OOFMAN29803/SoloconLM_50M

1import tensorflow as tf2from tensorflow.keras.layers import Embedding, MultiHeadAttention, Dense, Input, Dropout, LayerNormalization3from tensorflow.keras.models import Model, load_model4import numpy as np5import gradio as gr6 7# Define the Positional Encoding layer8class PositionalEncoding(tf.keras.layers.Layer):9    def __init__(self, position, d_model):10        super(PositionalEncoding, self).__init__()11        self.pos_encoding = self.positional_encoding(position, d_model)12 13    def get_angles(self, pos, i, d_model):14        angles = 1 / tf.pow(10000, (2 * (i // 2)) / tf.cast(d_model, tf.float32))15        return pos * angles16 17    def positional_encoding(self, position, d_model):18        angle_rads = self.get_angles(19            tf.range(position, dtype=tf.float32)[:, tf.newaxis],20            tf.range(d_model, dtype=tf.float32)[tf.newaxis, :],21            d_model22        )23        sines = tf.math.sin(angle_rads[:, 0::2])24        cosines = tf.math.cos(angle_rads[:, 1::2])25        pos_encoding = tf.concat([sines, cosines], axis=-1)26        pos_encoding = pos_encoding[tf.newaxis, ...]27        return tf.cast(pos_encoding, tf.float32)28 29    def call(self, inputs):30        return inputs + self.pos_encoding[:, :tf.shape(inputs)[1], :]31 32# Define the Transformer Encoder Layer33class TransformerEncoderLayer(tf.keras.layers.Layer):34    def __init__(self, d_model, num_heads, dff, rate=0.1):35        super().__init__()36        self.mha = MultiHeadAttention(key_dim=d_model, num_heads=num_heads)37        self.ffn = tf.keras.Sequential([38            Dense(dff, activation='relu'),39            Dense(d_model)40        ])41        self.layernorm1 = LayerNormalization(epsilon=1e-6)42        self.layernorm2 = LayerNormalization(epsilon=1e-6)43        self.layernorm3 = LayerNormalization(epsilon=1e-6)44        self.dropout1 = Dropout(rate)45        self.dropout2 = Dropout(rate)46 47    def call(self, x, training, mask=None):48        attn_output = self.mha(x, x, x, attention_mask=mask)49        attn_output = self.dropout1(attn_output, training=training)50        out1 = self.layernorm1(x + attn_output)51        ffn_output = self.ffn(out1)52        ffn_output = self.dropout2(ffn_output, training=training)53        out2 = self.layernorm3(ffn_output)54        return self.layernorm2(out1 + out2)55 56# Define tokenizer function for questions and answers57def qa_tokenizer(text, vocab_size=2000):58    # Dummy implementation, replace with actual tokenization logic59    return [ord(c) % vocab_size for c in text]60 61def preprocess_input(text, max_length=500, vocab_size=2000):62    tokenized_text = qa_tokenizer(text, vocab_size)63    padded_text = tf.keras.preprocessing.sequence.pad_sequences([tokenized_text], maxlen=max_length, padding='post')64    return padded_text65 66def postprocess_output(prediction, vocab_size=2000):67    # Convert the prediction to a sequence of token IDs68    predicted_sequence = np.argmax(prediction, axis=-1)[0]69    # Convert the token IDs to a numeric format70    numeric_sequence = ' '.join(str(id) for id in predicted_sequence)71    return numeric_sequence72 73# Load the trained model with custom objects74custom_objects = {75    'PositionalEncoding': PositionalEncoding,76    'TransformerEncoderLayer': TransformerEncoderLayer77}78model = load_model('Solocon DemoTest.h5', custom_objects=custom_objects)79 80# Function to make predictions81def predict(text):82    input_data = preprocess_input(text)83    predictions = model.predict(input_data)84    output = postprocess_output(predictions)85    return output86 87# Example usage88iface = gr.Interface(fn=predict, inputs="text", outputs="text")89 90# Launch the interface91iface.launch()92