!13114 retinanet performance improve

From: @chenmai1102
Reviewed-by: 
Signed-off-by:

model_zoo/official/cv/retinanet/README_CN.md

@@ -165,11 +165,11 @@ MSCOCO2017
 # 八卡并行训练示例：
 
 创建 RANK_TABLE_FILE
-sh run_distribute_train.sh DEVICE_NUM EPOCH_SIZE LR DATASET RANK_TABLE_FILE PRE_TRAINED(optional) PRE_TRAINED_EPOCH_SIZE(optional)
+sh run_distribute_train.sh DEVICE_NUM EPOCH_SIZE LR RANK_TABLE_FILE PRE_TRAINED(optional) PRE_TRAINED_EPOCH_SIZE(optional)
 
 # 单卡训练示例：
 
-sh run_distribute_train.sh DEVICE_ID EPOCH_SIZE LR DATASET PRE_TRAINED(optional) PRE_TRAINED_EPOCH_SIZE(optional)
+sh run_single_train.sh DEVICE_ID EPOCH_SIZE LR PRE_TRAINED(optional) PRE_TRAINED_EPOCH_SIZE(optional)
 
 ```
 
@@ -182,6 +182,9 @@ sh run_distribute_train.sh DEVICE_ID EPOCH_SIZE LR DATASET PRE_TRAINED(optional)
 ```运行
 # 训练示例
 
+  训练前，先创建MindRecord文件，以COCO数据集为例
+    python create_data.py --dataset coco
+
   python:
     data和存储mindrecord文件的路径在config里设置
 
@@ -193,12 +196,12 @@ sh run_distribute_train.sh DEVICE_ID EPOCH_SIZE LR DATASET PRE_TRAINED(optional)
 
       # 八卡并行训练示例(在retinanet目录下运行)：
 
-      sh scripts/run_distribute_train.sh 8 500 0.1 coco RANK_TABLE_FILE(创建的RANK_TABLE_FILE的地址) PRE_TRAINED(预训练checkpoint地址) PRE_TRAINED_EPOCH_SIZE（预训练EPOCH大小）
-      例如：sh scripts/run_distribute_train.sh 8 500 0.1 coco scripts/rank_table_8pcs.json /dataset/retinanet-322_458.ckpt 322
+      sh scripts/run_distribute_train.sh 8 500 0.1 RANK_TABLE_FILE(创建的RANK_TABLE_FILE的地址) PRE_TRAINED(预训练checkpoint地址) PRE_TRAINED_EPOCH_SIZE（预训练EPOCH大小）
+      例如：sh scripts/run_distribute_train.sh 8 500 0.1 scripts/rank_table_8pcs.json /dataset/retinanet-322_458.ckpt 322
 
       # 单卡训练示例(在retinanet目录下运行)：
 
-      sh scripts/run_single_train.sh 0 500 0.1 coco /dataset/retinanet-322_458.ckpt 322
+      sh scripts/run_single_train.sh 0 500 0.1 /dataset/retinanet-322_458.ckpt 322
 ```
 
 #### 结果

model_zoo/official/cv/retinanet/create_data.py

@@ -0,0 +1,25 @@
+# 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.
+# ============================================================================
+
+"""create mindrecord for training retinanet."""
+
+import argparse
+from src.dataset import create_mindrecord
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="retinanet dataset create")
+    parser.add_argument("--dataset", type=str, default="coco", help="Dataset, default is coco.")
+    args_opt = parser.parse_args()
+    mindrecord_file = create_mindrecord(args_opt.dataset, "retinanet.mindrecord", True)

model_zoo/official/cv/retinanet/export.py

@@ -0,0 +1,46 @@
+# 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.
+# ============================================================================
+"""export for retinanet"""
+import argparse
+import numpy as np
+import mindspore.common.dtype as mstype
+from mindspore import context, Tensor
+from mindspore.train.serialization import load_checkpoint, load_param_into_net, export
+from src.retinanet import  retinanet50, resnet50, retinanetInferWithDecoder
+from src.config import config
+from src.box_utils import default_boxes
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='retinanet evaluation')
+    parser.add_argument("--device_id", type=int, default=0, help="Device id, default is 0.")
+    parser.add_argument("--run_platform", type=str, default="Ascend", choices=("Ascend"),
+                        help="run platform, only support Ascend.")
+    parser.add_argument("--file_format", type=str, choices=["AIR", "MINDIR"], default="MINDIR", help="file format")
+    parser.add_argument("--batch_size", type=int, default=1, help="batch size")
+    parser.add_argument("--file_name", type=str, default="retinanet", help="output file name.")
+    args_opt = parser.parse_args()
+    context.set_context(mode=context.GRAPH_MODE, device_target=args_opt.run_platform, device_id=args_opt.device_id)
+
+    backbone = resnet50(config.num_classes)
+    net = retinanet50(backbone, config)
+    net = retinanetInferWithDecoder(net, Tensor(default_boxes), config)
+    param_dict = load_checkpoint(config.checkpoint_path)
+    net.init_parameters_data()
+    load_param_into_net(net, param_dict)
+    net.set_train(False)
+    shape = [args_opt.batch_size, 3] + config.img_shape
+    input_data = Tensor(np.zeros(shape), mstype.float32)
+    export(net, input_data, file_name=args_opt.file_name, file_format=args_opt.file_format)

model_zoo/official/cv/retinanet/scripts/run_distribute_train.sh

@@ -21,27 +21,24 @@ echo "for example: sh run_distribute_train.sh 8 500 0.1 coco /data/hccl.json /op
 echo "It is better to use absolute path."
 echo "================================================================================================================="
 
-if [ $# != 5 ] && [ $# != 7 ]
+if [ $# != 4 ] && [ $# != 6 ]
 then
-    echo "Usage: sh run_distribute_train.sh [DEVICE_NUM] [EPOCH_SIZE] [LR] [DATASET] \
+    echo "Usage: sh run_distribute_train.sh [DEVICE_NUM] [EPOCH_SIZE] [LR] \
 [RANK_TABLE_FILE] [PRE_TRAINED](optional) [PRE_TRAINED_EPOCH_SIZE](optional)"
     exit 1
 fi
 
-# Before start distribute train, first create mindrecord files.
-BASE_PATH=$(cd "`dirname $0`" || exit; pwd)
-cd $BASE_PATH/../ || exit
-python train.py --only_create_dataset=True
+core_num=`cat /proc/cpuinfo |grep "processor"|wc -l`
+process_cores=$(($core_num/8))
 
 echo "After running the script, the network runs in the background. The log will be generated in LOGx/log.txt"
 
 export RANK_SIZE=$1
 EPOCH_SIZE=$2
 LR=$3
-DATASET=$4
-PRE_TRAINED=$6
-PRE_TRAINED_EPOCH_SIZE=$7
-export RANK_TABLE_FILE=$5
+PRE_TRAINED=$5
+PRE_TRAINED_EPOCH_SIZE=$6
+export RANK_TABLE_FILE=$4
 
 for((i=0;i<RANK_SIZE;i++))
 do
@@ -51,27 +48,30 @@ do
     cp ./*.py ./LOG$i
     cp -r ./src ./LOG$i
     cp -r ./scripts ./LOG$i
+    start=`expr $i \* $process_cores`
+    end=`expr $start \+ $(($process_cores-1))`
+    cmdopt=$start"-"$end
     cd ./LOG$i || exit
     export RANK_ID=$i
     echo "start training for rank $i, device $DEVICE_ID"
     env > env.log
-    if [ $# == 5 ]
+    if [ $# == 4 ]
     then
-        python train.py  \
+        taskset -c $cmdopt python train.py  \
+        --workers=$process_cores  \
         --distribute=True  \
         --lr=$LR \
-        --dataset=$DATASET \
         --device_num=$RANK_SIZE  \
         --device_id=$DEVICE_ID  \
         --epoch_size=$EPOCH_SIZE > log.txt 2>&1 &
     fi
 
-    if [ $# == 7 ]
+    if [ $# == 6 ]
     then
-        python train.py  \
+        taskset -c $cmdopt python train.py  \
+        --workers=$process_cores  \
         --distribute=True  \
         --lr=$LR \
-        --dataset=$DATASET \
         --device_num=$RANK_SIZE  \
         --device_id=$DEVICE_ID  \
         --pre_trained=$PRE_TRAINED \

model_zoo/official/cv/retinanet/scripts/run_single_train.sh

@@ -8,6 +8,7 @@
 # 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.
@@ -20,9 +21,9 @@ echo "for example: sh run_single_train.sh 0 500 0.1 coco /opt/retinanet-500_458.
 echo "It is better to use absolute path."
 echo "================================================================================================================="
 
-if [ $# != 4 ] && [ $# != 6 ]
+if [ $# != 3 ] && [ $# != 5 ]
 then
-    echo "Usage: sh run_single_train.sh [DEVICE_ID] [EPOCH_SIZE] [LR] [DATASET] \
+    echo "Usage: sh run_single_train.sh [DEVICE_ID] [EPOCH_SIZE] [LR] \
 [PRE_TRAINED](optional) [PRE_TRAINED_EPOCH_SIZE](optional)"
     exit 1
 fi
@@ -37,9 +38,8 @@ echo "After running the script, the network runs in the background. The log will
 export DEVICE_ID=$1
 EPOCH_SIZE=$2
 LR=$3
-DATASET=$4
-PRE_TRAINED=$5
-PRE_TRAINED_EPOCH_SIZE=$6
+PRE_TRAINED=$4
+PRE_TRAINED_EPOCH_SIZE=$5
 
 rm -rf LOG$1
 mkdir ./LOG$1
@@ -48,23 +48,21 @@ cp -r ./src ./LOG$1
 cd ./LOG$1 || exit
 echo "start training for device $1"
 env > env.log
-if [ $# == 4 ]
+if [ $# == 3 ]
 then
     python train.py  \
     --distribute=False  \
     --lr=$LR \
-    --dataset=$DATASET \
     --device_num=1  \
     --device_id=$DEVICE_ID  \
     --epoch_size=$EPOCH_SIZE > log.txt 2>&1 &
 fi
 
-if [ $# == 6 ]
+if [ $# == 5 ]
 then
     python train,py  \
     --distribute=False  \
     --lr=$LR \
-    --dataset=$DATASET \
     --device_num=1  \
     --device_id=$DEVICE_ID  \
     --pre_trained=$PRE_TRAINED \

model_zoo/official/cv/retinanet/src/dataset.py

@@ -389,7 +389,7 @@ def data_to_mindrecord_byte_image(dataset="coco", is_training=True, prefix="reti
 
 
 def create_retinanet_dataset(mindrecord_file, batch_size, repeat_num, device_num=1, rank=0,
-                             is_training=True, num_parallel_workers=64):
+                             is_training=True, num_parallel_workers=24):
     """Creatr retinanet dataset with MindDataset."""
     ds = de.MindDataset(mindrecord_file, columns_list=["img_id", "image", "annotation"], num_shards=device_num,
                         shard_id=rank, num_parallel_workers=num_parallel_workers, shuffle=is_training)

model_zoo/official/cv/retinanet/src/retinanet.py

@@ -251,9 +251,15 @@ class retinanetWithLossCell(nn.Cell):
         self.expand_dims = P.ExpandDims()
         self.class_loss = SigmoidFocalClassificationLoss(config.gamma, config.alpha)
         self.loc_loss = nn.SmoothL1Loss()
+        self.cast = P.Cast()
+
+        self.network.to_float(mstype.float16)
 
     def construct(self, x, gt_loc, gt_label, num_matched_boxes):
         pred_loc, pred_label = self.network(x)
+        pred_loc = self.cast(pred_loc, mstype.float32)
+        pred_label = self.cast(pred_label, mstype.float32)
+
         mask = F.cast(self.less(0, gt_label), mstype.float32)
         num_matched_boxes = self.reduce_sum(F.cast(num_matched_boxes, mstype.float32))
 

model_zoo/official/cv/retinanet/train.py

@@ -18,7 +18,6 @@
 import os
 import argparse
 import ast
-import mindspore
 import mindspore.nn as nn
 from mindspore import context, Tensor
 from mindspore.communication.management import init, get_rank
@@ -29,7 +28,7 @@ from mindspore.train.serialization import load_checkpoint, load_param_into_net
 from mindspore.common import set_seed
 from src.retinanet import  retinanetWithLossCell, TrainingWrapper, retinanet50, resnet50
 from src.config import config
-from src.dataset import create_retinanet_dataset, create_mindrecord
+from src.dataset import create_retinanet_dataset
 from src.lr_schedule import get_lr
 from src.init_params import init_net_param, filter_checkpoint_parameter
 
@@ -59,15 +58,14 @@ class Monitor(Callback):
 
 def main():
     parser = argparse.ArgumentParser(description="retinanet training")
-    parser.add_argument("--only_create_dataset", type=ast.literal_eval, default=False,
-                        help="If set it true, only create Mindrecord, default is False.")
+
     parser.add_argument("--distribute", type=ast.literal_eval, default=False,
                         help="Run distribute, default is False.")
+    parser.add_argument("--workers", type=int, default=24, help="Num parallel workers.")
     parser.add_argument("--device_id", type=int, default=0, help="Device id, default is 0.")
     parser.add_argument("--device_num", type=int, default=1, help="Use device nums, default is 1.")
     parser.add_argument("--lr", type=float, default=0.1, help="Learning rate, default is 0.1.")
     parser.add_argument("--mode", type=str, default="sink", help="Run sink mode or not, default is sink.")
-    parser.add_argument("--dataset", type=str, default="coco", help="Dataset, default is coco.")
     parser.add_argument("--epoch_size", type=int, default=500, help="Epoch size, default is 500.")
     parser.add_argument("--batch_size", type=int, default=32, help="Batch size, default is 32.")
     parser.add_argument("--pre_trained", type=str, default=None, help="Pretrained Checkpoint file path.")
@@ -98,56 +96,55 @@ def main():
     else:
         raise ValueError("Unsupported platform.")
 
-    mindrecord_file = create_mindrecord(args_opt.dataset, "retinanet.mindrecord", True)
-
-    if not args_opt.only_create_dataset:
-        loss_scale = float(args_opt.loss_scale)
-
-        # When create MindDataset, using the fitst mindrecord file, such as retinanet.mindrecord0.
-        dataset = create_retinanet_dataset(mindrecord_file, repeat_num=1,
-                                           batch_size=args_opt.batch_size, device_num=device_num, rank=rank)
-
-        dataset_size = dataset.get_dataset_size()
-        print("Create dataset done!")
-
-
-        backbone = resnet50(config.num_classes)
-        retinanet = retinanet50(backbone, config)
-        net = retinanetWithLossCell(retinanet, config)
-        net.to_float(mindspore.float16)
-        init_net_param(net)
-
-        if args_opt.pre_trained:
-            if args_opt.pre_trained_epoch_size <= 0:
-                raise KeyError("pre_trained_epoch_size must be greater than 0.")
-            param_dict = load_checkpoint(args_opt.pre_trained)
-            if args_opt.filter_weight:
-                filter_checkpoint_parameter(param_dict)
-            load_param_into_net(net, param_dict)
-
-        lr = Tensor(get_lr(global_step=config.global_step,
-                           lr_init=config.lr_init, lr_end=config.lr_end_rate * args_opt.lr, lr_max=args_opt.lr,
-                           warmup_epochs1=config.warmup_epochs1, warmup_epochs2=config.warmup_epochs2,
-                           warmup_epochs3=config.warmup_epochs3, warmup_epochs4=config.warmup_epochs4,
-                           warmup_epochs5=config.warmup_epochs5, total_epochs=args_opt.epoch_size,
-                           steps_per_epoch=dataset_size))
-        opt = nn.Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), lr,
-                          config.momentum, config.weight_decay, loss_scale)
-        net = TrainingWrapper(net, opt, loss_scale)
-        model = Model(net)
-        print("Start train retinanet, the first epoch will be slower because of the graph compilation.")
-        cb = [TimeMonitor(), LossMonitor()]
-        cb += [Monitor(lr_init=lr.asnumpy())]
-        config_ck = CheckpointConfig(save_checkpoint_steps=dataset_size * args_opt.save_checkpoint_epochs,
-                                     keep_checkpoint_max=config.keep_checkpoint_max)
-        ckpt_cb = ModelCheckpoint(prefix="retinanet", directory=config.save_checkpoint_path, config=config_ck)
-        if args_opt.distribute:
-            if rank == 0:
-                cb += [ckpt_cb]
-            model.train(args_opt.epoch_size, dataset, callbacks=cb, dataset_sink_mode=True)
-        else:
+    mindrecord_file = os.path.join(config.mindrecord_dir, "retinanet.mindrecord0")
+
+    loss_scale = float(args_opt.loss_scale)
+
+    # When create MindDataset, using the fitst mindrecord file, such as retinanet.mindrecord0.
+    dataset = create_retinanet_dataset(mindrecord_file, repeat_num=1,
+                                       num_parallel_workers=args_opt.workers,
+                                       batch_size=args_opt.batch_size, device_num=device_num, rank=rank)
+
+    dataset_size = dataset.get_dataset_size()
+    print("Create dataset done!")
+
+
+    backbone = resnet50(config.num_classes)
+    retinanet = retinanet50(backbone, config)
+    net = retinanetWithLossCell(retinanet, config)
+    init_net_param(net)
+
+    if args_opt.pre_trained:
+        if args_opt.pre_trained_epoch_size <= 0:
+            raise KeyError("pre_trained_epoch_size must be greater than 0.")
+        param_dict = load_checkpoint(args_opt.pre_trained)
+        if args_opt.filter_weight:
+            filter_checkpoint_parameter(param_dict)
+        load_param_into_net(net, param_dict)
+
+    lr = Tensor(get_lr(global_step=config.global_step,
+                       lr_init=config.lr_init, lr_end=config.lr_end_rate * args_opt.lr, lr_max=args_opt.lr,
+                       warmup_epochs1=config.warmup_epochs1, warmup_epochs2=config.warmup_epochs2,
+                       warmup_epochs3=config.warmup_epochs3, warmup_epochs4=config.warmup_epochs4,
+                       warmup_epochs5=config.warmup_epochs5, total_epochs=args_opt.epoch_size,
+                       steps_per_epoch=dataset_size))
+    opt = nn.Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), lr,
+                      config.momentum, config.weight_decay, loss_scale)
+    net = TrainingWrapper(net, opt, loss_scale)
+    model = Model(net)
+    print("Start train retinanet, the first epoch will be slower because of the graph compilation.")
+    cb = [TimeMonitor(), LossMonitor()]
+    cb += [Monitor(lr_init=lr.asnumpy())]
+    config_ck = CheckpointConfig(save_checkpoint_steps=dataset_size * args_opt.save_checkpoint_epochs,
+                                 keep_checkpoint_max=config.keep_checkpoint_max)
+    ckpt_cb = ModelCheckpoint(prefix="retinanet", directory=config.save_checkpoint_path, config=config_ck)
+    if args_opt.distribute:
+        if rank == 0:
             cb += [ckpt_cb]
-            model.train(args_opt.epoch_size, dataset, callbacks=cb, dataset_sink_mode=True)
+        model.train(args_opt.epoch_size, dataset, callbacks=cb, dataset_sink_mode=True)
+    else:
+        cb += [ckpt_cb]
+        model.train(args_opt.epoch_size, dataset, callbacks=cb, dataset_sink_mode=True)
 
 if __name__ == '__main__':
     main()