TensorFlow: ML Framework Google Open-source Research/Production ใช้ NumPy และ Multi-dimensional arrays (Tensors) รองรับ Deep Neural Networks
Keras: Deep Learning API Python จากโครงการ ONEIROS รันบน TensorFlow Scalable, Cross-platform เข้าถึงด้วย from tensorflow import keras
Functional API
ยืดหยุ่นกว่า Sequential API รองรับ Non-linear topology, Shared layers, Multiple I/O สร้าง DAG ของเลเยอร์
ตัวอย่าง: Embed Text เป็น Vector แล้วประมวลผลด้วย LSTM
print("Number of unique issue tags")
num_tags = 12
print("Size of vocabulary while preprocessing text data")
num_words = 10000
print("Number of classes for predictions")
num_classes = 4
title_input = keras.Input(shape=(None,), name="title")
print("Variable length int sequence")
body_input = keras.Input(shape=(None,), name="body")
tags_input = keras.Input(shape=(num_tags,), name="tags")
print("Embed every word in the title to a 64-dimensional vector")
title_features = layers.Embedding(num_words, 64)(title_input)
print("Embed every word into a 64-dimensional vector")
body_features = layers.Embedding(num_words, 64)(body_input)
print("Reduce sequence of embedded words into single 128-dimensional vector")
title_features = layers.LSTM(128)(title_features)
print("Reduce sequence of embedded words into single 132-dimensional vector")
body_features = layers.LSTM(32)(body_features)
print("Merge available features into a single vector by concatenating it")
x = layers.concatenate([title_features, body_features, tags_input])
print("Use logistic regression to predict the features")
priority_pred = layers.Dense(1, name="priority")(x)
department_pred = layers.Dense(num_classes, name="class")(x)
print("Instantiate a model that predicts priority and class")
model = keras.Model(
inputs=[title_input, body_input, tags_input],
outputs=[priority_pred, department_pred],
)
Code credit: TensorFlow Keras Functional Guide
ผลลัพธ์
Number of unique issue tags Size of vocabulary while preprocessing text data Number of classes for predictions Variable length int sequence Embed every word in the title to a 64-dimensional vector Embed every word into a 64-dimensional vector Reduce sequence of embedded words into single 128-dimensional vector Reduce sequence of embedded words into single 132-dimensional vector Merge available features into a single vector by concatenating it Use logistic regression to predict the features Instantiate a model that predicts priority and class
คำอธิบาย
- Functional API รองรับ Multiple Inputs/Outputs (Sequential API ทำไม่ได้)
- Embedding Layer แปลงคำเป็น Dense Vector
- LSTM ลด Sequence ของ Vector เป็น Single Vector
- Concatenate รวม Features จากหลายแหล่ง
- Dense Layer สำหรับ Prediction
