def inference_to_get_feats(cfg, model, val_loader, num_query, dataset):
device = cfg.MODEL.DEVICE
model.to(device)
logger = logging.getLogger("reid_baseline.inference")
logger.info("Enter inferencing")
metric = evaluator(num_query, dataset, cfg, max_rank=50)
debug = True
model.eval()
start = time.time()
with torch.no_grad():
for batch in val_loader:
data, pid, camid, img_path = batch
data = data.cuda()
feats = model(data)
if cfg.TEST.FLIP_TEST:
data_flip = data.flip(dims=[3]) # NCHW
feats_flip = model(data_flip)
feats = (feats + feats_flip) / 2
output = [feats, pid, camid, img_path]
metric.update(output)
if debug:
print(type(metric.feats))
feats = torch.cat(metric.feats, dim=0)
if metric.feat_norm:
feats = torch.nn.functional.normalize(feats, dim=1, p=2)
# feats = torch.nn.functional.normalize(metric.feats, dim=1, p=2) if metric.feat_norm else metric.feats
return metric.img_paths, feats
def inference(cfg, model, val_loader, num_query, dataset):
device = cfg.MODEL.DEVICE
model.to(device)
logger = logging.getLogger("reid_baseline.inference")
logger.info("Enter inferencing")
metric = evaluator(num_query, dataset, cfg, max_rank=100)
model.eval()
start = time.time()
with torch.no_grad():
for batch in val_loader:
data, pid, camid, img_path = batch
data = data.cuda()
feats = model(data)
if cfg.TEST.FLIP_TEST:
data_flip = data.flip(dims=[3]) # NCHW
feats_flip = model(data_flip)
feats = (feats + feats_flip) / 2
output = [feats, pid, camid, img_path]
metric.update(output)
end = time.time()
logger.info("inference takes {:.3f}s".format((end - start)))
torch.cuda.empty_cache()
cmc, mAP, indices_np = metric.compute()
logger.info('Validation Results')
logger.info("mAP: {:.1%}".format(mAP))
for r in [1, 5, 10]:
logger.info("CMC curve, Rank-{:<3}:{:.1%}".format(r, cmc[r - 1]))
return indices_np
def inference(cfg, model, val_loader, num_query, dataset):
device = cfg.MODEL.DEVICE
model.to(device) #model转GPU
logger = logging.getLogger("reid_baseline.inference")
logger.info("Enter inferencing")
metric = evaluator(num_query, dataset, cfg,
max_rank=50) #reid评估指标算子 /lib/utils/reid_eval.py
model.eval() #固定参数
start = time.time()
with torch.no_grad():
for batch in val_loader: #分batch取出测试数据
data, pid, camid, img_path = batch #一个batch包含这些信息
data = data.cuda() #一个batch的img转GPU
feats = model(data) #前向计算，得到feats特征
if cfg.TEST.FLIP_TEST: #测试技巧，flip图像后再次计算特征，与原图特征求平均
data_flip = data.flip(dims=[3]) # NCHW 水平flip？
feats_flip = model(data_flip)
feats = (feats + feats_flip) / 2 #求平均
#完成一个batch的特征计算
output = [feats, pid, camid, img_path] #只用于计算评估指标
metric.update(output) #更新reid评估指标算子
#结束一个batch
#结束所有batch
end = time.time()
logger.info("inference takes {:.3f}s".format((end - start)))
torch.cuda.empty_cache()
cmc, mAP, indices_np = metric.compute(
) #计算reid评估指标，计算距离矩阵和特征并保存，其中indices_np是距离矩阵的列排序index，间接表示哪个gallery距离每一行的query最近
logger.info('Validation Results')
logger.info("mAP: {:.1%}".format(mAP))
for r in [1, 5, 10]:
logger.info("CMC curve, Rank-{:<3}:{:.1%}".format(r, cmc[r - 1]))
return indices_np
def validate(model, dataset, val_loader, num_query, epoch, cfg, logger):
metric = evaluator(num_query, dataset, cfg, max_rank=50)
model.eval()
with torch.no_grad():
for batch in val_loader:
data, pid, camid, img_path = batch
data = data.cuda()
feats = model(data)
output = [feats, pid, camid, img_path]
metric.update(output)
cmc, mAP, _ = metric.compute()
logger.info("Validation Results - Epoch: {}".format(epoch))
logger.info("mAP: {:.1%}".format(mAP))
for r in [1, 5, 10]:
logger.info("CMC curve, Rank-{:<3}:{:.1%}".format(r, cmc[r - 1]))
return mAP