Incubator-mxnet: one_hot usage does not match documentation.

[jmacglashan]

The new one_hot usage does not match the usage described by the documentation.

Documentation claims:

one_hot([1,0,2,0], 3) = [[ 0.  1.  0.]
                         [ 1.  0.  0.]
                         [ 0.  0.  1.]
                         [ 1.  0.  0.]]

Using the same code yields the following error:

mxnet.base.MXNetError: Invalid Parameter format for depth expect int but value='[1, 0, 2, 0]', in operator one_hot(name="", depth="[1, 0, 2, 0]", on_value="3")

Somehow indices is being passed to depth and depth to on_value. Trying to manually specify indices makes it claim it doesn't know that argument:

mx.nd.one_hot(indices=[1,0,2,0], depth=3)

yeilds error

mxnet.base.MXNetError: Cannot find argument 'indices', Possible Arguments:
----------------
depth : int, required
    The dimension size at dim = axis.
on_value : double, optional, default=1
    The value assigned to the locations represented by indices.
off_value : double, optional, default=0
    The value assigned to the locations not represented by indices.
dtype : {'float16', 'float32', 'float64', 'int32', 'uint8'},optional, default='float32'
    DType of the output
, in operator one_hot(name="", depth="3", indices="[1, 0, 2, 0]")

Making the indices argument an ndarray does not help.

Also, the old onehot_encode function has been deprecated making this somewhat problematic.

[piiswrong]

input needs to be an ndarray, not list

[jmacglashan]

As I noted at the end, that still doesn't work, unfortunately. Though it is a slightly different error:

mx.nd.one_hot(mx.nd.array([1,0,2,0]), 3)

yields

mxnet.base.MXNetError: Type inconsistent, Provided=0, inferred type=4

[EdenMacdonald]

I get the same message when using mx.nd.one_hot(). Try casting your indices array to np.int32:

import mxnet as mx
import numpy as np

idxs = mx.nd.array([1,0,2,0])
mx.nd.one_hot(idxs.astype(np.int32), 3)

This seems to solve the problem for me, even though the error message doesn't say anything about types. Perhaps a cast is needed inside mx.nd.one_hot() in order to support the functionality from the docs? Or an assert indicies.dtype == np.int32 so the error is more clear.

Edit: I just found the same problem in mx.nd.batch_take(). Casting to np.int32 worked there as well.

[jmacglashan]

Ah yes, that will get it through on my end too. Alternatively, you can set the type in the constructor.

mx.nd.one_hot(mx.nd.array([1,0,2,0], dtype=np.int32), 3)

However, I will second then that the operator should do an astype for you if the input type is not int32.

[yajiedesign]

This issue is closed due to lack of activity in the last 90 days. Feel free to reopen if this is still an active issue. Thanks!