Se usó la API de Cloud Translation para traducir esta página.
Switch to English

Entrene y sirva un modelo de TensorFlow con TensorFlow Serving

Esta guía entrena un modelo de red neuronal para clasificar imágenes de ropa, como zapatillas y camisetas , guarda el modelo entrenado y luego lo muestra con TensorFlow Serving . La atención se centra en TensorFlow Serving, en lugar del modelado y el entrenamiento en TensorFlow, por lo que para obtener un ejemplo completo que se centra en el modelado y el entrenamiento, consulte el ejemplo de Clasificación básica .

Esta guía usa tf.keras , una API de alto nivel para crear y entrenar modelos en TensorFlow.

import sys

# Confirm that we're using Python 3
assert sys.version_info.major is 3, 'Oops, not running Python 3. Use Runtime > Change runtime type'
# TensorFlow and tf.keras
print("Installing dependencies for Colab environment")
!pip install -Uq grpcio==1.26.0

import tensorflow as tf
from tensorflow import keras

# Helper libraries
import numpy as np
import matplotlib.pyplot as plt
import os
import subprocess

print('TensorFlow version: {}'.format(tf.__version__))
Installing dependencies for Colab environment
[K     |████████████████████████████████| 2.4MB 4.6MB/s 
[?25hInstalling TensorFlow
TensorFlow 2.x selected.
TensorFlow version: 2.1.0-rc1

Crea tu modelo

Importar el conjunto de datos Fashion MNIST

Esta guía utiliza el conjunto de datos Fashion MNIST que contiene 70.000 imágenes en escala de grises en 10 categorías. Las imágenes muestran prendas de vestir individuales a baja resolución (28 por 28 píxeles), como se ve aquí:

Sprite de moda MNIST
Figura 1. Muestras de moda-MNIST (por Zalando, licencia del MIT).

Fashion MNIST está pensado como un reemplazo directo del clásico conjunto de datos MNIST , que a menudo se usa como el "Hola, mundo" de los programas de aprendizaje automático para la visión por computadora. Puede acceder a Fashion MNIST directamente desde TensorFlow, solo importe y cargue los datos.

fashion_mnist = keras.datasets.fashion_mnist
(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()

# scale the values to 0.0 to 1.0
train_images = train_images / 255.0
test_images = test_images / 255.0

# reshape for feeding into the model
train_images = train_images.reshape(train_images.shape[0], 28, 28, 1)
test_images = test_images.reshape(test_images.shape[0], 28, 28, 1)

class_names = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat',
               'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']

print('\ntrain_images.shape: {}, of {}'.format(train_images.shape, train_images.dtype))
print('test_images.shape: {}, of {}'.format(test_images.shape, test_images.dtype))
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-labels-idx1-ubyte.gz
32768/29515 [=================================] - 0s 0us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-images-idx3-ubyte.gz
26427392/26421880 [==============================] - 0s 0us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-labels-idx1-ubyte.gz
8192/5148 [===============================================] - 0s 0us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-images-idx3-ubyte.gz
4423680/4422102 [==============================] - 0s 0us/step

train_images.shape: (60000, 28, 28, 1), of float64
test_images.shape: (10000, 28, 28, 1), of float64

Entrena y evalúa tu modelo

Usemos la CNN más simple posible, ya que no estamos enfocados en la parte del modelado.

model = keras.Sequential([
  keras.layers.Conv2D(input_shape=(28,28,1), filters=8, kernel_size=3, 
                      strides=2, activation='relu', name='Conv1'),
  keras.layers.Flatten(),
  keras.layers.Dense(10, activation=tf.nn.softmax, name='Softmax')
])
model.summary()

testing = False
epochs = 5

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

test_loss, test_acc = model.evaluate(test_images, test_labels)
print('\nTest accuracy: {}'.format(test_acc))
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
Conv1 (Conv2D)               (None, 13, 13, 8)         80        
_________________________________________________________________
flatten (Flatten)            (None, 1352)              0         
_________________________________________________________________
Softmax (Dense)              (None, 10)                13530     
=================================================================
Total params: 13,610
Trainable params: 13,610
Non-trainable params: 0
_________________________________________________________________
Train on 60000 samples
Epoch 1/5
60000/60000 [==============================] - 11s 185us/sample - loss: 0.5466 - accuracy: 0.8087
Epoch 2/5
60000/60000 [==============================] - 5s 79us/sample - loss: 0.4032 - accuracy: 0.8580
Epoch 3/5
60000/60000 [==============================] - 5s 76us/sample - loss: 0.3613 - accuracy: 0.8712
Epoch 4/5
60000/60000 [==============================] - 5s 75us/sample - loss: 0.3406 - accuracy: 0.8797
Epoch 5/5
60000/60000 [==============================] - 4s 75us/sample - loss: 0.3247 - accuracy: 0.8848
10000/10000 [==============================] - 1s 73us/sample - loss: 0.3510 - accuracy: 0.8747

Test accuracy: 0.8747000098228455

Guarda tu modelo

Para cargar nuestro modelo entrenado en TensorFlow Serving, primero debemos guardarlo en formato SavedModel . Esto creará un archivo protobuf en una jerarquía de directorios bien definida e incluirá un número de versión. TensorFlow Serving nos permite seleccionar qué versión de un modelo o "servible" queremos usar cuando hacemos solicitudes de inferencia. Cada versión se exportará a un subdirectorio diferente en la ruta indicada.

# Fetch the Keras session and save the model
# The signature definition is defined by the input and output tensors,
# and stored with the default serving key
import tempfile

MODEL_DIR = tempfile.gettempdir()
version = 1
export_path = os.path.join(MODEL_DIR, str(version))
print('export_path = {}\n'.format(export_path))

tf.keras.models.save_model(
    model,
    export_path,
    overwrite=True,
    include_optimizer=True,
    save_format=None,
    signatures=None,
    options=None
)

print('\nSaved model:')
!ls -l {export_path}
export_path = /tmp/1

Warning:tensorflow:From /tensorflow-2.1.0/python3.6/tensorflow_core/python/ops/resource_variable_ops.py:1786: calling BaseResourceVariable.__init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version.
Instructions for updating:
If using Keras pass *_constraint arguments to layers.
INFO:tensorflow:Assets written to: /tmp/1/assets

Saved model:
total 84
drwxr-xr-x 2 root root  4096 Jan  7 23:15 assets
-rw-r--r-- 1 root root 74086 Jan  7 23:15 saved_model.pb
drwxr-xr-x 2 root root  4096 Jan  7 23:15 variables

Examina tu modelo guardado

Usaremos la utilidad de línea de comando saved_model_cli para ver MetaGraphDefs (los modelos) y SignatureDefs (los métodos que puede llamar) en nuestro SavedModel. Consulta esta discusión sobre la CLI del modelo guardado en la Guía de TensorFlow.

saved_model_cli show --dir {export_path} --all

MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:

signature_def['__saved_model_init_op']:
  The given SavedModel SignatureDef contains the following input(s):
  The given SavedModel SignatureDef contains the following output(s):
    outputs['__saved_model_init_op'] tensor_info:
        dtype: DT_INVALID
        shape: unknown_rank
        name: NoOp
  Method name is: 

signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['Conv1_input'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 28, 28, 1)
        name: serving_default_Conv1_input:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['Softmax'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 10)
        name: StatefulPartitionedCall:0
  Method name is: tensorflow/serving/predict
WARNING:tensorflow:From /tensorflow-2.1.0/python3.6/tensorflow_core/python/ops/resource_variable_ops.py:1786: calling BaseResourceVariable.__init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version.
Instructions for updating:
If using Keras pass *_constraint arguments to layers.

Defined Functions:
  Function Name: '__call__'
    Option #1
      Callable with:
        Argument #1
          Conv1_input: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='Conv1_input')
        Argument #2
          DType: bool
          Value: True
        Argument #3
          DType: NoneType
          Value: None
    Option #2
      Callable with:
        Argument #1
          inputs: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='inputs')
        Argument #2
          DType: bool
          Value: True
        Argument #3
          DType: NoneType
          Value: None
    Option #3
      Callable with:
        Argument #1
          inputs: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='inputs')
        Argument #2
          DType: bool
          Value: False
        Argument #3
          DType: NoneType
          Value: None
    Option #4
      Callable with:
        Argument #1
          Conv1_input: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='Conv1_input')
        Argument #2
          DType: bool
          Value: False
        Argument #3
          DType: NoneType
          Value: None

  Function Name: '_default_save_signature'
    Option #1
      Callable with:
        Argument #1
          Conv1_input: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='Conv1_input')

  Function Name: 'call_and_return_all_conditional_losses'
    Option #1
      Callable with:
        Argument #1
          inputs: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='inputs')
        Argument #2
          DType: bool
          Value: True
        Argument #3
          DType: NoneType
          Value: None
    Option #2
      Callable with:
        Argument #1
          inputs: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='inputs')
        Argument #2
          DType: bool
          Value: False
        Argument #3
          DType: NoneType
          Value: None
    Option #3
      Callable with:
        Argument #1
          Conv1_input: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='Conv1_input')
        Argument #2
          DType: bool
          Value: False
        Argument #3
          DType: NoneType
          Value: None
    Option #4
      Callable with:
        Argument #1
          Conv1_input: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='Conv1_input')
        Argument #2
          DType: bool
          Value: True
        Argument #3
          DType: NoneType
          Value: None

¡Eso nos dice mucho sobre nuestro modelo! En este caso, acabamos de entrenar nuestro modelo, por lo que ya conocemos las entradas y salidas, pero si no lo hiciéramos, esta sería información importante. No nos dice todo, como el hecho de que se trata de datos de imagen en escala de grises, por ejemplo, pero es un gran comienzo.

Sirva su modelo con TensorFlow Serving

Agrega el URI de distribución de TensorFlow Serving como fuente del paquete:

Nos estamos preparando para instalar TensorFlow Serving con Aptitude, ya que este Colab se ejecuta en un entorno Debian. tensorflow-model-server el tensorflow-model-server a la lista de paquetes que conoce Aptitude. Tenga en cuenta que estamos ejecutando como root.

# This is the same as you would do from your command line, but without the [arch=amd64], and no sudo
# You would instead do:
# echo "deb [arch=amd64] http://storage.googleapis.com/tensorflow-serving-apt stable tensorflow-model-server tensorflow-model-server-universal" | sudo tee /etc/apt/sources.list.d/tensorflow-serving.list && \
# curl https://storage.googleapis.com/tensorflow-serving-apt/tensorflow-serving.release.pub.gpg | sudo apt-key add -

!echo "deb http://storage.googleapis.com/tensorflow-serving-apt stable tensorflow-model-server tensorflow-model-server-universal" | tee /etc/apt/sources.list.d/tensorflow-serving.list && \
curl https://storage.googleapis.com/tensorflow-serving-apt/tensorflow-serving.release.pub.gpg | apt-key add -
!apt update
deb http://storage.googleapis.com/tensorflow-serving-apt stable tensorflow-model-server tensorflow-model-server-universal
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100  2943  100  2943    0     0  11496      0 --:--:-- --:--:-- --:--:-- 11496
OK
Get:1 http://storage.googleapis.com/tensorflow-serving-apt stable InRelease [3,012 B]
Get:2 https://cloud.r-project.org/bin/linux/ubuntu bionic-cran35/ InRelease [3,626 B]
Ign:3 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64  InRelease
Ign:4 https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  InRelease
Hit:5 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64  Release
Get:6 https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  Release [564 B]
Get:7 https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  Release.gpg [833 B]
Hit:8 http://ppa.launchpad.net/graphics-drivers/ppa/ubuntu bionic InRelease
Hit:9 http://archive.ubuntu.com/ubuntu bionic InRelease
Get:10 http://security.ubuntu.com/ubuntu bionic-security InRelease [88.7 kB]
Get:11 http://storage.googleapis.com/tensorflow-serving-apt stable/tensorflow-model-server amd64 Packages [354 B]
Get:12 https://cloud.r-project.org/bin/linux/ubuntu bionic-cran35/ Packages [81.6 kB]
Get:13 http://archive.ubuntu.com/ubuntu bionic-updates InRelease [88.7 kB]
Get:14 http://storage.googleapis.com/tensorflow-serving-apt stable/tensorflow-model-server-universal amd64 Packages [364 B]
Get:15 http://ppa.launchpad.net/marutter/c2d4u3.5/ubuntu bionic InRelease [15.4 kB]
Get:17 https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  Packages [30.4 kB]
Get:18 http://archive.ubuntu.com/ubuntu bionic-backports InRelease [74.6 kB]
Get:19 http://ppa.launchpad.net/marutter/c2d4u3.5/ubuntu bionic/main Sources [1,749 kB]
Get:20 http://security.ubuntu.com/ubuntu bionic-security/universe amd64 Packages [796 kB]
Get:21 http://archive.ubuntu.com/ubuntu bionic-updates/main amd64 Packages [1,073 kB]
Get:22 http://security.ubuntu.com/ubuntu bionic-security/main amd64 Packages [776 kB]
Get:23 http://security.ubuntu.com/ubuntu bionic-security/restricted amd64 Packages [21.3 kB]
Get:24 http://archive.ubuntu.com/ubuntu bionic-updates/multiverse amd64 Packages [10.8 kB]
Get:25 http://archive.ubuntu.com/ubuntu bionic-updates/universe amd64 Packages [1,324 kB]
Get:26 http://archive.ubuntu.com/ubuntu bionic-updates/restricted amd64 Packages [35.5 kB]
Get:27 http://ppa.launchpad.net/marutter/c2d4u3.5/ubuntu bionic/main amd64 Packages [844 kB]
Fetched 7,019 kB in 4s (1,913 kB/s)
Reading package lists... Done
Building dependency tree       
Reading state information... Done
21 packages can be upgraded. Run 'apt list --upgradable' to see them.

Instalar TensorFlow Serving

Esto es todo lo que necesita: ¡una línea de comando!

apt-get install tensorflow-model-server
Reading package lists... Done
Building dependency tree       
Reading state information... Done
The following package was automatically installed and is no longer required:
  libnvidia-common-430
Use 'apt autoremove' to remove it.
The following NEW packages will be installed:
  tensorflow-model-server
0 upgraded, 1 newly installed, 0 to remove and 21 not upgraded.
Need to get 140 MB of archives.
After this operation, 0 B of additional disk space will be used.
Get:1 http://storage.googleapis.com/tensorflow-serving-apt stable/tensorflow-model-server amd64 tensorflow-model-server all 2.0.0 [140 MB]
Fetched 140 MB in 2s (78.8 MB/s)
Selecting previously unselected package tensorflow-model-server.
(Reading database ... 145674 files and directories currently installed.)
Preparing to unpack .../tensorflow-model-server_2.0.0_all.deb ...
Unpacking tensorflow-model-server (2.0.0) ...
Setting up tensorflow-model-server (2.0.0) ...

Comience a ejecutar TensorFlow Serving

Aquí es donde comenzamos a ejecutar TensorFlow Serving y cargamos nuestro modelo. Después de que se cargue, podemos comenzar a realizar solicitudes de inferencia usando REST. Hay algunos parámetros importantes:

  • rest_api_port : el puerto que usará para las solicitudes REST.
  • model_name : lo usará en la URL de las solicitudes REST. Puede ser cualquier cosa.
  • model_base_path : esta es la ruta al directorio donde ha guardado su modelo.
os.environ["MODEL_DIR"] = MODEL_DIR
nohup tensorflow_model_server \
  --rest_api_port=8501 \
  --model_name=fashion_model \
  --model_base_path="${MODEL_DIR}" >server.log 2>&1
Starting job # 0 in a separate thread.

tail server.log
[warn] getaddrinfo: address family for nodename not supported
[evhttp_server.cc : 238] NET_LOG: Entering the event loop ...

Realiza una solicitud a tu modelo en TensorFlow Serving

Primero, echemos un vistazo a un ejemplo aleatorio de nuestros datos de prueba.

def show(idx, title):
  plt.figure()
  plt.imshow(test_images[idx].reshape(28,28))
  plt.axis('off')
  plt.title('\n\n{}'.format(title), fontdict={'size': 16})

import random
rando = random.randint(0,len(test_images)-1)
show(rando, 'An Example Image: {}'.format(class_names[test_labels[rando]]))

png

Ok, eso parece interesante. ¿Qué tan difícil es para ti reconocerlo? Ahora creemos el objeto JSON para un lote de tres solicitudes de inferencia y veamos qué tan bien reconoce nuestro modelo las cosas:

import json
data = json.dumps({"signature_name": "serving_default", "instances": test_images[0:3].tolist()})
print('Data: {} ... {}'.format(data[:50], data[len(data)-52:]))
Data: {"signature_name": "serving_default", "instances": ...  [0.0], [0.0], [0.0], [0.0], [0.0], [0.0], [0.0]]]]}

Realizar solicitudes de REST

Versión más nueva del servidor

Enviaremos una solicitud de predicción como POST al punto final REST de nuestro servidor y le pasaremos tres ejemplos. Le pediremos a nuestro servidor que nos proporcione la última versión de nuestro servidor sin especificar una versión en particular.

!pip install -q requests

import requests
headers = {"content-type": "application/json"}
json_response = requests.post('http://localhost:8501/v1/models/fashion_model:predict', data=data, headers=headers)
predictions = json.loads(json_response.text)['predictions']

show(0, 'The model thought this was a {} (class {}), and it was actually a {} (class {})'.format(
  class_names[np.argmax(predictions[0])], np.argmax(predictions[0]), class_names[test_labels[0]], test_labels[0]))

png

Una versión particular del servidor

Ahora especifiquemos una versión particular de nuestro servidor. Como solo tenemos uno, seleccionemos la versión 1. También veremos los tres resultados.

headers = {"content-type": "application/json"}
json_response = requests.post('http://localhost:8501/v1/models/fashion_model/versions/1:predict', data=data, headers=headers)
predictions = json.loads(json_response.text)['predictions']

for i in range(0,3):
  show(i, 'The model thought this was a {} (class {}), and it was actually a {} (class {})'.format(
    class_names[np.argmax(predictions[i])], np.argmax(predictions[i]), class_names[test_labels[i]], test_labels[i]))

png

png

png