def baseline_one_model_train_no_background():
cfg_dir = '../config_alcohol/cascade_rcnn_r50_fpn_1x'
cfg_names = [
'baseline.py',
]
# watch train effects using different base cfg
ratios = [1]
ns = ratios
cfgs = []
for i, n in enumerate(ns):
cfg = mmcv.Config.fromfile(os.path.join(cfg_dir, cfg_names[0]))
cfg.data['imgs_per_gpu'] = 2
cfg.optimizer['lr'] = cfg.optimizer['lr'] / 8 * (
cfg.data['imgs_per_gpu'] / 2)
cfg.cfg_name = 'baseline_one_model'
cfg.uid = 'background=No'
cfg.work_dir = os.path.join(cfg.work_dir, cfg.cfg_name,
'baseline_one_model,background=No')
cfg.resume_from = os.path.join(cfg.work_dir, 'latest.pth')
if not os.path.exists(cfg.resume_from):
cfg.resume_from = None
cfgs.append(cfg)
batch_train(cfgs, sleep_time=60 * 2)
from batch_test import batch_test
save_path = os.path.join(cfg_dir,
'baseline_one_model_test,background=No,.txt')
batch_test(cfgs, save_path, 60 * 2, mode='val')
def anchor_scales_cluster_train():
from tricks.data_cluster import box_cluster
cfg_dir = '../config_alcohol/cascade_rcnn_r50_fpn_1x'
cfg_names = [DATA_NAME + '.py', ]
cfg = mmcv.Config.fromfile(os.path.join(cfg_dir, cfg_names[0]))
# new added
boxes = box_cluster(cfg.data['train']['ann_file'], n=5)
anchor_scales = np.sqrt(boxes[0][0] * boxes[0][1])
anchor_scales = min(anchor_scales * 1333 / 2446, anchor_scales * 800 / 1000) / 4
anchor_scales = int(anchor_scales)
cfg.model['rpn_head']['anchor_scales'] = list([anchor_scales])
cfg.data['imgs_per_gpu'] = 2
cfg.optimizer['lr'] = cfg.optimizer['lr'] / 8 * (cfg.data['imgs_per_gpu'] / 2)
cfg.cfg_name = DATA_NAME + '_baseline'
if anchor_scales == 8:
cfg.uid = 'mode=baseline'
print('This trick is the same with baseline model.')
else:
cfg.uid = 'anchor_scales_cluster=6'
cfg.work_dir = os.path.join(cfg.work_dir, cfg.cfg_name, cfg.cfg_name + ',' + cfg.uid)
cfg.resume_from = os.path.join(cfg.work_dir, 'latest.pth')
if not os.path.exists(cfg.resume_from):
cfg.resume_from = None
cfgs = [cfg]
batch_train(cfgs, sleep_time=0 * 60 * 2)
from batch_test import batch_test
save_path = os.path.join(cfg_dir, DATA_NAME + '_test.txt')
batch_test(cfgs, save_path, 60 * 2, mode=DATA_MODE)
def anchor_ratios_cluster_train():
cfg_dir = '../config_alcohol/cascade_rcnn_r50_fpn_1x'
cfg_names = [DATA_NAME + '.py', ]
cfg = mmcv.Config.fromfile(os.path.join(cfg_dir, cfg_names[0]))
# new added
from tricks.data_cluster import anchor_cluster
anchor_ratios = anchor_cluster(cfg.data['train']['ann_file'], n=6)
cfg.model['rpn_head']['anchor_ratios'] = list(anchor_ratios)
cfg.data['imgs_per_gpu'] = 2
cfg.optimizer['lr'] = cfg.optimizer['lr'] / 8 * (cfg.data['imgs_per_gpu'] / 2)
cfg.cfg_name = DATA_NAME + '_baseline'
cfg.uid = 'anchor_cluster=6'
cfg.work_dir = os.path.join(cfg.work_dir, cfg.cfg_name, cfg.cfg_name + ',' + cfg.uid)
cfg.resume_from = os.path.join(cfg.work_dir, 'latest.pth')
if not os.path.exists(cfg.resume_from):
cfg.resume_from = None
cfgs = [cfg]
batch_train(cfgs, sleep_time=0 * 60 * 2)
from batch_test import batch_test
save_path = os.path.join(cfg_dir, DATA_NAME + '_test.txt')
batch_test(cfgs, save_path, 60 * 2, mode=DATA_MODE)
def iou_thr_train(iou_thrs=[0.5, 0.6, 0.7]):
cfg_dir = '../config_alcohol/cascade_rcnn_r50_fpn_1x'
cfg_names = [DATA_NAME + '.py', ]
cfg = mmcv.Config.fromfile(os.path.join(cfg_dir, cfg_names[0]))
for i, rcnn in enumerate(cfg.train_cfg['rcnn']):
rcnn['assigner']['pos_iou_thr'] = iou_thrs[i]
rcnn['assigner']['neg_iou_thr'] = iou_thrs[i]
rcnn['assigner']['min_pos_iou'] = iou_thrs[i]
cfg.data['imgs_per_gpu'] = 2
cfg.optimizer['lr'] = cfg.optimizer['lr'] / 8 * (cfg.data['imgs_per_gpu'] / 2)
cfg.cfg_name = DATA_NAME + '_baseline'
cfg.uid = 'rcnn_iou_thrs={}'.format(str(iou_thrs))
cfg.work_dir = os.path.join(cfg.work_dir, cfg.cfg_name, cfg.cfg_name + ',' + cfg.uid)
cfg.resume_from = os.path.join(cfg.work_dir, 'latest.pth')
if not os.path.exists(cfg.resume_from):
cfg.resume_from = None
cfgs = [cfg]
batch_train(cfgs, sleep_time=0 * 60 * 2)
from batch_test import batch_test
save_path = os.path.join(cfg_dir, DATA_NAME + '_test.txt')
batch_test(cfgs, save_path, 60 * 2, mode=DATA_MODE)
def backbone_dcn_train():
cfg_dir = '../config_alcohol/cascade_rcnn_r50_fpn_1x'
cfg_names = [DATA_NAME + '.py', ]
cfg = mmcv.Config.fromfile(os.path.join(cfg_dir, cfg_names[0]))
cfg.model['backbone']['dcn'] = dict( # 在最后三个block加入可变形卷积
modulated=False, deformable_groups=1, fallback_on_stride=False)
cfg.model['backbone']['stage_with_dcn'] = (False, True, True, True)
cfg.data['imgs_per_gpu'] = 2
cfg.optimizer['lr'] = cfg.optimizer['lr'] / 8 * (cfg.data['imgs_per_gpu'] / 2)
cfg.cfg_name = DATA_NAME + '_baseline'
cfg.uid = 'backbone_dcn=True'
cfg.work_dir = os.path.join(cfg.work_dir, cfg.cfg_name, cfg.cfg_name + ',' + cfg.uid)
cfg.resume_from = os.path.join(cfg.work_dir, 'latest.pth')
if not os.path.exists(cfg.resume_from):
cfg.resume_from = None
cfgs = [cfg]
batch_train(cfgs, sleep_time=0 * 60 * 2)
from batch_test import batch_test
save_path = os.path.join(cfg_dir, DATA_NAME + '_test.txt')
batch_test(cfgs, save_path, 60 * 2, mode=DATA_MODE)
def multi_scale_train(img_scale=None, multiscale_mode='range', ratio_range=None, keep_ratio=True):
cfg_dir = '../config_alcohol/cascade_rcnn_r50_fpn_1x'
cfg_names = [DATA_NAME + '.py', ]
cfg = mmcv.Config.fromfile(os.path.join(cfg_dir, cfg_names[0]))
cfg.train_pipeline[2] = mmcv.ConfigDict(
type='Resize', img_scale=img_scale, ratio_range=ratio_range,
multiscale_mode=multiscale_mode, keep_ratio=keep_ratio)
sx = int(np.mean([v[0] for v in img_scale]))
sy = int(np.mean([v[1] for v in img_scale]))
cfg.test_pipeline[1]['img_scale'] = [(sx, sy)]
cfg.data['train']['pipeline'] = cfg.train_pipeline
cfg.data['val']['pipeline'] = cfg.test_pipeline
cfg.data['test']['pipeline'] = cfg.test_pipeline
cfg.data['imgs_per_gpu'] = 2
cfg.optimizer['lr'] = cfg.optimizer['lr'] / 8 * (cfg.data['imgs_per_gpu'] / 2)
cfg.cfg_name = DATA_NAME + '_baseline'
cfg.uid = 'img_scale={},multiscale_mode={},ratio_range={}' \
.format(str(img_scale), str(multiscale_mode), str(ratio_range))
cfg.work_dir = os.path.join(cfg.work_dir, cfg.cfg_name, cfg.cfg_name + ',' + cfg.uid)
cfg.resume_from = os.path.join(cfg.work_dir, 'latest.pth')
if not os.path.exists(cfg.resume_from):
cfg.resume_from = None
cfgs = [cfg]
batch_train(cfgs, sleep_time=0 * 60 * 2)
from batch_test import batch_test
save_path = os.path.join(cfg_dir, DATA_NAME + '_test.txt')
batch_test(cfgs, save_path, 60 * 2, mode=DATA_MODE)
def OHEMSampler_train():
cfg_dir = '../config_alcohol/cascade_rcnn_r50_fpn_1x'
cfg_names = [DATA_NAME + '.py', ]
cfg = mmcv.Config.fromfile(os.path.join(cfg_dir, cfg_names[0]))
for rcnn in cfg.train_cfg['rcnn']:
rcnn['sampler'] = dict(
type='OHEMSampler',
num=512,
pos_fraction=0.25,
neg_pos_ub=-1,
add_gt_as_proposals=True)
cfg.data['imgs_per_gpu'] = 2
cfg.optimizer['lr'] = cfg.optimizer['lr'] / 8 * (cfg.data['imgs_per_gpu'] / 2)
cfg.cfg_name = DATA_NAME + '_baseline'
cfg.uid = 'sampler=OHEMSampler'
cfg.work_dir = os.path.join(cfg.work_dir, cfg.cfg_name, cfg.cfg_name + ',' + cfg.uid)
cfg.resume_from = os.path.join(cfg.work_dir, 'latest.pth')
if not os.path.exists(cfg.resume_from):
cfg.resume_from = None
cfgs = [cfg]
batch_train(cfgs, sleep_time=0 * 60 * 2)
from batch_test import batch_test
save_path = os.path.join(cfg_dir, DATA_NAME + '_test.txt')
batch_test(cfgs, save_path, 60 * 2, mode=DATA_MODE)
def joint_train():
cfg_dir = '../config_alcohol/cascade_rcnn_r50_fpn_1x'
cfg_names = [DATA_NAME + '.py', ]
cfg = mmcv.Config.fromfile(os.path.join(cfg_dir, cfg_names[0]))
# 0.810 (img_scale = [1920, 1080]) ==> (img_scale = [(1920, 1080), (1333, 800)], multiscale_mode='range')
img_scale = [(1920, 1080), (1333, 800)]
cfg.train_pipeline[2] = mmcv.ConfigDict(
type='Resize', img_scale=img_scale, ratio_range=None,
multiscale_mode='range', keep_ratio=True)
sx = int(max([v[0] for v in img_scale]))
sy = int(max([v[1] for v in img_scale]))
cfg.test_pipeline[1]['img_scale'] = [(sx, sy)]
cfg.data['train'] = cfg.train_pipeline
cfg.data['val'] = cfg.test_pipeline
cfg.data['test'] = cfg.test_pipeline
# 0.819
cfg.model['backbone']['dcn'] = dict( # 在最后三个block加入可变形卷积
modulated=False, deformable_groups=1, fallback_on_stride=False)
cfg.model['backbone']['stage_with_dcn'] = (False, True, True, True)
# 0.822
# global context
cfg.model['bbox_roi_extractor']['global_context'] = True
# # 0.746
# from tricks.data_cluster import anchor_cluster
# anchor_ratios = anchor_cluster(cfg.data['train']['ann_file'], n=6)
# cfg.model['rpn_head']['anchor_ratios'] = list(anchor_ratios)
# 0.???
# focal loss for rcnn
# for head in cfg.model['bbox_head']:
# head['loss_cls'] = dict(type='FocalLoss', use_sigmoid=True, loss_weight=1.0)
cfg.data['imgs_per_gpu'] = 2
cfg.optimizer['lr'] = cfg.optimizer['lr'] / 8 * (cfg.data['imgs_per_gpu'] / 2)
cfg.cfg_name = DATA_NAME + '_baseline'
cfg.uid = 'mode=joint_train+multiscale_mode=range'
cfg.work_dir = os.path.join(cfg.work_dir, cfg.cfg_name, cfg.cfg_name + ',' + cfg.uid)
cfg.resume_from = os.path.join(cfg.work_dir, 'latest.pth')
if not os.path.exists(cfg.resume_from):
cfg.resume_from = None
cfgs = [cfg]
batch_train(cfgs, sleep_time=0 * 60 * 2)
from batch_test import batch_test
save_path = os.path.join(cfg_dir, DATA_NAME + '_test.txt')
batch_test(cfgs, save_path, 60 * 2, mode=DATA_MODE)
from batch_train import batch_infer
batch_infer(cfgs)
def baseline_train():
cfg_dir = '../config_alcohol/cascade_rcnn_r50_fpn_1x'
cfg_names = [DATA_NAME + '.py', ]
cfg = mmcv.Config.fromfile(os.path.join(cfg_dir, cfg_names[0]))
cfg.data['imgs_per_gpu'] = 2
cfg.optimizer['lr'] = cfg.optimizer['lr'] / 8 * (cfg.data['imgs_per_gpu'] / 2)
cfg.cfg_name = DATA_NAME + '_baseline'
cfg.uid = 'mode=baseline'
cfg.work_dir = os.path.join(cfg.work_dir, cfg.cfg_name, cfg.cfg_name + ',' + cfg.uid)
cfg.resume_from = os.path.join(cfg.work_dir, 'latest.pth')
if not os.path.exists(cfg.resume_from):
cfg.resume_from = None
cfgs = [cfg]
batch_train(cfgs, sleep_time=0 * 60 * 2)
from batch_test import batch_test
save_path = os.path.join(cfg_dir, DATA_NAME + '_test.txt')
batch_test(cfgs, save_path, 60 * 2, mode=DATA_MODE)
def SWA_train():
import torch
import os
import mmcv
from mmdet.models import build_detector
def get_model(config, model_dir):
model = build_detector(config.model, test_cfg=config.test_cfg)
checkpoint = torch.load(model_dir)
state_dict = checkpoint['state_dict']
model.load_state_dict(state_dict, strict=True)
return model
def model_average(modelA, modelB, alpha):
# modelB占比 alpha
for A_param, B_param in zip(modelA.parameters(), modelB.parameters()):
A_param.data = A_param.data * (1 - alpha) + alpha * B_param.data
return modelA
def swa(cfg, epoch_inds, alpha=0.7):
###########################注意，此py文件没有更新batchnorm层，所以只有在mmdetection默认冻住BN情况下使用，如果训练时BN层被解冻，不应该使用此py ＃＃＃＃＃
#########逻辑上会 score 会高一点不会太多，需要指定的参数是 [config_dir , epoch_indices , alpha] ######################
if isinstance(cfg, str):
config = mmcv.Config.fromfile(cfg)
else:
config = cfg
work_dir = config.work_dir
model_dir_list = [os.path.join(work_dir, 'epoch_{}.pth'.format(epoch)) for epoch in epoch_inds]
model_ensemble = None
for model_dir in model_dir_list:
if model_ensemble is None:
model_ensemble = get_model(config, model_dir)
else:
model_fusion = get_model(config, model_dir)
model_ensemble = model_average(model_ensemble, model_fusion, alpha)
checkpoint = torch.load(model_dir_list[-1])
checkpoint['state_dict'] = model_ensemble.state_dict()
ensemble_path = os.path.join(work_dir, 'epoch_ensemble.pth')
torch.save(checkpoint, ensemble_path)
return ensemble_path
cfg_dir = '../config_alcohol/cascade_rcnn_r50_fpn_1x'
cfg_names = [DATA_NAME + '.py', ]
cfg = mmcv.Config.fromfile(os.path.join(cfg_dir, cfg_names[0]))
cfg.data['imgs_per_gpu'] = 2
cfg.optimizer['lr'] = cfg.optimizer['lr'] / 8 * (cfg.data['imgs_per_gpu'] / 2)
cfg.cfg_name = DATA_NAME + '_baseline'
cfg.uid = 'SWA=[10,11,12]'
cfg.work_dir = os.path.join(cfg.work_dir, cfg.cfg_name, cfg.cfg_name + ',' + 'mode=baseline')
ensemble_path = swa(cfg, [10, 11, 12])
cfg.resume_from = ensemble_path
if not os.path.exists(cfg.resume_from):
cfg.resume_from = None
cfgs = [cfg]
# batch_train(cfgs, sleep_time=0 * 60 * 2)
from batch_test import batch_test
save_path = os.path.join(cfg_dir, DATA_NAME + '_test.txt')
batch_test(cfgs, save_path, 60 * 2, mode=DATA_MODE)