training_nli.py

"""
This examples trains a CrossEncoder for the NLI task. A CrossEncoder takes a sentence pair
as input and outputs a label. Here, it learns to predict the labels: "contradiction": 0, "entailment": 1, "neutral": 2.

It does NOT produce a sentence embedding and does NOT work for individual sentences.

Usage:
python training_nli.py
"""

import csv
import gzip
import logging
import math
import os
from datetime import datetime

from torch.utils.data import DataLoader

from sentence_transformers import LoggingHandler, util
from sentence_transformers.cross_encoder import CrossEncoder
from sentence_transformers.cross_encoder.evaluation import CEF1Evaluator, CESoftmaxAccuracyEvaluator
from sentence_transformers.evaluation import SequentialEvaluator
from sentence_transformers.readers import InputExample

#### Just some code to print debug information to stdout
logging.basicConfig(
    format="%(asctime)s - %(message)s", datefmt="%Y-%m-%d %H:%M:%S", level=logging.INFO, handlers=[LoggingHandler()]
)
logger = logging.getLogger(__name__)
#### /print debug information to stdout


# As dataset, we use SNLI + MultiNLI
# Check if dataset exists. If not, download and extract  it
nli_dataset_path = "datasets/AllNLI.tsv.gz"

if not os.path.exists(nli_dataset_path):
    util.http_get("https://sbert.net/datasets/AllNLI.tsv.gz", nli_dataset_path)


# Read the AllNLI.tsv.gz file and create the training dataset
logger.info("Read AllNLI train dataset")

label2int = {"contradiction": 0, "entailment": 1, "neutral": 2}
train_samples = []
dev_samples = []
with gzip.open(nli_dataset_path, "rt", encoding="utf8") as fIn:
    reader = csv.DictReader(fIn, delimiter="\t", quoting=csv.QUOTE_NONE)
    for row in reader:
        label_id = label2int[row["label"]]
        if row["split"] == "train":
            train_samples.append(InputExample(texts=[row["sentence1"], row["sentence2"]], label=label_id))
        else:
            dev_samples.append(InputExample(texts=[row["sentence1"], row["sentence2"]], label=label_id))


train_batch_size = 16
num_epochs = 4
model_save_path = "output/training_allnli-" + datetime.now().strftime("%Y-%m-%d_%H-%M-%S")

# Define our CrossEncoder model. We use distilroberta-base as basis and setup it up to predict 3 labels
model = CrossEncoder("distilroberta-base", num_labels=len(label2int))

# We wrap train_samples, which is a list of InputExample, in a pytorch DataLoader
train_dataloader = DataLoader(train_samples, shuffle=True, batch_size=train_batch_size)

# During training, we use CESoftmaxAccuracyEvaluator and CEF1Evaluator to measure the performance on the dev set
accuracy_evaluator = CESoftmaxAccuracyEvaluator.from_input_examples(dev_samples, name="AllNLI-dev")
f1_evaluator = CEF1Evaluator.from_input_examples(dev_samples, name="AllNLI-dev")
evaluator = SequentialEvaluator([accuracy_evaluator, f1_evaluator])

warmup_steps = math.ceil(len(train_dataloader) * num_epochs * 0.1)  # 10% of train data for warm-up
logger.info("Warmup-steps: {}".format(warmup_steps))


# Train the model
model.fit(
    train_dataloader=train_dataloader,
    evaluator=evaluator,
    epochs=num_epochs,
    evaluation_steps=10000,
    warmup_steps=warmup_steps,
    output_path=model_save_path,
)