情况是这样的，我将模型训练好，得到参数dnn_params1.tar.gz，然后用这个模型打分也是正常的。

但是，如果我将这个dnn_params1.tar.gz用tar -xzvf 命令解压缩到目录temp/下，然后再跳转到temp目录下，用命令tar -zcvf 将所有文件压缩成dnn_params2.tar.gz。

然后我加载dnn_params2.tar.gz预测，结果就报错。

Traceback (most recent call last):

  File "dnn_attention.py", line 153, in

    dnn_infer("model_params/dnn_params_pass0.tar.gz")

  File "dnn_attention.py", line 118, in dnn_infer

    parameters = paddle.parameters.Parameters.from_tar(gzip.open(param_file_name))

  File "/home/forum/paddle_local/python27-gcc482/lib/python2.7/site-packages/paddle/v2/parameters.py", line 382, in from_tar

    f = tar.extractfile(param_name)

  File "/home/forum/paddle_local/python27-gcc482/lib/python2.7/tarfile.py", line 2136, in extractfile

    tarinfo = self.getmember(member)

  File "/home/forum/paddle_local/python27-gcc482/lib/python2.7/tarfile.py", line 1821, in getmember

    raise KeyError("filename %r not found" % name)

KeyError: "filename u'_attention_layer.w0' not found"