TensorFlow is an open source library for numerical computation, specializing in machine learning applications.

What you will build

In this codelab, you will learn how to run TensorFlow on a single machine, and will train a simple classifier to classify images of flowers.

Image CC-BY by Retinafunk

daisy (score = 0.99071)
sunflowers (score = 0.00595)
dandelion (score = 0.00252)
roses (score = 0.00049)
tulips (score = 0.00032)

We will be using transfer learning, which means we are starting with a model that has been already trained on another problem. We will then be retraining it on a similar problem. Deep learning from scratch can take days, but transfer learning can be done in short order.

We are going to use a model trained on the ImageNet Large Visual Recognition Challenge dataset. These models can differentiate between 1,000 different classes, like Dalmatian or dishwasher. You will have a choice of model architectures, so you can determine the right tradeoff between speed, size and accuracy for your problem.

We will use this same model, but retrain it to tell apart a small number of classes based on our own examples.

What you'll Learn

What you need

Install TensorFlow

Before we can begin the tutorial you need to install tensorflow.

Clone the git repository

All the code used in this codelab is contained in this git repository. Clone the repository and cd into it. This is where we will be working.

git clone https://github.com/googlecodelabs/tensorflow-for-poets-2

cd tensorflow-for-poets-2

Before you start any training, you'll need a set of images to teach the model about the new classes you want to recognize. We've created an archive of creative-commons licensed flower photos to use initially. Download the photos (218 MB) by invoking the following two commands:

curl http://download.tensorflow.org/example_images/flower_photos.tgz \
    | tar xz -C tf_files

You should now have a copy of the flower photos in your working directory. Confirm the contents of your working directory by issuing the following command:

ls tf_files/flower_photos

The preceding command should display the following objects:


Configure your MobileNet

The retrain script can retrain either Inception V3 model or a MobileNet. In this exercise, we will use a MobileNet. The principal difference is that Inception V3 is optimized for accuracy, while the MobileNets are optimized to be small and efficient, at the cost of some accuracy.

Inception V3 has a first-choice accuracy of 78% on ImageNet, but is the model is 85MB, and requires many times more processing than even the largest MobileNet configuration, which achieves 70.5% accuracy, with just a 19MB download.

Pick the following configuration options:

With the recommended settings, it typically takes only a couple of minutes to retrain on a laptop. You will pass the settings inside Linux shell variables. Set those shell variables as follows:


The graph below shows the first-choice-accuracies of these configurations (y-axis), vs the number of calculations required (x-axis), and the size of the model (circle area).

16 points are shown for mobilenet. For each of the 4 model sizes (circle area in the figure) there is one point for each image resolution setting. The 128px image size models are represented by the lower-left point in each set, while the 224px models are in the upper right.

Other notable architectures are also included for reference. "GoogleNet" in this figure is "Inception V1" in this table. An extended version of this figure is available in slides 84-89 here.

Start TensorBoard

Before starting the training, launch tensorboard in the background. TensorBoard is a monitoring and inspection tool included with tensorflow. You will use it to monitor the training progress.

tensorboard --logdir tf_files/training_summaries &

Investigate the retraining script

The retrain script is from the tensorflow hub repo, but it is not installed as part of the pip package. So for simplicity I've included it in the codelab repository. You can run the script using the python command. Take a minute to skim its "help".

python -m scripts.retrain -h

Run the training

As noted in the introduction, Imagenet models are networks with millions of parameters that can differentiate a large number of classes. We're only training the final layer of that network, so training will end in a reasonable amount of time.

Start your retraining with one big command (note the --summaries_dir option, sending training progress reports to the directory that tensorboard is monitoring) :

python -m scripts.retrain \
  --bottleneck_dir=tf_files/bottlenecks \
  --how_many_training_steps=500 \
  --model_dir=tf_files/models/ \
  --summaries_dir=tf_files/training_summaries/"${ARCHITECTURE}" \
  --output_graph=tf_files/retrained_graph.pb \
  --output_labels=tf_files/retrained_labels.txt \
  --architecture="${ARCHITECTURE}" \

This script downloads the pre-trained model, adds a new final layer, and trains that layer on the flower photos you've downloaded.

ImageNet does not include any of these flower species we're training on here. However, the kinds of information that make it possible for ImageNet to differentiate among 1,000 classes are also useful for distinguishing other objects. By using this pre-trained network, we are using that information as input to the final classification layer that distinguishes our flower classes.

Optional: I'm NOT in a hurry!

The first retraining command iterates only 500 times. You can very likely get improved results (i.e. higher accuracy) by training for longer. To get this improvement, remove the parameter --how_many_training_steps to use the default 4,000 iterations.

python -m scripts.retrain \
  --bottleneck_dir=tf_files/bottlenecks \
  --model_dir=tf_files/models/"${ARCHITECTURE}" \
  --summaries_dir=tf_files/training_summaries/"${ARCHITECTURE}" \
  --output_graph=tf_files/retrained_graph.pb \
  --output_labels=tf_files/retrained_labels.txt \
  --architecture="${ARCHITECTURE}" \

More about Bottlenecks

This section and the next provide background on how this retraining process works.

The first phase analyzes all the images on disk and calculates the bottleneck values for each of them. What's a bottleneck?

These ImageNet models are made up of many layers stacked on top of each other, a simplified picture of Inception V3 from TensorBoard, is shown above (all the details are available in this paper, with a complete picture on page 6). These layers are pre-trained and are already very valuable at finding and summarizing information that will help classify most images. For this codelab, you are training only the last layer (final_training_ops in the figure below). While all the previous layers retain their already-trained state.

In the above figure, the node labeled "softmax", on the left side, is the output layer of the original model. While all the nodes to the right of the "softmax" were added by the retraining script.

A bottleneck is an informal term we often use for the layer just before the final output layer that actually does the classification. "Bottelneck" is not used to imply that the layer is slowing down the network. We use the term bottleneck because near the output, the representation is much more compact than in the main body of the network.

Every image is reused multiple times during training. Calculating the layers behind the bottleneck for each image takes a significant amount of time. Since these lower layers of the network are not being modified their outputs can be cached and reused.

So the script is running the constant part of the network, everything below the node labeled Bottlene... above, and caching the results.

The command you ran saves these files to the bottlenecks/ directory. If you rerun the script, they'll be reused, so you don't have to wait for this part again.

More about Training

Once the script finishes generating all the bottleneck files, the actual training of the final layer of the network begins.

The training operates efficiently by feeding the cached value for each image into the Bottleneck layer. The true label for each image is also fed into the node labeled GroundTruth. Just these two inputs are enough to calculate the classification probabilities, training updates, and the various performance metrics.

As it trains you'll see a series of step outputs, each one showing training accuracy, validation accuracy, and the cross entropy:

The figures below show an example of the progress of the model's accuracy and cross entropy as it trains. If your model has finished generating the bottleneck files you can check your model's progress by opening TensorBoard, and clicking on the figure's name to show them. TensorBoard may print out warnings to your command line. These can safely be ignored.

A true measure of the performance of the network is to measure its performance on a data set that is not in the training data. This performance is measured using the validation accuracy. If the training accuracy is high but the validation accuracy remains low, that means the network is overfitting, and the network is memorizing particular features in the training images that don't help it classify images more generally.

The training's objective is to make the cross entropy as small as possible, so you can tell if the learning is working by keeping an eye on whether the loss keeps trending downwards, ignoring the short-term noise.

By default, this script runs 4,000 training steps. Each step chooses 10 images at random from the training set, finds their bottlenecks from the cache, and feeds them into the final layer to get predictions. Those predictions are then compared against the actual labels to update the final layer's weights through a backpropagation process.

As the process continues, you should see the reported accuracy improve. After all the training steps are complete, the script runs a final test accuracy evaluation on a set of images that are kept separate from the training and validation pictures. This test evaluation provides the best estimate of how the trained model will perform on the classification task.

You should see an accuracy value of between 85% and 99%, though the exact value will vary from run to run since there's randomness in the training process. (If you are only training on two classes, you should expect higher accuracy.) This number value indicates the percentage of the images in the test set that are given the correct label after the model is fully trained.

The retraining script writes data to the following two files:

Classifying an image

The codelab repo also contains a copy of tensorflow's label_image.py example, which you can use to test your network. Take a minute to read the help for this script:

python -m  scripts.label_image -h

As you can see, this Python program takes quite a few arguments. The defaults are all set for this project, but if you used a MobileNet architecture with a different image size you will need to set the --input_size argument using the variable you created earlier: --input_size=${IMAGE_SIZE}.

Now, let's run the script on this image of a daisy:


Image CC-BY by Retinafunk

python -m scripts.label_image \
    --graph=tf_files/retrained_graph.pb  \

Each execution will print a list of flower labels, in most cases with the correct flower on top (though each retrained model may be slightly different).

You might get results like this for a daisy photo:

daisy (score = 0.99071)
sunflowers (score = 0.00595)
dandelion (score = 0.00252)
roses (score = 0.00049)
tulips (score = 0.00032)

This indicates a high confidence (~99%) that the image is a daisy, and low confidence for any other label.

You can use label_image.py to classify any image file you choose, either from your downloaded collection, or new ones. You just have to change the --image file name argument to the script.


Image CC-BY by Lori Branham

python -m scripts.label_image \
    --graph=tf_files/retrained_graph.pb  \

The retraining script has several other command line options you can use.

You can read about these options in the help for the retrain script:

python -m scripts.retrain -h

Try adjusting some of these options to see if you can increase the final validation accuracy.

For example, the --learning_rate parameter controls the magnitude of the updates to the final layer during training. So far we have left it out, so the program has used the default learning_rate value of 0.01. If you specify a small learning_rate, like 0.005, the training will take longer, but the overall precision might increase. Higher values of learning_rate, like 1.0, could train faster, but typically reduces precision, or even makes training unstable.

You need to experiment carefully to see what works for your case.

After you see the script working on the flower example images, you can start looking at teaching the network to recognize different categories.

In theory, all you need to do is run the tool, specifying a particular set of sub-folders. Each sub-folder is named after one of your categories and contains only images from that category.

If you complete this step and pass the root folder of the subdirectories as the argument for the --image_dir parameter, the script should train the images that you've provided, just like it did for the flowers.

The classification script uses the folder names as label names, and the images inside each folder should be pictures that correspond to that label, as you can see in the flower archive:

Collect as many pictures of each label as you can and try it out!

Congratulations, you've taken your first steps into a larger world of deep learning!

You can see more about using TensorFlow at the TensorFlow website or the TensorFlow Github project. There are lots of other resources available for TensorFlow, including a discussion group and whitepaper.

If you make a trained model that you want to run in production, you should also check out TensorFlow Serving, an open source project that makes it easier to manage TensorFlow projects.

If you're interested in running TensorFlow on mobile devices try the second part of this tutorial which will show you how to optimize your model to run on Android.

Or just go have some fun in the TensorFlow Playground!

This codelab is based on Pete Warden's TensorFlow for Poets blog post and this retraining tutorial.