When I run this code https://github.com/erezposner/Pose2Seg
And I made all steps in this tutorial https://towardsdatascience.com/detection-free-human-instance-segmentation-using-pose2seg-and-pytorch-72f48dc4d23e
but I have this error in cuda:
RuntimeError: CUDA out of memory. Tried to allocate 128.00 MiB (GPU 0; 4.00 GiB total capacity; 2.57 GiB already allocated; 74.77 MiB free; 2.85 GiB reserved in total by PyTorch) (malloc at ..\c10\cuda\CUDACachingAllocator.cpp:289) (no backtrace available)
How can I solve this?
(base) C:\Users\ASUS\Pose2Seg>python train.py
06-23 07:30:01 ===========> loading model <===========
total params in model is 334, in pretrained model is 336, init 334
06-23 07:30:03 ===========> loading data <===========
loading annotations into memory...
Done (t=4.56s)
creating index...
index created!
06-23 07:30:08 ===========> set optimizer <===========
06-23 07:30:08 ===========> training <===========
C:\Users\ASUS\Anaconda3\Anaconda\lib\site-packages\torch\nn\functional.py:2796: UserWarning: nn.functional.upsample is deprecated. Use nn.functional.interpolate instead.
warnings.warn("nn.functional.upsample is deprecated. Use nn.functional.interpolate instead.")
C:\Users\ASUS\Anaconda3\Anaconda\lib\site-packages\torch\nn\functional.py:2973: UserWarning: Default upsampling behavior when mode=bilinear is changed to align_corners=False since 0.4.0. Please specify align_corners=True if the old behavior is desired. See the documentation of nn.Upsample for details.
"See the documentation of nn.Upsample for details.".format(mode))
C:\Users\ASUS\Anaconda3\Anaconda\lib\site-packages\torch\nn\functional.py:3289: UserWarning: Default grid_sample and affine_grid behavior has changed to align_corners=False since 1.3.0. Please specify align_corners=True if the old behavior is desired. See the documentation of grid_sample for details.
warnings.warn("Default grid_sample and affine_grid behavior has changed "
C:\Users\ASUS\Anaconda3\Anaconda\lib\site-packages\torch\nn\functional.py:3226: UserWarning: Default grid_sample and affine_grid behavior has changed to align_corners=False since 1.3.0. Please specify align_corners=True if the old behavior is desired. See the documentation of grid_sample for details.
warnings.warn("Default grid_sample and affine_grid behavior has changed "
06-23 07:30:13 Epoch: [0][0/56599] Lr: [6.68e-05] Time 4.228 (4.228) Data 0.028 (0.028) loss 0.85738 (0.85738)
06-23 07:30:22 Epoch: [0][10/56599] Lr: [6.813333333333334e-05] Time 0.847 (1.280) Data 0.012 (0.051) loss 0.44195 (0.71130)
06-23 07:30:33 Epoch: [0][20/56599] Lr: [6.946666666666667e-05] Time 0.882 (1.180) Data 0.045 (0.037) loss 0.41523 (0.60743)
Traceback (most recent call last):
File "train.py", line 157, in <module>
optimizer, epoch, iteration)
File "train.py", line 74, in train
loss.backward()
File "C:\Users\ASUS\Anaconda3\Anaconda\lib\site-packages\torch\tensor.py", line 198, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph)
File "C:\Users\ASUS\Anaconda3\Anaconda\lib\site-packages\torch\autograd\__init__.py", line 100, in backward
allow_unreachable=True) # allow_unreachable flag
RuntimeError: CUDA out of memory. Tried to allocate 128.00 MiB (GPU 0; 4.00 GiB total capacity; 2.57 GiB already allocated; 74.77 MiB free; 2.85 GiB reserved in total by PyTorch) (malloc at ..\c10\cuda\CUDACachingAllocator.cpp:289)
(no backtrace available)
cudatoolkit == 10.1.243
python3.6.5
The version of libs:
>>> import tensorflow
2020-06-23 09:45:01.840827: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudart64_101.dll
>>> tensorflow.__version__
'2.2.0'
>>> import keras
Using TensorFlow backend.
>>> keras.__version__
'2.3.1'
>>> import torch
>>> torch.__version__
'1.5.1'
>>> import torchvision
>>> torchvision.__version__
'0.6.1'
>>> import pycocotools
train.py code
import os
import sys
import time
import logging
import argparse
import numpy as np
from tqdm import tqdm
import torch
import torch.utils.data
from lib.averageMeter import AverageMeters
from lib.logger import colorlogger
from lib.timer import Timers
from lib.averageMeter import AverageMeters
from lib.torch_utils import adjust_learning_rate
import os
from modeling.build_model import Pose2Seg
from datasets.CocoDatasetInfo import CocoDatasetInfo, annToMask
from test import test
NAME = "release_base"
# Set `LOG_DIR` and `SNAPSHOT_DIR`
def setup_logdir():
timestamp = time.strftime("%Y-%m-%d_%H_%M_%S", time.localtime())
LOGDIR = os.path.join(os.getcwd(), 'logs', '%s_%s' % (NAME, timestamp))
SNAPSHOTDIR = os.path.join(
os.getcwd(), 'snapshot', '%s_%s' % (NAME, timestamp))
if not os.path.exists(LOGDIR):
os.makedirs(LOGDIR)
if not os.path.exists(SNAPSHOTDIR):
os.makedirs(SNAPSHOTDIR)
return LOGDIR, SNAPSHOTDIR
LOGDIR, SNAPSHOTDIR = setup_logdir()
# Set logging
logger = colorlogger(log_dir=LOGDIR, log_name='train_logs.txt')
# Set Global Timer
timers = Timers()
# Set Global AverageMeter
averMeters = AverageMeters()
def train(model, dataloader, optimizer, epoch, iteration):
# switch to train mode
model.train()
averMeters.clear()
end = time.time()
for i, inputs in enumerate(dataloader):
averMeters['data_time'].update(time.time() - end)
iteration += 1
lr = adjust_learning_rate(optimizer, iteration, BASE_LR=0.0002,
WARM_UP_FACTOR=1.0/3, WARM_UP_ITERS=1000,
STEPS=(0, 14150*15, 14150*20), GAMMA=0.1)
# forward
outputs = model(**inputs)
# loss
loss = outputs
# backward
averMeters['loss'].update(loss.data.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
# measure elapsed time
averMeters['batch_time'].update(time.time() - end)
end = time.time()
if i % 10 == 0:
logger.info('Epoch: [{0}][{1}/{2}]\t'
'Lr: [{3}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'loss {loss.val:.5f} ({loss.avg:.5f})\t'
.format(
epoch, i, len(dataloader), lr,
batch_time=averMeters['batch_time'], data_time=averMeters['data_time'],
loss=averMeters['loss'])
)
if i % 10000 == 0:
torch.save(model.state_dict(), os.path.join(
SNAPSHOTDIR, '%d_%d.pkl' % (epoch, i)))
torch.save(model.state_dict(), os.path.join(
SNAPSHOTDIR, 'last.pkl'))
return iteration
class Dataset():
def __init__(self):
ImageRoot = r'C:\Users\ASUS\Pose2Seg\data\coco2017\train2017'
AnnoFile = r'C:\Users\ASUS\Pose2Seg\data\coco2017\annotations\person_keypoints_train2017_pose2seg.json'
self.datainfos = CocoDatasetInfo(
ImageRoot, AnnoFile, onlyperson=True, loadimg=True)
def __len__(self):
return len(self.datainfos)
def __getitem__(self, idx):
rawdata = self.datainfos[idx]
img = rawdata['data']
image_id = rawdata['id']
height, width = img.shape[0:2]
gt_kpts = np.float32(rawdata['gt_keypoints']).transpose(
0, 2, 1) # (N, 17, 3)
gt_segms = rawdata['segms']
gt_masks = np.array([annToMask(segm, height, width)
for segm in gt_segms])
return {'img': img, 'kpts': gt_kpts, 'masks': gt_masks}
def collate_fn(self, batch):
batchimgs = [data['img'] for data in batch]
batchkpts = [data['kpts'] for data in batch]
batchmasks = [data['masks'] for data in batch]
return {'batchimgs': batchimgs, 'batchkpts': batchkpts, 'batchmasks': batchmasks}
if __name__ == '__main__':
logger.info('===========> loading model <===========')
model = Pose2Seg().cuda()
# model.init("")
model.train()
logger.info('===========> loading data <===========')
datasetTrain = Dataset()
dataloaderTrain = torch.utils.data.DataLoader(datasetTrain, batch_size=1, shuffle=True,
num_workers=0, pin_memory=False,
collate_fn=datasetTrain.collate_fn)
logger.info('===========> set optimizer <===========')
''' set your optimizer like this. Normally is Adam/SGD. '''
#optimizer = torch.optim.SGD(model.parameters(), 0.0002, momentum=0.9, weight_decay=0.0005)
optimizer = torch.optim.Adam(
model.parameters(), 0.0002, weight_decay=0.0000)
iteration = 0
epoch = 0
try:
while iteration < 14150*25:
logger.info('===========> training <===========')
iteration = train(model, dataloaderTrain,
optimizer, epoch, iteration)
epoch += 1
logger.info('===========> testing <===========')
test(model, dataset='cocoVal', logger=logger.info)
test(model, dataset='OCHumanVal', logger=logger.info)
except (KeyboardInterrupt):
logger.info('Save ckpt on exception ...')
torch.save(model.state_dict(), os.path.join(
SNAPSHOTDIR, 'interrupt_%d_%d.pkl' % (epoch, iteration)))
logger.info('Save ckpt done.')
