Create the lambda function for image classification

Then, we'll create the lambda function that downloads images from S3 and inferences with TensorFlow Lite.

1. Click "Functions" in the sidebar of the Lambda console.

2. Click "Create function".

3. Choose "Author from scatch". Fill all the following attributes and click "Create function":

• "Function name": "image-classification"

• "Runtime": "Python 3.9"

• "Permissions": click "Change default execution role". Choose "Use an existing role" and select "LabRole".

4. Then it will jump to the page of "image-classification".

5. Finish the following 2 steps. [Step 1] Create a bucket in S3 for uploading the two files below. The bucket should be named as "cslab-resources-{your_student_id}". For details about bucket creation and file uploading, refer to Section 2 and Section 7 of this lab, respectively.

4MB

mobilenet_v1_1.0_224_quant.tflite

10KB

labels_mobilenet_quant_v1_224.txt

[Step 2] Go back to the page of "image-classification". Enter the following code in "Function code" with a replacement of {your_student_id}:

import logging
import os
import sys
import tempfile
import boto3
from PIL import Image
import numpy as np
import tflite_runtime.interpreter as tflite

S3_MODEL_PATH = "s3://cslab-resources-{your_student_id}/mobilenet_v1_1.0_224_quant[1].tflite"
S3_LABEL_PATH = "s3://cslab-resources-{your_student_id}/labels_mobilenet_quant_v1_224[1].txt"

logging.basicConfig(level=logging.INFO)

# Create a image classification model
def create_model(model_file_path):
    interpreter = tflite.Interpreter(
        model_path=model_file_path, num_threads=1)
    interpreter.allocate_tensors()
    return interpreter


def load_labels(label_file_path):
    with open(label_file_path, 'r') as f:
        return [line.strip() for line in f.readlines()]

def download_s3_file(s3_client, s3_path):
    path_parts = s3_path.replace("s3://", "").split("/")
    bucket = path_parts.pop(0)
    key = "/".join(path_parts)
    tmp = tempfile.NamedTemporaryFile(delete=False)
    s3_client.download_file(bucket, key, tmp.name)
    return tmp.name

logging.info('Initializing the image classification model')
dynamodb = boto3.client('dynamodb')
s3 = boto3.client('s3')
model_path = download_s3_file(s3, S3_MODEL_PATH)
label_path = download_s3_file(s3, S3_LABEL_PATH)
model = create_model(model_path)
input_details = model.get_input_details()
output_details = model.get_output_details()
labels = load_labels(label_path)
tablename = os.environ["dynamodb_table_name"]


# logic for inference a image with a local image path,
# return the image class with the probability
def inference(image_path):
    image = Image.open(image_path).resize((224,224))
    image = np.expand_dims(image, axis=0)
    model.set_tensor(input_details[0]['index'], image)
    model.invoke()
    inference_ret = np.squeeze(model.get_tensor(output_details[0]['index']))
    label_index = inference_ret.argsort()[-1]
    image_class_name = labels[label_index]
    probability = inference_ret[label_index]
    return (image_class_name, probability)

def write_result_to_table(image_path, image_class_name):
    dynamodb.put_item(TableName=tablename,
                      Item={'image_name': {'S': image_path},
                            'image_class_name': {'S': image_class_name}})


# Lambda handler, it download the image from S3 and do the inference
def lambda_handler(event, context):
    logging.info('The image classification lambda model is triggered by S3')
    for record in event['Records']:
        bucket = record['s3']['bucket']['name']
        key = record['s3']['object']['key']
        logging.info(f'Processing image from S3 path: {bucket}/{key}')

        try:
            # load image
            with tempfile.NamedTemporaryFile() as tmp:
                s3.download_file(bucket, key, tmp.name)

                (prediction_label, prediction_prob) = inference(tmp.name)

                write_result_to_table(f"{bucket}/{key}", prediction_label)
        except Exception as e:
            logging.error(f"Fail to process image with path: {bucket}/{key}",e)


# Main for local testing
if __name__ == "__main__":
    print(download_s3_file(s3, S3_LABEL_PATH))

6. Click "Deploy" to deploy the code to the platform.