摘要：本案例代码是FCOS论文复现的体验案例，此模型为FCOS论文中所提出算法在ModelArts + PyTorch框架下的实现。本代码支持FCOS + ResNet-101在MS-COCO数据集上完整的训练和测试流程
本文分享自华为云社区《通用物体检测算法 FCOS(目标检测/Pytorch)》，作者： HWCloudAI 。
FCOS:Fully Convolutional One-Stage Object Detection
本案例代码是FCOS论文复现的体验案例
此模型为FCOS论文中所提出算法在ModelArts + PyTorch框架下的实现。该算法使用MS-COCO公共数据集进行训练和评估。本代码支持FCOS + ResNet-101在MS-COCO数据集上完整的训练和测试流程
具体的算法介绍：https://marketplace.huaweicloud.com/markets/aihub/modelhub/detail/?id=ce7acc40-0540-45c9-a0c6-e2fda8d1ac7e
注意事项：
1.本案例使用框架: PyTorch1.0.0
2.本案例使用硬件: GPU
3.运行代码方法: 点击本页面顶部菜单栏的三角形运行按钮或按Ctrl+Enter键 运行每个方块中的代码
1.数据和代码下载
import os
import moxing as mox
# 数据代码下载
mox.file.copy_parallel('obs://obs-aigallery-zc/algorithm/FCOS.zip','FCOS.zip')
# 解压缩
os.system('unzip FCOS.zip -d ./')
2.模型训练
2.1依赖库安装及加载
"""
Basic training script for PyTorch
"""
# Set up custom environment before nearly anything else is imported
# NOTE: this should be the first import (no not reorder)
import os
import argparse
import torch
import shutil
src_dir = './FCOS/'
os.chdir(src_dir)
os.system('pip install -r ./pip-requirements.txt')
os.system('python -m pip install ./trained_model/model/framework-2.0-cp36-cp36m-linux_x86_64.whl')
os.system('python setup.py build develop')
from framework.utils.env import setup_environment
from framework.config import cfg
from framework.data import make_data_loader
from framework.solver import make_lr_scheduler
from framework.solver import make_optimizer
from framework.engine.inference import inference
from framework.engine.trainer import do_train
from framework.modeling.detector import build_detection_model
from framework.utils.checkpoint import DetectronCheckpointer
from framework.utils.collect_env import collect_env_info
from framework.utils.comm import synchronize, \
get_rank, is_pytorch_1_1_0_or_later
from framework.utils.logger import setup_logger
from framework.utils.miscellaneous import mkdir
2.2训练函数
def train(cfg, local_rank, distributed, new_iteration=False):
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
if cfg.MODEL.USE_SYNCBN:
assert is_pytorch_1_1_0_or_later(), \
"SyncBatchNorm is only available in pytorch >= 1.1.0"
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[local_rank], output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(
cfg, model, optimizer, scheduler, output_dir, save_to_disk
)
print(cfg.MODEL.WEIGHT)
extra_checkpoint_data = checkpointer.load_from_file(cfg.MODEL.WEIGHT)
print(extra_checkpoint_data)
arguments.update(extra_checkpoint_data)
if new_iteration:
arguments["iteration"] = 0
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
do_train(
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
arguments,
)
return model
2.3设置参数，开始训练
def main():
parser = argparse.ArgumentParser(description="PyTorch Object Detection Training")
parser.add_argument(
'--train_url',
default='./outputs',
type=str,
help='the path to save training outputs'
)
parser.add_argument(
"--config-file",
default="./trained_model/model/fcos_resnet_101_fpn_2x.yaml",
metavar="FILE",
help="path to config file",
type=str,
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument('--train_iterations', default=0, type=int)
parser.add_argument('--warmup_iterations', default=500, type=int)
parser.add_argument('--train_batch_size', default=8, type=int)
parser.add_argument('--solver_lr', default=0.01, type=float)
parser.add_argument('--decay_steps', default='120000,160000', type=str)
parser.add_argument('--new_iteration',default=False, action='store_true')
args, unknown = parser.parse_known_args()
cfg.merge_from_file(args.config_file)
# load the model trained on MS-COCO
if args.train_iterations > 0:
cfg.SOLVER.MAX_ITER = args.train_iterations
if args.warmup_iterations > 0:
cfg.SOLVER.WARMUP_ITERS = args.warmup_iterations
if args.train_batch_size > 0:
cfg.SOLVER.IMS_PER_BATCH = args.train_batch_size
if args.solver_lr > 0:
cfg.SOLVER.BASE_LR = args.solver_lr
if len(args.decay_steps) > 0:
steps = args.decay_steps.replace(' ', ',')
steps = steps.replace(';', ',')
steps = steps.replace('；', ',')
steps = steps.replace('，', ',')
steps = steps.split(',')
steps = tuple([int(x) for x in steps])
cfg.SOLVER.STEPS = steps
cfg.freeze()
num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
args.distributed = num_gpus > 1
if args.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(
backend="nccl", init_method="env://"
)
synchronize()
output_dir = args.train_url
if output_dir:
mkdir(output_dir)
logger = setup_logger("framework", output_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(args)
logger.info("Loaded configuration file {}".format(args.config_file))
train(cfg, args.local_rank, args.distributed, args.new_iteration)
if __name__ == "__main__":
main()
3.模型测试
3.1预测函数
from framework.engine.predictor import Predictor
from PIL import Image,ImageDraw
import numpy as np
import matplotlib.pyplot as plt
def predict(img_path,model_path):
config_file = "./trained_model/model/fcos_resnet_101_fpn_2x.yaml"
cfg.merge_from_file(config_file)
cfg.defrost()
cfg.MODEL.WEIGHT = model_path
cfg.OUTPUT_DIR = None
cfg.freeze()
predictor = Predictor(cfg=cfg, min_image_size=800)
src_img = Image.open(img_path)
img = src_img.convert('RGB')
img = np.array(img)
img = img[:, :, ::-1]
predictions = predictor.compute_prediction(img)
top_predictions = predictor.select_top_predictions(predictions)
bboxes = top_predictions.bbox.int().numpy().tolist()
bboxes = [[x[1], x[0], x[3], x[2]] for x in bboxes]
scores = top_predictions.get_field("scores").numpy().tolist()
scores = [round(x, 4) for x in scores]
labels = top_predictions.get_field("labels").numpy().tolist()
labels = [predictor.CATEGORIES[x] for x in labels]
draw = ImageDraw.Draw(src_img)
for i,bbox in enumerate(bboxes):
draw.text((bbox[1],bbox[0]),labels[i] + ':'+str(scores[i]),fill=(255,0,0))
draw.rectangle([bbox[1],bbox[0],bbox[3],bbox[2]],fill=None,outline=(255,0,0))
return src_img
3.2开始预测
if __name__ == "__main__":
model_path = "./outputs/weights/fcos_resnet_101_fpn_2x/model_final.pth" # 训练得到的模型
image_path = "./trained_model/model/demo_image.jpg" # 预测的图像
img = predict(image_path,model_path)
plt.figure(figsize=(10,10)) #设置窗口大小
plt.imshow(img)
plt.show()
2021-06-09 15:33:15,362 framework.utils.checkpoint INFO: Loading checkpoint from ./outputs/weights/fcos_resnet_101_fpn_2x/model_final.pth
