¡El Día de la Comunidad de ML es el 9 de noviembre! Únase a nosotros para recibir actualizaciones de TensorFlow, JAX, y más Más información

Entrene y sirva un modelo de TensorFlow con TensorFlow Serving

Esta guía forma a un modelo de red neuronal para clasificar imágenes de prendas de vestir, como camisas y zapatillas de deporte , guarda el modelo entrenado, y luego se sirve con TensorFlow Servir . La atención se centra en TensorFlow Servir, en lugar de la modelización y la formación en TensorFlow, por lo que para un ejemplo completo que se centra en el modelado y la formación ver el ejemplo básico de clasificación .

Esta guía utiliza tf.keras , una API de alto nivel para los modelos de construcción y de tren 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__))

Crea tu modelo

Importar el conjunto de datos Fashion MNIST

Esta guía utiliza la moda MNIST conjunto de datos 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. Las muestras de la manera-MNIST (por Zalando, Licencia MIT).

Moda MNIST pretende ser una gota en el reemplazo para el clásico MNIST utilizado conjunto de datos-a menudo como el "Hola, mundo" de los programas de aprendizaje automático para la visión por ordenador. 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 [==============================] - 1s 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, name='Dense')
])
model.summary()

testing = False
epochs = 5

model.compile(optimizer='adam', 
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=[keras.metrics.SparseCategoricalAccuracy()])
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         
_________________________________________________________________
Dense (Dense)                (None, 10)                13530     
=================================================================
Total params: 13,610
Trainable params: 13,610
Non-trainable params: 0
_________________________________________________________________
Epoch 1/5
1875/1875 [==============================] - 10s 2ms/step - loss: 0.5482 - sparse_categorical_accuracy: 0.8089
Epoch 2/5
1875/1875 [==============================] - 3s 2ms/step - loss: 0.4066 - sparse_categorical_accuracy: 0.8566
Epoch 3/5
1875/1875 [==============================] - 3s 2ms/step - loss: 0.3674 - sparse_categorical_accuracy: 0.8698
Epoch 4/5
1875/1875 [==============================] - 3s 2ms/step - loss: 0.3431 - sparse_categorical_accuracy: 0.8790
Epoch 5/5
1875/1875 [==============================] - 3s 2ms/step - loss: 0.3268 - sparse_categorical_accuracy: 0.8832
313/313 [==============================] - 1s 1ms/step - loss: 0.3571 - sparse_categorical_accuracy: 0.8737

Test accuracy: 0.8737000226974487

Guarda tu modelo

Para cargar nuestro modelo entrenado en TensorFlow servir en primer lugar hay que guardarla en SavedModel formato. Esto creará un archivo protobuf en una jerarquía de directorios bien definida e incluirá un número de versión. TensorFlow Porción nos permite seleccionar qué versión de un modelo, o "servable" queremos usar cuando hacemos peticiones 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:absl:Function `_wrapped_model` contains input name(s) Conv1_input with unsupported characters which will be renamed to conv1_input in the SavedModel.
2021-09-30 02:27:59.771453: W tensorflow/python/util/util.cc:348] Sets are not currently considered sequences, but this may change in the future, so consider avoiding using them.
INFO:tensorflow:Assets written to: /tmp/1/assets
INFO:tensorflow:Assets written to: /tmp/1/assets
Saved model:
total 96
drwxr-xr-x 2 kbuilder kbuilder  4096 Sep 30 02:28 assets
-rw-rw-r-- 1 kbuilder kbuilder  7981 Sep 30 02:28 keras_metadata.pb
-rw-rw-r-- 1 kbuilder kbuilder 80657 Sep 30 02:28 saved_model.pb
drwxr-xr-x 2 kbuilder kbuilder  4096 Sep 30 02:28 variables

Examina tu modelo guardado

Vamos a utilizar la utilidad de línea de comandos saved_model_cli mirar las MetaGraphDefs (los modelos) y SignatureDefs (los métodos que se pueden llamar) en nuestra SavedModel. Ver esta discusión de la SavedModel CLI en la Guía 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['Dense'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 10)
        name: StatefulPartitionedCall:0
  Method name is: tensorflow/serving/predict

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: False
        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
          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 #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

  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: True
        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

¡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 Sirviendo usando Aptitud ya que este Colab ejecuta en un entorno de Debian. Vamos a añadir el tensorflow-model-server paquete a la lista de paquetes que Aptitud conoce. Tenga en cuenta que estamos ejecutando como root.

import sys
# We need sudo prefix if not on a Google Colab.
if 'google.colab' not in sys.modules:
  SUDO_IF_NEEDED = 'sudo'
else:
  SUDO_IF_NEEDED = ''
# 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" | {SUDO_IF_NEEDED} tee /etc/apt/sources.list.d/tensorflow-serving.list && \
curl https://storage.googleapis.com/tensorflow-serving-apt/tensorflow-serving.release.pub.gpg | {SUDO_IF_NEEDED} apt-key add -
!{SUDO_IF_NEEDED} 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  14569      0 --:--:-- --:--:-- --:--:-- 14569
OK
Hit:1 http://asia-east1.gce.archive.ubuntu.com/ubuntu bionic InRelease
Get:2 http://asia-east1.gce.archive.ubuntu.com/ubuntu bionic-updates InRelease [88.7 kB]
Get:3 http://asia-east1.gce.archive.ubuntu.com/ubuntu bionic-backports InRelease [74.6 kB]
Hit:4 https://nvidia.github.io/libnvidia-container/stable/ubuntu18.04/amd64  InRelease
Get:5 https://nvidia.github.io/nvidia-container-runtime/ubuntu18.04/amd64  InRelease [1481 B]
Hit:6 https://nvidia.github.io/nvidia-docker/ubuntu18.04/amd64  InRelease
Get:7 http://storage.googleapis.com/tensorflow-serving-apt stable InRelease [3012 B]
Get:8 http://asia-east1.gce.archive.ubuntu.com/ubuntu bionic-updates/main amd64 Packages [2250 kB]
Get:9 http://asia-east1.gce.archive.ubuntu.com/ubuntu bionic-updates/universe amd64 Packages [1754 kB]
Ign:10 http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  InRelease
Get:11 http://packages.cloud.google.com/apt google-cloud-logging-wheezy InRelease [5483 B]
Hit:12 http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  Release
Get:13 http://security.ubuntu.com/ubuntu bionic-security InRelease [88.7 kB]
Hit:14 http://archive.canonical.com/ubuntu bionic InRelease
Get:15 https://packages.cloud.google.com/apt eip-cloud-bionic InRelease [5419 B]
Err:11 http://packages.cloud.google.com/apt google-cloud-logging-wheezy InRelease
  The following signatures couldn't be verified because the public key is not available: NO_PUBKEY FEEA9169307EA071 NO_PUBKEY 8B57C5C2836F4BEB
Get:17 http://storage.googleapis.com/tensorflow-serving-apt stable/tensorflow-model-server amd64 Packages [341 B]
Err:15 https://packages.cloud.google.com/apt eip-cloud-bionic InRelease
  The following signatures couldn't be verified because the public key is not available: NO_PUBKEY FEEA9169307EA071 NO_PUBKEY 8B57C5C2836F4BEB
Get:18 http://storage.googleapis.com/tensorflow-serving-apt stable/tensorflow-model-server-universal amd64 Packages [348 B]
Fetched 4272 kB in 1s (4172 kB/s)



112 packages can be upgraded. Run 'apt list --upgradable' to see them.
W: An error occurred during the signature verification. The repository is not updated and the previous index files will be used. GPG error: http://packages.cloud.google.com/apt google-cloud-logging-wheezy InRelease: The following signatures couldn't be verified because the public key is not available: NO_PUBKEY FEEA9169307EA071 NO_PUBKEY 8B57C5C2836F4BEB
W: An error occurred during the signature verification. The repository is not updated and the previous index files will be used. GPG error: https://packages.cloud.google.com/apt eip-cloud-bionic InRelease: The following signatures couldn't be verified because the public key is not available: NO_PUBKEY FEEA9169307EA071 NO_PUBKEY 8B57C5C2836F4BEB
W: Failed to fetch https://packages.cloud.google.com/apt/dists/eip-cloud-bionic/InRelease  The following signatures couldn't be verified because the public key is not available: NO_PUBKEY FEEA9169307EA071 NO_PUBKEY 8B57C5C2836F4BEB
W: Failed to fetch http://packages.cloud.google.com/apt/dists/google-cloud-logging-wheezy/InRelease  The following signatures couldn't be verified because the public key is not available: NO_PUBKEY FEEA9169307EA071 NO_PUBKEY 8B57C5C2836F4BEB
W: Some index files failed to download. They have been ignored, or old ones used instead.

Instalar publicación de TensorFlow

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

{SUDO_IF_NEEDED} apt-get install tensorflow-model-server
The following packages were automatically installed and are no longer required:
  linux-gcp-5.4-headers-5.4.0-1040 linux-gcp-5.4-headers-5.4.0-1043
  linux-gcp-5.4-headers-5.4.0-1044 linux-gcp-5.4-headers-5.4.0-1049
  linux-headers-5.4.0-1049-gcp linux-image-5.4.0-1049-gcp
  linux-modules-5.4.0-1049-gcp linux-modules-extra-5.4.0-1049-gcp
Use 'sudo apt autoremove' to remove them.
The following NEW packages will be installed:
  tensorflow-model-server
0 upgraded, 1 newly installed, 0 to remove and 112 not upgraded.
Need to get 347 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 2.6.0 [347 MB]
Fetched 347 MB in 6s (58.6 MB/s)
Selecting previously unselected package tensorflow-model-server.
(Reading database ... 282213 files and directories currently installed.)
Preparing to unpack .../tensorflow-model-server_2.6.0_all.deb ...
Unpacking tensorflow-model-server (2.6.0) ...
Setting up tensorflow-model-server (2.6.0) ...

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 va a utilizar para las solicitudes REST.
  • model_name : Vamos a usar esto en la URL de solicitudes REST. Puede ser cualquier cosa.
  • model_base_path : Esta es la ruta de acceso al directorio en el que 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
tail server.log

Realice una solicitud a su 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]))

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]))