def valid(val_loader, val_cls_list, val_attrs, model_dir):
# model = resnet50(pretrained=False, cut_at_pooling=False, num_features=1024, norm=False, dropout=0, num_classes=30)
# model = get_network(num_classes=30, depth=50)
model = ResNet50M(num_classes=30)
model = nn.DataParallel(model).cuda()
checkpoint = load_checkpoint(
osp.join(model_dir, 'checkpoint_64_33.pth.tar'))
model.module.load_state_dict(checkpoint['state_dict'])
model.eval()
feat, label = [], []
for i, d in enumerate(val_loader):
imgs, _, l = d
inputs = Variable(imgs)
_, outputs = model(inputs)
feat.append(outputs.data.cpu().numpy())
label.extend(l)
feat = np.vstack(feat)
# name = name.hstack(name)
dist = compute_dist(feat, val_attrs, 'cosine')
result = []
for i, v in enumerate(dist):
max = v.max()
v = list(v)
index = v.index(max)
result.append((int(label[i]), val_cls_list[index]))
n = 0
for tar, pre in result:
if pre == tar:
n = n + 1
print('the acc is {}/{}'.format(n, len(result)))
return n
def test(test_loader, test_cls_list, test_attrs, model_dir):
model = resnet50(pretrained=False, cut_at_pooling=False, num_features=1024, norm=False, dropout=0, num_classes=30)
model = nn.DataParallel(model).cuda()
checkpoint = load_checkpoint(osp.join(model_dir, 'checkpoint.pth.tar'))
model.module.load_state_dict(checkpoint['state_dict'])
model.eval()
feat, name = [], []
for i, d in enumerate(test_loader):
imgs, fnames, _ = d
inputs = Variable(imgs)
_, outputs = model(inputs)
# outputs = F.sigmoid(outputs)
feat.append(outputs.data.cpu().numpy())
name.extend(fnames)
feat = np.vstack(feat)
# name = name.hstack(name)
dist = compute_dist(feat, test_attrs, 'cosine')
result = []
for i, v in enumerate(dist):
max = v.max()
v = list(v)
index = v.index(max)
result.append((int(name[i][:3]), test_cls_list[index]))
n = 0
for pre, tar in result:
if pre == tar:
n = n+1
print('the acc is {}/{}'.format(n, len(result)))