diff --git a/official/cv/resnext/README_CN.md b/official/cv/resnext/README_CN.md
index 65bc48c3a32ab0aa5d5aecd0077ba31376c27eaf..ce817653d4cf12d1b915dd4c9966a79a5183856b 100644
--- a/official/cv/resnext/README_CN.md
+++ b/official/cv/resnext/README_CN.md
@@ -203,6 +203,11 @@ GPU:
bash run_distribute_train_for_gpu.sh DATA_PATH
# 鍗曟満璁粌
bash run_standalone_train_for_gpu.sh DEVICE_ID DATA_PATH
+GPU_ResNext101:
+ # 鍒嗗竷寮忚缁冪ず渚嬶紙8鍗★級
+ bash scripts/run_distribute_train_for_gpu_resnext101.sh DATA_PATH CONFIG_PATH
+ # 鍗曟満璁粌
+ bash scripts/run_standalone_train_for_gpu_resnext101.sh DEVICE_ID DATA_PATH CONFIG_PATH
```
## 杩佺Щ璁粌杩囩▼
@@ -232,6 +237,8 @@ python eval.py --data_path ~/imagenet/val/ --device_target Ascend --checkpoint_f
鎴栭€氳繃shell鑴氭湰寮€濮嬭缁冿細
```shell
+# GPU_ResNext101璇勪及
+bash scripts/run_eval_for_gpu_resnext101.sh DEVICE_ID DATA_PATH CHECKPOINT_FILE_PATH CONFIG_PATH
# 璇勪及
bash scripts/run_eval.sh DEVICE_ID DATA_PATH CHECKPOINT_FILE_PATH PLATFORM
```
@@ -243,6 +250,8 @@ DEVICE_TARGET is Ascend or GPU, default is Ascend.
```shell
# 妫€鏌ョ偣璇勪及
bash scripts/run_eval.sh 0 /opt/npu/datasets/classification/val /ResNeXt_100.ckpt Ascend
+# GPU_ResNext101璇勪及
+bash scripts/run_eval_for_gpu_resnext101.sh 0 ~/ImageNet2012/val ~/outputs_demo/best_acc.ckpt ./resnext101_config.yaml
```
#### 缁撴灉
@@ -348,6 +357,20 @@ Total data:50000, top1 accuracy:0.79858, top5 accuracy:0.94716
| 鎬绘椂闀� | 7.8灏忔椂 锛�8鍗★級 | 21.5灏忔椂 锛�8鍗★級 |
| 璋冧紭妫€鏌ョ偣 | 192 M锛�.ckpt鏂囦欢锛� | 192 M锛�.ckpt鏂囦欢锛� |
+| 鍙傛暟 | ResNeXt101 |
+| ------------- | ------------------------------------------------- |
+| 璧勬簮 | GeForce RTX 3090锛汣PU锛�3.50GHz锛�64鏍革紱鍐呭瓨锛�264GB |
+| 涓婁紶鏃ユ湡 | 2022-6-30 |
+| MindSpore鐗堟湰 | 1.7.0 |
+| 鏁版嵁闆� | ImageNet |
+| 璁粌鍙傛暟 | src/config.py |
+| 浼樺寲鍣� | Momentum |
+| 鎹熷け鍑芥暟 | Softmax浜ゅ弶鐔� |
+| 鎹熷け | 1.2427 |
+| 鍑嗙‘鐜� | 78.04%(TOP1) |
+| 鎬绘椂闀� | 50灏忔椂 锛�8鍗★級 |
+| 璋冧紭妫€鏌ョ偣 | 638 M锛�.ckpt鏂囦欢锛� |
+
#### 鎺ㄧ悊鎬ц兘
| 鍙傛暟 |ResNeXt50 | | |
@@ -360,15 +383,15 @@ Total data:50000, top1 accuracy:0.79858, top5 accuracy:0.94716
| 杈撳嚭 | 姒傜巼 | 姒傜巼 | 姒傜巼 |
| 鍑嗙‘鐜� | acc=78.16%(TOP1) | acc=78.05%(TOP1) | |
-| 鍙傛暟 | ResNeXt101 |
-| ------------------- | --------------------------- |
-| 璧勬簮 | Ascend 310; OS Euler2.8 |
-| 涓婁紶鏃ユ湡 | 06/22/2021 (month/day/year) |
-| MindSpore鐗堟湰 | 1.2.0 |
-| 鏁版嵁闆� | ImageNet |
-| batch_size | 1 |
-| 杈撳嚭 | 姒傜巼 |
-| 鍑嗙‘鐜� | TOP1: 79.85%, TOP5: 94.71% |
+| 鍙傛暟 | ResNeXt101 | |
+| ------------------- | --------------------------- | --------------------------- |
+| 璧勬簮 | Ascend 310; OS Euler2.8 | GeForce RTX 3090 |
+| 涓婁紶鏃ユ湡 | 06/22/2021 (month/day/year) | 2022-6-30 |
+| MindSpore鐗堟湰 | 1.2.0 | 1.7.0 |
+| 鏁版嵁闆� | ImageNet | ImageNet |
+| batch_size | 1 | 1 |
+| 杈撳嚭 | 姒傜巼 | 姒傜巼 |
+| 鍑嗙‘鐜� | TOP1: 79.85%, TOP5: 94.71% | 78.04%(TOP1) |
# 闅忔満鎯呭喌璇存槑
diff --git a/official/cv/resnext/resnext101_config.yaml b/official/cv/resnext/resnext101_config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9b5ed9069c6817fe171391fead01243686e627e6
--- /dev/null
+++ b/official/cv/resnext/resnext101_config.yaml
@@ -0,0 +1,79 @@
+# Builtin Configurations(DO NOT CHANGE THESE CONFIGURATIONS unless you know exactly what you are doing)
+enable_modelarts: False
+network: "resnext101"
+# Url for modelarts
+data_url: ""
+train_url: ""
+checkpoint_url: ""
+# Path for local
+run_distribute: False
+enable_profiling: False
+data_path: "/cache/data"
+eval_data_path: "/cache/data"
+output_path: ""
+load_path: ""
+device_target: "GPU"
+checkpoint_path: "./checkpoint/"
+checkpoint_file_path: ""
+run_eval: True
+eval_per_batch_size: 64
+eval_interval: 1
+eval_start_epoch: 120
+save_best_ckpt: True
+ckpt_path: 'outputs_demo/'
+# ==============================================================================
+# Training options
+image_size: [224,224]
+num_classes: 1000
+batch_size: 1
+
+lr: 0.05
+lr_scheduler: "cosine_annealing"
+lr_epochs: [30,60,90,120]
+lr_gamma: 0.1
+eta_min: 0
+T_max: 150
+max_epoch: 150
+warmup_epochs: 1
+
+weight_decay: 0.0001
+momentum: 0.9
+is_dynamic_loss_scale: 0
+loss_scale: 1024
+label_smooth: 1
+label_smooth_factor: 0.1
+per_batch_size: 96
+
+ckpt_interval: 5
+ckpt_save_max: 5
+is_save_on_master: 1
+rank_save_ckpt_flag: 0
+outputs_dir: ""
+log_path: "./output_log"
+
+# Export options
+device_id: 0
+width: 224
+height: 224
+file_name: "resnext101"
+file_format: 'MINDIR'
+result_path: ""
+label_path: ""
+
+---
+# Help description for each configuration
+enable_modelarts: "Whether training on modelarts, default: False"
+data_url: "Dataset url for obs"
+train_url: "Training output url for obs"
+checkpoint_url: "The location of checkpoint for obs"
+data_path: "Dataset path for local"
+output_path: "Training output path for local"
+load_path: "The location of checkpoint for obs"
+device_target: "Target device type, available: [Ascend, GPU, CPU]"
+enable_profiling: "Whether enable profiling while training, default: False"
+num_classes: "Class for dataset"
+batch_size: "Batch size for training and evaluation"
+epoch_size: "Total training epochs."
+keep_checkpoint_max: "keep the last keep_checkpoint_max checkpoint"
+checkpoint_path: "The location of the checkpoint file."
+checkpoint_file_path: "The location of the checkpoint file."
diff --git a/official/cv/resnext/scripts/run_distribute_train_for_gpu_resnext101.sh b/official/cv/resnext/scripts/run_distribute_train_for_gpu_resnext101.sh
new file mode 100755
index 0000000000000000000000000000000000000000..c3ce9d1e1d4c0e20847223b96f74c63776971606
--- /dev/null
+++ b/official/cv/resnext/scripts/run_distribute_train_for_gpu_resnext101.sh
@@ -0,0 +1,55 @@
+#!/bin/bash
+# Copyright 2022 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 [ $# != 2 ]
+then
+ echo "Usage: bash run_distribute_train_for_gpu_resnext101.sh [DATA_PATH] [CONFIG_PATH]."
+ exit 1
+fi
+
+get_real_path(){
+ if [ "${1:0:1}" == "/" ]; then
+ echo "$1"
+ else
+ echo "$(realpath -m $PWD/$1)"
+ fi
+}
+
+if [ ! -d "$(get_real_path $1)" ]
+then
+ echo "error: DATA_PATH=$1 is not a directory"
+ echo "Usage: bash run_distribute_train_for_gpu_resnext101.sh [DATA_PATH] [CONFIG_PATH]."
+exit 1
+fi
+
+if [ ! -f "$(get_real_path $2)" ]
+then
+ echo "error: CONFIG_PATH=$2 is not a file"
+ echo "Usage: bash run_distribute_train_for_gpu_resnext101.sh [DATA_PATH] [CONFIG_PATH]."
+exit 1
+fi
+
+DATA_PATH=$(get_real_path $1)
+CONFIG_PATH=$(get_real_path $2)
+
+TRAIN_DATA_PATH="${DATA_PATH}/train"
+VALID_DATA_PATH="${DATA_PATH}/val"
+
+rm -rf logs
+mkdir logs
+cp -r ./train.py ./resnext101_config.yaml ./src ./logs
+cd ./logs
+echo "start training"
+mpirun --allow-run-as-root -n 8 python ./train.py --data_path=${TRAIN_DATA_PATH} --eval_data_path=${VALID_DATA_PATH} --run_distribute=True --config_path=${CONFIG_PATH} > train.log 2>&1 &
diff --git a/official/cv/resnext/scripts/run_eval_for_gpu_resnext101.sh b/official/cv/resnext/scripts/run_eval_for_gpu_resnext101.sh
new file mode 100755
index 0000000000000000000000000000000000000000..424f1c42af5b43e9b597de3042ac055d51988091
--- /dev/null
+++ b/official/cv/resnext/scripts/run_eval_for_gpu_resnext101.sh
@@ -0,0 +1,51 @@
+#!/bin/bash
+# Copyright 2022 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 [ $# != 4 ]
+then
+ echo "Usage: bash run_eval_for_gpu_resnext101.sh [DEVICE_ID] [EVAL_DATA_DIR] [CHECKPOINTPATH] [CONFIG_PATH]."
+ exit 1
+fi
+
+if [ ! -d "$(get_real_path $2)" ]
+then
+ echo "error: EVAL_DATA_DIR=$2 is not a directory"
+ echo "Usage: bash run_eval_for_gpu_resnext101.sh [DEVICE_ID] [EVAL_DATA_DIR] [CHECKPOINTPATH] [CONFIG_PATH]."
+exit 1
+fi
+
+if [ ! -f "$(get_real_path $3)" ]
+then
+ echo "error: CHECKPOINTPATH=$3 is not a file"
+ echo "Usage: bash run_eval_for_gpu_resnext101.sh [DEVICE_ID] [EVAL_DATA_DIR] [CHECKPOINTPATH] [CONFIG_PATH]."
+exit 1
+fi
+
+if [ ! -f "$(get_real_path $4)" ]
+then
+ echo "error: CONFIG_PATH=$4 is not a file"
+ echo "Usage: bash run_eval_for_gpu_resnext101.sh [DEVICE_ID] [EVAL_DATA_DIR] [CHECKPOINTPATH] [CONFIG_PATH]."
+exit 1
+fi
+
+export CUDA_VISIBLE_DEVICES=$1
+DATA_DIR=$(get_real_path $2)
+PATH_CHECKPOINT=$(get_real_path $3)
+CONFIG_PATH=$(get_real_path $4)
+python eval.py \
+ --checkpoint_file_path=$PATH_CHECKPOINT \
+ --data_path=$DATA_DIR \
+ --config_path=$CONFIG_PATH > eval.log 2>&1 &
+
diff --git a/official/cv/resnext/scripts/run_standalone_train_for_gpu_resnext101.sh b/official/cv/resnext/scripts/run_standalone_train_for_gpu_resnext101.sh
new file mode 100755
index 0000000000000000000000000000000000000000..fab81e872c478e2a6689f1327c02f730741ffb6f
--- /dev/null
+++ b/official/cv/resnext/scripts/run_standalone_train_for_gpu_resnext101.sh
@@ -0,0 +1,53 @@
+#!/bin/bash
+# Copyright 2022 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 [ $# != 3 ]
+then
+ echo "Usage: bash run_standalone_train_for_gpu_resnext101.sh [DEVICE_ID] [DATA_PATH] [CONFIG_PATH]."
+ exit 1
+fi
+
+get_real_path(){
+ if [ "${1:0:1}" == "/" ]; then
+ echo "$1"
+ else
+ echo "$(realpath -m $PWD/$1)"
+ fi
+}
+
+if [ ! -d "$(get_real_path $2)" ]
+then
+ echo "error: DATA_PATH=$2 is not a directory"
+ echo "Usage: bash run_standalone_train_for_gpu_resnext101.sh [DEVICE_ID] [DATA_PATH] [CONFIG_PATH]."
+exit 1
+fi
+
+if [ ! -f "$(get_real_path $3)" ]
+then
+ echo "error: CONFIG_PATH=$3 is not a file"
+ echo "Usage: bash run_standalone_train_for_gpu_resnext101.sh [DEVICE_ID] [DATA_PATH] [CONFIG_PATH]."
+exit 1
+fi
+
+export CUDA_VISIBLE_DEVICES=$1
+DATA_DIR=$(get_real_path $2)
+CONFIG_PATH=$(get_real_path $3)
+TRAIN_DATA_PATH="${DATA_DIR}/train"
+EVAL_DATA_PATH="${DATA_DIR}/val"
+
+python train.py \
+ --data_path=${TRAIN_DATA_PATH} \
+ --eval_data_path=${EVAL_DATA_PATH} \
+ --config_path=${CONFIG_PATH} > train.log 2>&1 &
diff --git a/official/cv/resnext/src/backbone/resnet.py b/official/cv/resnext/src/backbone/resnet.py
index cbb6c544ff0cb775c49fe23a9a731a3bd0e5ca2e..be76782ffe7b4fa40bd15991990d725e1187595d 100644
--- a/official/cv/resnext/src/backbone/resnet.py
+++ b/official/cv/resnext/src/backbone/resnet.py
@@ -1,4 +1,4 @@
-# Copyright 2020 Huawei Technologies Co., Ltd
+# Copyright 2020-2022 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.
@@ -16,7 +16,7 @@
ResNet based ResNext
"""
import mindspore.nn as nn
-from mindspore.ops.operations import Add, Split, Concat
+from mindspore.ops.operations import TensorAdd, Split, Concat
from mindspore.ops import operations as P
from mindspore.common.initializer import TruncatedNormal
@@ -105,7 +105,7 @@ class BasicBlock(nn.Cell):
self.down_sample = down_sample
self.down_sample_flag = True
- self.add = Add()
+ self.add = TensorAdd()
def construct(self, x):
identity = x
@@ -176,7 +176,7 @@ class Bottleneck(nn.Cell):
self.down_sample_flag = True
self.cast = P.Cast()
- self.add = Add()
+ self.add = TensorAdd()
def construct(self, x):
identity = x
@@ -279,4 +279,6 @@ def resnext50(platform="Ascend"):
return ResNet(Bottleneck, [3, 4, 6, 3], width_per_group=4, groups=32, platform=platform)
def resnext101(platform="Ascend"):
- return ResNet(Bottleneck, [3, 4, 23, 3], width_per_group=4, groups=32, platform=platform)
+ if platform == "Ascend":
+ return ResNet(Bottleneck, [3, 4, 23, 3], width_per_group=4, groups=32, platform=platform)
+ return ResNet(Bottleneck, [3, 4, 23, 3], width_per_group=4, groups=64, platform=platform)
diff --git a/official/cv/resnext/src/eval_callback.py b/official/cv/resnext/src/eval_callback.py
new file mode 100644
index 0000000000000000000000000000000000000000..d1db389440e57f76515fdccde38382750c52cdea
--- /dev/null
+++ b/official/cv/resnext/src/eval_callback.py
@@ -0,0 +1,158 @@
+# Copyright 2022 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.
+# ============================================================================
+"""Evaluation callback when training"""
+
+import os
+import stat
+import time
+from mindspore import save_checkpoint
+from mindspore import log as logger
+from mindspore.train.callback import Callback
+
+
+class EvalCallBack(Callback):
+ """
+ Evaluation callback when training.
+
+ Args:
+ eval_function (function): evaluation function.
+ eval_param_dict (dict): evaluation parameters' configure dict.
+ interval (int): run evaluation interval, default is 1.
+ eval_start_epoch (int): evaluation start epoch, default is 1.
+ save_best_ckpt (bool): Whether to save best checkpoint, default is True.
+ best_ckpt_name (str): bast checkpoint name, default is `best.ckpt`.
+ metrics_name (str): evaluation metrics name, default is `acc`.
+
+ Returns:
+ None
+
+ Examples:
+ >>> EvalCallBack(eval_function, eval_param_dict)
+ """
+
+ def __init__(self, eval_function, eval_param_dict, interval=1, eval_start_epoch=1, save_best_ckpt=True,
+ ckpt_directory="./", best_ckpt_name="best.ckpt", metrics_name="acc"):
+ super(EvalCallBack, self).__init__()
+ self.eval_param_dict = eval_param_dict
+ self.eval_function = eval_function
+ self.eval_start_epoch = eval_start_epoch
+ if interval < 1:
+ raise ValueError("interval should >= 1.")
+ self.interval = interval
+ self.save_best_ckpt = save_best_ckpt
+ self.best_res = 0
+ self.best_epoch = 0
+ if not os.path.isdir(ckpt_directory):
+ os.makedirs(ckpt_directory)
+ self.bast_ckpt_path = os.path.join(ckpt_directory, best_ckpt_name)
+ self.metrics_name = metrics_name
+
+ def remove_ckpoint_file(self, file_name):
+ """Remove the specified checkpoint file from this checkpoint manager and also from the directory."""
+ try:
+ os.chmod(file_name, stat.S_IWRITE)
+ os.remove(file_name)
+ except OSError:
+ logger.warning("OSError, failed to remove the older ckpt file %s.", file_name)
+ except ValueError:
+ logger.warning("ValueError, failed to remove the older ckpt file %s.", file_name)
+
+ def epoch_end(self, run_context):
+ """Callback when epoch end."""
+ cb_params = run_context.original_args()
+ cur_epoch = cb_params.cur_epoch_num
+ if cur_epoch >= self.eval_start_epoch and (cur_epoch - self.eval_start_epoch) % self.interval == 0:
+ eval_start = time.time()
+ res = self.eval_function(self.eval_param_dict)
+ eval_cost = time.time() - eval_start
+ print("epoch: {}, {}: {}, eval_cost:{:.2f}".format(cur_epoch, self.metrics_name, res, eval_cost),
+ flush=True)
+ if res >= self.best_res:
+ self.best_res = res
+ self.best_epoch = cur_epoch
+ print("update best result: {}".format(res), flush=True)
+ if self.save_best_ckpt:
+ if os.path.exists(self.bast_ckpt_path):
+ self.remove_ckpoint_file(self.bast_ckpt_path)
+ save_checkpoint(cb_params.train_network, self.bast_ckpt_path)
+ print("update best checkpoint at: {}".format(self.bast_ckpt_path), flush=True)
+
+ def end(self, run_context):
+ print("End training, the best {0} is: {1}, the best {0} epoch is {2}".format(self.metrics_name,
+ self.best_res,
+ self.best_epoch), flush=True)
+
+class ProgressMonitor(Callback):
+ """monitor loss and time"""
+ def __init__(self, args):
+ super(ProgressMonitor, self).__init__()
+ self.me_epoch_start_time = 0
+ self.me_epoch_start_step_num = 0
+ self.args = args
+ self.ckpt_history = []
+
+ def begin(self, run_context):
+ self.args.logger.info('start network train...')
+
+ def epoch_begin(self, run_context):
+ pass
+
+ def epoch_end(self, run_context, *me_args):
+ """describe network construct"""
+ cb_params = run_context.original_args()
+ me_step = cb_params.cur_step_num - 1
+
+ real_epoch = me_step // self.args.steps_per_epoch
+ time_used = time.time() - self.me_epoch_start_time
+ fps_mean = (self.args.per_batch_size * (me_step-self.me_epoch_start_step_num))
+ fps_mean = fps_mean * self.args.group_size
+ fps_mean = fps_mean / time_used
+ self.args.logger.info('epoch[{}], iter[{}], loss:{}, '
+ 'mean_fps:{:.2f}'
+ 'imgs/sec'.format(real_epoch,
+ me_step,
+ cb_params.net_outputs,
+ fps_mean))
+
+ if self.args.rank_save_ckpt_flag:
+ import glob
+ ckpts = glob.glob(os.path.join(self.args.outputs_dir, '*.ckpt'))
+ for ckpt in ckpts:
+ ckpt_fn = os.path.basename(ckpt)
+ if not ckpt_fn.startswith('{}-'.format(self.args.rank)):
+ continue
+ if ckpt in self.ckpt_history:
+ continue
+ self.ckpt_history.append(ckpt)
+ self.args.logger.info('epoch[{}], iter[{}], loss:{}, '
+ 'ckpt:{},'
+ 'ckpt_fn:{}'.format(real_epoch,
+ me_step,
+ cb_params.net_outputs,
+ ckpt,
+ ckpt_fn))
+
+
+ self.me_epoch_start_step_num = me_step
+ self.me_epoch_start_time = time.time()
+
+ def step_begin(self, run_context):
+ pass
+
+ def step_end(self, run_context, *me_args):
+ pass
+
+ def end(self, run_context):
+ self.args.logger.info('end network train...')
diff --git a/official/cv/resnext/train.py b/official/cv/resnext/train.py
index cb0704f108d74b6e25acb6ffc808d26a5970b2c8..bf8645064632a9dd026501c84f65fcde092d6481 100644
--- a/official/cv/resnext/train.py
+++ b/official/cv/resnext/train.py
@@ -1,4 +1,4 @@
-# Copyright 2020 Huawei Technologies Co., Ltd
+# Copyright 2020-2022 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.
@@ -35,6 +35,7 @@ from src.utils.var_init import load_pretrain_model
from src.image_classification import get_network
from src.model_utils.config import config
from src.model_utils.moxing_adapter import moxing_wrapper
+from src.eval_callback import EvalCallBack, ProgressMonitor
set_seed(1)
@@ -84,13 +85,23 @@ def set_graph_kernel_context(device_target):
if device_target == "GPU":
context.set_context(enable_graph_kernel=True)
+def apply_eval(eval_param):
+ eval_model = eval_param["model"]
+ eval_ds = eval_param["dataset"]
+ metrics_name = eval_param["metrics_name"]
+ res = eval_model.eval(eval_ds)
+ return res[metrics_name]
+
@moxing_wrapper()
def train():
"""training process"""
set_parameters()
- if os.getenv('DEVICE_ID', "not_set").isdigit():
- context.set_context(device_id=int(os.getenv('DEVICE_ID')))
- set_graph_kernel_context(config.device_target)
+ if config.device_target == "Ascend":
+ if os.getenv('DEVICE_ID', "not_set").isdigit():
+ context.set_context(device_id=int(os.getenv('DEVICE_ID')))
+ elif config.device_target == "GPU":
+ if os.getenv('CUDA_VISIBLE_DEVICES', "not_set").isdigit():
+ context.set_context(device_id=0)
# init distributed
if config.run_distribute:
@@ -101,6 +112,7 @@ def train():
de_dataset = classification_dataset(config.data_path, config.image_size,
config.per_batch_size, 1,
config.rank, config.group_size, num_parallel_workers=8)
+ de_dataset.map_model = 4 # !!!important
config.steps_per_epoch = de_dataset.get_dataset_size()
config.logger.save_args(config)
@@ -133,11 +145,37 @@ def train():
else:
loss_scale_manager = FixedLossScaleManager(config.loss_scale, drop_overflow_update=False)
- model = Model(network, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale_manager,
- metrics={'acc'}, amp_level="O3")
-
- # checkpoint save
- callbacks = [TimeMonitor(data_size=config.steps_per_epoch), LossMonitor()]
+ if config.network == "resnext101":
+ model = Model(network, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale_manager,
+ metrics={'acc'}, amp_level="O2")
+ # checkpoint save
+ callbacks = [ProgressMonitor(config), LossMonitor()]
+ else:
+ model = Model(network, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale_manager,
+ metrics={'acc'}, amp_level="O3")
+ callbacks = [TimeMonitor(data_size=config.steps_per_epoch), LossMonitor()]
+
+ if config.run_eval and config.rank_save_ckpt_flag:
+ if config.eval_data_path is None or (not os.path.isdir(config.eval_data_path)):
+ raise ValueError("{} is not a existing path.".format(config.eval_data_path))
+ eval_de_dataset = classification_dataset(config.eval_data_path,
+ image_size=config.image_size,
+ per_batch_size=config.eval_per_batch_size,
+ max_epoch=1,
+ rank=config.rank,
+ group_size=config.group_size,
+ mode='eval')
+ eval_param_dict = {"model": model, "dataset": eval_de_dataset, "metrics_name": "acc"}
+ eval_callback = EvalCallBack(apply_eval,
+ eval_param_dict,
+ interval=config.eval_interval,
+ eval_start_epoch=config.eval_start_epoch,
+ save_best_ckpt=config.save_best_ckpt,
+ ckpt_directory=config.ckpt_path,
+ best_ckpt_name="best_acc.ckpt",
+ metrics_name="acc"
+ )
+ callbacks.append(eval_callback)
if config.rank_save_ckpt_flag:
ckpt_config = CheckpointConfig(save_checkpoint_steps=config.ckpt_interval * config.steps_per_epoch,
keep_checkpoint_max=config.ckpt_save_max)
@@ -146,7 +184,6 @@ def train():
directory=save_ckpt_path,
prefix='{}'.format(config.rank))
callbacks.append(ckpt_cb)
-
model.train(config.max_epoch, de_dataset, callbacks=callbacks, dataset_sink_mode=True)