tf2_image_retraining.ipynb

Kernel: Python 3

Copyright 2021 The TensorFlow Hub Authors.

Licensed under the Apache License, Version 2.0 (the "License");

# Copyright 2021 The TensorFlow Hub Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================

Reentrenamiento de un clasificador de imágenes

Ver en TensorFlow.org

Ejecutar en Google Colab

Ver en GitHub

Descargar el bloc de notas

Ver modelos de TF Hub

Introducción

Los modelos de clasificación de imágenes tienen millones de parámetros. Entrenarlos desde cero requiere una gran cantidad de datos de entrenamiento etiquetados y mucha potencia de cálculo. El aprendizaje por transferencia es una técnica que acorta mucho el proceso al tomar una parte de un modelo que ya ha sido entrenado en una tarea relacionada y reutilizarla en un modelo nuevo.

En este Colab se enseña cómo construir un modelo de Keras para clasificar cinco especies de flores. Usaremos un TF2 SavedModel previamente entrenado de TensorFlow Hub para la extracción de características de imágenes, entrenado en el conjunto de datos ImageNet, que es mucho más grande y general. De forma opcional, el extractor de características se puede entrenar ("ajustar") junto con el clasificador recién agregado.

¿Prefiere una herramienta?

Este es un tutorial de codificación de TensorFlow. Si desea una herramienta que simplemente genere el modelo TensorFlow o TFLite, échele un vistazo a la herramienta de línea de comandos make_image_classifier que se instala mediante el paquete tensorflow-hub[make_image_classifier] de PIP o en este Colab de TFLite.

Preparación

import itertools
import os

import matplotlib.pylab as plt
import numpy as np

import tensorflow as tf
import tensorflow_hub as hub

print("TF version:", tf.__version__)
print("Hub version:", hub.__version__)
print("GPU is", "available" if tf.config.list_physical_devices('GPU') else "NOT AVAILABLE")

Seleccione el módulo TF2 SavedModel que usará

Para empezar, use https://tfhub.dev/google/imagenet/mobilenet_v2_100_224/feature_vector/4. Se puede usar la misma URL en el código para identificar el SavedModel y en su navegador para mostrar su documentación. (Tenga en cuenta que los modelos en formato TF1 Hub no funcionarán).

Puede encontrar más modelos TF2 que generan vectores de características de imagen aquí.

Hay varios modelos posibles para probar. Todo lo que necesita hacer es seleccionar uno diferente en la celda a continuación y seguir con el cuaderno.

#@title

model_name = "efficientnetv2-xl-21k" # @param ['efficientnetv2-s', 'efficientnetv2-m', 'efficientnetv2-l', 'efficientnetv2-s-21k', 'efficientnetv2-m-21k', 'efficientnetv2-l-21k', 'efficientnetv2-xl-21k', 'efficientnetv2-b0-21k', 'efficientnetv2-b1-21k', 'efficientnetv2-b2-21k', 'efficientnetv2-b3-21k', 'efficientnetv2-s-21k-ft1k', 'efficientnetv2-m-21k-ft1k', 'efficientnetv2-l-21k-ft1k', 'efficientnetv2-xl-21k-ft1k', 'efficientnetv2-b0-21k-ft1k', 'efficientnetv2-b1-21k-ft1k', 'efficientnetv2-b2-21k-ft1k', 'efficientnetv2-b3-21k-ft1k', 'efficientnetv2-b0', 'efficientnetv2-b1', 'efficientnetv2-b2', 'efficientnetv2-b3', 'efficientnet_b0', 'efficientnet_b1', 'efficientnet_b2', 'efficientnet_b3', 'efficientnet_b4', 'efficientnet_b5', 'efficientnet_b6', 'efficientnet_b7', 'bit_s-r50x1', 'inception_v3', 'inception_resnet_v2', 'resnet_v1_50', 'resnet_v1_101', 'resnet_v1_152', 'resnet_v2_50', 'resnet_v2_101', 'resnet_v2_152', 'nasnet_large', 'nasnet_mobile', 'pnasnet_large', 'mobilenet_v2_100_224', 'mobilenet_v2_130_224', 'mobilenet_v2_140_224', 'mobilenet_v3_small_100_224', 'mobilenet_v3_small_075_224', 'mobilenet_v3_large_100_224', 'mobilenet_v3_large_075_224']

model_handle_map = {
  "efficientnetv2-s": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet1k_s/feature_vector/2",
  "efficientnetv2-m": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet1k_m/feature_vector/2",
  "efficientnetv2-l": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet1k_l/feature_vector/2",
  "efficientnetv2-s-21k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_s/feature_vector/2",
  "efficientnetv2-m-21k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_m/feature_vector/2",
  "efficientnetv2-l-21k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_l/feature_vector/2",
  "efficientnetv2-xl-21k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_xl/feature_vector/2",
  "efficientnetv2-b0-21k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_b0/feature_vector/2",
  "efficientnetv2-b1-21k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_b1/feature_vector/2",
  "efficientnetv2-b2-21k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_b2/feature_vector/2",
  "efficientnetv2-b3-21k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_b3/feature_vector/2",
  "efficientnetv2-s-21k-ft1k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_ft1k_s/feature_vector/2",
  "efficientnetv2-m-21k-ft1k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_ft1k_m/feature_vector/2",
  "efficientnetv2-l-21k-ft1k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_ft1k_l/feature_vector/2",
  "efficientnetv2-xl-21k-ft1k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_ft1k_xl/feature_vector/2",
  "efficientnetv2-b0-21k-ft1k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_ft1k_b0/feature_vector/2",
  "efficientnetv2-b1-21k-ft1k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_ft1k_b1/feature_vector/2",
  "efficientnetv2-b2-21k-ft1k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_ft1k_b2/feature_vector/2",
  "efficientnetv2-b3-21k-ft1k": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet21k_ft1k_b3/feature_vector/2",
  "efficientnetv2-b0": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet1k_b0/feature_vector/2",
  "efficientnetv2-b1": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet1k_b1/feature_vector/2",
  "efficientnetv2-b2": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet1k_b2/feature_vector/2",
  "efficientnetv2-b3": "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet1k_b3/feature_vector/2",
  "efficientnet_b0": "https://tfhub.dev/tensorflow/efficientnet/b0/feature-vector/1",
  "efficientnet_b1": "https://tfhub.dev/tensorflow/efficientnet/b1/feature-vector/1",
  "efficientnet_b2": "https://tfhub.dev/tensorflow/efficientnet/b2/feature-vector/1",
  "efficientnet_b3": "https://tfhub.dev/tensorflow/efficientnet/b3/feature-vector/1",
  "efficientnet_b4": "https://tfhub.dev/tensorflow/efficientnet/b4/feature-vector/1",
  "efficientnet_b5": "https://tfhub.dev/tensorflow/efficientnet/b5/feature-vector/1",
  "efficientnet_b6": "https://tfhub.dev/tensorflow/efficientnet/b6/feature-vector/1",
  "efficientnet_b7": "https://tfhub.dev/tensorflow/efficientnet/b7/feature-vector/1",
  "bit_s-r50x1": "https://tfhub.dev/google/bit/s-r50x1/1",
  "inception_v3": "https://tfhub.dev/google/imagenet/inception_v3/feature-vector/4",
  "inception_resnet_v2": "https://tfhub.dev/google/imagenet/inception_resnet_v2/feature-vector/4",
  "resnet_v1_50": "https://tfhub.dev/google/imagenet/resnet_v1_50/feature-vector/4",
  "resnet_v1_101": "https://tfhub.dev/google/imagenet/resnet_v1_101/feature-vector/4",
  "resnet_v1_152": "https://tfhub.dev/google/imagenet/resnet_v1_152/feature-vector/4",
  "resnet_v2_50": "https://tfhub.dev/google/imagenet/resnet_v2_50/feature-vector/4",
  "resnet_v2_101": "https://tfhub.dev/google/imagenet/resnet_v2_101/feature-vector/4",
  "resnet_v2_152": "https://tfhub.dev/google/imagenet/resnet_v2_152/feature-vector/4",
  "nasnet_large": "https://tfhub.dev/google/imagenet/nasnet_large/feature_vector/4",
  "nasnet_mobile": "https://tfhub.dev/google/imagenet/nasnet_mobile/feature_vector/4",
  "pnasnet_large": "https://tfhub.dev/google/imagenet/pnasnet_large/feature_vector/4",
  "mobilenet_v2_100_224": "https://tfhub.dev/google/imagenet/mobilenet_v2_100_224/feature_vector/4",
  "mobilenet_v2_130_224": "https://tfhub.dev/google/imagenet/mobilenet_v2_130_224/feature_vector/4",
  "mobilenet_v2_140_224": "https://tfhub.dev/google/imagenet/mobilenet_v2_140_224/feature_vector/4",
  "mobilenet_v3_small_100_224": "https://tfhub.dev/google/imagenet/mobilenet_v3_small_100_224/feature_vector/5",
  "mobilenet_v3_small_075_224": "https://tfhub.dev/google/imagenet/mobilenet_v3_small_075_224/feature_vector/5",
  "mobilenet_v3_large_100_224": "https://tfhub.dev/google/imagenet/mobilenet_v3_large_100_224/feature_vector/5",
  "mobilenet_v3_large_075_224": "https://tfhub.dev/google/imagenet/mobilenet_v3_large_075_224/feature_vector/5",
}

model_image_size_map = {
  "efficientnetv2-s": 384,
  "efficientnetv2-m": 480,
  "efficientnetv2-l": 480,
  "efficientnetv2-b0": 224,
  "efficientnetv2-b1": 240,
  "efficientnetv2-b2": 260,
  "efficientnetv2-b3": 300,
  "efficientnetv2-s-21k": 384,
  "efficientnetv2-m-21k": 480,
  "efficientnetv2-l-21k": 480,
  "efficientnetv2-xl-21k": 512,
  "efficientnetv2-b0-21k": 224,
  "efficientnetv2-b1-21k": 240,
  "efficientnetv2-b2-21k": 260,
  "efficientnetv2-b3-21k": 300,
  "efficientnetv2-s-21k-ft1k": 384,
  "efficientnetv2-m-21k-ft1k": 480,
  "efficientnetv2-l-21k-ft1k": 480,
  "efficientnetv2-xl-21k-ft1k": 512,
  "efficientnetv2-b0-21k-ft1k": 224,
  "efficientnetv2-b1-21k-ft1k": 240,
  "efficientnetv2-b2-21k-ft1k": 260,
  "efficientnetv2-b3-21k-ft1k": 300, 
  "efficientnet_b0": 224,
  "efficientnet_b1": 240,
  "efficientnet_b2": 260,
  "efficientnet_b3": 300,
  "efficientnet_b4": 380,
  "efficientnet_b5": 456,
  "efficientnet_b6": 528,
  "efficientnet_b7": 600,
  "inception_v3": 299,
  "inception_resnet_v2": 299,
  "nasnet_large": 331,
  "pnasnet_large": 331,
}

model_handle = model_handle_map.get(model_name)
pixels = model_image_size_map.get(model_name, 224)

print(f"Selected model: {model_name} : {model_handle}")

IMAGE_SIZE = (pixels, pixels)
print(f"Input size {IMAGE_SIZE}")

BATCH_SIZE = 16#@param {type:"integer"}

Preparar el conjunto de datos de Flores

Las entradas se redimensionan adecuadamente para el módulo seleccionado. El aumento del conjunto de datos (es decir, las distorsiones aleatorias de una imagen cada vez que se lee) mejora el entrenamiento, especialmente. cuando se ajusta.

data_dir = tf.keras.utils.get_file(
    'flower_photos',
    'https://storage.googleapis.com/download.tensorflow.org/example_images/flower_photos.tgz',
    untar=True)

def build_dataset(subset):
  return tf.keras.preprocessing.image_dataset_from_directory(
      data_dir,
      validation_split=.20,
      subset=subset,
      label_mode="categorical",
      # Seed needs to provided when using validation_split and shuffle = True.
      # A fixed seed is used so that the validation set is stable across runs.
      seed=123,
      image_size=IMAGE_SIZE,
      batch_size=1)

train_ds = build_dataset("training")
class_names = tuple(train_ds.class_names)
train_size = train_ds.cardinality().numpy()
train_ds = train_ds.unbatch().batch(BATCH_SIZE)
train_ds = train_ds.repeat()

normalization_layer = tf.keras.layers.Rescaling(1. / 255)
preprocessing_model = tf.keras.Sequential([normalization_layer])
do_data_augmentation = False #@param {type:"boolean"}
if do_data_augmentation:
  preprocessing_model.add(
      tf.keras.layers.RandomRotation(40))
  preprocessing_model.add(
      tf.keras.layers.RandomTranslation(0, 0.2))
  preprocessing_model.add(
      tf.keras.layers.RandomTranslation(0.2, 0))
  # Like the old tf.keras.preprocessing.image.ImageDataGenerator(),
  # image sizes are fixed when reading, and then a random zoom is applied.
  # If all training inputs are larger than image_size, one could also use
  # RandomCrop with a batch size of 1 and rebatch later.
  preprocessing_model.add(
      tf.keras.layers.RandomZoom(0.2, 0.2))
  preprocessing_model.add(
      tf.keras.layers.RandomFlip(mode="horizontal"))
train_ds = train_ds.map(lambda images, labels:
                        (preprocessing_model(images), labels))

val_ds = build_dataset("validation")
valid_size = val_ds.cardinality().numpy()
val_ds = val_ds.unbatch().batch(BATCH_SIZE)
val_ds = val_ds.map(lambda images, labels:
                    (normalization_layer(images), labels))

Definir el modelo

Solo hay que colocar un clasificador lineal sobre el feature_extractor_layer con el módulo Hub.

Para mejorar la velocidad, comenzamos con una feature_extractor_layer no entrenable, pero también se puede habilitar el ajuste para una mayor precisión.

do_fine_tuning = False #@param {type:"boolean"}

print("Building model with", model_handle)
model = tf.keras.Sequential([
    # Explicitly define the input shape so the model can be properly
    # loaded by the TFLiteConverter
    tf.keras.layers.InputLayer(input_shape=IMAGE_SIZE + (3,)),
    hub.KerasLayer(model_handle, trainable=do_fine_tuning),
    tf.keras.layers.Dropout(rate=0.2),
    tf.keras.layers.Dense(len(class_names),
                          kernel_regularizer=tf.keras.regularizers.l2(0.0001))
])
model.build((None,)+IMAGE_SIZE+(3,))
model.summary()

Entrenar el modelo

model.compile(
  optimizer=tf.keras.optimizers.SGD(learning_rate=0.005, momentum=0.9), 
  loss=tf.keras.losses.CategoricalCrossentropy(from_logits=True, label_smoothing=0.1),
  metrics=['accuracy'])

steps_per_epoch = train_size // BATCH_SIZE
validation_steps = valid_size // BATCH_SIZE
hist = model.fit(
    train_ds,
    epochs=5, steps_per_epoch=steps_per_epoch,
    validation_data=val_ds,
    validation_steps=validation_steps).history

plt.figure()
plt.ylabel("Loss (training and validation)")
plt.xlabel("Training Steps")
plt.ylim([0,2])
plt.plot(hist["loss"])
plt.plot(hist["val_loss"])

plt.figure()
plt.ylabel("Accuracy (training and validation)")
plt.xlabel("Training Steps")
plt.ylim([0,1])
plt.plot(hist["accuracy"])
plt.plot(hist["val_accuracy"])

Pruebe el modelo en una imagen de los datos de validación:

x, y = next(iter(val_ds))
image = x[0, :, :, :]
true_index = np.argmax(y[0])
plt.imshow(image)
plt.axis('off')
plt.show()

# Expand the validation image to (1, 224, 224, 3) before predicting the label
prediction_scores = model.predict(np.expand_dims(image, axis=0))
predicted_index = np.argmax(prediction_scores)
print("True label: " + class_names[true_index])
print("Predicted label: " + class_names[predicted_index])

Finalmente, el modelo entrenado se puede guardar para implementarlo en TF Serving o TFLite (en dispositivos móviles) de la siguiente manera.

saved_model_path = f"/tmp/saved_flowers_model_{model_name}"
tf.saved_model.save(model, saved_model_path)

Opcional: implementación en TensorFlow Lite

TensorFlow Lite le permite implementar modelos de TensorFlow en dispositivos móviles y de IoT. El siguiente código muestra cómo convertir el modelo entrenado a TFLite y aplicar herramientas posteriores al entrenamiento del kit de herramientas de optimización de modelos de TensorFlow. Finalmente, lo ejecuta en TFLite Interpreter para examinar la calidad resultante.

• La conversión sin optimización proporciona los mismos resultados que antes (hasta un error de redondeo).

• La conversión con optimización sin ningún dato cuantifica los pesos del modelo a 8 bits, pero la inferencia aún usa el cálculo de punto flotante para las activaciones de la red neuronal. Esto reduce el tamaño del modelo casi en un factor de 4 y mejora la latencia de la CPU en dispositivos móviles.

• Además, el cálculo de las activaciones de la red neuronal también se puede cuantificar en números enteros de 8 bits si se proporciona un pequeño conjunto de datos de referencia para calibrar el rango de cuantificación. En un dispositivo móvil, esto acelera aún más la inferencia y permite ejecutar aceleradores como Edge TPU.

#@title Optimization settings
optimize_lite_model = False  #@param {type:"boolean"}
#@markdown Setting a value greater than zero enables quantization of neural network activations. A few dozen is already a useful amount.
num_calibration_examples = 60  #@param {type:"slider", min:0, max:1000, step:1}
representative_dataset = None
if optimize_lite_model and num_calibration_examples:
  # Use a bounded number of training examples without labels for calibration.
  # TFLiteConverter expects a list of input tensors, each with batch size 1.
  representative_dataset = lambda: itertools.islice(
      ([image[None, ...]] for batch, _ in train_ds for image in batch),
      num_calibration_examples)

converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_path)
if optimize_lite_model:
  converter.optimizations = [tf.lite.Optimize.DEFAULT]
  if representative_dataset:  # This is optional, see above.
    converter.representative_dataset = representative_dataset
lite_model_content = converter.convert()

with open(f"/tmp/lite_flowers_model_{model_name}.tflite", "wb") as f:
  f.write(lite_model_content)
print("Wrote %sTFLite model of %d bytes." %
      ("optimized " if optimize_lite_model else "", len(lite_model_content)))

interpreter = tf.lite.Interpreter(model_content=lite_model_content)
# This little helper wraps the TFLite Interpreter as a numpy-to-numpy function.
def lite_model(images):
  interpreter.allocate_tensors()
  interpreter.set_tensor(interpreter.get_input_details()[0]['index'], images)
  interpreter.invoke()
  return interpreter.get_tensor(interpreter.get_output_details()[0]['index'])

#@markdown For rapid experimentation, start with a moderate number of examples.
num_eval_examples = 50  #@param {type:"slider", min:0, max:700}
eval_dataset = ((image, label)  # TFLite expects batch size 1.
                for batch in train_ds
                for (image, label) in zip(*batch))
count = 0
count_lite_tf_agree = 0
count_lite_correct = 0
for image, label in eval_dataset:
  probs_lite = lite_model(image[None, ...])[0]
  probs_tf = model(image[None, ...]).numpy()[0]
  y_lite = np.argmax(probs_lite)
  y_tf = np.argmax(probs_tf)
  y_true = np.argmax(label)
  count +=1
  if y_lite == y_tf: count_lite_tf_agree += 1
  if y_lite == y_true: count_lite_correct += 1
  if count >= num_eval_examples: break
print("TFLite model agrees with original model on %d of %d examples (%g%%)." %
      (count_lite_tf_agree, count, 100.0 * count_lite_tf_agree / count))
print("TFLite model is accurate on %d of %d examples (%g%%)." %
      (count_lite_correct, count, 100.0 * count_lite_correct / count))