Add Timeseries forecasting for column-major data, and introduce Timeseries output feature #3212
Conversation
for more information, see https://pre-commit.ci
tgaddair
changed the title
for more information, see https://pre-commit.ci
| next_series[feature.column] = pd.Series(preds[key].iloc[-1]) | ||
|
|
||
| next_preds = pd.DataFrame(next_series) | ||
| dataset = pd.concat([dataset, next_preds], axis=0).reset_index(drop=True) |
There was a problem hiding this comment.
this appraoch is totally fine, but obviously we are preprocessing over and over again datapoints we have already processed, so there's something to be optimzied.
Also from a model perspective, I belive HF does something smart for caching activations for the next steps of text generation, which is kinda similar here potentially, so we can look into that roo.
There was a problem hiding this comment.
Nice, yeah I think we should definitely think about improvements here. But it definitely depends on how far back / forward we want to forecast. If we're looking at +/- 100 samples this should be fine.
| @@ -146,7 +148,7 @@ def input_shape(self): | |||
| return self.dense.input_shape | |||
|
|
|||
| def forward(self, inputs, **kwargs): | |||
| values = self.activation(self.dense(inputs)) | |||
| values = self.activation(self.dense(inputs)) * self.multiplier | |||
There was a problem hiding this comment.
I actually forgot we already had the activation! that was probably for doing softmax / sigmoid for vector features :)
ludwig/encoders/sequence_encoders.py
Outdated
| while len(input_sequence.shape) < 3: | ||
| input_sequence = input_sequence.unsqueeze(-1) | ||
| hidden = self.reduce_sequence(input_sequence) | ||
|
|
||
| # Output may have [batch_size, s, 1] shape, so ensure it is [batch_size, s] | ||
| hidden = hidden.reshape((batch_size, -1)) |
There was a problem hiding this comment.
not sure what we decided for sequence and text (probalby to remove the passthrough encoder) but this may not work in some scnarios (happy to elaborate)
There was a problem hiding this comment.
Hmm, seems so from the tests. Our docs are misleading, then. What do you think is a good workaround in this case?
for more information, see https://pre-commit.ci
|
I'm trying to use the example provided in examples/forecasting which uses 'timeseries' as the output feature type. However, I encounter a ConfigValidationError indicating that 'timeseries' is not a supported output feature type. Here is the error message: ludwig.error.ConfigValidationError: Output feature Seattle_next uses an invalid/unsupported output type 'timeseries'. Supported output features: ['binary', 'category', 'number', 'set', 'vector', 'sequence', 'text']. I cloned the Ludwig repository and installed it in a Python 3.8 environment. As this is one of the provided examples, I expect it to run without any configuration errors. The 'timeseries' output feature should be supported, according to the example. I'd appreciate guidance on how to resolve this issue. Could you confirm whether 'timeseries' is a supported output feature in the current Ludwig version? If it is, what am I doing wrong? Thank you for your assistance. |
|
For anyone else experiencing the same issue in the future, the solution was to clone the current repository state (v0.8dev) instead of the current release. |
|
Glad you got it working @MihailMiller! v0.8 should be released in the next couple of weeks with this feature officially supported. |
Follow-up to the work done in #1131, but for PyTorch instead of TensorFlow.
This PR adds the following:
preprocessing.window_sizeto form the row-major timeseries input.preprocessing.horizonis set.LudwigModel.forecastandludwig forecastcommand that allows forecasting with a trained model.passthroughanddenseencoders with timeseries inputs of constant length.huberloss as an alternative to MAE / MSE (default for timeseries output).mean_absolute_percentage_error(MAPE) loss and metricmultiplierto apply in theprojectordecoder to the output after activation.Example:
ludwig train --config config.yaml --dataset temperature.csvludwig forecast -n 10 --model_path results/experiment_run/model --dataset temperature.csv