Running target operation for training in C++ about hello_tf_c_api

[Romit-Maulik]

Yup! So for a pretrained and frozen graph you would use something like

// Inputs and dimensions
float input_vals[num_samples][num_inputs]; // populate this with inputs
const std::vector<std::int64_t> input_dims = {num_samples, num_inputs}; //dimensions
// Set up tensors
TF_Tensor* output_tensor_ = nullptr;
TF_Tensor* input_tensor_ = tf_utils::CreateTensor(TF_FLOAT,
                                          input_dims, input_dims.size(),
                                          &input_vals, num_samples*num_inputs*sizeof(float));
// Arrays of tensors
TF_Tensor* input_tensors_[1] = {input_tensor_}; // Array of all the inputs to the model
TF_Tensor* output_tensors_[1] = {output_tensor_}; // Array of all the outputs from the model
// Arrays of operations
TF_Output inputs[1] = {input_ph_}; // The input placeholder(s) 
TF_Output outputs[1] = {output_}; // The output operation

TF_SessionRun(sess_,
                nullptr, // Run options.
                inputs, input_tensors_, 1, // Input tensor ops, input tensor values, number of inputs.
                outputs, output_tensors_, 1, // Output tensor ops, output tensor values, number of outputs.
                nullptr, 0, // *No* target operations, number of target ops.
                nullptr, // Run metadata.
                status_ // Output status.
);

[Romit-Maulik]

I see - thanks! That's what I started out with and am now toying with the idea of moving more functionality into C++.

PS - I found a stackoverflow answer that details (a bit about how one may save a graph to a checkpoint in C++) - I'll look into it.

[Neargye]

Try this

...
TF_Operation* train_ = TF_GraphOperationByName(graph_, "train_step"); 
std::vector<TF_Operation*> target_opers = {train_ };
...
TF_SessionRun(sess_, nullptr, // Run options.
	                &input_op_, &input_tensor_, 1, // Input tensors, input tensor values, number of inputs.
	                &out_op_, &output_tensor_, 1, // Output tensors, output tensor values, number of outputs.
	                target_opers.data(), target_opers.size(), // Target operations, number of targets.
	                nullptr, // Run metadata.
	                status_ // Output status.
	                );

[Neargye]

TF_SessionRun takes an array of target, so the type const TF_Operation* const*.

[Romit-Maulik]

Hi Neargye - thanks! I managed to fix the issue by doing the following:

TF_Operation* train_ = TF_GraphOperationByName(graph_, "train_step");
TF_SessionRun(sess_, nullptr, // Run options.
	                &input_op_, &input_tensor_, 1, // Input tensors, input tensor values, number of inputs.
	                &out_op_, &output_tensor_, 1, // Output tensors, output tensor values, number of outputs.
	                &train_, 1, // Target operations, number of targets.
	                nullptr, // Run metadata.
	                status_ // Output status.
	                );

which I suspect allowed a safe cast.

While I have you here - I was trying to use the above code to train an unfrozen graph in C++ (I had defined the graph through the tensorflow python API). However, I noticed that unfrozen graphs cannot be used for any sort of output- (for instance to print the loss). Have you had any experience with this issue and is there a work around for it?

I suspect I could dump a checkpoint graph to disk and assess on the side - but in-situ would be nice. Do you have an example that shows how to save a graph to disk from C++? Thanks!

[Neargye]

Unfortunately, I did not work with ungrozen graph. I usually use python to train, and use on c ++ then to run.

[carlos-vaz]

What is target operations? I thought that the session (which contains the graph) included all operations. Why would TF_SessionRun take the session as the first argument AS WELL AS a list of all the target operations?

[Romit-Maulik]

The target operations here correspond to those you want to run during the session on the unfrozen graph. In the previous snippet they were those related to optimization of the graph. This gist has all the code you'll need to understand this!

https://gist.github.com/asimshankar/7c9f8a9b04323e93bb217109da8c7ad2

[carlos-vaz]

Thanks Romit! So if I understand correctly, if you want to include all graph operations you leave Target ops as NULL? I am using a frozen graph of Deeplab pre-trained on the Cityscapes dataset, so I believe it is fully optimized and ready to predict.

[youngallien]

Hi ,
i want to do gradient descent in tensorflow-C_api ,but can't find any example tell me how to code,
such as :
how to set and initialize a variable ?
how to set optimizer and train?
@Romit-Maulik @Neargye Thanks for your attention and hope to receive your reply.

[Neargye]

@yangjituan you can also check https://gist.github.com/asimshankar/7c9f8a9b04323e93bb217109da8c7ad2.

[youngallien]

@Neargye Thanks very much, It's very useful,