Question

我训练了我的网络，并将最终的权重保存为这样的摘要：

# ...
weights_summaries.append(tf.summary.tensor_summary('out-weights', weights['out'])) 
# ... write summary

现在我想使用我的分类器。我试图从摘要中加载权重：

for e in tf.train.summary_iterator(summary_file):
    for v in e.summary.value:
        # ...
        # found the node
        elif v.node_name == 'out-weights':
            weights['out'] = tf.Variable(v.tensor)
            # it doesn't work!
            weights['out'] = tf.Variable.from_proto(v)
            # assert!
            weights['out'] = tf.Variable.from_proto(v.tensor)
            # assert!
            weights['out'] = tf.Variable(tf.Tensor.from_proto(v.tensor))
            # Tensor.from_proto is not defined!

那么，我应该如何加载重量？我知道＆＃34;全球＆＃34;模型保护程序，但我宁愿只保存我需要的数据。

提前致谢，亚历山大

Answer 1

如果您想要该模型的子集，我建议您使用var_list constructor中的tf.train.Saver参数。

Answer 2

最后，我找到了解决方案：

def tensor_summary_value_to_variable(value):
    fb = numpy.frombuffer(v.tensor.tensor_content, dtype = numpy.float32)
    shape = []
    for d in v.tensor.tensor_shape.dim:
        shape.append(d.size)
    fb = fb.reshape(shape)
    var = tf.Variable(fb)
    return var

如何将张量摘要恢复为变量

2 个答案: