머신러닝 엔지니어 실무-Section1

Smiling Sammy·2022년 1월 25일

mlops 머신러닝 엔지니어 실무

머신러닝 엔지니어 실무

목록 보기

2/5

인프런 Chris Song님 강의 정리

Section1: 머신러닝 프로젝트 실험관리

Lecture

실험관리

실험관리가 필요한 이유

어떤 데이터로 학습시켰는지 까먹음
가장 성능 좋은 모델이 무엇인지 파악해야됨
- Metric에 따른 기록 및 navigation을 해야됨

Weights and Biases

실험관리 기능을 기본으로 탑재
다양한 python framework에 적용 가능
- Tensorflow, Pytorch, Keras, Scikit-learn ..
GPU 사용률 파악
실시간으로 퍼포먼스 디버깅을 할 수 있음

실습

실습 코드

import random

import numpy as np
import tensorflow as tf
from wandb.keras import WandbCallback

# Simple Keras Model

# Launch 20 experiments, trying different dropout rates
for run in range(20):
  # Start a run, tracking hyperparameters
  wandb.init(
      project="keras-intro",
      # Set entity to specify your username or team name
      # ex: entity="wandb",
      config={
          "layer_1": 512,
          "activation_1": "relu",
          "dropout": random.uniform(0.01, 0.80),
          "layer_2": 10,
          "activation_2": "softmax",
          "optimizer": "sgd",
          "loss": "sparse_categorical_crossentropy",
          "metric": "accuracy",
          "epoch": 6,
          "batch_size": 256
      })
  config = wandb.config

  # Get the data
  mnist = tf.keras.datasets.mnist
  (x_train, y_train), (x_test, y_test) = mnist.load_data()
  x_train, x_test = x_train / 255.0, x_test / 255.0
  x_train, y_train = x_train[::5], y_train[::5]  # Subset data for a faster demo
  x_test, y_test = x_test[::20], y_test[::20]
  labels = [str(digit) for digit in range(np.max(y_train) + 1)]

  # Build a model
  model = tf.keras.models.Sequential([
      tf.keras.layers.Flatten(input_shape=(28, 28)),
      tf.keras.layers.Dense(config.layer_1, activation=config.activation_1),
      tf.keras.layers.Dropout(config.dropout),
      tf.keras.layers.Dense(config.layer_2, activation=config.activation_2)
      ])

  model.compile(optimizer=config.optimizer,
                loss=config.loss,
                metrics=[config.metric]
                )

  # WandbCallback auto-saves all metrics from model.fit(), plus predictions on validation_data
  logging_callback = WandbCallback(log_evaluation=True)

  history = model.fit(x=x_train, y=y_train,
                      epochs=config.epoch,
                      batch_size=config.batch_size,
                      validation_data=(x_test, y_test),
                      callbacks=[logging_callback]
                      )

  # Mark the run as finished
  wandb.finish()

하이퍼 파라미터 최적화

Hyper Parameter Optimization(HPO)
모델 성능 개선에 도움이 됨
머신러닝의 반복업무이므로 자동화를 해야함

W&B Sweeps

Weights and Biases의 하이퍼 파라미터를 자동으로 최적화해주는 도구

실습1

# %load train_lib.py

def train():
    import numpy as np
    import tensorflow as tf
    import wandb
    config_defaults = {
        'layers': 128
    }
    wandb.init(config=config_defaults, magic=True)

    fashion_mnist = tf.keras.datasets.fashion_mnist
    (train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()
    class_names = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat',
                   'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']

    train_images.shape
    train_images = train_images / 255.0
    test_images = test_images / 255.0

    model = tf.keras.Sequential([
        tf.keras.layers.Flatten(input_shape=(28, 28)),
        tf.keras.layers.Dense(wandb.config.layers, activation=tf.nn.relu),
        tf.keras.layers.Dense(10, activation=tf.nn.softmax)
    ])

    model.compile(optimizer='adam',
                  loss='sparse_categorical_crossentropy',
                  metrics=['accuracy'])
    
    model.fit(train_images, train_labels, epochs=5,
                  validation_data=(test_images, test_labels))

sweep_config = {
    'method': 'grid',
    'parameters': {
        'layers': {
            'values': [32, 64, 96, 128, 256]
        }
    }
}

import wandb
sweep_id = wandb.sweep(sweep_config)
wandb.agent(sweep_id, function=train)

실습2

# %load train_lib.py

def train():
    import numpy as np
    import tensorflow as tf
    import wandb

    config_defaults = {
        'layer1_size': 128,
        'layer1_activation': 'relu',
        'dropout_rate': 0.2,
        'optimizer': 'adam',
        'learning_rate': 0.01        
    }

    wandb.init(
        project='sweep-practice',
        config=config_defaults, 
        magic=True)

    config = wandb.config

    fashion_mnist = tf.keras.datasets.fashion_mnist
    (train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()
    class_names = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat',
                   'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']

    train_images.shape
    train_images = train_images / 255.0
    test_images = test_images / 255.0

    model = tf.keras.Sequential([
        tf.keras.layers.Flatten(input_shape=(28, 28)),
        tf.keras.layers.Dense(config.layer1_size, activation=config.layer1_activation),
        tf.keras.layers.Dropout(config.dropout_rate),
        tf.keras.layers.Dense(10, activation=tf.nn.softmax)
    ])

    if config.optimizer == "rmsprop":
      opt = tf.keras.optimizers.RMSprop(learning_rate=config.learning_rate)
    else:
      opt = tf.keras.optimizers.Adam(learning_rate=config.learning_rate)

    model.compile(optimizer=opt,
                  loss='sparse_categorical_crossentropy',
                  metrics=['accuracy'])
    
    model.fit(train_images, train_labels, epochs=5,
                  validation_data=(test_images, test_labels))

sweep_config = {
    'method': 'grid',
    'parameters': {
        'layer1_size': {
            'values': [32, 64, 96, 128]
        },
        'layer1_activation': {
            'values': ['relu', 'sigmoid']
        },
        'dropout_rate': {
            'values': [0.1, 0.2, 0.3, 0.4, 0.5]
        },
        'optimizer': {
            'values': ['adam', 'rmsprop']
        },
        'learning_rate': {
            'values': [0.1, 0.01, 0.001]
        }        
    }
}

sweep_id = wandb.sweep(sweep_config, project='sweep-practice')
wandb.agent(sweep_id, function=train)