Aureliano/electra-if

ELECTRA for IF

ELECTRA is a method for self-supervised language representation learning. They are trained to distinguish "real" input tokens vs "fake" input tokens generated by another neural network, similar to the discriminator of a GAN.

For a detailed description and experimental results, please refer to the original paper ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators.

This repository contains a small ELECTRA discriminator finetuned on a corpus of interactive fiction commands labelled with the WordNet synset offset of the verb in the sentence. The original dataset has been collected from the list of action in the walkthroughs for the game included in the Jericho framework and manually annotated. For more information visit https://github.com/aporporato/electra and https://github.com/aporporato/jericho-corpora.

How to use the discriminator in transformers

(Heavily based on: https://github.com/huggingface/notebooks/blob/master/examples/text_classification-tf.ipynb)

import math
import numpy as np

import tensorflow as tf
from datasets import load_metric, Dataset, DatasetDict
from transformers import TFAutoModelForSequenceClassification, AutoTokenizer, DataCollatorWithPadding, create_optimizer
from transformers.keras_callbacks import KerasMetricCallback

# This example shows how this model can be used:
#  you should finetune the model of your specific corpus if commands, bigger than this
dict_train = {
    "idx": ["0", "1", "2", "3", "4", "5", "6", "7", "8", "9", "10", "11", "12", "13", "14", "15", "16", "17", "18",
            "19", "20"],
    "sentence": ["e", "get pen", "drop book", "x paper", "i", "south", "get paper", "drop the pen", "x book",
                 "inventory", "n", "get the book", "drop paper", "look at Pen", "inv", "g", "s", "get sandwich",
                 "drop sandwich", "x sandwich", "agin"],
    "label": ["travel.v.01", "take.v.04", "drop.v.01", "examine.v.02", "inventory.v.01", "travel.v.01", "take.v.04",
              "drop.v.01", "examine.v.02", "inventory.v.01", "travel.v.01", "take.v.04", "drop.v.01", "examine.v.02",
              "inventory.v.01", "repeat.v.01", "travel.v.01", "take.v.04", "drop.v.01", "examine.v.02", "repeat.v.01"]
}
dict_val = {
    "idx": ["0", "1", "2", "3", "4", "5"],
    "sentence": ["w", "get shield", "drop sword", "x spikes", "i", "repeat"],
    "label": ["travel.v.01", "take.v.04", "drop.v.01", "examine.v.02", "inventory.v.01", "repeat.v.01"]
}

raw_train_dataset = Dataset.from_dict(dict_train)
raw_val_dataset = Dataset.from_dict(dict_val)
raw_dataset = DatasetDict()
raw_dataset["train"] = raw_train_dataset
raw_dataset["val"] = raw_val_dataset
raw_dataset = raw_dataset.class_encode_column("label")
print(raw_dataset)
print(raw_dataset["train"].features)
print(raw_dataset["val"].features)
print(raw_dataset["train"][1])
label2id = {}
id2label = {}
for i, l in enumerate(raw_dataset["train"].features["label"].names):
    label2id[l] = i
    id2label[i] = l

discriminator = TFAutoModelForSequenceClassification.from_pretrained("Aureliano/electra-if",
                                                                     label2id=label2id,
                                                                     id2label=id2label)
tokenizer = AutoTokenizer.from_pretrained("Aureliano/electra-if")

tokenize_function = lambda example: tokenizer(example["sentence"], truncation=True)

pre_tokenizer_columns = set(raw_dataset["train"].features)
encoded_dataset = raw_dataset.map(tokenize_function, batched=True)
tokenizer_columns = list(set(encoded_dataset["train"].features) - pre_tokenizer_columns)

data_collator = DataCollatorWithPadding(tokenizer=tokenizer, return_tensors="tf")

batch_size = len(encoded_dataset["train"])
tf_train_dataset = encoded_dataset["train"].to_tf_dataset(
    columns=tokenizer_columns,
    label_cols=["labels"],
    shuffle=True,
    batch_size=batch_size,
    collate_fn=data_collator
)
tf_validation_dataset = encoded_dataset["val"].to_tf_dataset(
    columns=tokenizer_columns,
    label_cols=["labels"],
    shuffle=False,
    batch_size=batch_size,
    collate_fn=data_collator
)

loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
num_epochs = 25
batches_per_epoch = math.ceil(len(encoded_dataset["train"]) / batch_size)
total_train_steps = int(batches_per_epoch * num_epochs)

optimizer, schedule = create_optimizer(
    init_lr=5e-5, num_warmup_steps=total_train_steps // 5, num_train_steps=total_train_steps
)

metric = load_metric("accuracy")


def compute_metrics(eval_predictions):
    logits, labels = eval_predictions
    predictions = np.argmax(logits, axis=-1)
    return metric.compute(predictions=predictions, references=labels)


metric_callback = KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=tf_validation_dataset)
callbacks = [metric_callback]

discriminator.compile(optimizer=optimizer, loss=loss, metrics=["sparse_categorical_accuracy"])
discriminator.fit(
    tf_train_dataset,
    epochs=num_epochs,
    validation_data=tf_validation_dataset,
    callbacks=callbacks
)

print("Evaluate on test data")
results = discriminator.evaluate(tf_validation_dataset)
print("test loss, test acc:", results)

text = "i"
encoded_input = tokenizer(text, return_tensors='tf')
output = discriminator(encoded_input)
prediction = tf.nn.softmax(output["logits"][0], -1)
label = id2label[tf.math.argmax(prediction).numpy()]
print("\n", text, ":", label,
      "\n")  # ideally 'inventory.v.01' (-> "make or include in an itemized record or report"), but probably only with a better finetuning dataset

text = "get lamp"
encoded_input = tokenizer(text, return_tensors='tf')
output = discriminator(encoded_input)
prediction = tf.nn.softmax(output["logits"][0], -1)
label = id2label[tf.math.argmax(prediction).numpy()]
print("\n", text, ":", label,
      "\n")  # ideally 'take.v.04' (-> "get into one's hands, take physically"), but probably only with a better finetuning dataset

text = "w"
encoded_input = tokenizer(text, return_tensors='tf')
output = discriminator(encoded_input)
prediction = tf.nn.softmax(output["logits"][0], -1)
label = id2label[tf.math.argmax(prediction).numpy()]
print("\n", text, ":", label,
      "\n")  # ideally 'travel.v.01' (-> "change location; move, travel, or proceed, also metaphorically"), but probably only with a better finetuning dataset