Calculate metrics in mpc

[qwertyuiop-s]

Good afternoon!

How I can calculate loss in this example?

I tried next:

• In training process add loss = exe.run(feed=sample, fetch_list=[avg_loss]) and try to decrypt by load_decrypt_data(loss_file, (1,), decrypt_file_loss) Have next:

• Also I tried to calculate loss during the infer ( or testing) by adding prediction, loss = exe.run(program=infer_program, feed=sample, fetch_list=[softmax, avg_loss]) and I get Error:

`NotFoundError: Input variable(mean_0.tmp_0) cannot be found in scope for operator 'Fetch'.Confirm that you have used the fetch `Variable` format instead of the string literal('mean_0.tmp_0') in `fetch_list` parameter when using `executor.run` method. In other words, the format of `executor.run(fetch_list=[fetch_var])`(fetch_var is a Variable) is recommended.
  [Hint: fetch_var should not be null.] at (/paddle/paddle/fluid/operators/controlflow/fetch_op.cc:82)
  [operator < fetch > error]`

What is the right way ?

Best wishes!

[jhjiangcs]

Good afternoon!

How I can calculate loss in this example?

I tried next:

• In training process add loss = exe.run(feed=sample, fetch_list=[avg_loss]) and try to decrypt by load_decrypt_data(loss_file, (1,), decrypt_file_loss) Have next:
• Also I tried to calculate loss during the infer ( or testing) by adding prediction, loss = exe.run(program=infer_program, feed=sample, fetch_list=[softmax, avg_loss]) and I get Error:

`NotFoundError: Input variable(mean_0.tmp_0) cannot be found in scope for operator 'Fetch'.Confirm that you have used the fetch `Variable` format instead of the string literal('mean_0.tmp_0') in `fetch_list` parameter when using `executor.run` method. In other words, the format of `executor.run(fetch_list=[fetch_var])`(fetch_var is a Variable) is recommended.
  [Hint: fetch_var should not be null.] at (/paddle/paddle/fluid/operators/controlflow/fetch_op.cc:82)
  [operator < fetch > error]`

What is the right way ?

Best wishes!

This example uses sigmoid_cross_entropy_with_logits (for efficiency, we do not calculate loss in current version).

sigmoid_cross_entropy_with_logits: forward: out = sigmoid(x) backward: dx = sigmoid(x) - label