 Ver em TensorFlow.org
|
 Executar no Google Colab
|
 Ver código fonte no GitHub
|
 Baixar notebook
|
O TensorFlow Hub é uma maneira de compartilhar componentes de modelo pré-treinados. Consulte o TensorFlow Module Hub para obter uma lista pesquisável de modelos pré-treinados. Este tutorial demonstra:
- Como usar o TensorFlow Hub com o
tf.keras. - Como fazer a classificação da imagem usando o TensorFlow Hub.
- Como fazer um simples aprendizado de transferência.
Configuração
from __future__ import absolute_import, division, print_function, unicode_literals
import matplotlib.pylab as plt
try:
# %tensorflow_version only exists in Colab.
!pip install -q tf-nightly
except Exception:
pass
import tensorflow as tf
!pip install -q -U tf-hub-nightly
!pip install -q tfds-nightly
import tensorflow_hub as hub
from tensorflow.keras import layers
Um Classificador ImageNet
Baixar o classificador
Use hub.module para carregar uma mobilenet e tf.keras.layers.Lambda para envolvê-la como uma camada keras. Qualquer URL do classificador de imagem compatível com TensorFlow 2 do tfhub.dev funcionará aqui.
classifier_url ="https://tfhub.dev/google/tf2-preview/mobilenet_v2/classification/2"
IMAGE_SHAPE = (224, 224)
classifier = tf.keras.Sequential([
hub.KerasLayer(classifier_url, input_shape=IMAGE_SHAPE+(3,))
])
Execute-o em uma única imagem
Faça o download de uma única imagem para experimentar o modelo.
import numpy as np
import PIL.Image as Image
grace_hopper = tf.keras.utils.get_file('image.jpg','https://storage.googleapis.com/download.tensorflow.org/example_images/grace_hopper.jpg')
grace_hopper = Image.open(grace_hopper).resize(IMAGE_SHAPE)
grace_hopper
Downloading data from https://storage.googleapis.com/download.tensorflow.org/example_images/grace_hopper.jpg 8192/61306 [===>..........................] - ETA: 0s 65536/61306 [================================] - 0s 0us/step
grace_hopper = np.array(grace_hopper)/255.0
grace_hopper.shape
(224, 224, 3)
Adicione uma dimensão em batch e passe a imagem para o modelo.
result = classifier.predict(grace_hopper[np.newaxis, ...])
result.shape
(1, 1001)
O resultado é um vetor de 1001 elementos de logits, classificando a probabilidade de cada classe para a imagem.
Portanto, o ID da classe superior pode ser encontrado com argmax:
predicted_class = np.argmax(result[0], axis=-1)
predicted_class
653
Decodificar as previsões
Temos o ID da classe previsto,
Busque as etiquetas ImageNet e decodifique as previsões
labels_path = tf.keras.utils.get_file('ImageNetLabels.txt','https://storage.googleapis.com/download.tensorflow.org/data/ImageNetLabels.txt')
imagenet_labels = np.array(open(labels_path).read().splitlines())
Downloading data from https://storage.googleapis.com/download.tensorflow.org/data/ImageNetLabels.txt 8192/10484 [======================>.......] - ETA: 0s 16384/10484 [==============================================] - 0s 0us/step
plt.imshow(grace_hopper)
plt.axis('off')
predicted_class_name = imagenet_labels[predicted_class]
_ = plt.title("Prediction: " + predicted_class_name.title())
Transferência de aprendizado simples
Usando o TF Hub, é simples treinar novamente a camada superior do modelo para reconhecer as classes em nosso conjunto de dados.
Conjunto de Dados
Neste exemplo, você usará o conjunto de dados de flores TensorFlow:
data_root = tf.keras.utils.get_file(
'flower_photos','https://storage.googleapis.com/download.tensorflow.org/example_images/flower_photos.tgz',
untar=True)
Downloading data from https://storage.googleapis.com/download.tensorflow.org/example_images/flower_photos.tgz 8192/228813984 [..............................] - ETA: 2s 5578752/228813984 [..............................] - ETA: 2s 8396800/228813984 [>.............................] - ETA: 48s 22552576/228813984 [=>............................] - ETA: 17s 38133760/228813984 [===>..........................] - ETA: 9s 41951232/228813984 [====>.........................] - ETA: 11s 57114624/228813984 [======>.......................] - ETA: 8s 72818688/228813984 [========>.....................] - ETA: 5s 81092608/228813984 [=========>....................] - ETA: 5s 96215040/228813984 [===========>..................] - ETA: 3s 111460352/228813984 [=============>................] - ETA: 3s 127205376/228813984 [===============>..............] - ETA: 2s 134225920/228813984 [================>.............] - ETA: 2s 149045248/228813984 [==================>...........] - ETA: 1s 164954112/228813984 [====================>.........] - ETA: 1s 180723712/228813984 [======================>.......] - ETA: 1s 196534272/228813984 [========================>.....] - ETA: 0s 212320256/228813984 [==========================>...] - ETA: 0s 227868672/228813984 [============================>.] - ETA: 0s 228818944/228813984 [==============================] - 4s 0us/step
A maneira mais simples de carregar esses dados em nosso modelo é usando tf.keras.preprocessing.image.ImageDataGenerator,
Todos os módulos de imagem do TensorFlow Hub esperam entradas flutuantes na faixa [0,1]. Use o parâmetro rescale do ImageDataGenerator para conseguir isso.
O tamanho da imagem será tratado posteriormente.
image_generator = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1/255)
image_data = image_generator.flow_from_directory(str(data_root), target_size=IMAGE_SHAPE)
Found 3670 images belonging to 5 classes.
O objeto resultante é um iterador que retorna os pares image_batch, label_batch.
for image_batch, label_batch in image_data:
print("Image batch shape: ", image_batch.shape)
print("Label batch shape: ", label_batch.shape)
break
Image batch shape: (32, 224, 224, 3) Label batch shape: (32, 5)
Rode o classificador em um lote de imagens
Agora, execute o classificador em um lote de imagens.
result_batch = classifier.predict(image_batch)
result_batch.shape
(32, 1001)
predicted_class_names = imagenet_labels[np.argmax(result_batch, axis=-1)]
predicted_class_names
array(['mushroom', 'sea urchin', 'green lizard', 'daisy', 'cardoon',
'plow', 'daisy', 'greenhouse', 'daisy', 'velvet', 'pot', 'daisy',
'daisy', 'cardoon', 'daisy', 'daisy', 'coral fungus', 'daisy',
'tractor', 'daisy', 'daisy', 'vase', 'rapeseed',
"yellow lady's slipper", "yellow lady's slipper", 'vase', 'daisy',
'daisy', 'bee', 'daisy', 'daisy', 'hip'], dtype='<U30')
Agora verifique como essas previsões estão alinhadas com as imagens:
plt.figure(figsize=(10,9))
plt.subplots_adjust(hspace=0.5)
for n in range(30):
plt.subplot(6,5,n+1)
plt.imshow(image_batch[n])
plt.title(predicted_class_names[n])
plt.axis('off')
_ = plt.suptitle("ImageNet predictions")
Veja o arquivo LICENSE.txt para atribuições de imagem.
Os resultados estão longe de serem perfeitos, mas razoáveis, considerando que essas não são as classes para as quais o modelo foi treinado (exceto "daisy").
Faça o download do modelo sem cabeça
O TensorFlow Hub também distribui modelos sem a camada de classificação superior. Eles podem ser usados para transferir facilmente o aprendizado.
Qualquer URL do vetor de recurso de imagem compatível com Tensorflow 2 do tfhub.dev funcionará aqui.
feature_extractor_url = "https://tfhub.dev/google/tf2-preview/mobilenet_v2/feature_vector/2"
Crie o extrator de características.
feature_extractor_layer = hub.KerasLayer(feature_extractor_url,
input_shape=(224,224,3))
Isto retorna um vetor de tamanho 1280 para cada imagem:
feature_batch = feature_extractor_layer(image_batch)
print(feature_batch.shape)
(32, 1280)
Congele as variáveis na camada extrator de característica, para que o treinamento modifique apenas a nova camada do classificador.
feature_extractor_layer.trainable = False
Anexar um cabeçalho de classificação
Agora envolva a camada do hub em um modelo tf.keras.Sequential e adicione uma nova camada de classificação.
model = tf.keras.Sequential([
feature_extractor_layer,
layers.Dense(image_data.num_classes, activation='softmax')
])
model.summary()
Model: "sequential_1" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= keras_layer_1 (KerasLayer) (None, 1280) 2257984 _________________________________________________________________ dense (Dense) (None, 5) 6405 ================================================================= Total params: 2,264,389 Trainable params: 6,405 Non-trainable params: 2,257,984 _________________________________________________________________
predictions = model(image_batch)
predictions.shape
TensorShape([32, 5])
Treine o Modelo
Use compile para configurar o processo de treinamento:
model.compile(
optimizer=tf.keras.optimizers.Adam(),
loss='categorical_crossentropy',
metrics=['acc'])
Agora use o método .fit para treinar o modelo.
Para manter este exemplo, treine apenas duas épocas. Para visualizar o progresso do treinamento, use um retorno de chamada personalizado para registrar a perda e a acurácia de cada lote individualmente, em vez da média da época.
class CollectBatchStats(tf.keras.callbacks.Callback):
def __init__(self):
self.batch_losses = []
self.batch_acc = []
def on_train_batch_end(self, batch, logs=None):
self.batch_losses.append(logs['loss'])
self.batch_acc.append(logs['acc'])
self.model.reset_metrics()
steps_per_epoch = np.ceil(image_data.samples/image_data.batch_size)
batch_stats_callback = CollectBatchStats()
history = model.fit_generator(image_data, epochs=2,
steps_per_epoch=steps_per_epoch,
callbacks = [batch_stats_callback])
/tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/keras/engine/training.py:1799: UserWarning: `Model.fit_generator` is deprecated and will be removed in a future version. Please use `Model.fit`, which supports generators.
warnings.warn('`Model.fit_generator` is deprecated and '
Epoch 1/2
1/115 [..............................] - ETA: 5:26 - loss: 2.2827 - acc: 0.1562
2/115 [..............................] - ETA: 40s - loss: 2.0535 - acc: 0.1875
3/115 [..............................] - ETA: 39s - loss: 1.8939 - acc: 0.2396
4/115 [>.............................] - ETA: 38s - loss: 1.8444 - acc: 0.2422
5/115 [>.............................] - ETA: 37s - loss: 1.8182 - acc: 0.2500
6/115 [>.............................] - ETA: 37s - loss: 1.7846 - acc: 0.2656
7/115 [>.............................] - ETA: 37s - loss: 1.7527 - acc: 0.2589
8/115 [=>............................] - ETA: 36s - loss: 1.7308 - acc: 0.2812
9/115 [=>............................] - ETA: 36s - loss: 1.6836 - acc: 0.3090
10/115 [=>............................] - ETA: 35s - loss: 1.6612 - acc: 0.3187
11/115 [=>............................] - ETA: 35s - loss: 1.5988 - acc: 0.3551
12/115 [==>...........................] - ETA: 35s - loss: 1.5893 - acc: 0.3620
13/115 [==>...........................] - ETA: 34s - loss: 1.5614 - acc: 0.3774
14/115 [==>...........................] - ETA: 34s - loss: 1.5388 - acc: 0.3884
15/115 [==>...........................] - ETA: 33s - loss: 1.5029 - acc: 0.4042
16/115 [===>..........................] - ETA: 33s - loss: 1.4796 - acc: 0.4062
17/115 [===>..........................] - ETA: 33s - loss: 1.4504 - acc: 0.4228
18/115 [===>..........................] - ETA: 32s - loss: 1.4357 - acc: 0.4253
19/115 [===>..........................] - ETA: 32s - loss: 1.4107 - acc: 0.4375
20/115 [====>.........................] - ETA: 32s - loss: 1.3866 - acc: 0.4437
21/115 [====>.........................] - ETA: 31s - loss: 1.3642 - acc: 0.4524
22/115 [====>.........................] - ETA: 31s - loss: 1.3313 - acc: 0.4730
23/115 [=====>........................] - ETA: 31s - loss: 1.3094 - acc: 0.4851
24/115 [=====>........................] - ETA: 30s - loss: 1.2883 - acc: 0.4974
25/115 [=====>........................] - ETA: 30s - loss: 1.2626 - acc: 0.5088
26/115 [=====>........................] - ETA: 30s - loss: 1.2445 - acc: 0.5216
27/115 [======>.......................] - ETA: 29s - loss: 1.2252 - acc: 0.5289
28/115 [======>.......................] - ETA: 29s - loss: 1.2059 - acc: 0.5368
29/115 [======>.......................] - ETA: 29s - loss: 1.1867 - acc: 0.5474
30/115 [======>.......................] - ETA: 28s - loss: 1.1716 - acc: 0.5542
31/115 [=======>......................] - ETA: 28s - loss: 1.1553 - acc: 0.5615
32/115 [=======>......................] - ETA: 28s - loss: 1.1480 - acc: 0.5625
33/115 [=======>......................] - ETA: 27s - loss: 1.1356 - acc: 0.5691
34/115 [=======>......................] - ETA: 27s - loss: 1.1208 - acc: 0.5754
35/115 [========>.....................] - ETA: 27s - loss: 1.1059 - acc: 0.5804
36/115 [========>.....................] - ETA: 26s - loss: 1.0940 - acc: 0.5851
37/115 [========>.....................] - ETA: 26s - loss: 1.0783 - acc: 0.5929
38/115 [========>.....................] - ETA: 26s - loss: 1.0636 - acc: 0.5979
39/115 [=========>....................] - ETA: 25s - loss: 1.0490 - acc: 0.6066
40/115 [=========>....................] - ETA: 25s - loss: 1.0349 - acc: 0.6117
41/115 [=========>....................] - ETA: 25s - loss: 1.0271 - acc: 0.6159
42/115 [=========>....................] - ETA: 24s - loss: 1.0185 - acc: 0.6198
43/115 [==========>...................] - ETA: 24s - loss: 1.0078 - acc: 0.6257
44/115 [==========>...................] - ETA: 23s - loss: 1.0014 - acc: 0.6307
45/115 [==========>...................] - ETA: 23s - loss: 0.9896 - acc: 0.6354
46/115 [===========>..................] - ETA: 23s - loss: 0.9814 - acc: 0.6399
47/115 [===========>..................] - ETA: 22s - loss: 0.9736 - acc: 0.6430
48/115 [===========>..................] - ETA: 22s - loss: 0.9688 - acc: 0.6458
49/115 [===========>..................] - ETA: 22s - loss: 0.9596 - acc: 0.6505
50/115 [============>.................] - ETA: 21s - loss: 0.9528 - acc: 0.6538
51/115 [============>.................] - ETA: 21s - loss: 0.9425 - acc: 0.6575
52/115 [============>.................] - ETA: 21s - loss: 0.9360 - acc: 0.6587
53/115 [============>.................] - ETA: 20s - loss: 0.9226 - acc: 0.6645
54/115 [=============>................] - ETA: 20s - loss: 0.9123 - acc: 0.6684
55/115 [=============>................] - ETA: 20s - loss: 0.9071 - acc: 0.6710
56/115 [=============>................] - ETA: 19s - loss: 0.9000 - acc: 0.6741
57/115 [=============>................] - ETA: 19s - loss: 0.8930 - acc: 0.6771
58/115 [==============>...............] - ETA: 19s - loss: 0.8863 - acc: 0.6783
59/115 [==============>...............] - ETA: 18s - loss: 0.8804 - acc: 0.6806
60/115 [==============>...............] - ETA: 18s - loss: 0.8777 - acc: 0.6813
61/115 [==============>...............] - ETA: 18s - loss: 0.8731 - acc: 0.6824
62/115 [===============>..............] - ETA: 17s - loss: 0.8671 - acc: 0.6840
63/115 [===============>..............] - ETA: 17s - loss: 0.8611 - acc: 0.6870
64/115 [===============>..............] - ETA: 17s - loss: 0.8560 - acc: 0.6890
65/115 [===============>..............] - ETA: 16s - loss: 0.8498 - acc: 0.6913
66/115 [================>.............] - ETA: 16s - loss: 0.8413 - acc: 0.6951
67/115 [================>.............] - ETA: 16s - loss: 0.8351 - acc: 0.6978
68/115 [================>.............] - ETA: 15s - loss: 0.8295 - acc: 0.6999
69/115 [=================>............] - ETA: 15s - loss: 0.8259 - acc: 0.7006
70/115 [=================>............] - ETA: 15s - loss: 0.8211 - acc: 0.7031
71/115 [=================>............] - ETA: 14s - loss: 0.8169 - acc: 0.7042
72/115 [=================>............] - ETA: 14s - loss: 0.8126 - acc: 0.7057
73/115 [==================>...........] - ETA: 14s - loss: 0.8113 - acc: 0.7072
74/115 [==================>...........] - ETA: 13s - loss: 0.8069 - acc: 0.7086
75/115 [==================>...........] - ETA: 13s - loss: 0.8002 - acc: 0.7117
76/115 [==================>...........] - ETA: 13s - loss: 0.7962 - acc: 0.7130
77/115 [===================>..........] - ETA: 12s - loss: 0.7952 - acc: 0.7135
78/115 [===================>..........] - ETA: 12s - loss: 0.7923 - acc: 0.7147
79/115 [===================>..........] - ETA: 12s - loss: 0.7879 - acc: 0.7164
80/115 [===================>..........] - ETA: 11s - loss: 0.7842 - acc: 0.7180
81/115 [====================>.........] - ETA: 11s - loss: 0.7796 - acc: 0.7203
82/115 [====================>.........] - ETA: 11s - loss: 0.7802 - acc: 0.7191
83/115 [====================>.........] - ETA: 10s - loss: 0.7742 - acc: 0.7210
84/115 [====================>.........] - ETA: 10s - loss: 0.7715 - acc: 0.7221
85/115 [=====================>........] - ETA: 10s - loss: 0.7680 - acc: 0.7232
86/115 [=====================>........] - ETA: 9s - loss: 0.7635 - acc: 0.7253
87/115 [=====================>........] - ETA: 9s - loss: 0.7593 - acc: 0.7270
88/115 [=====================>........] - ETA: 9s - loss: 0.7567 - acc: 0.7283
89/115 [======================>.......] - ETA: 8s - loss: 0.7539 - acc: 0.7296
90/115 [======================>.......] - ETA: 8s - loss: 0.7519 - acc: 0.7302
91/115 [======================>.......] - ETA: 8s - loss: 0.7494 - acc: 0.7311
92/115 [=======================>......] - ETA: 7s - loss: 0.7448 - acc: 0.7327
93/115 [=======================>......] - ETA: 7s - loss: 0.7398 - acc: 0.7349
94/115 [=======================>......] - ETA: 7s - loss: 0.7387 - acc: 0.7344
95/115 [=======================>......] - ETA: 6s - loss: 0.7359 - acc: 0.7355
96/115 [========================>.....] - ETA: 6s - loss: 0.7337 - acc: 0.7367
97/115 [========================>.....] - ETA: 6s - loss: 0.7294 - acc: 0.7384
98/115 [========================>.....] - ETA: 5s - loss: 0.7250 - acc: 0.7408
99/115 [========================>.....] - ETA: 5s - loss: 0.7207 - acc: 0.7427
100/115 [=========================>....] - ETA: 5s - loss: 0.7192 - acc: 0.7434
101/115 [=========================>....] - ETA: 4s - loss: 0.7179 - acc: 0.7432
102/115 [=========================>....] - ETA: 4s - loss: 0.7155 - acc: 0.7442
103/115 [=========================>....] - ETA: 4s - loss: 0.7122 - acc: 0.7454
104/115 [==========================>...] - ETA: 3s - loss: 0.7094 - acc: 0.7458
105/115 [==========================>...] - ETA: 3s - loss: 0.7060 - acc: 0.7473
106/115 [==========================>...] - ETA: 3s - loss: 0.7049 - acc: 0.7482
107/115 [==========================>...] - ETA: 2s - loss: 0.7024 - acc: 0.7485
108/115 [===========================>..] - ETA: 2s - loss: 0.6997 - acc: 0.7494
109/115 [===========================>..] - ETA: 2s - loss: 0.6957 - acc: 0.7506
110/115 [===========================>..] - ETA: 1s - loss: 0.6943 - acc: 0.7511
111/115 [===========================>..] - ETA: 1s - loss: 0.6914 - acc: 0.7517
112/115 [============================>.] - ETA: 1s - loss: 0.6877 - acc: 0.7533
113/115 [============================>.] - ETA: 0s - loss: 0.6850 - acc: 0.7547
114/115 [============================>.] - ETA: 0s - loss: 0.6834 - acc: 0.7552
115/115 [==============================] - ETA: 0s - loss: 0.6819 - acc: 0.7562
115/115 [==============================] - 41s 334ms/step - loss: 0.6805 - acc: 0.7571
Epoch 2/2
1/115 [..............................] - ETA: 47s - loss: 0.1595 - acc: 1.0000
2/115 [..............................] - ETA: 39s - loss: 0.2053 - acc: 0.9688
3/115 [..............................] - ETA: 37s - loss: 0.2405 - acc: 0.9375
4/115 [>.............................] - ETA: 37s - loss: 0.2802 - acc: 0.9219
5/115 [>.............................] - ETA: 37s - loss: 0.2697 - acc: 0.9250
6/115 [>.............................] - ETA: 37s - loss: 0.3396 - acc: 0.8906
7/115 [>.............................] - ETA: 36s - loss: 0.3667 - acc: 0.8884
8/115 [=>............................] - ETA: 36s - loss: 0.3566 - acc: 0.8945
9/115 [=>............................] - ETA: 35s - loss: 0.3879 - acc: 0.8715
10/115 [=>............................] - ETA: 35s - loss: 0.4114 - acc: 0.8594
11/115 [=>............................] - ETA: 35s - loss: 0.4308 - acc: 0.8494
12/115 [==>...........................] - ETA: 34s - loss: 0.4158 - acc: 0.8542
13/115 [==>...........................] - ETA: 34s - loss: 0.4285 - acc: 0.8486
14/115 [==>...........................] - ETA: 34s - loss: 0.4164 - acc: 0.8527
15/115 [==>...........................] - ETA: 34s - loss: 0.4199 - acc: 0.8521
16/115 [===>..........................] - ETA: 33s - loss: 0.4212 - acc: 0.8555
17/115 [===>..........................] - ETA: 33s - loss: 0.4082 - acc: 0.8621
18/115 [===>..........................] - ETA: 32s - loss: 0.4053 - acc: 0.8628
19/115 [===>..........................] - ETA: 32s - loss: 0.3964 - acc: 0.8668
20/115 [====>.........................] - ETA: 32s - loss: 0.4023 - acc: 0.8625
21/115 [====>.........................] - ETA: 31s - loss: 0.3952 - acc: 0.8661
22/115 [====>.........................] - ETA: 31s - loss: 0.3905 - acc: 0.8693
23/115 [=====>........................] - ETA: 31s - loss: 0.3890 - acc: 0.8682
24/115 [=====>........................] - ETA: 30s - loss: 0.3933 - acc: 0.8672
25/115 [=====>........................] - ETA: 30s - loss: 0.3888 - acc: 0.8675
26/115 [=====>........................] - ETA: 30s - loss: 0.3894 - acc: 0.8690
27/115 [======>.......................] - ETA: 29s - loss: 0.3896 - acc: 0.8715
28/115 [======>.......................] - ETA: 29s - loss: 0.3857 - acc: 0.8739
29/115 [======>.......................] - ETA: 29s - loss: 0.3818 - acc: 0.8750
30/115 [======>.......................] - ETA: 28s - loss: 0.3824 - acc: 0.8771
31/115 [=======>......................] - ETA: 28s - loss: 0.3816 - acc: 0.8770
32/115 [=======>......................] - ETA: 28s - loss: 0.3837 - acc: 0.8770
33/115 [=======>......................] - ETA: 27s - loss: 0.3785 - acc: 0.8797
34/115 [=======>......................] - ETA: 27s - loss: 0.3756 - acc: 0.8787
35/115 [========>.....................] - ETA: 27s - loss: 0.3736 - acc: 0.8812
36/115 [========>.....................] - ETA: 26s - loss: 0.3709 - acc: 0.8811
37/115 [========>.....................] - ETA: 26s - loss: 0.3666 - acc: 0.8818
38/115 [========>.....................] - ETA: 25s - loss: 0.3789 - acc: 0.8750
39/115 [=========>....................] - ETA: 25s - loss: 0.3795 - acc: 0.8758
40/115 [=========>....................] - ETA: 25s - loss: 0.3807 - acc: 0.8766
41/115 [=========>....................] - ETA: 25s - loss: 0.3797 - acc: 0.8773
42/115 [=========>....................] - ETA: 24s - loss: 0.3759 - acc: 0.8787
43/115 [==========>...................] - ETA: 24s - loss: 0.3719 - acc: 0.8801
44/115 [==========>...................] - ETA: 24s - loss: 0.3718 - acc: 0.8800
45/115 [==========>...................] - ETA: 23s - loss: 0.3705 - acc: 0.8806
46/115 [===========>..................] - ETA: 23s - loss: 0.3684 - acc: 0.8811
47/115 [===========>..................] - ETA: 23s - loss: 0.3667 - acc: 0.8816
48/115 [===========>..................] - ETA: 22s - loss: 0.3660 - acc: 0.8815
49/115 [===========>..................] - ETA: 22s - loss: 0.3645 - acc: 0.8820
50/115 [============>.................] - ETA: 22s - loss: 0.3608 - acc: 0.8831
51/115 [============>.................] - ETA: 21s - loss: 0.3618 - acc: 0.8830
52/115 [============>.................] - ETA: 21s - loss: 0.3635 - acc: 0.8828
53/115 [============>.................] - ETA: 21s - loss: 0.3616 - acc: 0.8844
54/115 [=============>................] - ETA: 20s - loss: 0.3600 - acc: 0.8843
55/115 [=============>................] - ETA: 20s - loss: 0.3659 - acc: 0.8818
56/115 [=============>................] - ETA: 20s - loss: 0.3645 - acc: 0.8823
57/115 [=============>................] - ETA: 19s - loss: 0.3645 - acc: 0.8810
58/115 [==============>...............] - ETA: 19s - loss: 0.3614 - acc: 0.8825
59/115 [==============>...............] - ETA: 19s - loss: 0.3604 - acc: 0.8824
60/115 [==============>...............] - ETA: 18s - loss: 0.3610 - acc: 0.8828
61/115 [==============>...............] - ETA: 18s - loss: 0.3590 - acc: 0.8832
62/115 [===============>..............] - ETA: 18s - loss: 0.3612 - acc: 0.8821
63/115 [===============>..............] - ETA: 17s - loss: 0.3608 - acc: 0.8814
64/115 [===============>..............] - ETA: 17s - loss: 0.3576 - acc: 0.8828
65/115 [===============>..............] - ETA: 17s - loss: 0.3570 - acc: 0.8832
66/115 [================>.............] - ETA: 16s - loss: 0.3551 - acc: 0.8840
67/115 [================>.............] - ETA: 16s - loss: 0.3524 - acc: 0.8857
68/115 [================>.............] - ETA: 15s - loss: 0.3526 - acc: 0.8860
69/115 [=================>............] - ETA: 15s - loss: 0.3541 - acc: 0.8841
70/115 [=================>............] - ETA: 15s - loss: 0.3521 - acc: 0.8844
71/115 [=================>............] - ETA: 14s - loss: 0.3493 - acc: 0.8856
72/115 [=================>............] - ETA: 14s - loss: 0.3508 - acc: 0.8850
73/115 [==================>...........] - ETA: 14s - loss: 0.3512 - acc: 0.8848
74/115 [==================>...........] - ETA: 13s - loss: 0.3500 - acc: 0.8851
75/115 [==================>...........] - ETA: 13s - loss: 0.3534 - acc: 0.8842
76/115 [==================>...........] - ETA: 13s - loss: 0.3533 - acc: 0.8836
77/115 [===================>..........] - ETA: 12s - loss: 0.3533 - acc: 0.8835
78/115 [===================>..........] - ETA: 12s - loss: 0.3540 - acc: 0.8834
79/115 [===================>..........] - ETA: 12s - loss: 0.3531 - acc: 0.8833
80/115 [===================>..........] - ETA: 11s - loss: 0.3530 - acc: 0.8832
81/115 [====================>.........] - ETA: 11s - loss: 0.3536 - acc: 0.8827
82/115 [====================>.........] - ETA: 11s - loss: 0.3522 - acc: 0.8838
83/115 [====================>.........] - ETA: 10s - loss: 0.3518 - acc: 0.8833
84/115 [====================>.........] - ETA: 10s - loss: 0.3496 - acc: 0.8843
85/115 [=====================>........] - ETA: 10s - loss: 0.3494 - acc: 0.8846
86/115 [=====================>........] - ETA: 9s - loss: 0.3479 - acc: 0.8852
87/115 [=====================>........] - ETA: 9s - loss: 0.3465 - acc: 0.8858
88/115 [=====================>........] - ETA: 9s - loss: 0.3455 - acc: 0.8860
89/115 [======================>.......] - ETA: 8s - loss: 0.3441 - acc: 0.8862
90/115 [======================>.......] - ETA: 8s - loss: 0.3436 - acc: 0.8865
91/115 [======================>.......] - ETA: 8s - loss: 0.3419 - acc: 0.8874
92/115 [=======================>......] - ETA: 7s - loss: 0.3402 - acc: 0.8879
93/115 [=======================>......] - ETA: 7s - loss: 0.3394 - acc: 0.8878
94/115 [=======================>......] - ETA: 7s - loss: 0.3382 - acc: 0.8880
95/115 [=======================>......] - ETA: 6s - loss: 0.3390 - acc: 0.8882
96/115 [========================>.....] - ETA: 6s - loss: 0.3394 - acc: 0.8883
97/115 [========================>.....] - ETA: 6s - loss: 0.3421 - acc: 0.8872
98/115 [========================>.....] - ETA: 5s - loss: 0.3421 - acc: 0.8874
99/115 [========================>.....] - ETA: 5s - loss: 0.3420 - acc: 0.8867
100/115 [=========================>....] - ETA: 5s - loss: 0.3413 - acc: 0.8866
101/115 [=========================>....] - ETA: 4s - loss: 0.3404 - acc: 0.8874
102/115 [=========================>....] - ETA: 4s - loss: 0.3395 - acc: 0.8873
103/115 [=========================>....] - ETA: 4s - loss: 0.3392 - acc: 0.8871
104/115 [==========================>...] - ETA: 3s - loss: 0.3386 - acc: 0.8876
105/115 [==========================>...] - ETA: 3s - loss: 0.3374 - acc: 0.8881
106/115 [==========================>...] - ETA: 3s - loss: 0.3383 - acc: 0.8880
107/115 [==========================>...] - ETA: 2s - loss: 0.3369 - acc: 0.8887
108/115 [===========================>..] - ETA: 2s - loss: 0.3376 - acc: 0.8886
109/115 [===========================>..] - ETA: 2s - loss: 0.3377 - acc: 0.8888
110/115 [===========================>..] - ETA: 1s - loss: 0.3378 - acc: 0.8884
111/115 [===========================>..] - ETA: 1s - loss: 0.3375 - acc: 0.8885
112/115 [============================>.] - ETA: 1s - loss: 0.3371 - acc: 0.8887
113/115 [============================>.] - ETA: 0s - loss: 0.3393 - acc: 0.8880
114/115 [============================>.] - ETA: 0s - loss: 0.3388 - acc: 0.8879
115/115 [==============================] - ETA: 0s - loss: 0.3388 - acc: 0.8875
115/115 [==============================] - 39s 338ms/step - loss: 0.3387 - acc: 0.8871
Agora, depois de apenas algumas iterações de treinamento, já podemos ver que o modelo está progredindo na tarefa.
plt.figure()
plt.ylabel("Loss")
plt.xlabel("Training Steps")
plt.ylim([0,2])
plt.plot(batch_stats_callback.batch_losses)
[<matplotlib.lines.Line2D at 0x7f7b943b7e48>]
plt.figure()
plt.ylabel("Accuracy")
plt.xlabel("Training Steps")
plt.ylim([0,1])
plt.plot(batch_stats_callback.batch_acc)
[<matplotlib.lines.Line2D at 0x7f7b7a66f278>]
Verificando as previsões
Para refazer a plotagem de antes, primeiro obtenha a lista ordenada de nomes de classe:
class_names = sorted(image_data.class_indices.items(), key=lambda pair:pair[1])
class_names = np.array([key.title() for key, value in class_names])
class_names
array(['Daisy', 'Dandelion', 'Roses', 'Sunflowers', 'Tulips'],
dtype='<U10')
Execute o lote de imagens através do modelo e converta os índices em nomes de classe.
predicted_batch = model.predict(image_batch)
predicted_id = np.argmax(predicted_batch, axis=-1)
predicted_label_batch = class_names[predicted_id]
Plote o resultado
label_id = np.argmax(label_batch, axis=-1)
plt.figure(figsize=(10,9))
plt.subplots_adjust(hspace=0.5)
for n in range(30):
plt.subplot(6,5,n+1)
plt.imshow(image_batch[n])
color = "green" if predicted_id[n] == label_id[n] else "red"
plt.title(predicted_label_batch[n].title(), color=color)
plt.axis('off')
_ = plt.suptitle("Model predictions (green: correct, red: incorrect)")
Exporte seu modelo
Agora que você treinou o modelo, exporte-o como um modelo salvo:
import time
t = time.time()
export_path = "/tmp/saved_models/{}".format(int(t))
model.save(export_path, save_format='tf')
export_path
/tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/keras/engine/training.py:2289: UserWarning: `Model.state_updates` will be removed in a future version. This property should not be used in TensorFlow 2.0, as `updates` are applied automatically.
warnings.warn('`Model.state_updates` will be removed in a future version. '
/tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py:1377: UserWarning: `layer.updates` will be removed in a future version. This property should not be used in TensorFlow 2.0, as `updates` are applied automatically.
warnings.warn('`layer.updates` will be removed in a future version. '
INFO:tensorflow:Assets written to: /tmp/saved_models/1600797460/assets
INFO:tensorflow:Assets written to: /tmp/saved_models/1600797460/assets
'/tmp/saved_models/1600797460'
Agora confirme que podemos recarregá-lo e ainda dá os mesmos resultados:
reloaded = tf.keras.models.load_model(export_path)
result_batch = model.predict(image_batch)
reloaded_result_batch = reloaded.predict(image_batch)
abs(reloaded_result_batch - result_batch).max()
0.0
Este modelo salvo pode ser carregado para inferência posteriormente ou convertido para TFLite ou TFjs.