diff --git a/official/cv/centerface/src/centerface.py b/official/cv/centerface/src/centerface.py
index f51aaa2eb64edb67aa721fc0330963c8f8146bd7..405cb9e5f70ed87ac459a9e5615ab1e5accefa33 100644
--- a/official/cv/centerface/src/centerface.py
+++ b/official/cv/centerface/src/centerface.py
@@ -233,10 +233,8 @@ class TrainingWrapper(nn.TrainOneStepWithLossScaleCell):
if self.reducer_flag:
grads = self.grad_reducer(grads)
cond = self.get_overflow_status(status, grads)
- overflow = self.process_loss_scale(cond)
ret = (loss, cond, sens)
- if not overflow:
- self.optimizer(grads)
+ self.optimizer(grads)
return ret
diff --git a/official/cv/retinanet/src/lr_schedule.py b/official/cv/retinanet/src/lr_schedule.py
index 93ce8c4b653aa7bafc19ba66cb80c27a3283803e..65a846f14b7aa93cfa99aa26978a2045498857c0 100644
--- a/official/cv/retinanet/src/lr_schedule.py
+++ b/official/cv/retinanet/src/lr_schedule.py
@@ -45,7 +45,7 @@ def get_lr(global_step, lr_init, lr_end, lr_max, warmup_epochs1, warmup_epochs2,
warmup_steps4 = warmup_steps3 + steps_per_epoch * warmup_epochs4
warmup_steps5 = warmup_steps4 + steps_per_epoch * warmup_epochs5
step_radio = [1e-4, 1e-3, 1e-2, 0.1]
- if config.finetune:
+ if hasattr(config, finetune) and config.finetune:
step_radio = [1e-4, 1e-2, 0.1, 1]
for i in range(total_steps):
if i < warmup_steps1:
diff --git a/official/cv/retinanet/train.py b/official/cv/retinanet/train.py
index 99a74bd4926d2544740f561a881adf92d301874b..59475849a3a808e5130865a57167103c55bda2df 100644
--- a/official/cv/retinanet/train.py
+++ b/official/cv/retinanet/train.py
@@ -127,7 +127,8 @@ def main():
config.lr_end_rate = ast.literal_eval(config.lr_end_rate)
device_id = get_device_id()
if config.device_target == "Ascend":
- context.set_context(mempool_block_size="31GB")
+ if context.get_context("mode") == context.PYNATIVE_MODE:
+ context.set_context(mempool_block_size="31GB")
elif config.device_target == "GPU":
set_graph_kernel_context(config.device_target)
elif config.device_target == "CPU":
@@ -138,7 +139,7 @@ def main():
context.set_context(mode=context.GRAPH_MODE, device_target=config.device_target)
if config.distribute:
init()
- device_num = config.device_num
+ device_num = get_device_num()
rank = get_rank()
context.reset_auto_parallel_context()
context.set_auto_parallel_context(parallel_mode=ParallelMode.DATA_PARALLEL, gradients_mean=True)
@@ -163,7 +164,7 @@ def main():
retinanet = retinanet50(backbone, config)
net = retinanetWithLossCell(retinanet, config)
init_net_param(net)
- if config.finetune:
+ if hasattr(config, "finetune") and config.finetune:
init_net_param(net, initialize_mode='XavierUniform')
else:
init_net_param(net)
diff --git a/official/recommend/naml/script/run_distribute_train.sh b/official/recommend/naml/scripts/run_distribute_train.sh
similarity index 100%
rename from official/recommend/naml/script/run_distribute_train.sh
rename to official/recommend/naml/scripts/run_distribute_train.sh
diff --git a/official/recommend/naml/script/run_eval.sh b/official/recommend/naml/scripts/run_eval.sh
similarity index 100%
rename from official/recommend/naml/script/run_eval.sh
rename to official/recommend/naml/scripts/run_eval.sh
diff --git a/official/recommend/naml/script/run_infer_310.sh b/official/recommend/naml/scripts/run_infer_310.sh
similarity index 100%
rename from official/recommend/naml/script/run_infer_310.sh
rename to official/recommend/naml/scripts/run_infer_310.sh
diff --git a/official/recommend/naml/script/run_train.sh b/official/recommend/naml/scripts/run_train.sh
similarity index 100%
rename from official/recommend/naml/script/run_train.sh
rename to official/recommend/naml/scripts/run_train.sh
diff --git a/research/cv/glore_res/train.py b/research/cv/glore_res/train.py
index ca95099a9029b54d566af160d46772168602f22c..7628ed81ba0908c888946e02c643063de9061a1a 100644
--- a/research/cv/glore_res/train.py
+++ b/research/cv/glore_res/train.py
@@ -46,7 +46,7 @@ from src.save_callback import SaveCallback
if config.isModelArts:
import moxing as mox
-if config.net == 'resnet200' or config.net == 'resnet101':
+if config.net == 'resnet200' or config.net == 'resnet101' or config.net == 'resnet50':
if config.device_target == "GPU":
config.cast_fp16 = False
diff --git a/research/cv/yolov3_tiny/scripts/run_distribute_train.sh b/research/cv/yolov3_tiny/scripts/run_distribute_train.sh
index 2aab7edd3b1fa4f5d9617594e21a72e53b042a8a..4c97f89015a47619475347b3a9bb065b6d48bac2 100644
--- a/research/cv/yolov3_tiny/scripts/run_distribute_train.sh
+++ b/research/cv/yolov3_tiny/scripts/run_distribute_train.sh
@@ -80,6 +80,7 @@ do
--training_shape=640 \
--weight_decay=0.016 \
--loss_scale=1024 \
+ --num_parallel_workers=32 \
--lr_scheduler=cosine_annealing > log.txt 2>&1 &
cd ..
done
diff --git a/research/cv/yolov3_tiny/scripts/run_standalone_train.sh b/research/cv/yolov3_tiny/scripts/run_standalone_train.sh
index 29adf4b1e9ece4cffdccaea59a118c1b64067e5c..8dcd366584d8a7e91014dd7badd6614b16d41dbd 100644
--- a/research/cv/yolov3_tiny/scripts/run_standalone_train.sh
+++ b/research/cv/yolov3_tiny/scripts/run_standalone_train.sh
@@ -67,5 +67,6 @@ python train.py \
--training_shape=640 \
--per_batch_size=32 \
--weight_decay=0.016 \
+ --num_parallel_workers=32 \
--lr_scheduler=cosine_annealing > log.txt 2>&1 &
cd ..
diff --git a/research/cv/yolov3_tiny/src/yolo_dataset.py b/research/cv/yolov3_tiny/src/yolo_dataset.py
index fd0a25f2825e899beb7a8f0d80e1ae6dfcaf95e2..25d6afcd1e20e88e5aad3d8ed2efaf9c7f9ab793 100644
--- a/research/cv/yolov3_tiny/src/yolo_dataset.py
+++ b/research/cv/yolov3_tiny/src/yolo_dataset.py
@@ -316,9 +316,9 @@ def create_yolo_dataset(
CV.Normalize(mean, std),
hwc_to_chw
],
- num_parallel_workers=num_parallel_workers
+ num_parallel_workers=8
)
- ds = ds.batch(batch_size, num_parallel_workers=num_parallel_workers, drop_remainder=True)
+ ds = ds.batch(batch_size, num_parallel_workers=8, drop_remainder=True)
else:
ds = de.GeneratorDataset(
yolo_dataset,
@@ -331,9 +331,9 @@ def create_yolo_dataset(
input_columns=["image", "img_id"],
output_columns=["image", "image_shape", "img_id"],
column_order=["image", "image_shape", "img_id"],
- num_parallel_workers=num_parallel_workers
+ num_parallel_workers=8
)
- ds = ds.map(operations=hwc_to_chw, input_columns=["image"], num_parallel_workers=num_parallel_workers)
+ ds = ds.map(operations=hwc_to_chw, input_columns=["image"], num_parallel_workers=8)
ds = ds.batch(batch_size, drop_remainder=True)
ds = ds.repeat(max_epoch)
return ds, len(yolo_dataset)