 View on TensorFlow.org
|
 Run in Google Colab
|
 View source on GitHub
|
 Download notebook
|
In this tutorial you will learn how to classify cats vs dogs images by using transfer learning from a pre-trained network.
A pre-trained model is a saved network that was previously trained on a large dataset, typically on a large-scale image-classification task. You either use the pretrained model as it is, or use transfer learning to customize this model to a given task.
The intuition behind transfer learning is that if a model trained on a large and general enough dataset, this model will effectively serve as a generic model of the visual world. You can then take advantage of these learned feature maps without having to start from scratch training a large model on a large dataset.
In this notebook, you will try two ways to customize a pretrained model:
Feature Extraction: Use the representations learned by a previous network to extract meaningful features from new samples. You simply add a new classifier, which will be trained from scratch, on top of the pretrained model so that you can repurpose the feature maps learned previously for our dataset.
You do not need to (re)train the entire model. The base convolutional network already contains features that are generically useful for classifying pictures. However, the final, classification part of the pretrained model is specific to original classification task, and subsequently specific to the set of classes on which the model was trained.
Fine-Tuning: Unfreezing a few of the top layers of a frozen model base and jointly training both the newly-added classifier layers and the last layers of the base model. This allows us to "fine tune" the higher-order feature representations in the base model in order to make them more relevant for the specific task.
You will follow the general machine learning workflow.
- Examine and understand the data
- Build an input pipeline, in this case using Keras
ImageDataGenerator - Compose our model
- Load in our pretrained base model (and pretrained weights)
- Stack our classification layers on top
- Train our model
- Evaluate model
from __future__ import absolute_import, division, print_function, unicode_literals
import os
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
keras = tf.keras
Data preprocessing
Data download
Use TensorFlow Datasets to load the cats and dogs dataset.
This tfds package is the easiest way to load pre-defined data. If you have your own data, and are interested in importing using it with TensorFlow see loading image data
import tensorflow_datasets as tfds
tfds.disable_progress_bar()
The tfds.load method downloads and caches the data, and returns a tf.data.Dataset object. These objects provide powerful, efficient methods for manipulating data and piping it into your model.
Since "cats_vs_dog" doesn't define standard splits, use the subsplit feature to divide it into (train, validation, test) with 80%, 10%, 10% of the data respectively.
SPLIT_WEIGHTS = (8, 1, 1)
splits = tfds.Split.TRAIN.subsplit(weighted=SPLIT_WEIGHTS)
(raw_train, raw_validation, raw_test), metadata = tfds.load(
'cats_vs_dogs', split=list(splits),
with_info=True, as_supervised=True)
Downloading and preparing dataset cats_vs_dogs (786.68 MiB) to /home/kbuilder/tensorflow_datasets/cats_vs_dogs/2.0.1... /usr/lib/python3/dist-packages/urllib3/connectionpool.py:860: InsecureRequestWarning: Unverified HTTPS request is being made. Adding certificate verification is strongly advised. See: https://urllib3.readthedocs.io/en/latest/advanced-usage.html#ssl-warnings InsecureRequestWarning) WARNING:absl:1738 images were corrupted and were skipped WARNING:tensorflow:From /home/kbuilder/.local/lib/python3.6/site-packages/tensorflow_datasets/core/file_format_adapter.py:209: tf_record_iterator (from tensorflow.python.lib.io.tf_record) is deprecated and will be removed in a future version. Instructions for updating: Use eager execution and: `tf.data.TFRecordDataset(path)` WARNING:tensorflow:From /home/kbuilder/.local/lib/python3.6/site-packages/tensorflow_datasets/core/file_format_adapter.py:209: tf_record_iterator (from tensorflow.python.lib.io.tf_record) is deprecated and will be removed in a future version. Instructions for updating: Use eager execution and: `tf.data.TFRecordDataset(path)` Dataset cats_vs_dogs downloaded and prepared to /home/kbuilder/tensorflow_datasets/cats_vs_dogs/2.0.1. Subsequent calls will reuse this data.
The resulting tf.data.Dataset objects contain (image, label) pairs. Where the images have variable shape and 3 channels, and the label is a scalar.
print(raw_train)
print(raw_validation)
print(raw_test)
<_OptionsDataset shapes: ((None, None, 3), ()), types: (tf.uint8, tf.int64)> <_OptionsDataset shapes: ((None, None, 3), ()), types: (tf.uint8, tf.int64)> <_OptionsDataset shapes: ((None, None, 3), ()), types: (tf.uint8, tf.int64)>
Show the first two images and labels from the training set:
get_label_name = metadata.features['label'].int2str
for image, label in raw_train.take(2):
plt.figure()
plt.imshow(image)
plt.title(get_label_name(label))
Format the Data
Use the tf.image module to format the images for the task.
Resize the images to a fixes input size, and rescale the input channels to a range of [-1,1]
IMG_SIZE = 160 # All images will be resized to 160x160
def format_example(image, label):
image = tf.cast(image, tf.float32)
image = (image/127.5) - 1
image = tf.image.resize(image, (IMG_SIZE, IMG_SIZE))
return image, label
Apply this function to each item in the dataset using the map method:
train = raw_train.map(format_example)
validation = raw_validation.map(format_example)
test = raw_test.map(format_example)
Now shuffle and batch the data.
BATCH_SIZE = 32
SHUFFLE_BUFFER_SIZE = 1000
train_batches = train.shuffle(SHUFFLE_BUFFER_SIZE).batch(BATCH_SIZE)
validation_batches = validation.batch(BATCH_SIZE)
test_batches = test.batch(BATCH_SIZE)
Inspect a batch of data:
for image_batch, label_batch in train_batches.take(1):
pass
image_batch.shape
TensorShape([32, 160, 160, 3])
Create the base model from the pre-trained convnets
You will create the base model from the MobileNet V2 model developed at Google. This is pre-trained on the ImageNet dataset, a large dataset of 1.4M images and 1000 classes of web images. ImageNet has a fairly arbitrary research training dataset with categories like jackfruit and syringe, but this base of knowledge will help us tell apart cats and dogs from our specific dataset.
First, you need to pick which layer of MobileNet V2 you will use for feature extraction. Obviously, the very last classification layer (on "top", as most diagrams of machine learning models go from bottom to top) is not very useful. Instead, you will follow the common practice to instead depend on the very last layer before the flatten operation. This layer is called the "bottleneck layer". The bottleneck features retain much generality as compared to the final/top layer.
First, instantiate a MobileNet V2 model pre-loaded with weights trained on ImageNet. By specifying the include_top=False argument, you load a network that doesn't include the classification layers at the top, which is ideal for feature extraction.
IMG_SHAPE = (IMG_SIZE, IMG_SIZE, 3)
# Create the base model from the pre-trained model MobileNet V2
base_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE,
include_top=False,
weights='imagenet')
Downloading data from https://github.com/JonathanCMitchell/mobilenet_v2_keras/releases/download/v1.1/mobilenet_v2_weights_tf_dim_ordering_tf_kernels_1.0_160_no_top.h5 9412608/9406464 [==============================] - 2s 0us/step
This feature extractor converts each 160x160x3 image to a 5x5x1280 block of features. See what it does to the example batch of images:
feature_batch = base_model(image_batch)
print(feature_batch.shape)
(32, 5, 5, 1280)
Feature extraction
You will freeze the convolutional base created from the previous step and use that as a feature extractor, add a classifier on top of it and train the top-level classifier.
Freeze the convolutional base
It's important to freeze the convolutional based before you compile and train the model. By freezing (or setting layer.trainable = False), you prevent the weights in a given layer from being updated during training. MobileNet V2 has many layers, so setting the entire model's trainable flag to False will freeze all the layers.
base_model.trainable = False
# Let's take a look at the base model architecture
base_model.summary()
Model: "mobilenetv2_1.00_160"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 160, 160, 3) 0
__________________________________________________________________________________________________
Conv1_pad (ZeroPadding2D) (None, 161, 161, 3) 0 input_1[0][0]
__________________________________________________________________________________________________
Conv1 (Conv2D) (None, 80, 80, 32) 864 Conv1_pad[0][0]
__________________________________________________________________________________________________
bn_Conv1 (BatchNormalization) (None, 80, 80, 32) 128 Conv1[0][0]
__________________________________________________________________________________________________
Conv1_relu (ReLU) (None, 80, 80, 32) 0 bn_Conv1[0][0]
__________________________________________________________________________________________________
expanded_conv_depthwise (Depthw (None, 80, 80, 32) 288 Conv1_relu[0][0]
__________________________________________________________________________________________________
expanded_conv_depthwise_BN (Bat (None, 80, 80, 32) 128 expanded_conv_depthwise[0][0]
__________________________________________________________________________________________________
expanded_conv_depthwise_relu (R (None, 80, 80, 32) 0 expanded_conv_depthwise_BN[0][0]
__________________________________________________________________________________________________
expanded_conv_project (Conv2D) (None, 80, 80, 16) 512 expanded_conv_depthwise_relu[0][0
__________________________________________________________________________________________________
expanded_conv_project_BN (Batch (None, 80, 80, 16) 64 expanded_conv_project[0][0]
__________________________________________________________________________________________________
block_1_expand (Conv2D) (None, 80, 80, 96) 1536 expanded_conv_project_BN[0][0]
__________________________________________________________________________________________________
block_1_expand_BN (BatchNormali (None, 80, 80, 96) 384 block_1_expand[0][0]
__________________________________________________________________________________________________
block_1_expand_relu (ReLU) (None, 80, 80, 96) 0 block_1_expand_BN[0][0]
__________________________________________________________________________________________________
block_1_pad (ZeroPadding2D) (None, 81, 81, 96) 0 block_1_expand_relu[0][0]
__________________________________________________________________________________________________
block_1_depthwise (DepthwiseCon (None, 40, 40, 96) 864 block_1_pad[0][0]
__________________________________________________________________________________________________
block_1_depthwise_BN (BatchNorm (None, 40, 40, 96) 384 block_1_depthwise[0][0]
__________________________________________________________________________________________________
block_1_depthwise_relu (ReLU) (None, 40, 40, 96) 0 block_1_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_1_project (Conv2D) (None, 40, 40, 24) 2304 block_1_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_1_project_BN (BatchNormal (None, 40, 40, 24) 96 block_1_project[0][0]
__________________________________________________________________________________________________
block_2_expand (Conv2D) (None, 40, 40, 144) 3456 block_1_project_BN[0][0]
__________________________________________________________________________________________________
block_2_expand_BN (BatchNormali (None, 40, 40, 144) 576 block_2_expand[0][0]
__________________________________________________________________________________________________
block_2_expand_relu (ReLU) (None, 40, 40, 144) 0 block_2_expand_BN[0][0]
__________________________________________________________________________________________________
block_2_depthwise (DepthwiseCon (None, 40, 40, 144) 1296 block_2_expand_relu[0][0]
__________________________________________________________________________________________________
block_2_depthwise_BN (BatchNorm (None, 40, 40, 144) 576 block_2_depthwise[0][0]
__________________________________________________________________________________________________
block_2_depthwise_relu (ReLU) (None, 40, 40, 144) 0 block_2_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_2_project (Conv2D) (None, 40, 40, 24) 3456 block_2_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_2_project_BN (BatchNormal (None, 40, 40, 24) 96 block_2_project[0][0]
__________________________________________________________________________________________________
block_2_add (Add) (None, 40, 40, 24) 0 block_1_project_BN[0][0]
block_2_project_BN[0][0]
__________________________________________________________________________________________________
block_3_expand (Conv2D) (None, 40, 40, 144) 3456 block_2_add[0][0]
__________________________________________________________________________________________________
block_3_expand_BN (BatchNormali (None, 40, 40, 144) 576 block_3_expand[0][0]
__________________________________________________________________________________________________
block_3_expand_relu (ReLU) (None, 40, 40, 144) 0 block_3_expand_BN[0][0]
__________________________________________________________________________________________________
block_3_pad (ZeroPadding2D) (None, 41, 41, 144) 0 block_3_expand_relu[0][0]
__________________________________________________________________________________________________
block_3_depthwise (DepthwiseCon (None, 20, 20, 144) 1296 block_3_pad[0][0]
__________________________________________________________________________________________________
block_3_depthwise_BN (BatchNorm (None, 20, 20, 144) 576 block_3_depthwise[0][0]
__________________________________________________________________________________________________
block_3_depthwise_relu (ReLU) (None, 20, 20, 144) 0 block_3_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_3_project (Conv2D) (None, 20, 20, 32) 4608 block_3_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_3_project_BN (BatchNormal (None, 20, 20, 32) 128 block_3_project[0][0]
__________________________________________________________________________________________________
block_4_expand (Conv2D) (None, 20, 20, 192) 6144 block_3_project_BN[0][0]
__________________________________________________________________________________________________
block_4_expand_BN (BatchNormali (None, 20, 20, 192) 768 block_4_expand[0][0]
__________________________________________________________________________________________________
block_4_expand_relu (ReLU) (None, 20, 20, 192) 0 block_4_expand_BN[0][0]
__________________________________________________________________________________________________
block_4_depthwise (DepthwiseCon (None, 20, 20, 192) 1728 block_4_expand_relu[0][0]
__________________________________________________________________________________________________
block_4_depthwise_BN (BatchNorm (None, 20, 20, 192) 768 block_4_depthwise[0][0]
__________________________________________________________________________________________________
block_4_depthwise_relu (ReLU) (None, 20, 20, 192) 0 block_4_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_4_project (Conv2D) (None, 20, 20, 32) 6144 block_4_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_4_project_BN (BatchNormal (None, 20, 20, 32) 128 block_4_project[0][0]
__________________________________________________________________________________________________
block_4_add (Add) (None, 20, 20, 32) 0 block_3_project_BN[0][0]
block_4_project_BN[0][0]
__________________________________________________________________________________________________
block_5_expand (Conv2D) (None, 20, 20, 192) 6144 block_4_add[0][0]
__________________________________________________________________________________________________
block_5_expand_BN (BatchNormali (None, 20, 20, 192) 768 block_5_expand[0][0]
__________________________________________________________________________________________________
block_5_expand_relu (ReLU) (None, 20, 20, 192) 0 block_5_expand_BN[0][0]
__________________________________________________________________________________________________
block_5_depthwise (DepthwiseCon (None, 20, 20, 192) 1728 block_5_expand_relu[0][0]
__________________________________________________________________________________________________
block_5_depthwise_BN (BatchNorm (None, 20, 20, 192) 768 block_5_depthwise[0][0]
__________________________________________________________________________________________________
block_5_depthwise_relu (ReLU) (None, 20, 20, 192) 0 block_5_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_5_project (Conv2D) (None, 20, 20, 32) 6144 block_5_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_5_project_BN (BatchNormal (None, 20, 20, 32) 128 block_5_project[0][0]
__________________________________________________________________________________________________
block_5_add (Add) (None, 20, 20, 32) 0 block_4_add[0][0]
block_5_project_BN[0][0]
__________________________________________________________________________________________________
block_6_expand (Conv2D) (None, 20, 20, 192) 6144 block_5_add[0][0]
__________________________________________________________________________________________________
block_6_expand_BN (BatchNormali (None, 20, 20, 192) 768 block_6_expand[0][0]
__________________________________________________________________________________________________
block_6_expand_relu (ReLU) (None, 20, 20, 192) 0 block_6_expand_BN[0][0]
__________________________________________________________________________________________________
block_6_pad (ZeroPadding2D) (None, 21, 21, 192) 0 block_6_expand_relu[0][0]
__________________________________________________________________________________________________
block_6_depthwise (DepthwiseCon (None, 10, 10, 192) 1728 block_6_pad[0][0]
__________________________________________________________________________________________________
block_6_depthwise_BN (BatchNorm (None, 10, 10, 192) 768 block_6_depthwise[0][0]
__________________________________________________________________________________________________
block_6_depthwise_relu (ReLU) (None, 10, 10, 192) 0 block_6_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_6_project (Conv2D) (None, 10, 10, 64) 12288 block_6_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_6_project_BN (BatchNormal (None, 10, 10, 64) 256 block_6_project[0][0]
__________________________________________________________________________________________________
block_7_expand (Conv2D) (None, 10, 10, 384) 24576 block_6_project_BN[0][0]
__________________________________________________________________________________________________
block_7_expand_BN (BatchNormali (None, 10, 10, 384) 1536 block_7_expand[0][0]
__________________________________________________________________________________________________
block_7_expand_relu (ReLU) (None, 10, 10, 384) 0 block_7_expand_BN[0][0]
__________________________________________________________________________________________________
block_7_depthwise (DepthwiseCon (None, 10, 10, 384) 3456 block_7_expand_relu[0][0]
__________________________________________________________________________________________________
block_7_depthwise_BN (BatchNorm (None, 10, 10, 384) 1536 block_7_depthwise[0][0]
__________________________________________________________________________________________________
block_7_depthwise_relu (ReLU) (None, 10, 10, 384) 0 block_7_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_7_project (Conv2D) (None, 10, 10, 64) 24576 block_7_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_7_project_BN (BatchNormal (None, 10, 10, 64) 256 block_7_project[0][0]
__________________________________________________________________________________________________
block_7_add (Add) (None, 10, 10, 64) 0 block_6_project_BN[0][0]
block_7_project_BN[0][0]
__________________________________________________________________________________________________
block_8_expand (Conv2D) (None, 10, 10, 384) 24576 block_7_add[0][0]
__________________________________________________________________________________________________
block_8_expand_BN (BatchNormali (None, 10, 10, 384) 1536 block_8_expand[0][0]
__________________________________________________________________________________________________
block_8_expand_relu (ReLU) (None, 10, 10, 384) 0 block_8_expand_BN[0][0]
__________________________________________________________________________________________________
block_8_depthwise (DepthwiseCon (None, 10, 10, 384) 3456 block_8_expand_relu[0][0]
__________________________________________________________________________________________________
block_8_depthwise_BN (BatchNorm (None, 10, 10, 384) 1536 block_8_depthwise[0][0]
__________________________________________________________________________________________________
block_8_depthwise_relu (ReLU) (None, 10, 10, 384) 0 block_8_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_8_project (Conv2D) (None, 10, 10, 64) 24576 block_8_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_8_project_BN (BatchNormal (None, 10, 10, 64) 256 block_8_project[0][0]
__________________________________________________________________________________________________
block_8_add (Add) (None, 10, 10, 64) 0 block_7_add[0][0]
block_8_project_BN[0][0]
__________________________________________________________________________________________________
block_9_expand (Conv2D) (None, 10, 10, 384) 24576 block_8_add[0][0]
__________________________________________________________________________________________________
block_9_expand_BN (BatchNormali (None, 10, 10, 384) 1536 block_9_expand[0][0]
__________________________________________________________________________________________________
block_9_expand_relu (ReLU) (None, 10, 10, 384) 0 block_9_expand_BN[0][0]
__________________________________________________________________________________________________
block_9_depthwise (DepthwiseCon (None, 10, 10, 384) 3456 block_9_expand_relu[0][0]
__________________________________________________________________________________________________
block_9_depthwise_BN (BatchNorm (None, 10, 10, 384) 1536 block_9_depthwise[0][0]
__________________________________________________________________________________________________
block_9_depthwise_relu (ReLU) (None, 10, 10, 384) 0 block_9_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_9_project (Conv2D) (None, 10, 10, 64) 24576 block_9_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_9_project_BN (BatchNormal (None, 10, 10, 64) 256 block_9_project[0][0]
__________________________________________________________________________________________________
block_9_add (Add) (None, 10, 10, 64) 0 block_8_add[0][0]
block_9_project_BN[0][0]
__________________________________________________________________________________________________
block_10_expand (Conv2D) (None, 10, 10, 384) 24576 block_9_add[0][0]
__________________________________________________________________________________________________
block_10_expand_BN (BatchNormal (None, 10, 10, 384) 1536 block_10_expand[0][0]
__________________________________________________________________________________________________
block_10_expand_relu (ReLU) (None, 10, 10, 384) 0 block_10_expand_BN[0][0]
__________________________________________________________________________________________________
block_10_depthwise (DepthwiseCo (None, 10, 10, 384) 3456 block_10_expand_relu[0][0]
__________________________________________________________________________________________________
block_10_depthwise_BN (BatchNor (None, 10, 10, 384) 1536 block_10_depthwise[0][0]
__________________________________________________________________________________________________
block_10_depthwise_relu (ReLU) (None, 10, 10, 384) 0 block_10_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_10_project (Conv2D) (None, 10, 10, 96) 36864 block_10_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_10_project_BN (BatchNorma (None, 10, 10, 96) 384 block_10_project[0][0]
__________________________________________________________________________________________________
block_11_expand (Conv2D) (None, 10, 10, 576) 55296 block_10_project_BN[0][0]
__________________________________________________________________________________________________
block_11_expand_BN (BatchNormal (None, 10, 10, 576) 2304 block_11_expand[0][0]
__________________________________________________________________________________________________
block_11_expand_relu (ReLU) (None, 10, 10, 576) 0 block_11_expand_BN[0][0]
__________________________________________________________________________________________________
block_11_depthwise (DepthwiseCo (None, 10, 10, 576) 5184 block_11_expand_relu[0][0]
__________________________________________________________________________________________________
block_11_depthwise_BN (BatchNor (None, 10, 10, 576) 2304 block_11_depthwise[0][0]
__________________________________________________________________________________________________
block_11_depthwise_relu (ReLU) (None, 10, 10, 576) 0 block_11_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_11_project (Conv2D) (None, 10, 10, 96) 55296 block_11_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_11_project_BN (BatchNorma (None, 10, 10, 96) 384 block_11_project[0][0]
__________________________________________________________________________________________________
block_11_add (Add) (None, 10, 10, 96) 0 block_10_project_BN[0][0]
block_11_project_BN[0][0]
__________________________________________________________________________________________________
block_12_expand (Conv2D) (None, 10, 10, 576) 55296 block_11_add[0][0]
__________________________________________________________________________________________________
block_12_expand_BN (BatchNormal (None, 10, 10, 576) 2304 block_12_expand[0][0]
__________________________________________________________________________________________________
block_12_expand_relu (ReLU) (None, 10, 10, 576) 0 block_12_expand_BN[0][0]
__________________________________________________________________________________________________
block_12_depthwise (DepthwiseCo (None, 10, 10, 576) 5184 block_12_expand_relu[0][0]
__________________________________________________________________________________________________
block_12_depthwise_BN (BatchNor (None, 10, 10, 576) 2304 block_12_depthwise[0][0]
__________________________________________________________________________________________________
block_12_depthwise_relu (ReLU) (None, 10, 10, 576) 0 block_12_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_12_project (Conv2D) (None, 10, 10, 96) 55296 block_12_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_12_project_BN (BatchNorma (None, 10, 10, 96) 384 block_12_project[0][0]
__________________________________________________________________________________________________
block_12_add (Add) (None, 10, 10, 96) 0 block_11_add[0][0]
block_12_project_BN[0][0]
__________________________________________________________________________________________________
block_13_expand (Conv2D) (None, 10, 10, 576) 55296 block_12_add[0][0]
__________________________________________________________________________________________________
block_13_expand_BN (BatchNormal (None, 10, 10, 576) 2304 block_13_expand[0][0]
__________________________________________________________________________________________________
block_13_expand_relu (ReLU) (None, 10, 10, 576) 0 block_13_expand_BN[0][0]
__________________________________________________________________________________________________
block_13_pad (ZeroPadding2D) (None, 11, 11, 576) 0 block_13_expand_relu[0][0]
__________________________________________________________________________________________________
block_13_depthwise (DepthwiseCo (None, 5, 5, 576) 5184 block_13_pad[0][0]
__________________________________________________________________________________________________
block_13_depthwise_BN (BatchNor (None, 5, 5, 576) 2304 block_13_depthwise[0][0]
__________________________________________________________________________________________________
block_13_depthwise_relu (ReLU) (None, 5, 5, 576) 0 block_13_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_13_project (Conv2D) (None, 5, 5, 160) 92160 block_13_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_13_project_BN (BatchNorma (None, 5, 5, 160) 640 block_13_project[0][0]
__________________________________________________________________________________________________
block_14_expand (Conv2D) (None, 5, 5, 960) 153600 block_13_project_BN[0][0]
__________________________________________________________________________________________________
block_14_expand_BN (BatchNormal (None, 5, 5, 960) 3840 block_14_expand[0][0]
__________________________________________________________________________________________________
block_14_expand_relu (ReLU) (None, 5, 5, 960) 0 block_14_expand_BN[0][0]
__________________________________________________________________________________________________
block_14_depthwise (DepthwiseCo (None, 5, 5, 960) 8640 block_14_expand_relu[0][0]
__________________________________________________________________________________________________
block_14_depthwise_BN (BatchNor (None, 5, 5, 960) 3840 block_14_depthwise[0][0]
__________________________________________________________________________________________________
block_14_depthwise_relu (ReLU) (None, 5, 5, 960) 0 block_14_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_14_project (Conv2D) (None, 5, 5, 160) 153600 block_14_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_14_project_BN (BatchNorma (None, 5, 5, 160) 640 block_14_project[0][0]
__________________________________________________________________________________________________
block_14_add (Add) (None, 5, 5, 160) 0 block_13_project_BN[0][0]
block_14_project_BN[0][0]
__________________________________________________________________________________________________
block_15_expand (Conv2D) (None, 5, 5, 960) 153600 block_14_add[0][0]
__________________________________________________________________________________________________
block_15_expand_BN (BatchNormal (None, 5, 5, 960) 3840 block_15_expand[0][0]
__________________________________________________________________________________________________
block_15_expand_relu (ReLU) (None, 5, 5, 960) 0 block_15_expand_BN[0][0]
__________________________________________________________________________________________________
block_15_depthwise (DepthwiseCo (None, 5, 5, 960) 8640 block_15_expand_relu[0][0]
__________________________________________________________________________________________________
block_15_depthwise_BN (BatchNor (None, 5, 5, 960) 3840 block_15_depthwise[0][0]
__________________________________________________________________________________________________
block_15_depthwise_relu (ReLU) (None, 5, 5, 960) 0 block_15_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_15_project (Conv2D) (None, 5, 5, 160) 153600 block_15_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_15_project_BN (BatchNorma (None, 5, 5, 160) 640 block_15_project[0][0]
__________________________________________________________________________________________________
block_15_add (Add) (None, 5, 5, 160) 0 block_14_add[0][0]
block_15_project_BN[0][0]
__________________________________________________________________________________________________
block_16_expand (Conv2D) (None, 5, 5, 960) 153600 block_15_add[0][0]
__________________________________________________________________________________________________
block_16_expand_BN (BatchNormal (None, 5, 5, 960) 3840 block_16_expand[0][0]
__________________________________________________________________________________________________
block_16_expand_relu (ReLU) (None, 5, 5, 960) 0 block_16_expand_BN[0][0]
__________________________________________________________________________________________________
block_16_depthwise (DepthwiseCo (None, 5, 5, 960) 8640 block_16_expand_relu[0][0]
__________________________________________________________________________________________________
block_16_depthwise_BN (BatchNor (None, 5, 5, 960) 3840 block_16_depthwise[0][0]
__________________________________________________________________________________________________
block_16_depthwise_relu (ReLU) (None, 5, 5, 960) 0 block_16_depthwise_BN[0][0]
__________________________________________________________________________________________________
block_16_project (Conv2D) (None, 5, 5, 320) 307200 block_16_depthwise_relu[0][0]
__________________________________________________________________________________________________
block_16_project_BN (BatchNorma (None, 5, 5, 320) 1280 block_16_project[0][0]
__________________________________________________________________________________________________
Conv_1 (Conv2D) (None, 5, 5, 1280) 409600 block_16_project_BN[0][0]
__________________________________________________________________________________________________
Conv_1_bn (BatchNormalization) (None, 5, 5, 1280) 5120 Conv_1[0][0]
__________________________________________________________________________________________________
out_relu (ReLU) (None, 5, 5, 1280) 0 Conv_1_bn[0][0]
==================================================================================================
Total params: 2,257,984
Trainable params: 0
Non-trainable params: 2,257,984
__________________________________________________________________________________________________
Add a classification head
To generate predictions from the block of features, average over the spatial 5x5 spatial locations, using a tf.keras.layers.GlobalAveragePooling2D layer to convert the features to a single 1280-element vector per image.
global_average_layer = tf.keras.layers.GlobalAveragePooling2D()
feature_batch_average = global_average_layer(feature_batch)
print(feature_batch_average.shape)
(32, 1280)
Apply a tf.keras.layers.Dense layer to convert these features into a single prediction per image. You don't need an activation function here because this prediction will be treated as a logit, or a raw prediction value. Positive numbers predict class 1, negative numbers predict class 0.
prediction_layer = keras.layers.Dense(1)
prediction_batch = prediction_layer(feature_batch_average)
print(prediction_batch.shape)
(32, 1)
Now stack the feature extractor, and these two layers using a tf.keras.Sequential model:
model = tf.keras.Sequential([
base_model,
global_average_layer,
prediction_layer
])
Compile the model
You must compile the model before training it. Since there are two classes, use a binary cross-entropy loss.
base_learning_rate = 0.0001
model.compile(optimizer=tf.keras.optimizers.RMSprop(lr=base_learning_rate),
loss='binary_crossentropy',
metrics=['accuracy'])
model.summary()
Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= mobilenetv2_1.00_160 (Model) (None, 5, 5, 1280) 2257984 _________________________________________________________________ global_average_pooling2d (Gl (None, 1280) 0 _________________________________________________________________ dense (Dense) (None, 1) 1281 ================================================================= Total params: 2,259,265 Trainable params: 1,281 Non-trainable params: 2,257,984 _________________________________________________________________
The 2.5M parameters in MobileNet are frozen, but there are 1.2K trainable parameters in the Dense layer. These are divided between two tf.Variable objects, the weights and biases.
len(model.trainable_variables)
2
Train the model
After training for 10 epochs, you should see ~96% accuracy.
num_train, num_val, num_test = (
metadata.splits['train'].num_examples*weight/10
for weight in SPLIT_WEIGHTS
)
initial_epochs = 10
steps_per_epoch = round(num_train)//BATCH_SIZE
validation_steps = 20
loss0,accuracy0 = model.evaluate(validation_batches, steps = validation_steps)
20/20 [==============================] - 2s 105ms/step - loss: 9.0489 - accuracy: 0.2531
print("initial loss: {:.2f}".format(loss0))
print("initial accuracy: {:.2f}".format(accuracy0))
initial loss: 9.05 initial accuracy: 0.25
history = model.fit(train_batches,
epochs=initial_epochs,
validation_data=validation_batches)
Epoch 1/10 582/582 [==============================] - 37s 64ms/step - loss: 3.2584 - accuracy: 0.6458 - val_loss: 0.0000e+00 - val_accuracy: 0.0000e+00 Epoch 2/10 582/582 [==============================] - 33s 56ms/step - loss: 1.3147 - accuracy: 0.8339 - val_loss: 0.5798 - val_accuracy: 0.9172 Epoch 3/10 582/582 [==============================] - 32s 56ms/step - loss: 0.8641 - accuracy: 0.8856 - val_loss: 0.4718 - val_accuracy: 0.9371 Epoch 4/10 582/582 [==============================] - 33s 56ms/step - loss: 0.6746 - accuracy: 0.9093 - val_loss: 0.4199 - val_accuracy: 0.9466 Epoch 5/10 582/582 [==============================] - 32s 55ms/step - loss: 0.5987 - accuracy: 0.9192 - val_loss: 0.3361 - val_accuracy: 0.9591 Epoch 6/10 582/582 [==============================] - 32s 55ms/step - loss: 0.5393 - accuracy: 0.9293 - val_loss: 0.2801 - val_accuracy: 0.9638 Epoch 7/10 582/582 [==============================] - 32s 54ms/step - loss: 0.5333 - accuracy: 0.9305 - val_loss: 0.3249 - val_accuracy: 0.9629 Epoch 8/10 582/582 [==============================] - 32s 54ms/step - loss: 0.4814 - accuracy: 0.9356 - val_loss: 0.3518 - val_accuracy: 0.9573 Epoch 9/10 582/582 [==============================] - 32s 55ms/step - loss: 0.4681 - accuracy: 0.9386 - val_loss: 0.3607 - val_accuracy: 0.9591 Epoch 10/10 582/582 [==============================] - 33s 56ms/step - loss: 0.4303 - accuracy: 0.9404 - val_loss: 0.3004 - val_accuracy: 0.9659
Learning curves
Let's take a look at the learning curves of the training and validation accuracy/loss when using the MobileNet V2 base model as a fixed feature extractor.
acc = history.history['accuracy']
val_acc = history.history['val_accuracy']
loss = history.history['loss']
val_loss = history.history['val_loss']
plt.figure(figsize=(8, 8))
plt.subplot(2, 1, 1)
plt.plot(acc, label='Training Accuracy')
plt.plot(val_acc, label='Validation Accuracy')
plt.legend(loc='lower right')
plt.ylabel('Accuracy')
plt.ylim([min(plt.ylim()),1])
plt.title('Training and Validation Accuracy')
plt.subplot(2, 1, 2)
plt.plot(loss, label='Training Loss')
plt.plot(val_loss, label='Validation Loss')
plt.legend(loc='upper right')
plt.ylabel('Cross Entropy')
plt.ylim([0,1.0])
plt.title('Training and Validation Loss')
plt.xlabel('epoch')
plt.show()
To a lesser extent, it is also because training metrics report the average for an epoch, while validation metrics are evaluated after the epoch, so validation metrics see a model that has trained slightly longer.
Fine tuning
In our feature extraction experiment, you were only training a few layers on top of an MobileNet V2 base model. The weights of the pre-trained network were not updated during training.
One way to increase performance even further is to train (or "fine-tune") the weights of the top layers of the pre-trained model alongside the training of the classifier you added. The training process will force the weights to be tuned from generic features maps to features associated specifically to our dataset.
Also, you should try to fine-tune a small number of top layers rather than the whole MobileNet model. In most convolutional networks, the higher up a layer is, the more specialized it is. The first few layers learn very simple and generic features which generalize to almost all types of images. As you go higher up, the features are increasingly more specific to the dataset on which the model was trained. The goal of fine-tuning is to adapt these specialized features to work with the new dataset, rather than overwrite the generic learning.
Un-freeze the top layers of the model
All you need to do is unfreeze the base_model and set the bottom layers be un-trainable. Then, you should recompile the model (necessary for these changes to take effect), and resume training.
base_model.trainable = True
# Let's take a look to see how many layers are in the base model
print("Number of layers in the base model: ", len(base_model.layers))
# Fine tune from this layer onwards
fine_tune_at = 100
# Freeze all the layers before the `fine_tune_at` layer
for layer in base_model.layers[:fine_tune_at]:
layer.trainable = False
Number of layers in the base model: 155
Compile the model
Compile the model using a much lower training rate.
model.compile(loss='binary_crossentropy',
optimizer = tf.keras.optimizers.RMSprop(lr=base_learning_rate/10),
metrics=['accuracy'])
model.summary()
Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= mobilenetv2_1.00_160 (Model) (None, 5, 5, 1280) 2257984 _________________________________________________________________ global_average_pooling2d (Gl (None, 1280) 0 _________________________________________________________________ dense (Dense) (None, 1) 1281 ================================================================= Total params: 2,259,265 Trainable params: 1,863,873 Non-trainable params: 395,392 _________________________________________________________________
len(model.trainable_variables)
58
Continue Train the model
If you trained to convergence earlier, this will get you a few percent more accuracy.
fine_tune_epochs = 10
total_epochs = initial_epochs + fine_tune_epochs
history_fine = model.fit(train_batches,
epochs=total_epochs,
initial_epoch = history.epoch[-1],
validation_data=validation_batches)
Epoch 10/20 582/582 [==============================] - 42s 72ms/step - loss: 0.4440 - accuracy: 0.9455 - val_loss: 0.0000e+00 - val_accuracy: 0.0000e+00 Epoch 11/20 582/582 [==============================] - 35s 60ms/step - loss: 0.3359 - accuracy: 0.9581 - val_loss: 0.1736 - val_accuracy: 0.9741 Epoch 12/20 582/582 [==============================] - 35s 61ms/step - loss: 0.2897 - accuracy: 0.9626 - val_loss: 0.1788 - val_accuracy: 0.9772 Epoch 13/20 582/582 [==============================] - 35s 60ms/step - loss: 0.2607 - accuracy: 0.9683 - val_loss: 0.1773 - val_accuracy: 0.9759 Epoch 14/20 582/582 [==============================] - 34s 59ms/step - loss: 0.2177 - accuracy: 0.9714 - val_loss: 0.1905 - val_accuracy: 0.9767 Epoch 15/20 582/582 [==============================] - 35s 60ms/step - loss: 0.2030 - accuracy: 0.9759 - val_loss: 0.1901 - val_accuracy: 0.9754 Epoch 16/20 582/582 [==============================] - 36s 61ms/step - loss: 0.1739 - accuracy: 0.9782 - val_loss: 0.1843 - val_accuracy: 0.9754 Epoch 17/20 582/582 [==============================] - 36s 62ms/step - loss: 0.1429 - accuracy: 0.9811 - val_loss: 0.1730 - val_accuracy: 0.9772 Epoch 18/20 582/582 [==============================] - 37s 63ms/step - loss: 0.1207 - accuracy: 0.9832 - val_loss: 0.1619 - val_accuracy: 0.9793 Epoch 19/20 582/582 [==============================] - 36s 62ms/step - loss: 0.1008 - accuracy: 0.9869 - val_loss: 0.1851 - val_accuracy: 0.9767 Epoch 20/20 582/582 [==============================] - 36s 62ms/step - loss: 0.0922 - accuracy: 0.9881 - val_loss: 0.1683 - val_accuracy: 0.9780
Let's take a look at the learning curves of the training and validation accuracy/loss, when fine tuning the last few layers of the MobileNet V2 base model and training the classifier on top of it. The validation loss is much higher than the training loss, so you may get some overfitting.
You may also get some overfitting as the new training set is relatively small and similar to the original MobileNet V2 datasets.
After fine tuning the model nearly reaches 98% accuracy.
acc += history_fine.history['accuracy']
val_acc += history_fine.history['val_accuracy']
loss += history_fine.history['loss']
val_loss += history_fine.history['val_loss']
plt.figure(figsize=(8, 8))
plt.subplot(2, 1, 1)
plt.plot(acc, label='Training Accuracy')
plt.plot(val_acc, label='Validation Accuracy')
plt.ylim([0.8, 1])
plt.plot([initial_epochs-1,initial_epochs-1],
plt.ylim(), label='Start Fine Tuning')
plt.legend(loc='lower right')
plt.title('Training and Validation Accuracy')
plt.subplot(2, 1, 2)
plt.plot(loss, label='Training Loss')
plt.plot(val_loss, label='Validation Loss')
plt.ylim([0, 1.0])
plt.plot([initial_epochs-1,initial_epochs-1],
plt.ylim(), label='Start Fine Tuning')
plt.legend(loc='upper right')
plt.title('Training and Validation Loss')
plt.xlabel('epoch')
plt.show()
Summary:
Using a pre-trained model for feature extraction: When working with a small dataset, it is common to take advantage of features learned by a model trained on a larger dataset in the same domain. This is done by instantiating the pre-trained model and adding a fully-connected classifier on top. The pre-trained model is "frozen" and only the weights of the classifier get updated during training. In this case, the convolutional base extracted all the features associated with each image and you just trained a classifier that determines the image class given that set of extracted features.
Fine-tuning a pre-trained model: To further improve performance, one might want to repurpose the top-level layers of the pre-trained models to the new dataset via fine-tuning. In this case, you tuned your weights such that your model learned high-level features specific to the dataset. This technique is usually recommended when the training dataset is large and very similar to the original dataset that the pre-trained model was trained on.
#@title MIT License
#
# Copyright (c) 2017 François Chollet # IGNORE_COPYRIGHT: cleared by OSS licensing
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
# DEALINGS IN THE SOFTWARE.