Metadata not found when using Rasa NLU from Python

Question

I'm trying to load a model using Python 3.7 and Rasa 1.1.4 in the following way:

from rasa.nlu.model import Interpreter
interpreter = Interpreter.load("./models/generic")

The following error happens:

FileNotFoundError:
[Errno 2] No such file or directory: './models/generic/metadata.json'

The model has been trained with the following command:

rasa train nlu --config config.yml
               --nlu data/generic.md
               --out models
               --fixed-model-name generic/model

Which only produces the following file, no metadata file has been generated:

models/generic/model.tar.gz

What is the best step forward? Generate the metadata file, load the model differently?

Answer 1

You can untar the model.tar.gz to expose the metadata.json and the rest of the model files.

Using the sample model generated from rasa init, if I untar it:

$ ls models 
20190622-213707.tar.gz
$ cd models
$ mkdir 20190622-213707
$ tar xvf 20190622-213707.tar.gz -C 20190622-213707
$ tree 20190622-213707

I'll get:

20190622-213707
├── core
│   ├── ....
├── fingerprint.json
└── nlu
    ├── checkpoint
    ├── component_1_RegexFeaturizer.pkl
    ├── component_4_CountVectorsFeaturizer.pkl
    ├── component_5_EmbeddingIntentClassifier.ckpt.data-00000-of-00001
    ├── component_5_EmbeddingIntentClassifier.ckpt.index
    ├── component_5_EmbeddingIntentClassifier.ckpt.meta
    ├── component_5_EmbeddingIntentClassifier_encoded_all_intents.pkl
    ├── component_5_EmbeddingIntentClassifier_inv_intent_dict.pkl
    ├── metadata.json
    └── training_data.json

...which reveals the metadata.json under the nlu folder.
You can then use the full path to the actual model directory.

interpreter = Interpreter.load("./models/20190622-213707/nlu/")
print(interpreter.parse("hello"))
# {'intent': {'name': 'greet', 'confidence': 0.9470939636230469},  
#  'entities': [], 
#  'intent_ranking': [{'name': 'greet', 'confidence': 0.9470939636230469}, 
#                     {'name': 'deny', 'confidence': 0.17162932455539703}, 
#                     {'name': 'affirm', 'confidence': 0.05398404598236084}, 
#                     {'name': 'mood_great', 'confidence': 0.0}, 
#                     {'name': 'goodbye', 'confidence': 0.0}, 
#                     {'name': 'mood_unhappy', 'confidence': 0.0}], 
#  'text': 'hello'}

---

I am not that familiar with rasa, but you might be mixing command line usage with "manual" Python APIs for writing your own trainer/interpreter apps.

Check out the Trainer class in rasa.nlu.model. It has a persist method that saves your model to a directory, but not as a tar.gz. The result of the Trainer can then be used by the Interpreter.