Skip to content

fix for np.float32 serialization #589

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 8 commits into from
Apr 7, 2023
Merged

fix for np.float32 serialization #589

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
885cc59
Added truncation and padding to tokenizer
santiatpml Mar 22, 2023
6a38984
Added truncation and padding to tokenizer
santiatpml Mar 22, 2023
0f692f1
np.float32 serialization fix
santiatpml Apr 6, 2023
2521e26
Updated tests with HF pipelines
santiatpml Apr 6, 2023
a8d1b9c
Merge branch 'santi-llm-apis' of https://github.com/postgresml/postgr…
santiatpml Apr 6, 2023
a9a3805
Removed transforms from tests
santiatpml Apr 7, 2023
de68ce6
Added protobuf to dockerfile local for finbert support
santiatpml Apr 7, 2023
e8bc2bd
Removed protobuf
santiatpml Apr 7, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
90 changes: 90 additions & 0 deletions pgml-extension/examples/finetune.sql
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,90 @@
-- Exit on error (psql)
\set ON_ERROR_STOP true
\timing on


SELECT pgml.load_dataset('kde4', kwargs => '{"lang1": "en", "lang2": "es"}');
CREATE OR REPLACE VIEW kde4_en_to_es AS
SELECT translation->>'en' AS "en", translation->>'es' AS "es"
FROM pgml.kde4
LIMIT 10;
SELECT pgml.tune(
'Translate English to Spanish',
task => 'translation',
relation_name => 'kde4_en_to_es',
y_column_name => 'es', -- translate into spanish
model_name => 'Helsinki-NLP/opus-mt-en-es',
hyperparams => '{
"learning_rate": 2e-5,
"per_device_train_batch_size": 16,
"per_device_eval_batch_size": 16,
"num_train_epochs": 1,
"weight_decay": 0.01,
"max_length": 128
}',
test_size => 0.5,
test_sampling => 'last'
);

SELECT pgml.load_dataset('imdb');
SELECT pgml.tune(
'IMDB Review Sentiment',
task => 'text-classification',
relation_name => 'pgml.imdb',
y_column_name => 'label',
model_name => 'distilbert-base-uncased',
hyperparams => '{
"learning_rate": 2e-5,
"per_device_train_batch_size": 16,
"per_device_eval_batch_size": 16,
"num_train_epochs": 1,
"weight_decay": 0.01
}',
test_size => 0.5,
test_sampling => 'last'
);
SELECT pgml.predict('IMDB Review Sentiment', 'I love SQL');

SELECT pgml.load_dataset('squad_v2');
SELECT pgml.tune(
'SQuAD Q&A v2',
'question-answering',
'pgml.squad_v2',
'answers',
'deepset/roberta-base-squad2',
hyperparams => '{
"evaluation_strategy": "epoch",
"learning_rate": 2e-5,
"per_device_train_batch_size": 16,
"per_device_eval_batch_size": 16,
"num_train_epochs": 1,
"weight_decay": 0.01,
"max_length": 384,
"stride": 128
}',
test_size => 11873,
test_sampling => 'last'
);


SELECT pgml.load_dataset('billsum', kwargs => '{"split": "ca_test"}');
CREATE OR REPLACE VIEW billsum_training_data
AS SELECT title || '\n' || text AS text, summary FROM pgml.billsum;
SELECT pgml.tune(
'Legal Summarization',
task => 'summarization',
relation_name => 'billsum_training_data',
y_column_name => 'summary',
model_name => 'sshleifer/distilbart-xsum-12-1',
hyperparams => '{
"learning_rate": 2e-5,
"per_device_train_batch_size": 2,
"per_device_eval_batch_size": 2,
"num_train_epochs": 1,
"weight_decay": 0.01,
"max_input_length": 1024,
"max_summary_length": 128
}',
test_size => 0.01,
test_sampling => 'last'
);
129 changes: 50 additions & 79 deletions pgml-extension/examples/transformers.sql
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,89 +32,60 @@ SELECT pgml.transform(
'Dominic Cobb is the foremost practitioner of the artistic science of extraction, inserting oneself into a subject''s dreams to obtain hidden information without the subject knowing, a concept taught to him by his professor father-in-law, Dr. Stephen Miles. Dom''s associates are Miles'' former students, who Dom requires as he has given up being the dream architect for reasons he won''t disclose. Dom''s primary associate, Arthur, believes it has something to do with Dom''s deceased wife, Mal, who often figures prominently and violently in those dreams, or Dom''s want to "go home" (get back to his own reality, which includes two young children). Dom''s work is generally in corporate espionage. As the subjects don''t want the information to get into the wrong hands, the clients have zero tolerance for failure. Dom is also a wanted man, as many of his past subjects have learned what Dom has done to them. One of those subjects, Mr. Saito, offers Dom a job he can''t refuse: to take the concept one step further into inception, namely planting thoughts into the subject''s dreams without them knowing. Inception can fundamentally alter that person as a being. Saito''s target is Robert Michael Fischer, the heir to an energy business empire, which has the potential to rule the world if continued on the current trajectory. Beyond the complex logistics of the dream architecture of the case and some unknowns concerning Fischer, the biggest obstacles in success for the team become worrying about one aspect of inception which Cobb fails to disclose to the other team members prior to the job, and Cobb''s newest associate Ariadne''s belief that Cobb''s own subconscious, especially as it relates to Mal, may be taking over what happens in the dreams.'
]
);
SELECT pgml.transform(
inputs => ARRAY[
'I love how amazingly simple ML has become!',
'I hate doing mundane and thankless tasks. ☹️'
],
task => '{"task": "text-classification",
"model": "finiteautomata/bertweet-base-sentiment-analysis"
}'::JSONB
) AS positivity;

SELECT pgml.load_dataset('kde4', kwargs => '{"lang1": "en", "lang2": "es"}');
CREATE OR REPLACE VIEW kde4_en_to_es AS
SELECT translation->>'en' AS "en", translation->>'es' AS "es"
FROM pgml.kde4
LIMIT 10;
SELECT pgml.tune(
'Translate English to Spanish',
task => 'translation',
relation_name => 'kde4_en_to_es',
y_column_name => 'es', -- translate into spanish
model_name => 'Helsinki-NLP/opus-mt-en-es',
hyperparams => '{
"learning_rate": 2e-5,
"per_device_train_batch_size": 16,
"per_device_eval_batch_size": 16,
"num_train_epochs": 1,
"weight_decay": 0.01,
"max_length": 128
}',
test_size => 0.5,
test_sampling => 'last'
);
SELECT pgml.transform(
task => 'text-classification',
inputs => ARRAY[
'I love how amazingly simple ML has become!',
'I hate doing mundane and thankless tasks. ☹️'
]
) AS positivity;

SELECT pgml.transform(
inputs => ARRAY[
'Stocks rallied and the British pound gained.',
'Stocks making the biggest moves midday: Nvidia, Palantir and more'
],
task => '{"task": "text-classification",
"model": "ProsusAI/finbert"
}'::JSONB
) AS market_sentiment;

SELECT pgml.load_dataset('imdb');
SELECT pgml.tune(
'IMDB Review Sentiment',
task => 'text-classification',
relation_name => 'pgml.imdb',
y_column_name => 'label',
model_name => 'distilbert-base-uncased',
hyperparams => '{
"learning_rate": 2e-5,
"per_device_train_batch_size": 16,
"per_device_eval_batch_size": 16,
"num_train_epochs": 1,
"weight_decay": 0.01
}',
test_size => 0.5,
test_sampling => 'last'
SELECT pgml.transform(
inputs => ARRAY[
'I have a problem with my iphone that needs to be resolved asap!!'
],
task => '{"task": "zero-shot-classification",
"model": "roberta-large-mnli"
}'::JSONB,
args => '{"candidate_labels": ["urgent", "not urgent", "phone", "tablet", "computer"]
}'::JSONB
) AS zero_shot;

SELECT pgml.transform(
inputs => ARRAY[
'Hugging Face is a French company based in New York City.'
],
task => 'token-classification'
);
SELECT pgml.predict('IMDB Review Sentiment', 'I love SQL');

SELECT pgml.load_dataset('squad_v2');
SELECT pgml.tune(
'SQuAD Q&A v2',
SELECT pgml.transform(
'question-answering',
'pgml.squad_v2',
'answers',
'deepset/roberta-base-squad2',
hyperparams => '{
"evaluation_strategy": "epoch",
"learning_rate": 2e-5,
"per_device_train_batch_size": 16,
"per_device_eval_batch_size": 16,
"num_train_epochs": 1,
"weight_decay": 0.01,
"max_length": 384,
"stride": 128
}',
test_size => 11873,
test_sampling => 'last'
);
inputs => ARRAY[
'{
"question": "Am I dreaming?",
"context": "I got a good nights sleep last night and started a simple tutorial over my cup of morning coffee. The capabilities seem unreal, compared to what I came to expect from the simple SQL standard I studied so long ago. The answer is staring me in the face, and I feel the uncanny call from beyond the screen to check the results."
}'
]
) AS answer;


SELECT pgml.load_dataset('billsum', kwargs => '{"split": "ca_test"}');
CREATE OR REPLACE VIEW billsum_training_data
AS SELECT title || '\n' || text AS text, summary FROM pgml.billsum;
SELECT pgml.tune(
'Legal Summarization',
task => 'summarization',
relation_name => 'billsum_training_data',
y_column_name => 'summary',
model_name => 'sshleifer/distilbart-xsum-12-1',
hyperparams => '{
"learning_rate": 2e-5,
"per_device_train_batch_size": 2,
"per_device_eval_batch_size": 2,
"num_train_epochs": 1,
"weight_decay": 0.01,
"max_input_length": 1024,
"max_summary_length": 128
}',
test_size => 0.01,
test_sampling => 'last'
);
12 changes: 9 additions & 3 deletions pgml-extension/src/bindings/transformers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@
import math
import shutil
import time

import numpy as np

import datasets
from rouge import Rouge
Expand Down Expand Up @@ -40,6 +40,12 @@
__cache_transformer_by_model_id = {}
__cache_sentence_transformer_by_name = {}

class NumpyJSONEncoder(json.JSONEncoder):
def default(self, obj):
if isinstance(obj, np.float32):
return float(obj)
return super().default(obj)

def transform(task, args, inputs):
task = json.loads(task)
args = json.loads(args)
Expand All @@ -50,7 +56,7 @@ def transform(task, args, inputs):
if pipe.task == "question-answering":
inputs = [json.loads(input) for input in inputs]

return json.dumps(pipe(inputs, **args))
return json.dumps(pipe(inputs, **args), cls = NumpyJSONEncoder)

def embed(transformer, text, kwargs):
kwargs = json.loads(kwargs)
Expand Down Expand Up @@ -101,7 +107,7 @@ def tokenize_summarization(tokenizer, max_length, x, y):
return datasets.Dataset.from_dict(encoding.data)

def tokenize_text_generation(tokenizer, max_length, y):
encoding = tokenizer(y, max_length=max_length)
encoding = tokenizer(y, max_length=max_length, truncation=True, padding="max_length")
return datasets.Dataset.from_dict(encoding.data)

def tokenize_question_answering(tokenizer, max_length, x, y):
Expand Down
1 change: 1 addition & 0 deletions pgml-extension/tests/test.sql
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,3 +27,4 @@ SELECT pgml.load_dataset('wine');
\i examples/multi_classification.sql
\i examples/regression.sql
\i examples/vectors.sql