Ludwig: Using sequence combiner

[BenMacKenzie]

I'd like to be able to process a sequence of records in an RNN. The following seems to work:

data:
x,y,label
a b c, d e f, 1
z x y, t r q, 0
a b c, d e f, 1
z x y, t r q, 0
a b c, d e f, 1
...
i.e., the first sequence of records is [[a, d],[b, e], [c, f]] with label 1.

model:
input_features:
name: x
type: sequence
encoder: embed
reduce_output: null
name: y
type: sequence
encoder: embed
reduce_output: null

combiner:
type: sequence
main_sequence_feature: x
encoder: rnn

output_features:
name: label
type: binary

The following does not:

data:
x,y,label
3.4 5.6 6.9, a b c, 1
3.4 5.6 6.9, a b c, 1
3.4 5.6 6.9, a b c, 1
3.4 5.6 6.9, a b c, 1

model
input_features:
name: x
type: timeseries
encoder: embed
reduce_output: null
name: y
type: sequence
encoder: embed
reduce_output: null

combiner:
type: sequence
main_sequence_feature: x
encoder: rnn

output_features:
name: label
type: binary

I don't really want to embed the time series features, but it seems like i need to specify some kind of encoder. At any rate...it doesn't work. Is there any way to support combining categorical and numerical features together and running through an rnn?

thanks!

[BenMacKenzie]

update:
I created a passthrough encoder which does nothing except reshape the input_sequence and changed line 144 in timeseries_feature.py so that the type would be tf.float32 instead of tf.int32 and I am able to train with the data set referenced above. am not very familiar with code base, so don't really know if this makes sense. in particular, is line 144 a bug?

[w4nderlust]

Thanks for this, I was actually working EXACTLY on this issue locally, I just pushed that in this branch https://github.com/uber/ludwig/tree/timeseries_fix . Please try it out with your data and let me know if it works, as in case i will add it to master immediately.

[w4nderlust]

I actually created a PR from that branch: https://github.com/uber/ludwig/pull/179

[yanisIk]

Out of curiosity, for which type of use case is this data format used ?

I'm asking because I'm looking into doing classification for timeseries data where my columns are numerical and categorical but I can't see how to achieve this with ludwig (the doc only shows the input feature as a sequence of numbers separated by spaces).

Thanks

[BenMacKenzie]

hidden_size = input.sequence.shape[-1] rather than
hidden_size = input.sequence.shape()[-1] on line 72, but otherwise it seems to work!

thanks so much!!

[BenMacKenzie]

Out of curiosity, for which type of use case is this data format used ?

I'm asking because I'm looking into doing classification for timeseries data where my columns are numerical and categorical but I can't see how to achieve this with ludwig (the doc only shows the input feature as a sequence of numbers separated by spaces).

Thanks

it's a pattern that we use for credit card or other transaction-oriented fraud detection

[w4nderlust]

@yanisIk sequnce is a sequence of categories while timeseries is a sequence of numbers, you can combine them with the sequence combiner as long ad they are of the same length. Read the documentation here.

@BenMacKenzie i'm updating and merging the PR. Thank you! Also closing the thread,