition
+        # TODO: uncomment below line or resolve root cause of deadlock issue if multi-GPUs need to be supported.
+        # time.sleep(2)  # Prevent possible multi-gpu deadlock during epoch transition
         for i, data_batch in enumerate(self.data_loader):
             self._inner_iter = i
             self.call_hook('before_train_iter')

external/model-preparation-algorithm/configs/classification/configuration.yaml

@@ -22,6 +22,7 @@ learning_parameters:
     warning:
       Increasing this value may cause the system to use more memory than available,
       potentially causing out of memory errors, please update with caution.
+    auto_hpo_state: NOT_POSSIBLE
   description: Learning Parameters
   header: Learning Parameters
   learning_rate:
@@ -42,6 +43,7 @@ learning_parameters:
       type: UI_RULES
     visible_in_ui: true
     warning: null
+    auto_hpo_state: NOT_POSSIBLE
   max_num_epochs:
     affects_outcome_of: TRAINING
     default_value: 200

external/model-preparation-algorithm/configs/classification/efficientnet_b0_cls_incr/hpo_config.yaml

@@ -1,16 +1,15 @@
-metric: mAP
-search_algorithm: smbo
-early_stop: median_stop
+metric: accuracy_top-1
+search_algorithm: asha
 hp_space:
   learning_parameters.learning_rate:
-    param_type: quniform
+    param_type: qloguniform
     range:
-      - 0.001
-      - 0.01
-      - 0.001
+      - 0.0003
+      - 0.1
+      - 0.0001
   learning_parameters.batch_size:
     param_type: qloguniform
     range:
-      - 8
-      - 64
+      - 32
+      - 128
       - 2

external/model-preparation-algorithm/configs/classification/efficientnet_b0_cls_incr/template.yaml

@@ -27,10 +27,12 @@ hyper_parameters:
     learning_parameters:
       batch_size:
         default_value: 32
+        auto_hpo_state: POSSIBLE
       num_workers:
         default_value: 4
       learning_rate:
         default_value: 0.007
+        auto_hpo_state: POSSIBLE
       num_iters:
         default_value: 20
     algo_backend:

external/model-preparation-algorithm/configs/classification/efficientnet_v2_s_cls_incr/hpo_config.yaml

@@ -1,16 +1,15 @@
-metric: mAP
-search_algorithm: smbo
-early_stop: median_stop
+metric: accuracy_top-1
+search_algorithm: asha
 hp_space:
   learning_parameters.learning_rate:
-    param_type: quniform
+    param_type: qloguniform
     range:
-      - 0.001
-      - 0.01
-      - 0.001
+      - 0.0014
+      - 0.035
+      - 0.0001
   learning_parameters.batch_size:
     param_type: qloguniform
     range:
-      - 8
-      - 64
+      - 20
+      - 48
       - 2

external/model-preparation-algorithm/configs/classification/efficientnet_v2_s_cls_incr/template.yaml

@@ -27,10 +27,12 @@ hyper_parameters:
     learning_parameters:
       batch_size:
         default_value: 32
+        auto_hpo_state: POSSIBLE
       num_workers:
         default_value: 4
       learning_rate:
         default_value: 0.007
+        auto_hpo_state: POSSIBLE
       num_iters:
         default_value: 20
     algo_backend:

external/model-preparation-algorithm/configs/classification/mobilenet_v3_large_075_cls_incr/hpo_config.yaml

@@ -1,16 +1,15 @@
-metric: mAP
-search_algorithm: smbo
-early_stop: median_stop
+metric: accuracy_top-1
+search_algorithm: asha
 hp_space:
   learning_parameters.learning_rate:
-    param_type: quniform
+    param_type: qloguniform
     range:
-      - 0.005
-      - 0.029
-      - 0.001
+      - 0.0032
+      - 0.08
+      - 0.0001
   learning_parameters.batch_size:
     param_type: qloguniform
     range:
-      - 8
-      - 64
+      - 20
+      - 48
       - 2

external/model-preparation-algorithm/configs/classification/mobilenet_v3_large_075_cls_incr/template_experiment.yaml

@@ -27,10 +27,12 @@ hyper_parameters:
     learning_parameters:
       batch_size:
         default_value: 32
+        auto_hpo_state: POSSIBLE
       num_workers:
         default_value: 4
       learning_rate:
         default_value: 0.016
+        auto_hpo_state: POSSIBLE
       learning_rate_warmup_iters:
         default_value: 100
       num_iters:

external/model-preparation-algorithm/configs/classification/mobilenet_v3_large_1_cls_incr/hpo_config.yaml

@@ -1,16 +1,15 @@
-metric: mAP
-search_algorithm: smbo
-early_stop: median_stop
+metric: accuracy_top-1
+search_algorithm: asha
 hp_space:
   learning_parameters.learning_rate:
-    param_type: quniform
+    param_type: qloguniform
     range:
-      - 0.005
-      - 0.029
-      - 0.001
+      - 0.0032
+      - 0.08
+      - 0.0001
   learning_parameters.batch_size:
     param_type: qloguniform
     range:
-      - 8
-      - 64
+      - 20
+      - 48
       - 2

external/model-preparation-algorithm/configs/classification/mobilenet_v3_large_1_cls_incr/template.yaml

@@ -27,10 +27,12 @@ hyper_parameters:
     learning_parameters:
       batch_size:
         default_value: 32
+        auto_hpo_state: POSSIBLE
       num_workers:
         default_value: 4
       learning_rate:
         default_value: 0.016
+        auto_hpo_state: POSSIBLE
       learning_rate_warmup_iters:
         default_value: 100
       num_iters:

external/model-preparation-algorithm/configs/classification/mobilenet_v3_small_cls_incr/hpo_config.yaml

@@ -1,16 +1,16 @@
-metric: mAP
-search_algorithm: smbo
-early_stop: median_stop
+metric: accuracy_top-1
+search_algorithm: asha
+early_stop: None
 hp_space:
   learning_parameters.learning_rate:
-    param_type: quniform
+    param_type: qloguniform
     range:
-      - 0.005
-      - 0.029
-      - 0.001
+      - 0.0032
+      - 0.08
+      - 0.0001
   learning_parameters.batch_size:
     param_type: qloguniform
     range:
-      - 8
-      - 64
+      - 20
+      - 48
       - 2

external/model-preparation-algorithm/configs/classification/mobilenet_v3_small_cls_incr/template_experiment.yaml

@@ -27,10 +27,12 @@ hyper_parameters:
     learning_parameters:
       batch_size:
         default_value: 32
+        auto_hpo_state: POSSIBLE
       num_workers:
         default_value: 4
       learning_rate:
         default_value: 0.016
+        auto_hpo_state: POSSIBLE
       learning_rate_warmup_iters:
         default_value: 100
       num_iters:

external/model-preparation-algorithm/configs/detection/configuration.yaml

@@ -23,6 +23,7 @@ learning_parameters:
     warning:
       Increasing this value may cause the system to use more memory than available,
       potentially causing out of memory errors, please update with caution.
+    auto_hpo_state: NOT_POSSIBLE
   description: Learning Parameters
   header: Learning Parameters
   learning_rate:
@@ -44,6 +45,7 @@ learning_parameters:
     value: 0.01
     visible_in_ui: true
     warning: null
+    auto_hpo_state: NOT_POSSIBLE
   learning_rate_warmup_iters:
     affects_outcome_of: TRAINING
     default_value: 100

external/model-preparation-algorithm/configs/detection/cspdarknet_yolox_cls_incr/hpo_config.yaml

@@ -1,16 +1,16 @@
 metric: mAP
-search_algorithm: smbo
+search_algorithm: asha
 early_stop: None
 hp_space:
   learning_parameters.learning_rate:
-    param_type: quniform
+    param_type: qloguniform
     range:
-      - 0.001
-      - 0.1
-      - 0.001
+      - 0.00002
+      - 0.0005
+      - 0.00001
   learning_parameters.batch_size:
     param_type: qloguniform
     range:
-      - 4
-      - 8
+      - 2
+      - 16
       - 2

external/model-preparation-algorithm/configs/detection/cspdarknet_yolox_cls_incr/template.yaml

@@ -27,8 +27,10 @@ hyper_parameters:
     learning_parameters:
       batch_size:
         default_value: 4
+        auto_hpo_state: POSSIBLE
       learning_rate:
         default_value: 0.0001
+        auto_hpo_state: POSSIBLE
       learning_rate_warmup_iters:
         default_value: 10
       num_iters:

external/model-preparation-algorithm/configs/detection/mobilenetv2_atss_cls_incr/hpo_config.yaml

@@ -1,16 +1,16 @@
 metric: mAP
-search_algorithm: smbo
+search_algorithm: asha
 early_stop: None
 hp_space:
   learning_parameters.learning_rate:
-    param_type: quniform
+    param_type: qloguniform
     range:
-      - 0.001
-      - 0.1
-      - 0.001
+      - 0.0001
+      - 0.01
+      - 0.0001
   learning_parameters.batch_size:
     param_type: qloguniform
     range:
-      - 4
-      - 8
+      - 2
+      - 16
       - 2

external/model-preparation-algorithm/configs/detection/mobilenetv2_atss_cls_incr/template.yaml

@@ -27,8 +27,10 @@ hyper_parameters:
     learning_parameters:
       batch_size:
         default_value: 4
+        auto_hpo_state: POSSIBLE
       learning_rate:
         default_value: 0.001
+        auto_hpo_state: POSSIBLE
       learning_rate_warmup_iters:
         default_value: 10
       num_iters:

external/model-preparation-algorithm/configs/detection/mobilenetv2_ssd_cls_incr/hpo_config.yaml

@@ -1,16 +1,16 @@
 metric: mAP
-search_algorithm: smbo
+search_algorithm: asha
 early_stop: None
 hp_space:
   learning_parameters.learning_rate:
-    param_type: quniform
+    param_type: qloguniform
     range:
-      - 0.001
-      - 0.1
+      - 0.0005
+      - 0.05
       - 0.001
   learning_parameters.batch_size:
     param_type: qloguniform
     range:
-      - 4
-      - 8
+      - 2
+      - 16
       - 2

external/model-preparation-algorithm/configs/detection/mobilenetv2_ssd_cls_incr/template.yaml

@@ -27,8 +27,10 @@ hyper_parameters:
     learning_parameters:
       batch_size:
         default_value: 4
+        auto_hpo_state: POSSIBLE
       learning_rate:
         default_value: 0.005
+        auto_hpo_state: POSSIBLE
       learning_rate_warmup_iters:
         default_value: 10
       num_iters:

external/model-preparation-algorithm/configs/instance-segmentation/configuration.yaml

@@ -23,6 +23,7 @@ learning_parameters:
     warning:
       Increasing this value may cause the system to use more memory than available,
       potentially causing out of memory errors, please update with caution.
+    auto_hpo_state: NOT_POSSIBLE
   description: Learning Parameters
   header: Learning Parameters
   learning_rate:
@@ -44,6 +45,7 @@ learning_parameters:
     value: 0.01
     visible_in_ui: true
     warning: null
+    auto_hpo_state: NOT_POSSIBLE
   learning_rate_warmup_iters:
     affects_outcome_of: TRAINING
     default_value: 100