r-and-d--try-gpt-bit-pair-100k-encoding

This may make a suitable base model, but may need further preprocessing:
import tensorflow as tf
from minbpe import GPT4Tokenizer 

from keras_nlp.models import GPT2Preprocessor

class TextEncoderLayer(tf.keras.layers.Layer):
    def __init__(self,
                 # tokenizer,
                 sequence_length = 100):
        super(TextEncoderLayer, self).__init__()
        tokenizer = GPT4Tokenizer()
        self.tokenizer = tokenizer
        self.sequence_length = sequence_length

    def call(self, text):
        _tokens = []
        for text_0 in text:
            tokens = self.tokenizer.encode(str(text_0), allowed_special="all")

            _tokens.append(tokens)
        # ragged_tokens = tf.ragged.constant(padded_tokens)
        # token_tensor = tf.constant(_tokens)
        padded_tokens =\
                tf.keras.preprocessing.sequence.pad_sequences(
                 _tokens, maxlen=self.sequence_length, padding='post')

        return tf.constant(padded_tokens) # ragged_tokens

# Usage example
text_1 = tf.constant(["<|endoftext|>hello world"], dtype=tf.string)
text = tf.constant(["<|endoftext|>hello world", "test 9"], dtype=tf.string)
        # tf.constant("<|endoftext|>hello world", dtype=tf.string),
            # tf.constant("test 9", dtype=tf.string)])
# tokenizer = GPT4Tokenizer()
text_encoder_layer = TextEncoderLayer() # tokenizer)

print("2 tensor: as layer:")
print(text_encoder_layer(text))

print("One tensor: as layer:")
print(text_encoder_layer(text_1))

# Check if compatible with preprocessor:

inp = tf.keras.layers.Input(shape=(), dtype=tf.string)
tokens_1 = TextEncoderLayer()(inp)
vocab_size = 100276
embedded = tf.keras.layers.Embedding(
    input_dim=vocab_size,
    output_dim=18,
    input_length=100)(tokens_1)
flat = tf.keras.layers.Flatten()(embedded)

m1 = tf.keras.Model(inputs=inp, outputs=flat)

result_1 = m1(text_1)
print("1 Tensor:")
print(result_1)

result = m1(text)

print("2 tensor:")
print(result)
david-thrower / cerebros-core-algorithm-alpha

r-and-d--try-gpt-bit-pair-100k-encoding #147
