CODEWITHSUNDEEP
deep-learningnumpy-ndarraymxnet
Jouline
Jouline

Reputation: 90

Adding <NDArray 4 @gpu(0)> and regular np.array

I have a loss variable that is returned after forward path, made of 4 "sub-losses" such as :

print(my_loss)

> Loss: 
 [0.37887186 0.4668851  0.4145702  0.506646  ]
 <NDArray 4 @gpu(0)>

I would like to sum all losses by epoch, while keeping the sub-losses split. I wanted to use numpy such as :

loss_to_save = np.zeros((4,))
loss_to_save += my_loss

However, this error is returned when trying the addition of arrays:

Traceback (most recent call last):
  File "train_schedule_copy.py", line 432, in <module>
    train(net, filename=cst.flname_weights, optimise="MCCExtent", resume=resumeFile)
  File "train_schedule_copy.py", line 292, in train
    loss_to_save += my_loss
  File ".local/lib/python3.8/site-packages/mxnet/ndarray/ndarray.py", line 291, in __radd__
    return self.__add__(other)
  File ".local/lib/python3.8/site-packages/mxnet/ndarray/ndarray.py", line 277, in __add__
    return add(self, other)
  File ".local/lib/python3.8/site-packages/mxnet/ndarray/ndarray.py", line 3634, in add
    return _ufunc_helper(
  File ".local/lib/python3.8/site-packages/mxnet/ndarray/ndarray.py", line 3578, in _ufunc_helper
    raise TypeError('type %s not supported' % str(type(rhs)))
TypeError: type <class 'numpy.ndarray'> not supported

From what I understand, the <NDArray 4 @gpu(0)> data type is not allowed to be added to regular numpy array ? How could achieve such operation ?

Upvotes: 0

Views: 20

Answers (2)

Jouline
Jouline

Reputation: 90

I managed to solve my issue by adding the asscalar() method, such as :

loss_to_save = np.zeros((4,), dtype=np.float64)
(...)
loss_ind_sums = []
for l in my_loss:
    summ = l.sum().asscalar()
    loss_ind_sums.append(summ)
loss_to_save += loss_ind_sums

Upvotes: 0

SickerDude43
SickerDude43

Reputation: 196

I don't know whether the problem is still actual or that you've solved it.

But the loss gets placed into an NDArray. NDArrays are built in arrays from the MXNet framework that are similar to numpy's NPArrays. But they are not the same!

To use the normal numpy methods you need to convert your ndArray to numpy.

For that you can use the .asnumpy() method. It will convert your NDArray to NPArray. After that the operations should work if they got the same shape.

Upvotes: 0

Related Questions