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

benchmark scripts

pull/4/head
xiaohangzhan 2020-06-17 01:31:59 +08:00
parent 766a816d3e
commit a1c322556a
30 changed files with 522 additions and 193 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

35
benchmarks/dist_test_cls.sh
View File

@ -1,35 +0,0 @@
#!/usr/bin/env bash
set -e
set -x
CFG=$1
EPOCH=$2
DATASET=$3 # imagenet or places205
GPUS=${GPUS:-1}
PORT=${PORT:-29500}
PY_ARGS=${@:4}
WORK_DIR=$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/
CHECKPOINT=$WORK_DIR/epoch_${EPOCH}.pth
WORK_DIR_EVAL=$WORK_DIR/${DATASET}_at_epoch_${EPOCH}/
# extract backbone
if [ ! -f "${CHECKPOINT::(-4)}_extracted.pth" ]; then
python tools/extract_backbone_weights.py $CHECKPOINT \
--save-path ${CHECKPOINT::(-4)}_extracted.pth
fi
# train
python -m torch.distributed.launch --nproc_per_node=$GPUS --master_port=$PORT \
tools/train.py \
configs/linear_classification/${DATASET}/r50_multihead.py \
--pretrained ${CHECKPOINT::(-4)}_extracted.pth \
--work_dir ${WORK_DIR_EVAL} --seed 0 --launcher="pytorch" ${PY_ARGS}
# test
python -m torch.distributed.launch --nproc_per_node=$GPUS --master_port=$PORT \
tools/test.py \
configs/linear_classification/${DATASET}/r50_multihead.py \
${WORK_DIR_EVAL}/latest.pth \
--work_dir ${WORK_DIR_EVAL} --launcher="pytorch"

26
benchmarks/dist_test_linear.sh 100755
View File

@ -0,0 +1,26 @@
#!/usr/bin/env bash
set -e
set -x
CFG=$1 # use cfgs under "configs/linear_classification/"
PRETRAIN=$2
PY_ARGS=${@:3}
GPUS=1 # in the standard setting, GPUS=1
PORT=${PORT:-29500}
WORK_DIR="$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/$(echo $PRETRAIN | rev | cut -d/ -f 1 | rev)"
# train
python -m torch.distributed.launch --nproc_per_node=$GPUS --master_port=$PORT \
tools/train.py \
$CFG \
--pretrained $PRETRAIN \
--work_dir $WORK_DIR --seed 0 --launcher="pytorch" ${PY_ARGS}
# test
python -m torch.distributed.launch --nproc_per_node=$GPUS --master_port=$PORT \
tools/test.py \
$CFG \
$WORK_DIR/latest.pth \
--work_dir $WORK_DIR --launcher="pytorch"

36
benchmarks/dist_test_semi.sh
View File

@ -3,37 +3,23 @@
set -e
set -x
CFG=$1
EPOCH=$2
PERCENT=$3
PY_ARGS=${@:4}
GPUS=${GPUS:-8}
CFG=$1 # use cfgs under "configs/classification/imagenet_*percent/"
PRETRAIN=$2
PY_ARGS=${@:3}
GPUS=8 # in the standard setting, GPUS=8
WORK_DIR=$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/
CHECKPOINT=$WORK_DIR/epoch_${EPOCH}.pth
WORK_DIR_EVAL=$WORK_DIR/imagenet_semi_${PERCENT}percent_at_epoch_${EPOCH}/
if [ ! "$PERCENT" == "1" ] && [ ! "$PERCENT" == 10 ]; then
echo "ERROR: PERCENT must in {1, 10}"
exit
fi
# extract backbone
if [ ! -f "${CHECKPOINT::(-4)}_extracted.pth" ]; then
python tools/extract_backbone_weights.py $CHECKPOINT \
--save-path ${CHECKPOINT::(-4)}_extracted.pth
fi
WORK_DIR="$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/$(echo $PRETRAIN | rev | cut -d/ -f 1 | rev)"
# train
python -m torch.distributed.launch --nproc_per_node=$GPUS \
tools/train.py \
configs/classification/imagenet_${PERCENT}percent/r50.py \
--pretrained ${CHECKPOINT::(-4)}_extracted.pth \
--work_dir ${WORK_DIR_EVAL} --seed 0 --launcher="pytorch" ${PY_ARGS}
$CFG \
--pretrained $PRETRAIN \
--work_dir $WORK_DIR --seed 0 --launcher="pytorch" ${PY_ARGS}
# test
python -m torch.distributed.launch --nproc_per_node=$GPUS \
tools/test.py \
configs/classification/imagenet_${PERCENT}percent/r50.py \
${WORK_DIR_EVAL}/latest.pth \
--work_dir ${WORK_DIR_EVAL} --launcher="pytorch"
$CFG \
$WORK_DIR/latest.pth \
--work_dir $WORK_DIR --launcher="pytorch"

15
benchmarks/dist_test_svm.sh
View File

@ -1,15 +0,0 @@
#!/bin/bash
set -e
set -x
CFG=$1
EPOCH=$2
FEAT_LIST=$3
GPUS=$4
WORK_DIR=$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/
bash tools/dist_extract.sh $CFG $WORK_DIR/epoch_${EPOCH}.pth $GPUS
bash benchmarks/eval_svm.sh $WORK_DIR $FEAT_LIST
bash benchmarks/eval_svm_lowshot.sh $WORK_DIR $FEAT_LIST

20
benchmarks/dist_test_svm_epoch.sh 100644
View File

@ -0,0 +1,20 @@
#!/bin/bash
set -e
set -x
CFG=$1
EPOCH=$2
FEAT_LIST=$3 # e.g.: "feat5", "feat4 feat5". If leave empty, the default is "feat5"
GPUS=${4:-8}
WORK_DIR=$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/
if [ ! -f $WORK_DIR/epoch_${EPOCH}.pth ]; then
echo "ERROR: File not exist: $WORK_DIR/epoch_${EPOCH}.pth"
exit
fi
bash tools/dist_extract.sh $CFG $GPUS --checkpoint $WORK_DIR/epoch_${EPOCH}.pth
bash benchmarks/svm_tools/eval_svm_full.sh $WORK_DIR $FEAT_LIST
bash benchmarks/svm_tools/eval_svm_lowshot.sh $WORK_DIR $FEAT_LIST

20
benchmarks/dist_test_svm_pretrain.sh 100644
View File

@ -0,0 +1,20 @@
#!/bin/bash
set -e
set -x
CFG=$1
PRETRAIN=$2 # pretrained model or "random" (random init)
FEAT_LIST=$3 # e.g.: "feat5", "feat4 feat5". If leave empty, the default is "feat5"
GPUS=${4:-8}
WORK_DIR=$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/
if [ ! -f $PRETRAIN ] and [ "$PRETRAIN" != "random" ]; then
echo "ERROR: PRETRAIN should be a file or a string \"random\", got: $PRETRAIN"
exit
fi
bash tools/dist_extract.sh $CFG $GPUS --pretrained $PRETRAIN
bash benchmarks/svm_tools/eval_svm_full.sh $WORK_DIR $FEAT_LIST
bash benchmarks/svm_tools/eval_svm_lowshot.sh $WORK_DIR $FEAT_LIST

56
benchmarks/srun_test_cls.sh
View File

@ -1,56 +0,0 @@
#!/usr/bin/env bash
set -e
set -x
PARTITION=$1
CFG=$2
EPOCH=$3
DATASET=$4 # imagenet or places205
PY_ARGS=${@:5}
JOB_NAME="openselfsup"
GPUS=${GPUS:-1}
GPUS_PER_NODE=${GPUS_PER_NODE:-1}
CPUS_PER_TASK=${CPUS_PER_TASK:-5}
SRUN_ARGS=${SRUN_ARGS:-""}
WORK_DIR=$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/
CHECKPOINT=$WORK_DIR/epoch_${EPOCH}.pth
WORK_DIR_EVAL=$WORK_DIR/${DATASET}_at_epoch_${EPOCH}/
# extract backbone
if [ ! -f "${CHECKPOINT::(-4)}_extracted.pth" ]; then
srun -p ${PARTITION} \
python tools/extract_backbone_weights.py $CHECKPOINT \
--save-path ${CHECKPOINT::(-4)}_extracted.pth
fi
# train
GLOG_vmodule=MemcachedClient=-1 \
srun -p ${PARTITION} \
--job-name=${JOB_NAME} \
--gres=gpu:${GPUS_PER_NODE} \
--ntasks=${GPUS} \
--ntasks-per-node=${GPUS_PER_NODE} \
--cpus-per-task=${CPUS_PER_TASK} \
--kill-on-bad-exit=1 \
${SRUN_ARGS} \
python -u tools/train.py \
configs/linear_classification/${DATASET}/r50_multihead.py \
--pretrained ${CHECKPOINT::(-4)}_extracted.pth \
--work_dir ${WORK_DIR_EVAL} --seed 0 --launcher="slurm" ${PY_ARGS}
# test
GLOG_vmodule=MemcachedClient=-1 \
srun -p ${PARTITION} \
--job-name=${JOB_NAME} \
--gres=gpu:${GPUS_PER_NODE} \
--ntasks=${GPUS} \
--ntasks-per-node=${GPUS_PER_NODE} \
--cpus-per-task=${CPUS_PER_TASK} \
--kill-on-bad-exit=1 \
${SRUN_ARGS} \
python -u tools/test.py \
configs/linear_classification/${DATASET}/r50_multihead.py \
${WORK_DIR_EVAL}/latest.pth \
--work_dir ${WORK_DIR_EVAL} --launcher="slurm"

46
benchmarks/srun_test_linear.sh 100644
View File

@ -0,0 +1,46 @@
#!/usr/bin/env bash
set -e
set -x
PARTITION=$1
CFG=$2
PRETRAIN=$3
PY_ARGS=${@:4}
JOB_NAME="openselfsup"
GPUS=1 # in the standard setting, GPUS=1
GPUS_PER_NODE=${GPUS_PER_NODE:-1}
CPUS_PER_TASK=${CPUS_PER_TASK:-5}
SRUN_ARGS=${SRUN_ARGS:-""}
WORK_DIR="$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/$(echo $PRETRAIN | rev | cut -d/ -f 1 | rev)"
# train
GLOG_vmodule=MemcachedClient=-1 \
srun -p ${PARTITION} \
--job-name=${JOB_NAME} \
--gres=gpu:${GPUS_PER_NODE} \
--ntasks=${GPUS} \
--ntasks-per-node=${GPUS_PER_NODE} \
--cpus-per-task=${CPUS_PER_TASK} \
--kill-on-bad-exit=1 \
${SRUN_ARGS} \
python -u tools/train.py \
$CFG \
--pretrained $PRETRAIN \
--work_dir $WORK_DIR --seed 0 --launcher="slurm" ${PY_ARGS}
# test
GLOG_vmodule=MemcachedClient=-1 \
srun -p ${PARTITION} \
--job-name=${JOB_NAME} \
--gres=gpu:${GPUS_PER_NODE} \
--ntasks=${GPUS} \
--ntasks-per-node=${GPUS_PER_NODE} \
--cpus-per-task=${CPUS_PER_TASK} \
--kill-on-bad-exit=1 \
${SRUN_ARGS} \
python -u tools/test.py \
$CFG \
$WORK_DIR/latest.pth \
--work_dir $WORK_DIR --launcher="slurm"

34
benchmarks/srun_test_semi.sh
View File

@ -5,29 +5,15 @@ set -x
PARTITION=$1
CFG=$2
EPOCH=$3
PERCENT=$4
PY_ARGS=${@:5}
PRETRAIN=$3
PY_ARGS=${@:4}
JOB_NAME="openselfsup"
GPUS=${GPUS:-8}
GPUS=8 # in the standard setting, GPUS=8
GPUS_PER_NODE=${GPUS_PER_NODE:-8}
CPUS_PER_TASK=${CPUS_PER_TASK:-5}
SRUN_ARGS=${SRUN_ARGS:-""}
WORK_DIR=$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/
CHECKPOINT=$WORK_DIR/epoch_${EPOCH}.pth
WORK_DIR_EVAL=$WORK_DIR/imagenet_semi_${PERCENT}percent_at_epoch_${EPOCH}/
if [ ! "$PERCENT" == "1" ] && [ ! "$PERCENT" == 10 ]; then
echo "ERROR: PERCENT must in {1, 10}"
exit
fi
# extract backbone
if [ ! -f "${CHECKPOINT::(-4)}_extracted.pth" ]; then
srun -p ${PARTITION} \
python tools/extract_backbone_weights.py $CHECKPOINT \
--save-path ${CHECKPOINT::(-4)}_extracted.pth
fi
WORK_DIR="$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/$(echo $PRETRAIN | rev | cut -d/ -f 1 | rev)"
# train
GLOG_vmodule=MemcachedClient=-1 \
@ -40,9 +26,9 @@ srun -p ${PARTITION} \
--kill-on-bad-exit=1 \
${SRUN_ARGS} \
python -u tools/train.py \
configs/classification/imagenet_${PERCENT}percent/r50.py \
--pretrained ${CHECKPOINT::(-4)}_extracted.pth \
--work_dir ${WORK_DIR_EVAL} --seed 0 --launcher="slurm" ${PY_ARGS}
$CFG \
--pretrained $PRETRAIN \
--work_dir $WORK_DIR --seed 0 --launcher="slurm" ${PY_ARGS}
# test
GLOG_vmodule=MemcachedClient=-1 \
@ -55,6 +41,6 @@ srun -p ${PARTITION} \
--kill-on-bad-exit=1 \
${SRUN_ARGS} \
python -u tools/test.py \
configs/classification/imagenet_${PERCENT}percent/r50.py \
${WORK_DIR_EVAL}/latest.pth \
--work_dir ${WORK_DIR_EVAL} --launcher="slurm"
$CFG \
$WORK_DIR/latest.pth \
--work_dir $WORK_DIR --launcher="slurm"

11
benchmarks/srun_test_svm.sh
View File

@ -4,12 +4,13 @@ set -x
PARTITION=$1
CFG=$2
EPOCH=$3
FEAT=$4
PRETRAIN=$3
FEAT_LIST=$4
GPUS=${5:-8}
WORK_DIR=$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/
bash tools/srun_extract.sh $PARTITION $CFG $WORK_DIR/epoch_${EPOCH}.pth
bash tools/srun_extract.sh $PARTITION $CFG $GPUS --pretrained $PRETRAIN
srun -p $PARTITION bash benchmarks/eval_svm.sh $WORK_DIR $FEAT
srun -p $PARTITION bash benchmarks/svm_tools/eval_svm_full.sh $WORK_DIR $FEAT_LIST
srun -p $PARTITION bash benchmarks/eval_svm.sh $WORK_DIR $FEAT
srun -p $PARTITION bash benchmarks/svm_tools/eval_svm_lowshot.sh $WORK_DIR $FEAT_LIST

21
benchmarks/srun_test_svm_epoch.sh 100644
View File

@ -0,0 +1,21 @@
#!/usr/bin/env bash
set -e
set -x
PARTITION=$1
CFG=$2
EPOCH=$3
FEAT_LIST=$4 # e.g.: "feat5", "feat4 feat5". If leave empty, the default is "feat5"
GPUS=${5:-8}
WORK_DIR=$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/
if [ ! -f $WORK_DIR/epoch_${EPOCH}.pth ]; then
echo "ERROR: File not exist: $WORK_DIR/epoch_${EPOCH}.pth"
exit
fi
bash tools/srun_extract.sh $PARTITION $CFG $GPUS --checkpoint $WORK_DIR/epoch_${EPOCH}.pth
srun -p $PARTITION bash benchmarks/svm_tools/eval_svm_full.sh $WORK_DIR $FEAT_LIST
srun -p $PARTITION bash benchmarks/svm_tools/eval_svm_lowshot.sh $WORK_DIR $FEAT_LIST

21
benchmarks/srun_test_svm_pretrain.sh 100644
View File

@ -0,0 +1,21 @@
#!/usr/bin/env bash
set -e
set -x
PARTITION=$1
CFG=$2
PRETRAIN=$3 # pretrained model or "random" (random init)
FEAT_LIST=$4 # e.g.: "feat5", "feat4 feat5". If leave empty, the default is "feat5"
GPUS=${5:-8}
WORK_DIR=$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/
if [ ! -f $PRETRAIN ] and [ "$PRETRAIN" != "random" ]; then
echo "ERROR: PRETRAIN should be a file or a string \"random\", got: $PRETRAIN"
exit
fi
bash tools/srun_extract.sh $PARTITION $CFG $GPUS --pretrained $PRETRAIN
srun -p $PARTITION bash benchmarks/svm_tools/eval_svm_full.sh $WORK_DIR $FEAT_LIST
srun -p $PARTITION bash benchmarks/svm_tools/eval_svm_lowshot.sh $WORK_DIR $FEAT_LIST

0
benchmarks/eval_svm.sh → benchmarks/svm_tools/eval_svm_full.sh
View File

0
benchmarks/eval_svm_lowshot.sh → benchmarks/svm_tools/eval_svm_lowshot.sh
View File

60
configs/classification/imagenet_10percent/r50_sobel.py 100644
View File

@ -0,0 +1,60 @@
_base_ = '../../base.py'
# model settings
model = dict(
type='Classification',
pretrained=None,
with_sobel=True,
backbone=dict(
type='ResNet',
depth=50,
in_channels=2,
out_indices=[4], # 0: conv-1, x: stage-x
norm_cfg=dict(type='SyncBN')),
head=dict(
type='ClsHead', with_avg_pool=True, in_channels=2048,
num_classes=1000))
# dataset settings
data_source_cfg = dict(
type='ImageNet',
memcached=True,
mclient_path='/mnt/lustre/share/memcached_client')
data_train_list = 'data/imagenet/meta/train_labeled_10percent.txt'
data_train_root = 'data/imagenet/train'
data_test_list = 'data/imagenet/meta/val_labeled.txt'
data_test_root = 'data/imagenet/val'
dataset_type = 'ClassificationDataset'
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),
]
data = dict(
imgs_per_gpu=32, # total 256
workers_per_gpu=2,
train=dict(
type=dataset_type,
data_source=dict(
list_file=data_train_list, root=data_train_root,
**data_source_cfg),
pipeline=train_pipeline),
val=dict(
type=dataset_type,
data_source=dict(
list_file=data_test_list, root=data_test_root, **data_source_cfg),
pipeline=test_pipeline))
# optimizer
optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005,
paramwise_options={'\Ahead.': dict(lr_mult=10)})
# learning policy
lr_config = dict(policy='step', step=[18, 24], gamma=0.2)
checkpoint_config = dict(interval=2)
# runtime settings
total_epochs = 30

60
configs/classification/imagenet_1percent/r50_sobel.py 100644
View File

@ -0,0 +1,60 @@
_base_ = '../../base.py'
# model settings
model = dict(
type='Classification',
pretrained=None,
with_sobel=True,
backbone=dict(
type='ResNet',
depth=50,
in_channels=2,
out_indices=[4], # 0: conv-1, x: stage-x
norm_cfg=dict(type='SyncBN')),
head=dict(
type='ClsHead', with_avg_pool=True, in_channels=2048,
num_classes=1000))
# dataset settings
data_source_cfg = dict(
type='ImageNet',
memcached=True,
mclient_path='/mnt/lustre/share/memcached_client')
data_train_list = 'data/imagenet/meta/train_labeled_1percent.txt'
data_train_root = 'data/imagenet/train'
data_test_list = 'data/imagenet/meta/val_labeled.txt'
data_test_root = 'data/imagenet/val'
dataset_type = 'ClassificationDataset'
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),
]
data = dict(
imgs_per_gpu=32, # total 256
workers_per_gpu=2,
train=dict(
type=dataset_type,
data_source=dict(
list_file=data_train_list, root=data_train_root,
**data_source_cfg),
pipeline=train_pipeline),
val=dict(
type=dataset_type,
data_source=dict(
list_file=data_test_list, root=data_test_root, **data_source_cfg),
pipeline=test_pipeline))
# optimizer
optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005,
paramwise_options={'\Ahead.': dict(lr_mult=100)})
# learning policy
lr_config = dict(policy='step', step=[12, 16], gamma=0.2)
checkpoint_config = dict(interval=2)
# runtime settings
total_epochs = 20

89
configs/linear_classification/imagenet/r50_multihead_sobel.py 100644
View File

@ -0,0 +1,89 @@
_base_ = '../../base.py'
# model settings
model = dict(
type='Classification',
pretrained=None,
frozen_backbone=True,
with_sobel=True,
backbone=dict(
type='ResNet',
depth=50,
in_channels=2,
out_indices=[0, 1, 2, 3, 4], # 0: conv-1, x: stage-x
norm_cfg=dict(type='BN')),
head=dict(
type='MultiClsHead',
pool_type='specified',
in_indices=[0, 1, 2, 3, 4],
with_last_layer_unpool=True,
backbone='resnet50',
norm_cfg=dict(type='BN', momentum=0.1, affine=False),
num_classes=1000))
# dataset settings
data_source_cfg = dict(
type='ImageNet',
memcached=False,
mclient_path='/mnt/lustre/share/memcached_client')
data_train_list = 'data/imagenet/meta/train_labeled.txt'
data_train_root = 'data/imagenet/train'
data_test_list = 'data/imagenet/meta/val_labeled.txt'
data_test_root = 'data/imagenet/val'
dataset_type = 'ClassificationDataset'
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='ColorJitter',
brightness=0.4,
contrast=0.4,
saturation=0.4,
hue=0.),
dict(type='ToTensor'),
dict(type='Lighting'),
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),
]
data = dict(
imgs_per_gpu=256, # total 256
workers_per_gpu=8,
train=dict(
type=dataset_type,
data_source=dict(
list_file=data_train_list, root=data_train_root,
**data_source_cfg),
pipeline=train_pipeline),
val=dict(
type=dataset_type,
data_source=dict(
list_file=data_test_list, root=data_test_root, **data_source_cfg),
pipeline=test_pipeline))
# additional hooks
custom_hooks = [
dict(
type='ValidateHook',
dataset=data['val'],
initial=True,
interval=10,
imgs_per_gpu=128,
workers_per_gpu=4,
eval_param=dict(topk=(1, )))
]
# optimizer
optimizer = dict(
type='SGD',
lr=0.01,
momentum=0.9,
weight_decay=0.0001,
paramwise_options=dict(norm_decay_mult=0.),
nesterov=True)
# learning policy
lr_config = dict(policy='step', step=[30, 60, 90])
checkpoint_config = dict(interval=10)
# runtime settings
total_epochs = 90

89
configs/linear_classification/places205/r50_multihead_sobel.py 100644
View File

@ -0,0 +1,89 @@
_base_ = '../../base.py'
# model settings
model = dict(
type='Classification',
pretrained=None,
frozen_backbone=True,
with_sobel=True,
backbone=dict(
type='ResNet',
depth=50,
in_channels=2,
out_indices=[0, 1, 2, 3, 4], # 0: conv-1, x: stage-x
norm_cfg=dict(type='BN')),
head=dict(
type='MultiClsHead',
pool_type='specified',
in_indices=[0, 1, 2, 3, 4],
with_last_layer_unpool=True,
backbone='resnet50',
norm_cfg=dict(type='BN', momentum=0.1, affine=False),
num_classes=205))
# dataset settings
data_source_cfg = dict(
type='Places205',
memcached=False,
mclient_path='/mnt/lustre/share/memcached_client')
data_train_list = 'data/places205/meta/train_labeled.txt'
data_train_root = 'data/places205/train'
data_test_list = 'data/places205/meta/val_labeled.txt'
data_test_root = 'data/places205/val'
dataset_type = 'ClassificationDataset'
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='ColorJitter',
brightness=0.4,
contrast=0.4,
saturation=0.4,
hue=0.),
dict(type='ToTensor'),
dict(type='Lighting'),
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),
]
data = dict(
imgs_per_gpu=256, # total 256
workers_per_gpu=8,
train=dict(
type=dataset_type,
data_source=dict(
list_file=data_train_list, root=data_train_root,
**data_source_cfg),
pipeline=train_pipeline),
val=dict(
type=dataset_type,
data_source=dict(
list_file=data_test_list, root=data_test_root, **data_source_cfg),
pipeline=test_pipeline))
# additional hooks
custom_hooks = [
dict(
type='ValidateHook',
dataset=data['val'],
initial=True,
interval=10,
imgs_per_gpu=128,
workers_per_gpu=4,
eval_param=dict(topk=(1, )))
]
# optimizer
optimizer = dict(
type='SGD',
lr=0.01,
momentum=0.9,
weight_decay=0.0001,
paramwise_options=dict(norm_decay_mult=0.),
nesterov=True)
# learning policy
lr_config = dict(policy='step', step=[30, 60, 90])
checkpoint_config = dict(interval=10)
# runtime settings
total_epochs = 90

2
configs/selfsup/deepcluster/r50.py
View File

@ -85,4 +85,4 @@ optimizer = dict(
lr_config = dict(policy='step', step=[400])
checkpoint_config = dict(interval=10)
# runtime settings
total_epochs = 480
total_epochs = 200

4
docs/GETTING_STARTED.md
View File

@ -86,7 +86,7 @@ Arguments:
1. First, extract backbone weights:
```shell
python tools/extract_backbone_weights.py ${CHECKPOINT} --save-path ${WEIGHT_FILE}
python tools/extract_backbone_weights.py ${CHECKPOINT} ${WEIGHT_FILE}
```
Arguments:
- `CHECKPOINTS`: the checkpoint file of a selfsup method named as `epoch_*.pth`.
@ -119,7 +119,7 @@ python tools/count_parameters.py ${CONFIG_FILE}
1. Extract the backbone weights as mentioned before. You don't have to extract it again if you've already done it in the benchmark step.
```shell
python tools/extract_backbone_weights.py ${CHECKPOINT} --save-path ${WEIGHT_FILE}
python tools/extract_backbone_weights.py ${CHECKPOINT} ${WEIGHT_FILE}
```
2. Compute the hash of the weight file and append the hash id to the filename.

2
openselfsup/datasets/contrastive.py
View File

@ -13,7 +13,7 @@ class ContrastiveDataset(BaseDataset):
super(ContrastiveDataset, self).__init__(data_source, pipeline)
def __getitem__(self, idx):
img, _ = self.data_source.get_sample(idx)
img = self.data_source.get_sample(idx)
img1 = self.pipeline(img)
img2 = self.pipeline(img)
img_cat = torch.cat((img1.unsqueeze(0), img2.unsqueeze(0)), dim=0)

7
openselfsup/datasets/data_sources/image_list.py
View File

@ -39,5 +39,8 @@ class ImageList(object):
else:
img = Image.open(self.fns[idx])
img = img.convert('RGB')
target = self.labels[idx] if self.has_labels else None
return img, target
if self.has_labels:
target = self.labels[idx]
return img, target
else:
return img

2
openselfsup/datasets/deepcluster.py
View File

@ -13,7 +13,7 @@ class DeepClusterDataset(BaseDataset):
self.labels = [-1 for _ in range(self.data_source.get_length())]
def __getitem__(self, idx):
img, _ = self.data_source.get_sample(idx)
img = self.data_source.get_sample(idx)
label = self.labels[idx]
img = self.pipeline(img)
return dict(img=img, pseudo_label=label, idx=idx)

2
openselfsup/datasets/npid.py
View File

@ -11,7 +11,7 @@ class NPIDDataset(BaseDataset):
super(NPIDDataset, self).__init__(data_source, pipeline)
def __getitem__(self, idx):
img, _ = self.data_source.get_sample(idx)
img = self.data_source.get_sample(idx)
img = self.pipeline(img)
return dict(img=img, idx=idx)

2
openselfsup/datasets/rotation_pred.py
View File

@ -25,7 +25,7 @@ class RotationPredDataset(BaseDataset):
super(RotationPredDataset, self).__init__(data_source, pipeline)
def __getitem__(self, idx):
img, _ = self.data_source.get_sample(idx)
img = self.data_source.get_sample(idx)
img = self.pipeline(img)
img = torch.stack(rotate(img), dim=0)
rotation_labels = torch.LongTensor([0, 1, 2, 3])

17
tools/dist_extract.sh
View File

@ -1,16 +1,13 @@
#!/usr/bin/env bash
set -x
PYTHON=${PYTHON:-"python"}
CFG=$1
CHECKPOINT=$2
GPUS=${3:-8}
GPUS=$2
PY_ARGS=${@:3} # "--checkpoint $CHECKPOINT --pretrained $PRETRAINED"
PORT=${PORT:-29500}
WORK_DIR=$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/
if [ "$CHECKPOINT" == "" ]; then
$PYTHON -m torch.distributed.launch --nproc_per_node=$GPUS --master_port=$PORT \
tools/extract.py $CFG --layer-ind "0,1,2,3,4" --work_dir $WORK_DIR --launcher pytorch
else
$PYTHON -m torch.distributed.launch --nproc_per_node=$GPUS --master_port=$PORT \
tools/extract.py $CFG --layer-ind "0,1,2,3,4" --checkpoint $CHECKPOINT \
--work_dir $WORK_DIR --launcher pytorch
fi
$PYTHON -m torch.distributed.launch --nproc_per_node=$GPUS --master_port=$PORT \
tools/extract.py $CFG --layer-ind "0,1,2,3,4" --work_dir $WORK_DIR \
--launcher pytorch ${PY_ARGS}

2
tools/dist_train.sh
View File

@ -3,8 +3,8 @@ PYTHON=${PYTHON:-"python"}
CFG=$1
GPUS=$2
PORT=${PORT:-29500}
PY_ARGS=${@:3}
PORT=${PORT:-29500}
WORK_DIR=$(echo ${CFG%.*} | sed -e "s/configs/work_dirs/g")/

19
tools/extract.py
View File

@ -52,6 +52,9 @@ def parse_args():
description='OpenSelfSup extract features of a model')
parser.add_argument('config', help='test config file path')
parser.add_argument('--checkpoint', default=None, help='checkpoint file')
parser.add_argument(
'--pretrained', default='random',
help='pretrained model file, exclusive to --checkpoint')
parser.add_argument(
'--dataset-config',
default='benchmarks/extract_info/voc07.py',
@ -89,9 +92,9 @@ def main():
layer_ind = [int(idx) for idx in args.layer_ind.split(',')]
cfg.model.backbone.out_indices = layer_ind
if args.checkpoint is None:
assert cfg.model.pretrained is not None, \
"Must have pretrain if no checkpoint is given."
# checkpoint and pretrained are exclusive
assert cfg.model.pretrained == "random" or args.checkpoint is None, \
"Checkpoint and pretrained are exclusive."
# check memcached package exists
if importlib.util.find_spec('mc') is None:
@ -106,6 +109,8 @@ def main():
distributed = True
init_dist(args.launcher, **cfg.dist_params)
# create work_dir
mmcv.mkdir_or_exist(osp.abspath(cfg.work_dir))
# logger
timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
log_file = osp.join(cfg.work_dir, 'extract_{}.log'.format(timestamp))
@ -124,7 +129,15 @@ def main():
# build the model and load checkpoint
model = build_model(cfg.model)
if args.checkpoint is not None:
logger.info("Use checkpoint: {} to extract features".format(
args.checkpoint))
load_checkpoint(model, args.checkpoint, map_location='cpu')
elif args.pretrained != "random":
logger.info('Use pretrained model: {} to extract features'.format(
args.pretrained))
else:
logger.info('No checkpoint or pretrained is give, use random init.')
if not distributed:
model = MMDataParallel(model, device_ids=[0])
else:

7
tools/extract_backbone_weights.py
View File

@ -7,15 +7,14 @@ def parse_args():
description='This script extracts backbone weights from a checkpoint')
parser.add_argument('checkpoint', help='checkpoint file')
parser.add_argument(
'--save-path', type=str, default=None, help='destination file name')
'output', type=str, help='destination file name')
args = parser.parse_args()
return args
def main():
args = parse_args()
if args.save_path is None:
args.save_path = args.checkpoint[:-4] + "_extracted.pth"
assert args.output.endswith(".pth")
ck = torch.load(args.checkpoint, map_location=torch.device('cpu'))
output_dict = dict(state_dict=dict(), author="OpenSelfSup")
has_backbone = False
@ -25,7 +24,7 @@ def main():
has_backbone = True
if not has_backbone:
raise Exception("Cannot find a backbone module in the checkpoint.")
torch.save(output_dict, args.save_path)
torch.save(output_dict, args.output)
if __name__ == '__main__':

10
tools/srun_extract.sh
View File

@ -1,12 +1,10 @@
#!/usr/bin/env bash
set -x
PARTITION=$1
CFG=$2
CHECKPOINT=$3
GPUS=${4:-8}
PY_ARGS=${@:5}
GPUS=$3
PY_ARGS=${@:4} # "--checkpoint $CHECKPOINT --pretrained $PRETRAINED"
JOB_NAME="openselfsup"
GPUS_PER_NODE=${GPUS_PER_NODE:-8}
CPUS_PER_TASK=${CPUS_PER_TASK:-5}
@ -23,5 +21,5 @@ srun -p ${PARTITION} \
--kill-on-bad-exit=1 \
${SRUN_ARGS} \
python -u tools/extract.py $CFG \
--layer-ind "0,1,2,3,4" --checkpoint $CHECKPOINT \
--work_dir $WORK_DIR --launcher="slurm" ${PY_ARGS}
--layer-ind "0,1,2,3,4" --work_dir $WORK_DIR \
--launcher="slurm" ${PY_ARGS}