Multi-output regression and classification

[ilblackdragon]

There are number of problems that require multiple outputs (including RNNs).
Support needs to be added across all models and data feeders.

[boscotsang]

Do you mean output an m by n matrix which m is the number of samples and n is the output which is used as multi classes classifications or feed into the next layer?

[ilblackdragon]

This specific issue is for multi-output classification and multi-output regressions (vs multi class, where there can be multiple classes for the same output).
As of #e443c73 there is multi output regression support, but still need mutli output classification support (especially for cases like decoder in RNN for words/characters).

The multi class case also may need some work. If you are interested into checking it out, I would start with trying to add an example to the examples folder and see what breaks.

[boscotsang]

For multiclass classification we can infer the number of class from the number of unique labels. Then we can transform the label to one hot code representing the class. Therefore we don't need to specify the number for class so the output_shape can be specified by the shape of y. By this way the output of binary or multi regression and classification can be represented by a tensor that the first axis is the number of batch_size and the last size is the number of output. How you do think about it?

[ilblackdragon]

This is actually works. Multioutput regressions are implemented. Multiclass classification works. Multioutput (multiclass) classification works as well - see example language modelling / neural translation.

[waleedka]

@ilblackdragon I've been searching for the multi-output sample you mentioned but I can't seem to find it. I checked the language modeling neural translation examples, but as far as I can tell, they're not using multi-output classification. I'm sure I'm missing something.

My understanding of multi-output classification is when you there are several outputs and each output can have multiple classes. For example, a model that takes pictures of animals in the wild and generates two outputs: (1) the animal type (cat, dot, bird, ...etc.) and (2) the weather in the image (sunny, cloudy, raining, ...etc.). Am I correct that this is what we refer to as multi-output? And if so, does skflow support it?

[wazzy]

For multi-output in RNN text classifier I am getting below error
File "build/bdist.linux-x86_64/egg/skflow/estimators/base.py", line 310, in predict return self._predict(X, axis=axis, batch_size=batch_size) File "build/bdist.linux-x86_64/egg/skflow/estimators/base.py", line 285, in _predict preds.append(predictions_for_batch.argmax(axis=axis)) ValueError: axis(=2) out of bounds

[terrytangyuan]

Please discuss this in StackOverflow since this is a usage question. This repo will be removed soon so you should ask question on TensorFlow's main repo. Thanks.