Fetching contributors…
Cannot retrieve contributors at this time
259 lines (171 sloc) 12.7 KB

Classifying your own images using transfer learning

Note: This example works (and is fun to play with), but uses TensorFlow v1.2. So it doesn't demonstrate current best practices.

The Google Vision API is great for identifying labels, or categories, for a given image. However, sometimes you want to further classify your own images, in more specialized categories that the Google Vision API hasn't been trained on.

This lab shows how we can use an existing NN model to do this, via transfer learning -- effectively bootstrapping an existing model to reduce the effort needed to learn something new.

Specifically, we will take an 'Inception v3' architecture model trained to classify images against 1000 different 'ImageNet' categories, and using its penultimate "bottleneck" layer, train a new top layer that can recognize other classes of images: your own classes.

We'll see that our new top layer does not need to be very complex, and that we typically don't need much data or much training of this new model, to get good results for our new image classifications.

Transfer learning

In addition to the transfer learning, this example shows off several other interesting aspects of TensorFlow and Cloud ML. It shows how to use Cloud Dataflow (Apache Beam) to do image preprocessing -- the Beam pipeline uses Inception v3 to generate the inputs to the new 'top layer' that we will train -- and how to save those preprocessing results in TFRecords for consumption by the training loop.

The example also includes a little "prediction web server" that shows how you can use the Cloud ML API for prediction once your trained model is serving.

(This example is based on the example here, but with a number of additional modifications).

The "hugs/not-hugs" image classification task

Just for fun, we'll show how we can train our NN to decide whether images are of 'huggable' or 'not huggable' things.

So, we'll use a training set of images that have been sorted into two categories -- whether or not one would want to hug the object in the photo. (Thanks to Julia Ferraioli for this dataset).

The 'hugs' does not have a large number of images, but as we will see, prediction on new images still works surprisingly well. This shows the power of 'bootstrapping' the pre-trained Inception model.

(This directory also includes some scripts that support training on a larger 'flowers classification' dataset too.)

1. Image Preprocessing

We start with a set of labeled images in a Google Cloud Storage bucket, and preprocess them to extract the image features from the "bottleneck" layer -- essentially, the penultimate layer -- of the Inception network. To do this, we load the saved Inception model and its variable values into TensorFlow, and run each image through that model. (This model has been open-sourced by Google).

More specifically, we process each image to produce its feature representation (also known as an embedding) in the form of a k-dimensional vector of floats (in our case, 2,048 dimensions). The preprocessing includes converting the image format, resizing images, and running the converted image through a pre-trained model to get the embeddings.

The reason this approach is so effective for bootstrapping new image classification is that these 'bottleneck' embeddings contain a lot of high-level feature information useful to Inception for its own image classification.

Although processing images in this manner can be reasonably expensive, each image can be processed independently and in parallel, making this task a great candidate for Cloud Dataflow.

Important: If you have not already, makes sure to follow these instructions to enable the Cloud Dataflow API.

1.1 Deploy the preprocessing job to Cloud Dataflow

We need to run preprocessing for both our training and evaluation images. We've defined a script,, to do this.

First, set the BUCKET variable to point to your GCS bucket (replacing your-bucket-name with the actual name):


You may need to change file permissions to allow execution:

chmod +755 *.sh

Then, run the pre-processing script. The script is already set up with links to the 'hugs' image data. The script will launch two non-blocking Cloud Dataflow jobs to do the preprocessing for the eval and training datasets. By default it only uses 3 workers for each job, but you can change this if you have larger quota.

(Setting the USER environment variable allows Dataflow to distinguish multiple user's jobs.)


This script will generate a timestamp-based GCS_PATH, that it will display in STDOUT. The pipelines will write the generated embeds into TFRecords files containing tf.train.Example protocol buffers, under $GCS_PATH/preproc.

You can see your pipeline jobs running in the Dataflow panel of the Cloud console. Before you use these generated embeds, you'll want to make sure that the Dataflow jobs have finished.

2. Modeling: Training the classifier

Once we've preprocessed our data, and have the generated image embeds, we can then train a simple classifier. The network will comprise a single fully- connected layer with RELU activations and with one output for each label in the dictionary to replace the original output layer. The final output is computed using the softmax function. In the training stages, we're using the dropout technique, which randomly ignores a subset of input weights to prevent over-fitting to the training dataset.

2.1 For the workshop, use pre-generated TFRecords for training

Because we have limited workshop time, we've saved a set of generated TFRecords. If you didn't do this during installation, copy them now to your own bucket as follows.

Set the BUCKET variable to point to your GCS bucket (replacing your-bucket-name with the actual name), then copy the records to your bucket. Then, set the GCS_PATH variable to the newly copied GCS subfolder:

gsutil cp -r gs://tf-ml-workshop/transfer_learning/hugs_preproc_tfrecords $BUCKET

(As indicated above, with more time, you could wait for your Dataflow preprocessing jobs to finish running, then point to your own generated image embeds instead).

2.2 Run the training script

Now, using the value of GCS_PATH that you set above, run your training job in the cloud:


This script will output summary and model checkpoint information under $GCS_PATH/training.

2.3 Monitor the training

As the training runs, you can view the logs with:

gcloud ml-engine jobs stream-logs "$JOB_ID"

We can also monitor the progress of the training using Tensorboard.

To do this, start up Tensorboard in a new shell (don't forget to activate your virtual environment), pointing it to the training logs in GCS:

tensorboard --logdir=$GCS_PATH/training

Then, visit http://localhost:6006.

3. Prediction: Using the trained model

For prediction, we don't want to separate the image preprocessing and inference into two separate steps, because we need to perform both in sequence for every image. Instead, we create a single TensorFlow graph that produces the image embedding and does the classification using the trained model in one step.

After training, the saved model information will be in $GCS_PATH/training/model.

Our next step is to tell Cloud ML that we want to use and serve that model. We do that via the following script, where v1 is our model version name, and hugs is our model name.

./ $GCS_PATH v1 hugs

The model is created first. This only needs to happen once, and is done as follows: gcloud ml-engine models create <model_name>

Then, we create a 'version' of that model, based on the data in our model directory ($GCS_PATH/training/model), and set that version as the default.

You can see what models and default versions we have in your project via:

gcloud ml-engine models list

It will take a minute or so for the model version to start "serving". Once our model is serving, we make prediction requests to it -- both from the command line and via the Cloud ML API.

3.1 Prediction from the command line using gcloud

To make a prediction request from the command line, we need to encode the image(s) we want to send it into a json format. See the script for the details. This command:

python -o request.json <image1> <image2> ...

results in a request.json file with the encoded image info. Then, run this command:

gcloud ml-engine predict --model $MODEL_NAME --json-instances request.json

You should see a result something along the lines of the following:

gcloud ml-engine predict --model hugs --json-instances request.json
KEY                             PREDICTION  SCORES
prediction_images/hedgehog.jpg  1           [4.091006485396065e-05, 0.9999591112136841, 1.8843516969013763e-08]

The prediction index (e.g. '1') corresponds to the label at that index in the 'label dict' used to construct the example set during preprocessing, and the score for each index is listed under SCORES. (The last element in the scores list is used for any example images that did not have an associated label).

The 'hugs' dataset label dict is here:

gsutil cat gs://oscon-tf-workshop-materials/transfer_learning/cloudml/hugs_photos/dict.txt

So, that means that index 0 is the 'hugs' label, and index 1 is 'not-hugs'. Therefore, the prediction above indicates that the hedgehog is 'not-hugs', with score 0.9999591112136841.

3.2 Prediction using the Cloud ML API: A prediction web server

We can also request predictions via the Cloud ML API, and the google api client libraries. We've included a little example web server that shows how to do this, and also makes it easy to see how a given image gets labeled.

The web app lets you upload an image, and then it generates a json request containing the image, and sends that to the Cloud ML API. In response, we get back prediction and score information, and display that on a result page.

Run the web server according to the instructions in its README.

Appendix: Running training locally

If you want to run the training session locally (this can be useful for debugging), you will need to point the GOOGLE_APPLICATION_CREDENTIALS environment variable to a local service account credentials file like this:

export GOOGLE_APPLICATION_CREDENTIALS=/path/to/credentials/file.json

Then initiate the local training like this, defining a local output path to use:

gcloud gcloud ml-engine local train --package-path trainer/ --module-name trainer.task \
    -- \
    --max-steps 1000 \
    --train_data_paths "$GCS_PATH/preproc/train*" \
    --eval_data_paths "$GCS_PATH/preproc/eval*" \
    --eval_set_size 19 \
    --output_path output/

Appendix: image sources

The source information for the 'hugs/no-hugs' images is here: gs://oscon-tf-workshop-materials/transfer_learning/hugs_photos_sources.csv.