AWS Sagemaker script example

[Medicmind]

If anyone is interested in running an AWS Sagemaker estimator job using keras_cv_attention_models here is a sample python script. To prepare your data you should create a directiory 'preprocessed' under your 'keras_cv_attention_models' and then place your data under this directory with a train.json file with all the classes like below which include two classes A and B:

{
"train": [
{
"image": "preprocessed/B/20051020_53178_0100_PP.jpg",
"label": 1
},
{
"image": "preprocessed/A/20051020_43832_0100_PP.jpg",
"label": 0
},
{
"image": "preprocessed/A/20051020_56592_0100_PP.jpg",
"label": 0
},
],
"test": [
{
"image": "preprocessed/B/20051020_44349_0100_PP.jpg",
"label": 1
},
{
"image": "preprocessed/A/20051020_62709_0100_PP.jpg",
"label": 0
},
{
"image": "preprocessed/B/20051020_57844_0100_PP.jpg",
"label": 1
},
{
"image": "preprocessed/B/20051205_35392_0400_PP.jpg",
"label": 1
},
],
"indices_2_labels": {
"1": "1",
"0": "0"
},
"info": {
"num_classes": 2
}
}

And then execute the following estimator code with your S3 bucket:

import sagemaker
from sagemaker import get_execution_role
from sagemaker.estimator import Estimator
import boto3
from sagemaker.tensorflow import TensorFlow
import sys
import time

sagemakerbucket='your-bucket'
sagemaker_session = sagemaker.Session() # Configure the hyper-parameters

role =get_execution_role()

entry_point='train_script.py'

hyperparameters = {
  's': "EfficientNetV2S",
  'm': "efficientnet.EfficientNetV2S",  
  'd': "preprocessed/train.json",
  'b': 16,
  'pretrained': "imagenet",
  "i": 384,
  'save_path': "/opt/ml/model",
  'final_model_dir': "/opt/ml/model",
}
 
jobname='test'

estimator = TensorFlow(
    entry_point=entry_point, #
    role=role,
    source_dir="keras_cv_attention_models", 
    instance_count=1, # number of instances
    instance_type='ml.g4dn.xlarge',
    framework_version="2.8", # TensorFlow 2.6
    py_version="py39",
    sagemaker_session=sagemaker_session,
    hyperparameters=hyperparameters,
	max_run=60*60,
    debugger_hook_config=False,
    # Training using SMDataParallel Distributed Training Framework
    distribution={"parameter_server": {"enabled": True}},
)# Submit SageMaker training job

estimator.fit( job_name=jobname) 

[leondgarse]

Beautiful!