diff --git a/research/cv/ProtoNet/README.md b/research/cv/ProtoNet/README.md
index 109618ffd5e13cdd2bc9e0bf0808f960c059d74e..84d821c65bfb810493c041f428bbe3e7f00fe777 100644
--- a/research/cv/ProtoNet/README.md
+++ b/research/cv/ProtoNet/README.md
@@ -60,7 +60,7 @@ python train.py
# [Environment Requirements](#contents)
-- Hardware(Ascend)
+- Hardware(Ascend&GPU)
- Prepare hardware environment with Ascend.
- Framework
- [MindSpore](https://www.mindspore.cn/install/en)
@@ -79,6 +79,12 @@ sh run_standalone_train_ascend.sh "../dataset" 1 60 500
sh run_standalone_eval_ascend.sh "../dataset" "./output/best_ck.ckpt" 1 5
# enter script dir, train ProtoNet distributed
sh run_distribution_ascend.sh "./rank_table.json" "../dataset" 60 500
+# enter script dir, train ProtoNet in GPU.
+sh run_standalone_train_gpu.sh "../dataset" 1 60 500
+# enter script dir, evaluate ProtoNet
+sh run_standalone_eval_gpu.sh "../dataset" "./output/best_ck.ckpt" 1 5
+# enter script dir, train ProtoNet distributed
+sh run_distribution_gpu.sh "../dataset" 60 500
```
## [Script and Sample Code](#contents)
@@ -89,6 +95,9 @@ sh run_distribution_ascend.sh "./rank_table.json" "../dataset" 60 500
鈹溾攢鈹€ requirements.txt
鈹溾攢鈹€ README.md // descriptions about lenet
鈹溾攢鈹€ scripts
+ 鈹� 鈹溾攢鈹€run_standalone_train_gpu.sh // train in GPU
+ 鈹� 鈹溾攢鈹€run_standalone_eval_gpu.sh // evaluate in GPU
+ 鈹� 鈹溾攢鈹€run_distribution_gpu.sh // distribution in GPU
鈹� 鈹溾攢鈹€run_standalone_train_ascend.sh // train in ascend
鈹� 鈹溾攢鈹€run_standalone_eval_ascend.sh // evaluate in ascend
鈹� 鈹溾攢鈹€run_distribution_ascend.sh // distribution in ascend
@@ -125,7 +134,8 @@ Major parameters in train.py and config.py as follows:
### Training
```bash
-sh run_standalone_train_ascend.sh "../dataset" 1 60 500
+bash run_standalone_train_ascend.sh "../dataset" 1 60 500
+bash run_standalone_train_gpu.sh "../dataset" 1 60 500
```
The model checkpoint will be saved in the current directory.
@@ -138,11 +148,13 @@ Before running the command below, please check the checkpoint path used for eval
```bash
sh run_standalone_eval_ascend.sh "../dataset" "./output/best_ck.ckpt" 1 5
+sh run_standalone_eval_gpu.sh "../dataset" "./output/best_ck.ckpt" 1 5
```
```shell
-Test Acc: 0.9954400658607483 Loss: 0.02102319709956646
+Test Acc in Ascend: 0.9954400658607483 Loss: 0.02102319709956646
+Test Acc in GPU: 0.996999979019165 Loss: 0.013885765336453915
```
# [Model Description](#contents)
@@ -151,21 +163,21 @@ Test Acc: 0.9954400658607483 Loss: 0.02102319709956646
### Evaluation Performance
-| Parameters | ProtoNet |
-| -------------------------- | ---------------------------------------------------------- |
-| Resource | CentOs 8.2; Ascend 910 ; CPU 2.60GHz锛�192cores锛汳emory 755G |
-| uploaded Date | 03/26/2021 (month/day/year) |
-| MindSpore Version | 1.1.1 |
-| Dataset | OMNIGLOT |
-| Training Parameters | episode=500, class_num = 5, lr=0.001, classes_per_it_tr=60, num_support_tr=5, num_query_tr=5, classes_per_it_val=20, num_support_val=5, num_query_val=15 |
-| Optimizer | Adam |
-| Loss Function | Prototypicalloss |
-| outputs | Accuracy |
-| Loss | 0.002 |
-| Speed | 215 ms/step |
-| Total time | 3 h 23m (8p) |
-| Checkpoint for Fine tuning | 440 KB (.ckpt file) |
-| Scripts | <https://gitee.com/mindspore/models/tree/master/research/cv/ProtoNet> |
+| Parameters | ProtoNet(Ascend) |ProtoNet(GPU) |
+| -------------------------- | ---------------------------------------------------------- | ---------------------------------------------------------- |
+| Resource | CentOs 8.2; Ascend 910 ; CPU 2.60GHz锛�192cores锛汳emory 755G | ubuntu 18.04; Tesla V100 ; CPU 2.60GHz |
+| uploaded Date | 03/26/2021 (month/day/year) | 08/28/2021 (month/day/year) |
+| MindSpore Version | 1.1.1 | 1.3.0 |
+| Dataset | OMNIGLOT |OMNIGLOT |
+| Training Parameters | episode=500, class_num = 5, lr=0.001, classes_per_it_tr=60, num_support_tr=5, num_query_tr=5, classes_per_it_val=20, num_support_val=5, num_query_val=15 | episode=500, class_num = 5, lr=0.001, classes_per_it_tr=60, num_support_tr=5, num_query_tr=5, classes_per_it_val=20, num_support_val=5, num_query_val=15 |
+| Optimizer | Adam | Adam |
+| Loss Function | Prototypicalloss | Prototypicalloss |
+| outputs | Accuracy | Accuracy |
+| Loss | 0.002 | 0.002 |
+| Speed | 215 ms/step | 144 ms/step |
+| Total time | 3 h 23m (8p) | 2 h 48m (8p) |
+| Checkpoint for Fine tuning | 440 KB (.ckpt file) | 441 KB (.ckpt file) |
+| Scripts | <https://gitee.com/mindspore/models/tree/master/research/cv/ProtoNet> |<https://gitee.com/mindspore/models/tree/master/research/cv/ProtoNet> |
# [ModelZoo Homepage](#contents)
diff --git a/research/cv/ProtoNet/scripts/run_distribution_gpu.sh b/research/cv/ProtoNet/scripts/run_distribution_gpu.sh
new file mode 100644
index 0000000000000000000000000000000000000000..bea9cc8b252aca3529065fd20226d94cf06a8538
--- /dev/null
+++ b/research/cv/ProtoNet/scripts/run_distribution_gpu.sh
@@ -0,0 +1,48 @@
+#!/bin/bash
+# Copyright 2021 Huawei Technologies Co., Ltd
+#
+# 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.
+# ============================================================================
+if [ $# != 4 ]
+then
+ echo "Usage: sh run_distribution_gpu.sh [DATA_PATH] [TRAIN_CLASS] [EPOCHS] [DEVICE_NUM]"
+exit 1
+fi
+
+get_real_path(){
+ if [ "${1:0:1}" == "/" ]; then
+ echo "$1"
+ else
+ echo "$(realpath -m $PWD/$1)"
+ fi
+}
+export DATA_PATH=$(get_real_path $1) # dataset path
+export TRAIN_CLASS=$2 # train class, propose 20
+export EPOCHS=$3 # num of epochs
+export DEVICE_NUM=$4 # device_num
+
+if [ ! -d $DATA_PATH ]
+then
+ echo "error: DATA_PATH=$DATA_PATH is not a directory"
+exit 1
+fi
+
+rm -rf distribute_output
+mkdir distribute_output
+
+mpirun --allow-run-as-root -n $DEVICE_NUM --output-filename log_output --merge-stderr-to-stdout \
+ python ../train.py --dataset_root=$DATA_PATH \
+ --device_target="GPU" \
+ --classes_per_it_tr=$TRAIN_CLASS \
+ --experiment_root="./distribute_output" \
+ --epochs=$EPOCHS > distribute_log 2>&1 &
diff --git a/research/cv/ProtoNet/scripts/run_standalone_eval_gpu.sh b/research/cv/ProtoNet/scripts/run_standalone_eval_gpu.sh
new file mode 100644
index 0000000000000000000000000000000000000000..81cf1a623211592d875fcb3896062933c6101636
--- /dev/null
+++ b/research/cv/ProtoNet/scripts/run_standalone_eval_gpu.sh
@@ -0,0 +1,49 @@
+#!/bin/bash
+# Copyright 2021 Huawei Technologies Co., Ltd
+#
+# 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.
+# ============================================================================
+# an simple tutorial as follows, more parameters can be setting
+if [ $# != 4 ]
+then
+ echo "Usage: sh run_standalone_eval_gpu.sh [DATA_PATH] [CKPT_PATH] [DEVICE_ID] [EVAL_CLASS]"
+exit 1
+fi
+
+get_real_path(){
+ if [ "${1:0:1}" == "/" ]; then
+ echo "$1"
+ else
+ echo "$(realpath -m $PWD/$1)"
+ fi
+}
+
+export DATA_PATH=$(get_real_path $1) # dataset path
+export CKPT_PATH=$2$(get_real_path $1) # path of ckpt file to load
+export DEVICE_ID=$3 # device id
+export EVAL_CLASS=$4 # class used in evaluation
+if [ ! -d $DATA_PATH ]
+then
+ echo "error: DATA_PATH=$DATA_PATH is not a directory"
+exit 1
+fi
+
+if [ ! -d CKPT_PATH ]
+then
+ echo "error: CKPT_PATH=$CKPT_PATH is not a directory"
+exit 1
+fi
+
+python ../eval.py --dataset_root=$DATA_PATH --experiment_root=$CKPT_PATH \
+ --device_id=$DEVICE_ID --device_target="GPU" \
+ --classes_per_it_val=$EVAL_CLASS > eval_log 2>&1 &
diff --git a/research/cv/ProtoNet/scripts/run_standalone_train_gpu.sh b/research/cv/ProtoNet/scripts/run_standalone_train_gpu.sh
new file mode 100644
index 0000000000000000000000000000000000000000..060fa1d22299bdeb06a60590d3594a468eb37a7f
--- /dev/null
+++ b/research/cv/ProtoNet/scripts/run_standalone_train_gpu.sh
@@ -0,0 +1,46 @@
+#!/bin/bash
+# Copyright 2021 Huawei Technologies Co., Ltd
+#
+# 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.
+# ============================================================================
+# an simple tutorial as follows, more parameters can be setting
+if [ $# != 4 ]
+then
+ echo "Usage: sh run_standalone_train_gpu.sh [DATA_PATH] [DEVICE_ID] [TRAIN_CLASS] [EPOCHS]"
+exit 1
+fi
+
+get_real_path(){
+ if [ "${1:0:1}" == "/" ]; then
+ echo "$1"
+ else
+ echo "$(realpath -m $PWD/$1)"
+ fi
+}
+
+export DATA_PATH=$(get_real_path $1) # dataset path
+export DEVICE_ID=$2 # device_id
+export TRAIN_CLASS=$3 # train class, propose 20
+export EPOCHS=$4 # num of epochs
+
+if [ ! -d $DATA_PATH ]
+then
+ echo "error: DATA_PATH=$DATA_PATH is not a directory"
+exit 1
+fi
+
+python ../train.py --dataset_root=$DATA_PATH \
+ --device_id=$DEVICE_ID --device_target="GPU" \
+ --classes_per_it_tr=$TRAIN_CLASS \
+ --experiment_root="./standalone_output" \
+ --epochs=$EPOCHS > log 2>&1 &
diff --git a/research/cv/ProtoNet/src/EvalCallBack.py b/research/cv/ProtoNet/src/EvalCallBack.py
index 08a297926b88709743494ebe0200d57716db07ff..08ab44565ef74355f469e572a7d2d6083b402685 100644
--- a/research/cv/ProtoNet/src/EvalCallBack.py
+++ b/research/cv/ProtoNet/src/EvalCallBack.py
@@ -26,7 +26,7 @@ class EvalCallBack(Callback):
"""
CallBack class
"""
- def __init__(self, options, net, eval_dataset, path):
+ def __init__(self, options, net, eval_dataset, path, rank_id=0):
self.net = net
self.eval_dataset = eval_dataset
self.path = path
@@ -34,6 +34,7 @@ class EvalCallBack(Callback):
self.avgloss = 0
self.bestacc = 0
self.options = options
+ self.rank_id = rank_id
def epoch_begin(self, run_context):
@@ -58,12 +59,12 @@ class EvalCallBack(Callback):
if self.avgacc > self.bestacc:
self.bestacc = self.avgacc
print('Epoch {}: Avg Accuracy: {}(best) Avg Loss:{}'.format(cur_epoch, self.avgacc, self.avgloss))
- best_path = os.path.join(self.path, 'best_ck.ckpt')
+ best_path = os.path.join(self.path, f'best_ck_{self.rank_id}.ckpt')
save_checkpoint(cur_net, best_path)
else:
print('Epoch {}: Avg Accuracy: {} Avg Loss:{}'.format(cur_epoch, self.avgacc, self.avgloss))
- last_path = os.path.join(self.path, 'last_ck.ckpt')
+ last_path = os.path.join(self.path, f'last_ck_{self.rank_id}.ckpt')
save_checkpoint(cur_net, last_path)
print("Best Acc:", self.bestacc)
print('=========EPOCH {} END========='.format(cur_epoch))
diff --git a/research/cv/ProtoNet/train.py b/research/cv/ProtoNet/train.py
index 9ba1873d1a6411731733931f0ee53e3314e6f73f..b2bbe5b319838eb49a2ecbff8996b86c5388e3eb 100644
--- a/research/cv/ProtoNet/train.py
+++ b/research/cv/ProtoNet/train.py
@@ -22,7 +22,7 @@ from mindspore.train import Model
from mindspore.train.callback import CheckpointConfig, ModelCheckpoint, TimeMonitor
from mindspore import dataset as ds
import mindspore.context as context
-from mindspore.communication.management import init
+from mindspore.communication.management import init, get_rank
from mindspore.context import ParallelMode
from src.EvalCallBack import EvalCallBack
from src.protonet import WithLossCell
@@ -35,7 +35,7 @@ local_data_url = './cache/data'
local_train_url = './cache/out'
-def train(opt, tr_dataloader, net, loss_fn, eval_loss_fn, optim, path, val_dataloader=None):
+def train(opt, tr_dataloader, net, loss_fn, eval_loss_fn, optim, path, rank_id, val_dataloader=None):
'''
train function
'''
@@ -47,11 +47,11 @@ def train(opt, tr_dataloader, net, loss_fn, eval_loss_fn, optim, path, val_datal
eval_data = ds.GeneratorDataset(val_dataloader, column_names=['data', 'label', 'classes'])
- eval_cb = EvalCallBack(opt, my_acc_cell, eval_data, path)
+ eval_cb = EvalCallBack(opt, my_acc_cell, eval_data, path, rank_id)
config = CheckpointConfig(save_checkpoint_steps=10,
keep_checkpoint_max=5,
saved_network=net)
- ckpoint_cb = ModelCheckpoint(prefix='protonet', directory=path, config=config)
+ ckpoint_cb = ModelCheckpoint(prefix=str(rank_id) + '_protonet', directory=path, config=config)
print('==========training test==========')
starttime = datetime.datetime.now()
@@ -68,24 +68,29 @@ def main():
global local_train_url
options = get_parser().parse_args()
+ device_num = int(os.environ.get("DEVICE_NUM", 1))
- if options.run_offline:
-
- device_num = int(os.environ.get("DEVICE_NUM", 1))
+ if options.device_target == "GPU":
+ rank_id = get_rank()
+ if options.run_offline:
if device_num > 1:
-
init()
context.reset_auto_parallel_context()
context.set_auto_parallel_context(device_num=device_num, parallel_mode=ParallelMode.DATA_PARALLEL,
gradients_mean=True)
- context.set_context(device_id=options.device_id)
+
+ if options.device_target == "Ascend":
+ context.set_context(device_id=options.device_id)
+
local_data_url = options.dataset_root
local_train_url = options.experiment_root
- if not os.path.exists(options.experiment_root):
- os.makedirs(options.experiment_root)
+ if device_num > 1 and options.device_target == "GPU":
+ pass
+ else:
+ if not os.path.exists(options.experiment_root):
+ os.makedirs(options.experiment_root)
else:
- device_num = int(os.environ.get("DEVICE_NUM", 1))
device_id = int(os.getenv("DEVICE_ID"))
import moxing as mox
@@ -114,7 +119,19 @@ def main():
Net = ProtoNet()
optim = nn.Adam(params=Net.trainable_params(), learning_rate=0.001)
- train(options, tr_dataloader, Net, loss_fn, eval_loss_fn, optim, local_train_url, val_dataloader)
+ if options.device_target == "Ascend":
+ train(
+ options,
+ tr_dataloader,
+ Net,
+ loss_fn,
+ eval_loss_fn,
+ optim,
+ local_train_url,
+ options.device_id,
+ val_dataloader)
+ elif options.device_target == "GPU":
+ train(options, tr_dataloader, Net, loss_fn, eval_loss_fn, optim, local_train_url, rank_id, val_dataloader)
if not options.run_offline:
mox.file.copy_parallel(src_url='./cache/out', dst_url=options.train_url)