mirrors
/
mmselfsup
mirror of https://github.com/open-mmlab/mmselfsup.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #76 from cxxgtxy/master 

Support data prefetching on the linear classification task
pull/79/head
Xiaohang Zhan 2020-12-16 17:49:02 +08:00 committed by GitHub
parent ed5000482b 15706cbed2
commit a21f1a4417
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
7 changed files with 54 additions and 28 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
configs/benchmarks/linear_classification/imagenet/r50_last.py
View File

@ -28,15 +28,16 @@ img_norm_cfg = dict(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
train_pipeline = [
dict(type='RandomResizedCrop', size=224),
dict(type='RandomHorizontalFlip'),
dict(type='ToTensor'),
dict(type='Normalize', **img_norm_cfg),
]
test_pipeline = [
dict(type='Resize', size=256),
dict(type='CenterCrop', size=224),
dict(type='ToTensor'),
dict(type='Normalize', **img_norm_cfg),
]
# prefetch
prefetch = False
if not prefetch:
train_pipeline.extend([dict(type='ToTensor'), dict(type='Normalize', **img_norm_cfg)])
test_pipeline.extend([dict(type='ToTensor'), dict(type='Normalize', **img_norm_cfg)])
data = dict(
imgs_per_gpu=32, # total 32*8=256, 8GPU linear cls
workers_per_gpu=5,
@ -45,12 +46,14 @@ data = dict(
data_source=dict(
list_file=data_train_list, root=data_train_root,
**data_source_cfg),
pipeline=train_pipeline),
pipeline=train_pipeline,
prefetch=prefetch),
val=dict(
type=dataset_type,
data_source=dict(
list_file=data_test_list, root=data_test_root, **data_source_cfg),
pipeline=test_pipeline))
pipeline=test_pipeline,
prefetch=prefetch))
# additional hooks
custom_hooks = [
dict(
@ -60,6 +63,8 @@ custom_hooks = [
interval=1,
imgs_per_gpu=128,
workers_per_gpu=4,
prefetch=prefetch,
img_norm_cfg=img_norm_cfg,
eval_param=dict(topk=(1, 5)))
]
# optimizer

17
configs/benchmarks/linear_classification/imagenet/r50_last_sobel.py
View File

@ -28,15 +28,16 @@ img_norm_cfg = dict(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
train_pipeline = [
dict(type='RandomResizedCrop', size=224),
dict(type='RandomHorizontalFlip'),
dict(type='ToTensor'),
dict(type='Normalize', **img_norm_cfg),
]
test_pipeline = [
dict(type='Resize', size=256),
dict(type='CenterCrop', size=224),
dict(type='ToTensor'),
dict(type='Normalize', **img_norm_cfg),
]
# prefetch
prefetch = False
if not prefetch:
train_pipeline.extend([dict(type='ToTensor'), dict(type='Normalize', **img_norm_cfg)])
test_pipeline.extend([dict(type='ToTensor'), dict(type='Normalize', **img_norm_cfg)])
data = dict(
imgs_per_gpu=32, # total 32*8=256, 8GPU linear cls
workers_per_gpu=5,
@ -45,12 +46,14 @@ data = dict(
data_source=dict(
list_file=data_train_list, root=data_train_root,
**data_source_cfg),
pipeline=train_pipeline),
pipeline=train_pipeline,
prefetch=prefetch),
val=dict(
type=dataset_type,
data_source=dict(
list_file=data_test_list, root=data_test_root, **data_source_cfg),
pipeline=test_pipeline))
pipeline=test_pipeline,
prefetch=prefetch))
# additional hooks
custom_hooks = [
dict(
@ -60,6 +63,8 @@ custom_hooks = [
interval=1,
imgs_per_gpu=128,
workers_per_gpu=4,
prefetch=prefetch,
img_norm_cfg=img_norm_cfg,
eval_param=dict(topk=(1, 5)))
]
# optimizer

17
configs/benchmarks/linear_classification/places205/r50_multihead.py
View File

@ -35,15 +35,16 @@ train_pipeline = [
dict(type='CenterCrop', size=256),
dict(type='RandomCrop', size=224),
dict(type='RandomHorizontalFlip'),
dict(type='ToTensor'),
dict(type='Normalize', **img_norm_cfg),
]
test_pipeline = [
dict(type='Resize', size=256),
dict(type='CenterCrop', size=224),
dict(type='ToTensor'),
dict(type='Normalize', **img_norm_cfg),
]
# prefetch
prefetch = False
if not prefetch:
train_pipeline.extend([dict(type='ToTensor'), dict(type='Normalize', **img_norm_cfg)])
test_pipeline.extend([dict(type='ToTensor'), dict(type='Normalize', **img_norm_cfg)])
data = dict(
imgs_per_gpu=32, # total 32x8=256
workers_per_gpu=4,
@ -52,12 +53,14 @@ data = dict(
data_source=dict(
list_file=data_train_list, root=data_train_root,
**data_source_cfg),
pipeline=train_pipeline),
pipeline=train_pipeline,
prefetch=prefetch),
val=dict(
type=dataset_type,
data_source=dict(
list_file=data_test_list, root=data_test_root, **data_source_cfg),
pipeline=test_pipeline))
pipeline=test_pipeline,
prefetch=prefetch))
# additional hooks
custom_hooks = [
dict(
@ -67,6 +70,8 @@ custom_hooks = [
interval=10,
imgs_per_gpu=32,
workers_per_gpu=4,
prefetch=prefetch,
img_norm_cfg=img_norm_cfg,
eval_param=dict(topk=(1, )))
]
# optimizer

17
configs/benchmarks/linear_classification/places205/r50_multihead_sobel.py
View File

@ -35,15 +35,16 @@ train_pipeline = [
dict(type='CenterCrop', size=256),
dict(type='RandomCrop', size=224),
dict(type='RandomHorizontalFlip'),
dict(type='ToTensor'),
dict(type='Normalize', **img_norm_cfg),
]
test_pipeline = [
dict(type='Resize', size=256),
dict(type='CenterCrop', size=224),
dict(type='ToTensor'),
dict(type='Normalize', **img_norm_cfg),
]
# prefetch
prefetch = False
if not prefetch:
train_pipeline.extend([dict(type='ToTensor'), dict(type='Normalize', **img_norm_cfg)])
test_pipeline.extend([dict(type='ToTensor'), dict(type='Normalize', **img_norm_cfg)])
data = dict(
imgs_per_gpu=32, # total 32x8=256
workers_per_gpu=4,
@ -52,12 +53,14 @@ data = dict(
data_source=dict(
list_file=data_train_list, root=data_train_root,
**data_source_cfg),
pipeline=train_pipeline),
pipeline=train_pipeline,
prefetch=prefetch),
val=dict(
type=dataset_type,
data_source=dict(
list_file=data_test_list, root=data_test_root, **data_source_cfg),
pipeline=test_pipeline))
pipeline=test_pipeline,
prefetch=prefetch))
# additional hooks
custom_hooks = [
dict(
@ -67,6 +70,8 @@ custom_hooks = [
interval=10,
imgs_per_gpu=32,
workers_per_gpu=4,
prefetch=prefetch,
img_norm_cfg=img_norm_cfg,
eval_param=dict(topk=(1, )))
]
# optimizer

2
docs/GETTING_STARTED.md
View File

@ -92,7 +92,7 @@ prefetch = True
3 . Replacing Pillow with Pillow-SIMD (https://github.com/uploadcare/pillow-simd.git) to make use of SIMD command sets with modern CPU.
```shell
pip uninstall pillow
pip install Pillow-SIMD
pip install Pillow-SIMD or CC="cc -mavx2" pip install -U --force-reinstall pillow-simd if AVX2 is available.
```
We test it using MoCoV2 using a total batch size of 256 on Tesla V100. The training time per step is decreased to 0.17s from 0.23s.
## Benchmarks

7
openselfsup/datasets/classification.py
View File

@ -4,6 +4,7 @@ from openselfsup.utils import print_log
from .registry import DATASETS
from .base import BaseDataset
from .utils import to_numpy
@DATASETS.register_module
@ -11,12 +12,14 @@ class ClassificationDataset(BaseDataset):
"""Dataset for classification.
"""
def __init__(self, data_source, pipeline):
super(ClassificationDataset, self).__init__(data_source, pipeline)
def __init__(self, data_source, pipeline, prefetch=False):
super(ClassificationDataset, self).__init__(data_source, pipeline, prefetch)
def __getitem__(self, idx):
img, target = self.data_source.get_sample(idx)
img = self.pipeline(img)
if self.prefetch:
img = torch.from_numpy(to_numpy(img))
return dict(img=img, gt_label=target)
def evaluate(self, scores, keyword, logger=None, topk=(1, 5)):

5
openselfsup/hooks/validate_hook.py
View File

@ -42,7 +42,10 @@ class ValidateHook(Hook):
eval_kwargs['imgs_per_gpu'],
eval_kwargs['workers_per_gpu'],
dist=dist_mode,
shuffle=False)
shuffle=False,
prefetch=eval_kwargs.get('prefetch', False),
img_norm_cfg=eval_kwargs.get('img_norm_cfg', dict()),
)
self.dist_mode = dist_mode
self.initial = initial
self.interval = interval