如何从 Python 训练代码中的 Vertex AI 托管数据集中加载图像？

Question

I am trying to create a custom training job in Vertex AI. I created a managed dataset stored in the same bucket I am exporting the training code to. I have a Python code that looks like this:

#Defining paths 
TRAIN_PATH = os.environ['AIP_TRAINING_DATA_URI']
VAL_PATH = os.environ['AIP_VALIDATION_DATA_URI']

#skipped model definition#

train_datagen = image.ImageDataGenerator(rescale = 1./255, shear_range = 0.2,zoom_range = 0.2, horizontal_flip = True)

test_dataset = image.ImageDataGenerator(rescale=1./255)

train_generator = train_datagen.flow_from_directory(
    TRAIN_PATH,
    target_size = (224,224),
    batch_size = 32,
    class_mode = 'binary')
validation_generator = test_dataset.flow_from_directory(
    VAL_PATH,
    target_size = (224,224),
    batch_size = 32,
    class_mode = 'binary')

hist_new = model.fit(
     train_generator, ...)

The question is, how do I load the images so the ImageDataGenerator can use them? The error I get when starting the training job is:

 No such file or directory: 'gs://(bucket name)/dataset-5820440723492700160-image_classification_multi_label-2022-05-29T10:53:33.245485Z/training-*'

Answer 1

It seems that the TRAIN_PATH and VAL_PATH should be valid local paths (this TF documentation does not mention other paths) and not GCS URIs. The data set is made available at the specified GCS URI but the training code should download the data set images from GCS to the local environment and then pass them to the ImageDataGenerator .

For information on downloading the data set from GCS, refer to this documentation .

Answer 2

In case you are using custom training container on Vertex, you can use gcs uris with FUSE filesystem , you don't have to do the mounting yourself, Vertex platform is taking care of that when you run a CustomJob simply read your paths as files:

'/gcs/(bucket name)/dataset-5820440723492700160-image_classification_multi_label-2022-05-29T10:53:33.245485Z/training-*'