Getting started question

[ianmuscat]

Hi there, I'm super new to ML and Ludwig, so I'm almost certainly missing something. I'm trying to complete an example in the getting started guide, but I'm stuck when it comes to getting predictions to work.

I'm using the epinions-1.csv (CSV file with a class and a text column) as suggested in the getting started guide and I'm running Ludwig using the official Docker container. The following is my configuration.

input_features:
- name: text
  type: text
  level: word
  encoder: parallel_cnn
output_features:
- name: class
  type: category

First off I start training using the experiment command. That seems to work and outputs a model. However when I try to use the following predict command on the same data with the class column removed (same CSV, just a single column), I don't get any CSVs with predictions as the docs seem to indicate I should. prediction.shapes.json contains an empty object and the NPY files also seem empty when loaded with numpy.load.

ludwig predict --dataset /data/epinions-1-text.csv --data_format csv --model_path /data/results/experiment_run/model --output_directory /data/predictions

Any idea what I'm doing wrong? Thanks! 🙏

[w4nderlust]

Hi @ianmuscat ,
from your description it seems like you are doing everything correctly.
I'm curious about the contents of epinions-1-text.csv, can you show the first couple lines?

Also, just for test purposes, you may want ot try one of the several text datasets provided in Ludwig to see if you encoutner issues with those too. https://ludwig-ai.github.io/ludwig-docs/user_guide/datasets/ The Reuters dataset should have a similar structure to yours (name of the columns may be different though).

[ianmuscat]

Hi @w4nderlust, thanks for your reply. Here's a screenshot of epinions-1.csv and epinions-1-text.csv.

I've also tried specifying epinions-1.csv as the dataset for prediction, but to no avail. Something I noticed is that even when using the experiment command, the prediction step is skipped entirely (not shown in the stdout output, unlike training) and all prediction-related files are empty.

I'll give the sample datasets a shot, thanks!

[ianmuscat]

Just tried this with the Reuters dataset (used this URL https://github.com/ludwig-ai/ludwig/blob/master/ludwig/datasets/reuters/config.yaml#L3) with the same file (column names happened to be the same). Same result as before -- when running experiment, training works, but predictions seem to be skipped (I see a number of warnings, but they seem GPU related?).

Running ludwig experiment

Training completes, but it seems like the predictions step is skipped (in some videos I've seen there is a "predicting" heading similar to "training")?

[w4nderlust]

So the evaluation is performed, which should create the predictions as a side effect.
Will look into this. Let e open an issue.

[jimthompson5802]

@ianmuscat Just to let you know I'm looking at the issue you wrote about. @w4nderlust created Issue #1234 to track the problem. Based on the information you provided, it looks like I am able to reproduce the situation you describe.

I suggest you follow that issue for any updates on this problem.