Cntk: How do I update the parameters?

[opza]

I am currently implementing Center loss using CTNK.
But is there a way to update the parameters for the center update?

I already tried the "C.assign" function, but I can not update on dynamic axes!

[KeDengMS]

C.assign only assigns to constants or parameters, so you may need to use unpack_batch to convert batch axis to static axis. Besides, note that you need to add the assign node to computation graph by doing something like:

a1 = C.assign(...)
loss = loss + 0*a1

[opza]

We have updated the parameter 'unpack_batch' input_variable to 'parameter' using 'C.assign'.
However, if the result is evaluated, the result is not displayed. Is it normal?

[KeDengMS]

What do you mean by "displayed"?

[opza]

---

import cntk as C
import numpy as np

data = np.array([[1,2,3]], np.float32)

param = C.Parameter((3))
inputs = C.input_variable((3))
un_inputs = C.unpack_batch(inputs)

compute = un_inputs + 3
param =C.assign(param, compute)

Error!

print(param.eval(data))

When I run the above code, I get the following error message.

Traceback (most recent call last):
File "", line 1, in
File "C:\Users\user\AppData\Local\Programs\Python\Python35\lib\site-packages\cntk\ops\functions.py", line 708, in eval
_, output_map = self.forward(arguments, outputs, device=device, as_numpy=as_numpy)
File "C:\Users\user\AppData\Local\Programs\Python\Python35\lib\site-packages\cntk\internal\swig_helper.py", line 69, in wrapper
result = f(args, *kwds)
File "C:\Users\user\AppData\Local\Programs\Python\Python35\lib\site-packages\cntk\ops\functions.py", line 842, in forward
keep_for_backward)
File "C:\Users\user\AppData\Local\Programs\Python\Python35\lib\site-packages\cntk\cntk_py.py", line 1962, in _forward
return _cntk_py.Function__forward(self, *args)
ValueError: AssignNode: All inputs should have same sample layout.

[CALL STACK]

std::enable_shared_from_this:: shared_from_this
- std::enable_shared_from_this:: shared_from_this (x3)
- CNTK::Internal:: UseSparseGradientAggregationInDataParallelSGD
- std::enable_shared_from_this:: shared_from_this
- CNTK::Internal:: UseSparseGradientAggregationInDataParallelSGD
- CNTK::Function:: Forward
- PyInit__cntk_py (x2)
- PyCFunction_Call
- PyEval_GetFuncDesc
- PyEval_EvalFrameEx (x2)
- PyEval_GetFuncDesc (x2)

[KeDengMS]

This is because your param has different shape than the value assigned to it. Change to param = C.parameter(data.shape) fixed the issue. NOTE that in your input data, the batch size is 2, and to have consistent param shape the batch size cannot be changed.

[opza]

Wow! The code works normally. Thank you!