Function feat_get

eval.py:13–85

(step, G, Cs, dataset_source, dataset_target, save_path,
             ova=True)

def feat_get(step, G, Cs, dataset_source, dataset_target, save_path,
             ova=True):
    G.eval()

    for batch_idx, data in enumerate(dataset_source):
        if batch_idx == 500:
            break
        with torch.no_grad():
            img_s = data[0]
            label_s = data[1]
            img_s, label_s = Variable(img_s.cuda()), \
                             Variable(label_s.cuda())
            feat_s = G(img_s)


            if batch_idx == 0:
                feat_all_s = feat_s.data.cpu().numpy()
                label_all_s = label_s.data.cpu().numpy()
            else:
                feat_s = feat_s.data.cpu().numpy()
                label_s = label_s.data.cpu().numpy()
                feat_all_s = np.r_[feat_all_s, feat_s]
                label_all_s = np.r_[label_all_s, label_s]
    for batch_idx, data in enumerate(dataset_target):
        if batch_idx == 500:
            break
        with torch.no_grad():
            img_t = data[0]
            label_t = data[1]
            img_t, label_t = Variable(img_t.cuda()), \
                             Variable(label_t.cuda())
            feat_t = G(img_t)

            out_t = Cs[0](feat_t)
            pred = out_t.data.max(1)[1]
            out_t = F.softmax(out_t)
            if ova:
                out_open = Cs[1](feat_t)
                out_open = F.softmax(out_open.view(out_t.size(0), 2, -1), 1)
                tmp_range = torch.range(0, out_t.size(0) - 1).long().cuda()
                pred_unk = out_open[tmp_range, 0, pred]
                weights_open = Cs[1].module.fc.weight.data.cpu().numpy()
            else:
                pred_unk = -torch.sum(out_t * torch.log(out_t), 1)

            if batch_idx == 0:
                feat_all = feat_t.data.cpu().numpy()
                label_all = label_t.data.cpu().numpy()
                unk_all = pred_unk.data.cpu().numpy()
                pred_all = pred.data.cpu().numpy()
                pred_all_soft = out_t.data.cpu().numpy()
            else:
                feat_t = feat_t.data.cpu().numpy()
                label_t = label_t.data.cpu().numpy()
                pred_unk = pred_unk.data.cpu().numpy()
                feat_all = np.r_[feat_all, feat_t]
                label_all = np.r_[label_all, label_t]
                unk_all = np.r_[unk_all, pred_unk]