Firstly, follow the setup guide to install the Google Cloud Machine Learning SDK. This will also ask you to install TensorFlow.
Set environment variables:
$> SRC_ROOT=wherever/you/cloned/the/files
Install prerequisites:
$> cd $SRC_ROOT
$> pip install -r requirements.txt
Download and prepare the data:
$> # NOTE: you need a GCP service account saved as $SRC_ROOT/tf/face_extract/vapi-acct.json
$> # to call Google Cloud Vision API
$> #
$> ./get_data.sh
$> #
$> # lots of output follows
For more details on the python scripts called by get_data.sh, see the "Getting and preparing the training data" section below.
Move the prepared data to /tmp, where the training code expects to find them by default:
$> mv data /tmp
When training the model, make sure to specify the correct number of training classes (--num_classes) and number of samples in your validation set (--valid_batch_size). This will differ depending on the number of files you've downloaded and how the data has been divided.
Check the training source for other flags you can specify.
$> cd tf
$> gcloud ml-engine local train \
--package-path=pubfig \
--module-name=pubfig.task \
-- \
--num_classes=<number_of_classes> \
--valid_batch_size=<number_of_validation_samples>
$> #
$> # lots of output follow
Your trained model will be exported to /tmp/model/ by default.
Install the Cloud ML SDK and prepare your environment per the setup guide.
Example:
$> gcloud ml-engine jobs submit training example_job123 \
--package-path=pubfig \
--module-name=pubfig.task \
--region=us-central1 \
--staging-bucket=gs://wwoo-train \
--runtime-version=1.2
The job output will be similar to the below. In this case, training terminates once 75% validation accuracy is reached.
14:13:29.600 Validating job requirements...
14:13:31.063 Job creation request has been successfully validated.
14:13:31.183 Job pubfig7 is queued.
14:13:36.350 Waiting for job to be provisioned.
14:20:35.038 Running command: python -m pubfig.train_log
14:20:36.915 Recursively copying from gs://wwoo-train/pubfig/out.tar.gz to /tmp/
14:20:45.528 get_image_label_list: read 4416 items
14:20:45.629 get_image_label_list: read 336 items
...
18:42:03.778 Step [6700] (valid): accuracy: 0.703125, loss: 3.42011
18:45:55.734 Step [6800] (train): accuracy: 0.729167, loss: 3.08664
18:46:04.288 Step [6800] (valid): accuracy: 0.708333, loss: 3.33873
18:49:49.276 Step [6900] (train): accuracy: 1.0, loss: 0.0014053
18:49:57.623 Step [6900] (valid): accuracy: 0.752604, loss: 2.98494
18:49:57.624 Step [6900] (complete)
18:49:57.866 Recursively copying from /tmp/logs to gs://wwoo-train/pubfig/export/
18:49:59.649 Recursively copying from /tmp/model to gs://wwoo-train/pubfig/export/
18:50:01.943 Module completed; cleaning up.
18:50:01.944 Clean up finished.
18:50:01.945 Task completed successfully.
18:50:25.382 Tearing down TensorFlow.
18:51:09.881 Finished tearing down TensorFlow.
18:51:16.318 Job completed successfully.
The model can be trained to 80% validation accuracy with 48 classes (face categories), using 4402 training and 336 validation samples. With the default hyperparameters, overfitting started to occur past ~1.2K steps using a learning rate of 0.01.
Orange = training set, Blue = validation set.
Cross entropy loss:
Classification accuracy:
This section describes the Python scripts called by get_data.sh. Provided everything completed successfully, you should not need to call these scripts directly to retrieve and process the data.
The model is trained using a subset of data from PubFig. PubFig provides a development set and evaluation set of images, with no people or sample overlaps between the two. For our face recognition use case, we will just use the evaluation dataset and split these further into training and validation.
PubFig provides only the links to images on the public web, not the images themselves. Therefore, it is necessary to download them separately using pubfig_get.py. Some images may not be downloaded successfully due to technical issues such as broken links, removed content, unreachable servers and so-on.
tf/face_extract/pubfig_get.py - Use this to download one of the PubFig datasets. I used data from the evaluation set. Save this as eval_urls.txt in the same directory as pubfig_get.py. pubfig_get.py will also crop faces using face vertices supplied in the PubFig metadata, but I encountered incorrect vertices in some cases. Set IMAGE_CROP = True if you want the script to crop out the faces using the supplied PubFig metadata.
Example invocation to read from eval_urls.txt and save to the ./data directory:
$> python tf/face_extract/pubfig_get.py tf/face_extract/eval_urls.txt ./data
pubfig_get.py generates a manifest.txt file, which is a list of local file paths to the downloaded files. You will likely see some duplicates, due to conflicting filenames that pubfig_get.py would have overwritten [TODO - fix]. Remove the duplicates using:
$> cat ./data/manifest.txt | sort | uniq > ./data/manifest_uniq.txt
tf/face_extract/crop_faces.py - Crop faces using Google Vision API, which is far more accurate than using the PubFig supplied metadata. You need to sign up for a Google Cloud Platform account to use the Vision API.
Follow the Vision API Quickstart to enable the Vision API on your Google Cloud Platform project. You will also need to generate a service account that crop_faces.py can use to call the Vision API with. Save the service account JSON as vapi-acct.json in the same directory as crop_faces.py.
Cropped files are saved in a crop directory in the same parent directory as the original file.
Example invocation to crop all files from the paths in manifest_uniq.txt, with the current working directory prepended to each file path (as paths are relative in ./data/manifest_uniq.txt):
$> python tf/face_extract/crop_faces.py ./data/manifest_uniq.txt $PWD
Note: Using Vision API will cost you money, though you can always sign up for a free Google Cloud Platform account with $300 in credits. It is your responsibility to manage your own usage.
tf/face_extract/split_data.py - Splits the cropped data into training and validation sets. You can adjust a number of factors in splitting the data - for example, the ratio of training to validation data (SPLIT_FACTOR), the minimum & maximum samples for a class to be included (MIN_SAMPLES and MAX_SAMPLES) and set a minimum/maximum skew (roll, pan and tilt) for a sample to be included [TODO: expose in more friendly way].
Example invocation to read from ./data/vision-manifest.txt, and write the training and validation dataset (as a set of paths) to train.txt and valid.txt.
$> python tf/face_extract/split_data.py ./data/vision-manifest.txt ./data/train.txt ./data/valid.txt
tf_face/web/ contains source which can be deployed to Google App Engine.
You will need to modify tf_face/web/main.py to match your project ID and model name. Also replace resources/vapi-acct.json.replaceme with your service account key.
Deploy using:
$> cd ${SRC_ROOT}/tf/web
$> gcloud app deploy
$>
$> # lots of output follows
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent