How to parallelize classification with Zero Shot Classification by Huggingface? - python-3.x

Question

I have around 70 categories (it can be 20 or 30 also) and I want to be able to parallelize the process using ray but I get an error:
import pandas as pd
import swifter
import json
import ray
from transformers import pipeline
classifier = pipeline("zero-shot-classification")
labels = ["vegetables", "potato", "bell pepper", "tomato", "onion", "carrot", "broccoli",
"lettuce", "cucumber", "celery", "corn", "garlic", "mashrooms", "cabbage", "spinach",
"beans", "cauliflower", "asparagus", "fruits", "bananas", "apples", "strawberries",
"grapes", "oranges", "lemons", "avocados", "peaches", "blueberries", "pineapple",
"cherries", "pears", "mangoe", "berries", "red meat", "beef", "pork", "mutton",
"veal", "lamb", "venison", "goat", "mince", "white meat", "chicken", "turkey",
"duck", "goose", "pheasant", "rabbit", "Processed meat", "sausages", "bacon",
"ham", "hot dogs", "frankfurters", "tinned meat", "salami", "pâtés", "beef jerky",
"chorizo", "pepperoni", "corned beef", "fish", "catfish", "cod", "pangasius", "pollock",
"tilapia", "tuna", "salmon", "seafood", "shrimp", "squid", "mussels", "scallop",
"octopus", "grains", "rice", "wheat", "bulgur", "corn", "oat", "quinoa", "buckwheat",
"meals", "salad", "soup", "steak", "pizza", "pie", "burger", "backery", "bread", "souce",
"pasta", "sandwich", "waffles", "barbecue", "roll", "wings", "ribs", "cookies"]
ray.init()
#ray.remote
def get_meal_category(seq, labels, n=3):
res_dict = classifier(seq, labels)
return list(zip([seq for i in range(n)], res_dict["labels"][0:n], res_dict["scores"][0:n]))
res_list = ray.get([get_meal_category.remote(merged_df["title"][i], labels) for i in range(10)])
Where merged_df is a big dataframe with meal names in it's labels column like:
['Cappuccino',
'Stove Top Stuffing Mix For Turkey (Kraft)',
'Stove Top Stuffing Mix For Turkey (Kraft)',
'Roasted Dark Turkey Meat',
'Roasted Dark Turkey Meat',
'Roasted Dark Turkey Meat',
'Cappuccino',
'Low Fat 2% Small Curd Cottage Cheese (Daisy)',
'Rice Cereal (Gerber)',
'Oranges']
Please advise how to avoid ray's error and parallelize the classification.
The error:
2021-02-17 16:54:51,689 WARNING worker.py:1107 -- Warning: The remote function __main__.get_meal_category has size 1630925709 when pickled. It will be stored in Redis, which could cause memory issues. This may mean that its definition uses a large array or other object.
---------------------------------------------------------------------------
ConnectionResetError Traceback (most recent call last)
~/.local/lib/python3.8/site-packages/redis/connection.py in send_packed_command(self, command, check_health)
705 for item in command:
--> 706 sendall(self._sock, item)
707 except socket.timeout:
~/.local/lib/python3.8/site-packages/redis/_compat.py in sendall(sock, *args, **kwargs)
8 def sendall(sock, *args, **kwargs):
----> 9 return sock.sendall(*args, **kwargs)
10
ConnectionResetError: [Errno 104] Connection reset by peer
During handling of the above exception, another exception occurred:
ConnectionError Traceback (most recent call last)
<ipython-input-9-1a5345832fba> in <module>
----> 1 res_list = ray.get([get_meal_category.remote(merged_df["title"][i], labels) for i in range(10)])
<ipython-input-9-1a5345832fba> in <listcomp>(.0)
----> 1 res_list = ray.get([get_meal_category.remote(merged_df["title"][i], labels) for i in range(10)])
~/.local/lib/python3.8/site-packages/ray/remote_function.py in _remote_proxy(*args, **kwargs)
99 #wraps(function)
100 def _remote_proxy(*args, **kwargs):
--> 101 return self._remote(args=args, kwargs=kwargs)
102
103 self.remote = _remote_proxy
~/.local/lib/python3.8/site-packages/ray/remote_function.py in _remote(self, args, kwargs, num_returns, num_cpus, num_gpus, memory, object_store_memory, accelerator_type, resources, max_retries, placement_group, placement_group_bundle_index, placement_group_capture_child_tasks, override_environment_variables, name)
205
206 self._last_export_session_and_job = worker.current_session_and_job
--> 207 worker.function_actor_manager.export(self)
208
209 kwargs = {} if kwargs is None else kwargs
~/.local/lib/python3.8/site-packages/ray/function_manager.py in export(self, remote_function)
142 key = (b"RemoteFunction:" + self._worker.current_job_id.binary() + b":"
143 + remote_function._function_descriptor.function_id.binary())
--> 144 self._worker.redis_client.hset(
145 key,
146 mapping={
~/.local/lib/python3.8/site-packages/redis/client.py in hset(self, name, key, value, mapping)
3048 items.extend(pair)
3049
-> 3050 return self.execute_command('HSET', name, *items)
3051
3052 def hsetnx(self, name, key, value):
~/.local/lib/python3.8/site-packages/redis/client.py in execute_command(self, *args, **options)
898 conn = self.connection or pool.get_connection(command_name, **options)
899 try:
--> 900 conn.send_command(*args)
901 return self.parse_response(conn, command_name, **options)
902 except (ConnectionError, TimeoutError) as e:
~/.local/lib/python3.8/site-packages/redis/connection.py in send_command(self, *args, **kwargs)
723 def send_command(self, *args, **kwargs):
724 "Pack and send a command to the Redis server"
--> 725 self.send_packed_command(self.pack_command(*args),
726 check_health=kwargs.get('check_health', True))
727
~/.local/lib/python3.8/site-packages/redis/connection.py in send_packed_command(self, command, check_health)
715 errno = e.args[0]
716 errmsg = e.args[1]
--> 717 raise ConnectionError("Error %s while writing to socket. %s." %
718 (errno, errmsg))
719 except BaseException:
ConnectionError: Error 104 while writing to socket. Connection reset by peer.

Answer 1

This error is happening because of sending large objects to redis. merged_df is a large dataframe and since you are calling get_meal_category 10 times, Ray will attempt to serialize merged_df 10 times. Instead if you put merged_df into the Ray object store just once, and then pass along a reference to the object, this should work.
EDIT: Since the classifier is also large, do something similar for that as well.
Can you try something like this:
ray.init()
df_ref = ray.put(merged_df)
model_ref = ray.put(classifier)
#ray.remote
def get_meal_category(classifier, df, i, labels, n=3):
seq = df["title"][i]
res_dict = classifier(seq, labels)
return list(zip([seq for i in range(n)], res_dict["labels"][0:n], res_dict["scores"][0:n]))
res_list = ray.get([get_meal_category.remote(model_ref, df_ref, i, labels) for i in range(10)])