example of train_input_fn method for Tensorflow Sagemaker Estimator

train_input_fn.py

def train_input_fn(training_dir, hyperparameters):
    
    #training directory is the path to our data in S3 bucket
    path=os.path.join(training_dir,'train.json')
    
    with open(path,'rb') as f:
        train=json.load(f)
    X_train=np.array(train['images'],dtype=np.float64)
    y_train=np.array(train['labels'],dtype=np.int64)    
        
    batch_size=hyperparameters.get('batch_size',32)
    
    # returns tuples of features and labels
    return tf.estimator.inputs.numpy_input_fn(
        x={"x": X_train},
        y=y_train,
        batch_size=batch_size,
        num_epochs=None,
        shuffle=True)()