Col_lin
/
mindspore2022
forked from huawei/mindspore2022
0
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

resnet acc mode update.

This commit is contained in:
linqingke 2021-08-30 20:11:58 +08:00
parent a6b9f3b611
commit a5586fd6a9
19 changed files with 77 additions and 46 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
mindspore/nn/acc/acc.py
View File

@ -28,7 +28,7 @@ _acc_config_level = {
"adasum": False},
"O1": {
"less_bn": True,
"grad_freeze": False,
"grad_freeze": True,
"adasum": False},
"O2": {
"less_bn": True,
@ -53,8 +53,8 @@ class AutoAcc:
self._gc_flag = True
self._param_groups = 10
self._freeze_type = 1
self._freeze_p = 0.5
self._total_steps = -1
self._freeze_p = 0.7
self._total_steps = 65536
self._gradient_groups = None
self._get_configuration(kwargs)
self._param_processer = ParameterProcess()

2
mindspore/nn/acc/grad_freeze.py
View File

@ -59,10 +59,12 @@ class FreezeOpt(Cell):
self.opt_class = type(opt.opt)
self.opt_init_args = opt.opt.init_args
self.lars_init_args = opt.init_args
self.parameters = opt.opt.parameters
else:
self.is_lars = False
self.opt_class = type(opt)
self.opt_init_args = opt.init_args
self.parameters = opt.parameters
self.opts = []
if train_parameter_groups is None:

10
mindspore/nn/wrap/grad_reducer.py
View File

@ -30,14 +30,14 @@ reduce_opt = C.MultitypeFuncGraph("reduce_opt")
def _init_allreduce_operators(length, split_indices, group=GlobalComm.WORLD_COMM_GROUP):
""" initialize allreduce communication operators"""
fusion_type = 2 ** 10
split = 1
split = 0
fusion = ()
for i in range(length):
fusion = fusion + (fusion_type,)
if split_indices[split - 1] <= i + 1:
if split >= len(split_indices):
continue
fusion_type += split
if split >= len(split_indices):
continue
if split_indices[split] <= i:
fusion_type += 1
split += 1
index = tuple(range(1, length + 1))
op_list = ()

2
model_zoo/official/cv/resnet/config/resnet101_imagenet2012_config.yaml
View File

@ -19,6 +19,7 @@ checkpoint_file_path: ""
optimizer: "Momentum"
infer_label: ""
class_num: 1001
train_image_size: 224
batch_size: 32
loss_scale: 1024
momentum: 0.9
@ -41,6 +42,7 @@ dataset: "imagenet2012"
device_num: 1
pre_trained: ""
run_eval: False
eval_image_size: 224
eval_dataset_path: ""
parameter_server: False
filter_weight: False

2
model_zoo/official/cv/resnet/config/resnet18_cifar10_config.yaml
View File

@ -19,6 +19,7 @@ checkpoint_file_path: ""
optimizer: "Momentum"
infer_label: ""
class_num: 10
train_image_size: 224
batch_size: 32
loss_scale: 1024
momentum: 0.9
@ -41,6 +42,7 @@ dataset: "cifar10"
device_num: 1
pre_trained: ""
run_eval: False
eval_image_size: 224
eval_dataset_path: ""
parameter_server: False
filter_weight: False

2
model_zoo/official/cv/resnet/config/resnet18_cifar10_config_gpu.yaml
View File

@ -19,6 +19,7 @@ checkpoint_file_path: ""
optimizer: "Momentum"
infer_label: ""
class_num: 10
train_image_size: 224
batch_size: 32
loss_scale: 1024
momentum: 0.9
@ -41,6 +42,7 @@ dataset: "cifar10"
device_num: 1
pre_trained: ""
run_eval: False
eval_image_size: 224
eval_dataset_path: ""
parameter_server: False
filter_weight: False

2
model_zoo/official/cv/resnet/config/resnet18_imagenet2012_config.yaml
View File

@ -19,6 +19,7 @@ checkpoint_file_path: ""
optimizer: "Momentum"
infer_label: ""
class_num: 1001
train_image_size: 224
batch_size: 256
loss_scale: 1024
momentum: 0.9
@ -43,6 +44,7 @@ dataset: "imagenet2012"
device_num: 1
pre_trained: ""
run_eval: False
eval_image_size: 224
eval_dataset_path: ""
parameter_server: False
filter_weight: False

2
model_zoo/official/cv/resnet/config/resnet18_imagenet2012_config_gpu.yaml
View File

@ -19,6 +19,7 @@ checkpoint_file_path: ""
optimizer: "Momentum"
infer_label: ""
class_num: 1001
train_image_size: 224
batch_size: 256
loss_scale: 1024
momentum: 0.9
@ -43,6 +44,7 @@ dataset: "imagenet2012"
device_num: 1
pre_trained: ""
run_eval: False
eval_image_size: 224
eval_dataset_path: ""
parameter_server: False
filter_weight: False

2
model_zoo/official/cv/resnet/config/resnet34_imagenet2012_config.yaml
View File

@ -19,6 +19,7 @@ checkpoint_file_path: ""
optimizer: "Momentum"
infer_label: ""
class_num: 1001
train_image_size: 224
batch_size: 256
loss_scale: 1024
momentum: 0.9
@ -43,6 +44,7 @@ dataset: "imagenet2012"
device_num: 1
pre_trained: ""
run_eval: False
eval_image_size: 224
eval_dataset_path: ""
parameter_server: False
filter_weight: False

2
model_zoo/official/cv/resnet/config/resnet50_cifar10_config.yaml
View File

@ -19,6 +19,7 @@ checkpoint_file_path: ""
optimizer: "Momentum"
infer_label: ""
class_num: 10
train_image_size: 224
batch_size: 32
loss_scale: 1024
momentum: 0.9
@ -41,6 +42,7 @@ dataset: "cifar10"
device_num: 1
pre_trained: ""
run_eval: False
eval_image_size: 224
eval_dataset_path: ""
parameter_server: False
filter_weight: False

24
model_zoo/official/cv/resnet/config/resnet50_imagenet2012_Acc_config.yaml
View File

@ -16,24 +16,25 @@ checkpoint_file_path: ""
# ==============================================================================
# Training options
optimizer: "Momentum"
optimizer: "LARS"
infer_label: ""
class_num: 1001
train_image_size: 192
batch_size: 256
loss_scale: 1024
momentum: 0.9
weight_decay: 0.0001
epoch_size: 90
momentum: 0.85
weight_decay: 5.0e-5
epoch_size: 38
pretrain_epoch_size: 0
save_checkpoint: True
save_checkpoint: False
save_checkpoint_epochs: 5
keep_checkpoint_max: 10
warmup_epochs: 5
lr_decay_mode: "cosine"
warmup_epochs: 1
lr_decay_mode: "poly"
use_label_smooth: True
label_smooth_factor: 0.1
lr_init: 0
lr_max: 0.8
lr_max: 13.96
lr_end: 0.0
lars_epsilon: 0.0
lars_coefficient: 0.001
@ -43,18 +44,19 @@ dataset: "imagenet2012"
device_num: 1
pre_trained: ""
run_eval: False
eval_image_size: 224
eval_dataset_path: ""
parameter_server: False
filter_weight: False
save_best_ckpt: True
eval_start_epoch: 40
eval_start_epoch: 1
eval_interval: 1
enable_cache: False
cache_session_id: ""
mode_name: "GRAPH"
acc_mode: "O1"
conv_init: "XavierUniform"
dense_init: "TruncatedNormal"
conv_init: "TruncatedNormal"
dense_init: "RandomNormal"
all_reduce_fusion_config:
- 85
- 160

2
model_zoo/official/cv/resnet/config/resnet50_imagenet2012_Ascend_Thor_config.yaml
View File

@ -19,6 +19,7 @@ checkpoint_file_path: ""
optimizer: "Thor"
infer_label: ""
class_num: 1001
train_image_size: 224
batch_size: 32
loss_scale: 128
momentum: 0.9
@ -42,6 +43,7 @@ dataset: "imagenet2012"
device_num: 1
pre_trained: ""
run_eval: False
eval_image_size: 224
eval_dataset_path: ""
parameter_server: False
filter_weight: False

2
model_zoo/official/cv/resnet/config/resnet50_imagenet2012_GPU_Thor_config.yaml
View File

@ -19,6 +19,7 @@ checkpoint_file_path: ""
optimizer: "Thor"
infer_label: ""
class_num: 1001
train_image_size: 224
batch_size: 32
loss_scale: 128
momentum: 0.9
@ -42,6 +43,7 @@ dataset: "imagenet2012"
device_num: 1
pre_trained: ""
run_eval: False
eval_image_size: 224
eval_dataset_path: ""
parameter_server: False
filter_weight: False

2
model_zoo/official/cv/resnet/config/resnet50_imagenet2012_config.yaml
View File

@ -19,6 +19,7 @@ checkpoint_file_path: ""
optimizer: "Momentum"
infer_label: ""
class_num: 1001
train_image_size: 224
batch_size: 256
loss_scale: 1024
momentum: 0.9
@ -43,6 +44,7 @@ dataset: "imagenet2012"
device_num: 1
pre_trained: ""
run_eval: False
eval_image_size: 224
eval_dataset_path: ""
parameter_server: False
filter_weight: False

2
model_zoo/official/cv/resnet/config/resnet_benchmark_GPU.yaml
View File

@ -18,10 +18,12 @@ checkpoint_file_path: ''
# Training options
optimizer: "Momentum"
infer_label: ""
train_image_size: 224
batch_size: 256
epoch_size: 2
print_per_steps: 20
eval: False
eval_image_size: 224
save_ckpt: False
mode_name: "GRAPH"
dtype: "fp16"

2
model_zoo/official/cv/resnet/config/se-resnet50_imagenet2012_config.yaml
View File

@ -19,6 +19,7 @@ checkpoint_file_path: ""
optimizer: "Momentum"
infer_label: ""
class_num: 1001
train_image_size: 224
batch_size: 32
loss_scale: 1024
momentum: 0.9
@ -44,6 +45,7 @@ dataset: "imagenet2012"
device_num: 1
pre_trained: ""
run_eval: False
eval_image_size: 256
eval_dataset_path: ""
parameter_server: False
filter_weight: False

4
model_zoo/official/cv/resnet/scripts/run_distribute_train.sh
View File

@ -121,8 +121,8 @@ do
if [ $# == 5 ]
then
taskset -c $cmdopt python train.py --run_distribute=True --device_num=$RANK_SIZE --data_path=$PATH2 \
--run_eval=$RUN_EVAL --eval_data_path=$EVAL_DATASET_PATH --enable_cache=True --cache_session_id=$CACHE_SESSION_ID \
--config_path=$CONFIG_FILE --output_path './output' &> log &
--run_eval=$RUN_EVAL --eval_dataset_path=$EVAL_DATASET_PATH --enable_cache=True \
--cache_session_id=$CACHE_SESSION_ID --config_path=$CONFIG_FILE --output_path './output' &> log &
if [ "x${RUN_EVAL}" == "xTrue" ]
then
echo -e "\nWhen training run is done, remember to shut down the cache server via \"cache_admin --stop\""

46
model_zoo/official/cv/resnet/src/dataset.py
View File

@ -25,8 +25,8 @@ from src.model_utils.config import config
from src.model_utils.device_adapter import get_device_num, get_rank_id
def create_dataset1(dataset_path, do_train, repeat_num=1, batch_size=32, target="Ascend", distribute=False,
enable_cache=False, cache_session_id=None):
def create_dataset1(dataset_path, do_train, repeat_num=1, batch_size=32, train_image_size=224, eval_image_size=224,
target="Ascend", distribute=False, enable_cache=False, cache_session_id=None):
"""
create a train or evaluate cifar10 dataset for resnet50
Args:
@ -67,7 +67,7 @@ def create_dataset1(dataset_path, do_train, repeat_num=1, batch_size=32, target=
]
trans += [
C.Resize((224, 224)),
C.Resize((train_image_size, train_image_size)),
C.Rescale(1.0 / 255.0, 0.0),
C.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010]),
C.HWC2CHW()
@ -95,8 +95,8 @@ def create_dataset1(dataset_path, do_train, repeat_num=1, batch_size=32, target=
return data_set
def create_dataset2(dataset_path, do_train, repeat_num=1, batch_size=32, target="Ascend", distribute=False,
enable_cache=False, cache_session_id=None):
def create_dataset2(dataset_path, do_train, repeat_num=1, batch_size=32, train_image_size=224, eval_image_size=224,
target="Ascend", distribute=False, enable_cache=False, cache_session_id=None):
"""
create a train or eval imagenet2012 dataset for resnet50
@ -130,14 +130,13 @@ def create_dataset2(dataset_path, do_train, repeat_num=1, batch_size=32, target=
data_set = ds.ImageFolderDataset(dataset_path, num_parallel_workers=12, shuffle=True,
num_shards=device_num, shard_id=rank_id)
image_size = 224
mean = [0.485 * 255, 0.456 * 255, 0.406 * 255]
std = [0.229 * 255, 0.224 * 255, 0.225 * 255]
# define map operations
if do_train:
trans = [
C.RandomCropDecodeResize(image_size, scale=(0.08, 1.0), ratio=(0.75, 1.333)),
C.RandomCropDecodeResize(train_image_size, scale=(0.08, 1.0), ratio=(0.75, 1.333)),
C.RandomHorizontalFlip(prob=0.5),
C.Normalize(mean=mean, std=std),
C.HWC2CHW()
@ -146,7 +145,7 @@ def create_dataset2(dataset_path, do_train, repeat_num=1, batch_size=32, target=
trans = [
C.Decode(),
C.Resize(256),
C.CenterCrop(image_size),
C.CenterCrop(eval_image_size),
C.Normalize(mean=mean, std=std),
C.HWC2CHW()
]
@ -174,8 +173,9 @@ def create_dataset2(dataset_path, do_train, repeat_num=1, batch_size=32, target=
return data_set
def create_dataset_pynative(dataset_path, do_train, repeat_num=1, batch_size=32, target="Ascend", distribute=False,
enable_cache=False, cache_session_id=None):
def create_dataset_pynative(dataset_path, do_train, repeat_num=1, batch_size=32, train_image_size=224,
eval_image_size=224, target="Ascend", distribute=False, enable_cache=False,
cache_session_id=None):
"""
create a train or eval imagenet2012 dataset for resnet50 benchmark
@ -209,14 +209,14 @@ def create_dataset_pynative(dataset_path, do_train, repeat_num=1, batch_size=32,
data_set = ds.ImageFolderDataset(dataset_path, num_parallel_workers=2, shuffle=True,
num_shards=device_num, shard_id=rank_id)
image_size = 224
mean = [0.485 * 255, 0.456 * 255, 0.406 * 255]
std = [0.229 * 255, 0.224 * 255, 0.225 * 255]
# define map operations
if do_train:
trans = [
C.RandomCropDecodeResize(image_size, scale=(0.08, 1.0), ratio=(0.75, 1.333)),
C.RandomCropDecodeResize(train_image_size, scale=(0.08, 1.0), ratio=(0.75, 1.333)),
C.RandomHorizontalFlip(prob=0.5),
C.Normalize(mean=mean, std=std),
C.HWC2CHW()
@ -225,7 +225,7 @@ def create_dataset_pynative(dataset_path, do_train, repeat_num=1, batch_size=32,
trans = [
C.Decode(),
C.Resize(256),
C.CenterCrop(image_size),
C.CenterCrop(eval_image_size),
C.Normalize(mean=mean, std=std),
C.HWC2CHW()
]
@ -253,8 +253,8 @@ def create_dataset_pynative(dataset_path, do_train, repeat_num=1, batch_size=32,
return data_set
def create_dataset3(dataset_path, do_train, repeat_num=1, batch_size=32, target="Ascend", distribute=False,
enable_cache=False, cache_session_id=None):
def create_dataset3(dataset_path, do_train, repeat_num=1, batch_size=32, train_image_size=224, eval_image_size=224,
target="Ascend", distribute=False, enable_cache=False, cache_session_id=None):
"""
create a train or eval imagenet2012 dataset for resnet101
Args:
@ -285,14 +285,14 @@ def create_dataset3(dataset_path, do_train, repeat_num=1, batch_size=32, target=
else:
data_set = ds.ImageFolderDataset(dataset_path, num_parallel_workers=8, shuffle=True,
num_shards=device_num, shard_id=rank_id)
image_size = 224
mean = [0.475 * 255, 0.451 * 255, 0.392 * 255]
std = [0.275 * 255, 0.267 * 255, 0.278 * 255]
# define map operations
if do_train:
trans = [
C.RandomCropDecodeResize(image_size, scale=(0.08, 1.0), ratio=(0.75, 1.333)),
C.RandomCropDecodeResize(train_image_size, scale=(0.08, 1.0), ratio=(0.75, 1.333)),
C.RandomHorizontalFlip(rank_id / (rank_id + 1)),
C.Normalize(mean=mean, std=std),
C.HWC2CHW()
@ -301,7 +301,7 @@ def create_dataset3(dataset_path, do_train, repeat_num=1, batch_size=32, target=
trans = [
C.Decode(),
C.Resize(256),
C.CenterCrop(image_size),
C.CenterCrop(eval_image_size),
C.Normalize(mean=mean, std=std),
C.HWC2CHW()
]
@ -329,8 +329,8 @@ def create_dataset3(dataset_path, do_train, repeat_num=1, batch_size=32, target=
return data_set
def create_dataset4(dataset_path, do_train, repeat_num=1, batch_size=32, target="Ascend", distribute=False,
enable_cache=False, cache_session_id=None):
def create_dataset4(dataset_path, do_train, repeat_num=1, batch_size=32, train_image_size=224, eval_image_size=224,
target="Ascend", distribute=False, enable_cache=False, cache_session_id=None):
"""
create a train or eval imagenet2012 dataset for se-resnet50
@ -362,14 +362,14 @@ def create_dataset4(dataset_path, do_train, repeat_num=1, batch_size=32, target=
else:
data_set = ds.ImageFolderDataset(dataset_path, num_parallel_workers=12, shuffle=True,
num_shards=device_num, shard_id=rank_id)
image_size = 224
mean = [123.68, 116.78, 103.94]
std = [1.0, 1.0, 1.0]
# define map operations
if do_train:
trans = [
C.RandomCropDecodeResize(image_size, scale=(0.08, 1.0), ratio=(0.75, 1.333)),
C.RandomCropDecodeResize(train_image_size, scale=(0.08, 1.0), ratio=(0.75, 1.333)),
C.RandomHorizontalFlip(prob=0.5),
C.Normalize(mean=mean, std=std),
C.HWC2CHW()
@ -378,7 +378,7 @@ def create_dataset4(dataset_path, do_train, repeat_num=1, batch_size=32, target=
trans = [
C.Decode(),
C.Resize(292),
C.CenterCrop(256),
C.CenterCrop(eval_image_size),
C.Normalize(mean=mean, std=std),
C.HWC2CHW()
]

7
model_zoo/official/cv/resnet/train.py
View File

@ -199,7 +199,9 @@ def run_eval(target, model, ckpt_save_dir, cb):
if config.eval_dataset_path is None or (not os.path.isdir(config.eval_dataset_path)):
raise ValueError("{} is not a existing path.".format(config.eval_dataset_path))
eval_dataset = create_dataset(dataset_path=config.eval_dataset_path, do_train=False,
batch_size=config.batch_size, target=target, enable_cache=config.enable_cache,
batch_size=config.batch_size, train_image_size=config.train_image_size,
eval_image_size=config.eval_image_size,
target=target, enable_cache=config.enable_cache,
cache_session_id=config.cache_session_id)
eval_param_dict = {"model": model, "dataset": eval_dataset, "metrics_name": "acc"}
eval_cb = EvalCallBack(apply_eval, eval_param_dict, interval=config.eval_interval,
@ -226,7 +228,8 @@ def train_net():
target = config.device_target
set_parameter()
dataset = create_dataset(dataset_path=config.data_path, do_train=True, repeat_num=1,
batch_size=config.batch_size, target=target,
batch_size=config.batch_size, train_image_size=config.train_image_size,
eval_image_size=config.eval_image_size, target=target,
distribute=config.run_distribute)
step_size = dataset.get_dataset_size()
net = resnet(class_num=config.class_num)