Loading an ML databunch gives attribute error

[harshikanu]

I am getting an error when I load a databunch and run the model with a Celery worker -

“AttributeError: Can’t get attribute ‘RobertaTextList’ on <module ‘celery.bin.celery’ from ‘/usr/local/lib/python3.7/site-packages/celery/bin/celery.py’>”

Do I have to import it in the Celery.py file as well? If yes, how?

I tried importing in celery file, but doesn’t seem to work.

Any pointers will be extremely helpful!

[schmidtbri]

It looks like you're trying to import a class that is not defined anywhere in your codebase. After doing a bit of research I found this blog post that contains a reference to a python class called "RobertaTextList": https://medium.com/analytics-vidhya/using-roberta-with-fastai-for-nlp-7ed3fed21f6c

It looks like the RobertaTextList class was defined in the code for that blog post so you'll need to copy and paste it into your model's codebase to be able to import it. You shouldn't need to modify the celery.py module to get this to work.

I would recommend that you isolate the model code from the task queue code by building a model package that receives a python dictionary and converts it into a "RobertaTextList" object in the model's predict() method. When you install the model into the task queue package and add it to the configuration, you'll interact with the model by sending it dictionaries instead of having to figure out how to serialize and deserialize "RobertaTextList" objects in the task queue (try to avoid this). If you tried to deploy the model as a web service, you might find that a RobertaTextList object doesn't serialize very well to JSON, so try to avoid using data structures that are not easy to transport across the network when deploying ML models.

If you isolate your model code from the task queue code, you'll import the "RobertaTextList" class in the module where the model's MLModel class lives. You'll then convert the "data" parameter into a RobertaTextList object and send it to the model. You should be able to get the model code to work without worrying about the task queue code at all. First get the model code to work in a separate package and then worry about installing the model code into the task queue. The task queue is designed so that running a model inside the task queue should be as easy as installing the model package and adding the model to the configuration.

In the "iris_model" package that is used in the task queue, you'll see that numpy is being imported at the top of the file: https://github.com/schmidtbri/ml-model-abc-improvements/blob/master/iris_model/iris_predict.py#L4 This is the place that you'll likely import the RobertaTextListClass since it is taking the place of the numpy package.

[harshikanu]

Thanks for the reply!

So I had the custom classes seperated from the actual prediction file and I was importing the classes to run it on production. Still doesn't work.

Also, the problem actually occurs when I am trying to load the databunch that is created with the attributes like RobertaTextList.
The databunch is a pickle file.

When I run it without Celery it works perfectly fine. But fails with Celery.
Can you help me with this?

[schmidtbri]

If you post the code and send me the steps you followed to create the issue I can take a look.

[harshikanu]

This is the file I am running. It needs training files to run (which I cant upload since they are large files). But it fails at line 276 where I load the databunch when connected with celery. Otherwise it works perfectly fine (I run it on Docker).

If you uncomment lines 384-388 you will be able to run with the dataset (csv file).

Roberta.zip

[schmidtbri]

Your code doesn't have any classes that derives from the MLModel base class. Read these blog posts and implement your model using the MLModel base class:

• https://medium.com/@brianschmidt_78145/a-simple-ml-model-base-class-ab40e2febf13
• https://medium.com/@brianschmidt_78145/improving-the-mlmodel-base-class-eded137629bd
• https://medium.com/@brianschmidt_78145/using-the-ml-model-base-class-7b984edf47c5

The task queue was designed to run model prediction code, not model training code.

[harshikanu]

I encountered another error now.

It gives me an error at the databunch -

Code -
data = RobertaTextList.from_df(df, “.”, cols=feat_cols, processor=processor)
.split_by_rand_pct(seed=config.seed)
.label_from_df(cols=label_cols, label_cls=CategoryList)
.databunch(bs=config.bs, pad_first=False, pad_idx=0)

RobertaTextList is a custom function -
class RobertaTextList(TextList):
_bunch = RobertaDataBunch
_label_cls = TextList

Error -

Traceback (most recent call last):
File “/usr/local/lib/python3.7/site-packages/celery/app/trace.py”, line 385, in trace_task
R = retval = fun(*args, **kwargs)
File “/usr/local/lib/python3.7/site-packages/celery/app/trace.py”, line 650, in protected_call
return self.run(*args, **kwargs)
File “run_analysis.py”, line 179, in run_alerts
alerts = alerts_roberta(alert_df)
File “/alerts_roberta.py”, line 166, in alerts_roberta
.label_from_df(cols=label_cols, label_cls=CategoryList)
File “/usr/local/lib/python3.7/site-packages/fastai/data_block.py”, line 484, in _inner
self.process()
File “/usr/local/lib/python3.7/site-packages/fastai/data_block.py”, line 538, in process
for ds,n in zip(self.lists, [‘train’,‘valid’,‘test’]): ds.process(xp, yp, name=n)
File “/usr/local/lib/python3.7/site-packages/fastai/data_block.py”, line 718, in process
self.x.process(xp)
File “/usr/local/lib/python3.7/site-packages/fastai/data_block.py”, line 84, in process
for p in self.processor: p.process(self)
File “/usr/local/lib/python3.7/site-packages/fastai/text/data.py”, line 297, in process
tokens += self.tokenizer.process_all(ds.items[i:i+self.chunksize])
File “/usr/local/lib/python3.7/site-packages/fastai/text/transform.py”, line 120, in process_all
return sum(e.map(self._process_all_1, partition_by_cores(texts, self.n_cpus)), [])
File “/usr/local/lib/python3.7/concurrent/futures/process.py”, line 671, in map
timeout=timeout)
File “/usr/local/lib/python3.7/concurrent/futures/_base.py”, line 587, in map
fs = [self.submit(fn, *args) for args in zip(*iterables)]
File “/usr/local/lib/python3.7/concurrent/futures/_base.py”, line 587, in
fs = [self.submit(fn, *args) for args in zip(*iterables)]
File “/usr/local/lib/python3.7/concurrent/futures/process.py”, line 641, in submit
self._start_queue_management_thread()
File “/usr/local/lib/python3.7/concurrent/futures/process.py”, line 583, in _start_queue_management_thread
self._adjust_process_count()
File “/usr/local/lib/python3.7/concurrent/futures/process.py”, line 607, in _adjust_process_count
p.start()
File “/usr/local/lib/python3.7/multiprocessing/process.py”, line 110, in start
‘daemonic processes are not allowed to have children’
AssertionError: daemonic processes are not allowed to have children

Could you help me with this please?

[schmidtbri]

You're not using any of the code in this repository for your project, this is the wrong place to ask this question.