使用python tensorflow API,可以执行以下操作:
a = tf.placeholder(tf.float32)
b = tf.placeholder(tf.float32)
adder_node = a + b
sess.run(adder_node, {a: [1,3], b: [2, 4]})
结果:[3. 7.]
有没有什么办法在C ++中通过一次调用Run
方法为模型提供多个输入?我试图使用feed_dicts的std :: vector
// prepare tensorflow inputs
std::vector<std::pair<std::string, tensorflow::Tensor>> feed_dict;
for(size_t i = 0; i < noutput_items; i++) {
tensorflow::TensorShape data_shape({1, d_vlen_in});
tensorflow::Tensor n_tensor(tensorflow::DT_FLOAT, data_shape);
auto n_data = n_tensor.flat<float>().data();
for(int j = 0 ; j < d_vlen_in ; j++) {
n_data[j] = in[j];
}
feed_dict.push_back(std::make_pair(d_layer_in, n_tensor));
in += d_vlen_in;
}
// prepare tensorflow outputs
std::vector<tensorflow::Tensor> outputs;
TF_CHECK_OK(d_session->Run(feed_dict, {d_layer_out}, {}, &outputs));
d_layer_in
和d_layer_out
是std::strings
,“input”是我的输入图层/占位符。
然而它失败了:
Non-OK-status: d_session->Run(feed_dict, {d_layer_out}, {}, &outputs) status: Invalid argument: Endpoint "input" fed more than once.
有人知道这样做的方法吗?我的主要目标是提高吞吐量。
所以我想出了答案,这很简单。 feed字典的几个元素用于设置几个输入变量,与python中的相同。然而,输入张量的第一(或零)维度是批量维度,其在某种情况下可以用作输入信号的时间维度。
// prepare tensorflow inputs
// dimension 0 is the batch dimension, i.e., time dimension
tensorflow::TensorShape data_shape({noutput_items, d_vlen_in});
tensorflow::Tensor in_tensor(tensorflow::DT_FLOAT, data_shape);
auto in_tensor_data = in_tensor.flat<float>().data();
for(size_t i = 0; i < noutput_items; i++) {
for(int j = 0 ; j < d_vlen_in ; j++) {
in_tensor_data[(i*d_vlen_in)+j] = in[(i*d_vlen_in)+j];
}
}
std::vector<std::pair<std::string, tensorflow::Tensor>> feed_dict = {
{ d_layer_in, in_tensor },
};
// prepare tensorflow outputs
std::vector<tensorflow::Tensor> outputs;
TF_CHECK_OK(d_session->Run(feed_dict, {d_layer_out}, {}, &outputs));
这当然会引入一些缓冲/延迟,但它使我的吞吐量增加了大约1000倍。