RuntimeError: CUDA out of memory
Opened this issue · 4 comments
Benybrahim commented
Thanks for your error report and we appreciate it a lot.
Checklist
- I have searched related issues but cannot get the expected help.
- The bug has not been fixed in the latest version.
Describe the bug
I'm trying to finetune a food segmentation model found here, on new dataset.
When trying to train the model. I got this error. The batch_size is set to 1.
Thank you in advance for any insights you can give.
Reproduction
-
Command
python train --config SETR_MLA_768x768_80k_jap_finetune.py --gpus 1
-
Configuration file
norm_cfg = dict(type='BN', requires_grad=True)
model = dict(
type='EncoderDecoder',
backbone=dict(
type='VIT_MLA',
model_name='vit_base_patch16_224',
img_size=768,
patch_size=16,
in_chans=3,
embed_dim=768,
depth=12,
num_heads=12,
num_classes=19,
drop_rate=0.0,
norm_cfg=dict(type='BN', requires_grad=True),
pos_embed_interp=True,
align_corners=False,
mla_channels=256,
mla_index=(5, 7, 9, 11)),
decode_head=dict(
type='VIT_MLAHead',
in_channels=1024,
channels=512,
img_size=768,
mla_channels=256,
mlahead_channels=128,
num_classes=104,
norm_cfg=dict(type='BN', requires_grad=True),
align_corners=False,
loss_decode=dict(
type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
auxiliary_head=[
dict(
type='VIT_MLA_AUXIHead',
in_channels=256,
channels=512,
in_index=0,
img_size=768,
num_classes=102,
align_corners=False,
loss_decode=dict(
type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
dict(
type='VIT_MLA_AUXIHead',
in_channels=256,
channels=512,
in_index=1,
img_size=768,
num_classes=102,
align_corners=False,
loss_decode=dict(
type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
dict(
type='VIT_MLA_AUXIHead',
in_channels=256,
channels=512,
in_index=2,
img_size=768,
num_classes=102,
align_corners=False,
loss_decode=dict(
type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
dict(
type='VIT_MLA_AUXIHead',
in_channels=256,
channels=512,
in_index=3,
img_size=768,
num_classes=102,
align_corners=False,
loss_decode=dict(
type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4))
])
train_cfg = dict()
test_cfg = dict(mode='slide', crop_size=(768, 768), stride=(512, 512))
dataset_type = 'CustomDataset'
data_root = './data/UECFOODPIXCOMPLETE'
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
crop_size = (768, 768)
train_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations'),
dict(type='Resize', img_scale=(2049, 1025), ratio_range=(0.5, 2.0)),
dict(type='RandomCrop', crop_size=(768, 768), cat_max_ratio=0.75),
dict(type='RandomFlip', prob=0.5),
dict(type='PhotoMetricDistortion'),
dict(
type='Normalize',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(type='Pad', size=(768, 768), pad_val=0, seg_pad_val=255),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_semantic_seg'])
]
test_pipeline = [
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(2049, 1025),
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(
type='Normalize',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img'])
])
]
data = dict(
samples_per_gpu=1,
workers_per_gpu=0,
train=dict(
type='CustomDataset',
data_root='./data/UECFOODPIXCOMPLETE/',
img_dir='img/train',
ann_dir='ann/train',
pipeline=[
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations'),
dict(
type='Resize', img_scale=(2049, 1025), ratio_range=(0.5, 2.0)),
dict(type='RandomCrop', crop_size=(768, 768), cat_max_ratio=0.75),
dict(type='RandomFlip', prob=0.5),
dict(type='PhotoMetricDistortion'),
dict(
type='Normalize',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(type='Pad', size=(768, 768), pad_val=0, seg_pad_val=255),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_semantic_seg'])
]),
val=dict(
type='CustomDataset',
data_root='./data/UECFOODPIXCOMPLETE',
img_dir='img/test',
ann_dir='ann/test',
pipeline=[
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(2049, 1025),
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(
type='Normalize',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img'])
])
]),
test=dict(
type='CustomDataset',
data_root='./data/UECFOODPIXCOMPLETE',
img_dir='img/test',
ann_dir='ann/test',
pipeline=[
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(2049, 1025),
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(
type='Normalize',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img'])
])
]))
log_config = dict(
interval=50, hooks=[dict(type='TextLoggerHook', by_epoch=False)])
dist_params = dict(backend='nccl')
log_level = 'INFO'
load_from = "./models/iter_80000.pth"
resume_from = None
workflow = [('train', 1)]
cudnn_benchmark = True
optimizer = dict(
type='SGD',
lr=0.002,
momentum=0.9,
weight_decay=0.0,
paramwise_cfg=dict(custom_keys=dict(head=dict(lr_mult=10.0))))
optimizer_config = dict()
lr_config = dict(policy='poly', power=0.9, min_lr=0.0001, by_epoch=False)
runner = dict(type='IterBasedRunner', max_iters=80000)
checkpoint_config = dict(by_epoch=False, interval=4000)
evaluation = dict(interval=4000, metric='mIoU')
find_unused_parameters = True
work_dir = 'train_results/'
gpu_ids = range(0, 1)
- Dataset
I used this japanese dataset for food segmentation: https://mm.cs.uec.ac.jp/uecfoodpix/
I'm trying to finetune it on the japanese data.
Environment
sys.platform: linux
Python: 3.7.7 (default, May 7 2020, 21:25:33) [GCC 7.3.0]
CUDA available: True
GPU 0,1,2,3: NVIDIA GeForce GTX 1080 Ti
CUDA_HOME: /usr/local/cuda
NVCC: Cuda compilation tools, release 10.1, V10.1.243
GCC: gcc (Ubuntu 7.4.0-1ubuntu1~18.04.1) 7.4.0
PyTorch: 1.6.0
PyTorch compiling details: PyTorch built with:
- GCC 7.3
- C++ Version: 201402
- Intel(R) Math Kernel Library Version 2020.0.1 Product Build 20200208 for Intel(R) 64 architecture applications
- Intel(R) MKL-DNN v1.5.0 (Git Hash e2ac1fac44c5078ca927cb9b90e1b3066a0b2ed0)
- OpenMP 201511 (a.k.a. OpenMP 4.5)
- NNPACK is enabled
- CPU capability usage: AVX2
- CUDA Runtime 10.1
- NVCC architecture flags: -gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_61,code=sm_61;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_37,code=compute_37
- CuDNN 7.6.3
- Magma 2.5.2
- Build settings: BLAS=MKL, BUILD_TYPE=Release, CXX_FLAGS= -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DNDEBUG -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DUSE_VULKAN_WRAPPER -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-unused-local-typedefs -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Wno-stringop-overflow, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, USE_CUDA=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON, USE_STATIC_DISPATCH=OFF,
TorchVision: 0.7.0
OpenCV: 4.5.3
MMCV: 1.3.0
MMCV Compiler: GCC 7.3
MMCV CUDA Compiler: 10.1
MMSegmentation: 0.11.0+6a60f26
Error traceback
RuntimeError: CUDA out of memory. Tried to allocate 230.00 MiB (GPU 0; 10.92 GiB total capacity; 9.71 GiB already allocated; 190.69 MiB free; 10.14 GiB reserved in total by PyTorch)
XiongweiWu commented
Hi, the SeTR/ViT-MLA version is quite memory consuming, and it takes nearly 20G GPU memory even with 1 batch size in my side. I suggest maybe you can (i) train the model on GPUs with more RAM (Tesla-V100 32G, etc); (ii) turn off the val set evaluation during training; or (iii) replace ViT-base backbone with lightweight backbone (ViT-small or ViT-tiny etc). You can find examples to add more backbones in vision_transformer
Benybrahim commented
Thank you @XiongweiWu .
For suggestions:
(i) I understand that the model need a stronger GPU, so no remark about this.
(ii) How can I turn off the val set? Removing the val from data dictionary doesn't work.
(iii) vit_small_patch16_224 can be used, but it return this error, probably related to difference in size. Any tips on how to solve it?
size mismatch for blocks.0.mlp.fc1.weight: copying a param with shape torch.Size([2304, 768]) from checkpoint, the shape in current model is torch.Size([3072, 768]).
size mismatch for blocks.0.mlp.fc1.bias: copying a param with shape torch.Size([2304]) from checkpoint, the shape in current model is torch.Size([3072]).
....
Another question, does the VIT_MLA have a with_cp parameters like the other backbones? I could not find this parameter.
XiongweiWu commented
You can change the iteration number for evaluation beyond the total training iterations (set as 80k, etc.) in config file. You can evaluate the checkpoints in other GPUs separately before the training process is finished.
For you second question I may take some time to check the case since I have not tried it before. I may update here when I finish.
Benybrahim commented
Thanks @XiongweiWu
I tried but I couldn't find how to turn off val set.
That would be great if you can pass me the variable to change in the configuration file.
Is it here?
runner = dict(type='IterBasedRunner', max_iters=80000)