This post aims to help you deploy GFPGAN on AWS using ECS Fargate.
What is GFPGAN?!
I've recently discovered this fantastic Machine Learning project trained and designed to restore faces.
The project's headline is:
GFPGAN aims at developing Practical Algorithms for Real-world Face Restoration.
In short, given a blurred, damaged, or low-quality photo of a person's face, the algorithm it's able to restore the face and improve its quality.
Example
GFPGAN is a new AI model for photo restoration. Nothing was manually adjusted in this example. 😮👍 https://t.co/AwNRxw6bZ3 #ai #ml #ArtificialIntelligence #MachineLearning #photography pic.twitter.com/EYS5c4FiWd
— Metin Seven (@metinse7en) June 17, 2021
The challenge
So, I decided to write this blog post because I recently deployed this tool on AWS, and there were a few things that I wish I had known before starting.
GFPGAN is written in Python, and its entry point assumes you're going to convert a batch of images, reading and writing those from disk.
Now, I am not sure about your use case, but this was a significant limitation for me.
If you're intimidated by Python, don't worry. The entry point is relatively simple to modify. I didn't have any Python experience before this.
My use case is to provide an HTTP API to a front-end application to call and retrieve the enhanced image.
Architecture
To provide such API, I came up with this architecture that is currently working fine in my side-project.
To better understand the flow, this is a sequence diagram:
How to Use it
Before you start getting your hands dirty with Terraform or CloudFormation code, I recommend trying out the Docker image locally.
Step 1 - Fork the repo
At the moment of writing this, the official repository has an open PR for adding the Dockerfile, shout out to Mostafa Elmenbawy for his valuable contribution!
I also want to personally thank Mostafa for helping me understand how GFPGAN Python entry point works!
I recommend forking his repo until the PR has been merged.
Step 2 - Build the image
docker build -t gfpgan .
Step 3 - Run the container
docker run \
--name gfpgan \
--rm \
--volume <PROJECT_ABSOLUTE_PATH>/inputs:/app/inputs \
--volume <PROJECT_ABSOLUTE_PATH>/results:/app/results \
gfpgan \
python3 inference_gfpgan.py --upscale 2 --test_path inputs --save_root results
The command above will start the container that, by running GFPGAN, will transform the default inputs photos and produce outputs in the results folder.
How to Deploy It
Before proceeding, please note that I've adapted the Python entry point to integrate with SQS to receive jobs, S3 to persist the enhanced image files, and DynamoDB as a temporary state to inform the web-app about the status of the job.
Step 1 - Customize the entry point
I have changed inference_gfpgan.py as follows:
import boto3
import cv2
import glob
import numpy as np
import os
import json
import requests
import logging
import time
import uuid
from pythonjsonlogger import jsonlogger
from basicsr.utils import imwrite
from pathlib import Path
from gfpgan import GFPGANer
sqs = boto3.resource("sqs")
queue = sqs.get_queue_by_name(QueueName=os.environ["SQS_QUEUE_NAME"])
dynamodb = boto3.resource("dynamodb")
table = dynamodb.Table(os.environ["DYNAMODB_TABLE_NAME"])
s3 = boto3.client("s3")
for name in ["boto", "urllib3", "s3transfer", "boto3", "botocore", "nose"]:
logging.getLogger(name).setLevel(logging.CRITICAL)
def setup_logging(log_level):
logger = logging.getLogger()
logger.setLevel(log_level)
json_handler = logging.StreamHandler()
formatter = jsonlogger.JsonFormatter(
fmt="%(asctime)s %(levelname)s %(name)s %(message)s"
)
json_handler.setFormatter(formatter)
logger.addHandler(json_handler)
def main(logger, message):
[f.unlink() for f in Path("./inputs").glob("*") if f.is_file()]
[f.unlink() for f in Path("./results").glob("*") if f.is_file()]
model_path = "experiments/pretrained_models/GFPGANCleanv1-NoCE-C2.pth"
upscale = 2
arch = "clean"
channel = 2
ext = "auto"
suffix = None
save_root = "results"
os.makedirs(save_root, exist_ok=True)
inputs_path = os.path.join("inputs", "input.jpg")
img_data = requests.get(url=message["imageUrl"]).content
logger.info(f"Downloading image {message['imageUrl']} to {inputs_path}")
try:
with open(inputs_path, "wb") as handler:
handler.write(img_data)
except:
logger.error('Cannot download file')
return
restorer = GFPGANer(
model_path=model_path,
upscale=upscale,
arch=arch,
channel_multiplier=channel,
bg_upsampler=None,
)
img_list = sorted(glob.glob(os.path.join("inputs", "*")))
for img_path in img_list:
# read image
img_name = os.path.basename(img_path)
logger.info(f"Processing {img_name} ...")
basename, ext = os.path.splitext(img_name)
input_img = cv2.imread(img_path, cv2.IMREAD_COLOR)
# restore faces and background if necessary
cropped_faces, restored_faces, restored_img = restorer.enhance(
input_img, has_aligned=False, only_center_face=True, paste_back=True
)
# save faces
for idx, (cropped_face, restored_face) in enumerate(
zip(cropped_faces, restored_faces)
):
# save cropped face
save_crop_path = os.path.join(
save_root, "cropped_faces", f"{basename}_{idx:02d}.png"
)
imwrite(cropped_face, save_crop_path)
# save restored face
if suffix is not None:
save_face_name = f"{basename}_{idx:02d}_{suffix}.png"
else:
save_face_name = f"{basename}_{idx:02d}.png"
save_restore_path = os.path.join(
save_root, "restored_faces", save_face_name
)
imwrite(restored_face, save_restore_path)
# save comparison image
cmp_img = np.concatenate((cropped_face, restored_face), axis=1)
imwrite(
cmp_img, os.path.join(save_root, "cmp", f"{basename}_{idx:02d}.png")
)
if restored_img is not None:
if ext == "auto":
extension = ext[1:]
else:
extension = ext
if suffix is not None:
save_restore_path = os.path.join(
save_root, "restored_imgs", f"{basename}_{suffix}.{extension}"
)
else:
save_restore_path = os.path.join(
save_root, "restored_imgs", f"{basename}.{extension}"
)
logger.info(f"Saving restored image to {save_restore_path} ...")
imwrite(restored_img, save_restore_path)
s3key = f'gfpgan_{message["orgId"]}_{uuid.uuid4().hex}.jpg'
s3.upload_file(
"results/restored_imgs/input..jpg", os.environ["S3_BUCKET_NAME"], s3key
)
new_record = message.copy()
new_record["enhancedImageS3Key"] = s3key
new_record["updatedAt"] = round(time.time())
new_record["timeTaken"] = new_record["updatedAt"] - new_record["createdAt"]
update_record(new_record)
logger.info(f"Result uploaded in S3: {s3key}")
def create_record(message):
response = table.put_item(
Item={
"enhancementId": message["enhancementId"],
"imageUrl": message["imageUrl"],
"createdAt": message["createdAt"],
}
)
return response
def update_record(message):
response = table.put_item(
Item={
"enhancementId": message["enhancementId"],
"imageUrl": message["imageUrl"],
"createdAt": message["createdAt"],
# Additional Props
"enhancedImageS3Key": message["enhancedImageS3Key"],
"updatedAt": message["updatedAt"],
"timeTaken": message["timeTaken"],
}
)
return response
def retrieve_messages():
setup_logging(logging.INFO)
logger = logging.getLogger()
logger.info("Starting retrieve_messages")
while True:
logger.debug("Fetching SQS messages")
messages = queue.receive_messages(MaxNumberOfMessages=1, WaitTimeSeconds=1)
for message in messages:
parsed_message = json.loads(message.body)
parsed_message["createdAt"] = round(time.time())
logger.info(parsed_message)
create_record(parsed_message)
main(logger, parsed_message)
message.delete()
time.sleep(5)
if __name__ == "__main__":
retrieve_messages()
Note: the code above pulls new messages from SQS every 5 seconds.
The SQS message should follow this schema:
{"enhancementId": "12345", "imageUrl": "https://example.com/photo.jpg"}
Step 2 - Update the Dockerfile
I've also updated the Dockerfile as follows. Again, shout out to Mostafa Elmenbawy for creating the initial Dockerfile! 🙏
FROM nvidia/cuda:10.0-cudnn7-devel
WORKDIR /app
RUN apt-get update && apt-get install -y --no-install-recommends --fix-missing \
# python
python3.8 python3-pip python3-setuptools python3-dev \
# OpenCV deps
libglib2.0-0 libsm6 libxext6 libxrender1 libgl1-mesa-glx \
# c++
g++ \
# others
wget unzip
# Ninja
RUN wget https://github.com/ninja-build/ninja/releases/download/v1.8.2/ninja-linux.zip && \
unzip ninja-linux.zip -d /usr/local/bin/ && \
update-alternatives --install /usr/bin/ninja ninja /usr/local/bin/ninja 1 --force
# basicsr facexlib
RUN python3 -m pip install --upgrade pip && \
pip3 install --no-cache-dir torch>=1.7 opencv-python>=4.5 && \
pip3 install --no-cache-dir basicsr facexlib realesrgan boto3 python-json-logger uuid
# weights
RUN wget https://github.com/TencentARC/GFPGAN/releases/download/v0.2.0/GFPGANCleanv1-NoCE-C2.pth \
-P experiments/pretrained_models &&\
wget https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/GFPGANv1.pth \
-P experiments/pretrained_models && \
wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth \
-P /usr/local/lib/python3.6/dist-packages/realesrgan/weights/ && \
wget https://github.com/xinntao/facexlib/releases/download/v0.1.0/detection_Resnet50_Final.pth \
-P /usr/local/lib/python3.6/dist-packages/facexlib/weights/
RUN rm -rf /var/cache/apt/* /var/lib/apt/lists/* && \
apt-get autoremove -y && apt-get clean
COPY requirements.txt .
RUN pip3 install --no-cache-dir -r requirements.txt
COPY . .
RUN pip3 install .
CMD ["python3", "inference_gfpgan.py"]
Step 3 - Publish the image to ECR
In your CI/CD, build the Dockerfile and publish the image to ECR.
Step 4 - Deploy the ECS task using Fargate
Using Terraform or CloudFormation, create and deploy the ECS task. Please note, this service doesn't need to be exposed to the internet via Route53 or an Application Load Balancer.
If you have deployed APIs or public-facing services before to ECS, you might be familiar with configuring the Target Group's healthcheck. The healthcheck detects if the main process running in the container is still health or the container needs to be restarted.
In this case, the container doesn't expose an HTTP endpoint for the Target Group to call, so we have to use container level healthchecks.
Instead of configuring the healthcheck in the Target Group, container-level healthchecks are defined in the ECS task definition, as follows:
...
"healthCheck": {
"command": [ "CMD-SHELL", "exit 0" ],
"startPeriod": 20
},
...
Please note, the CMD-SHELL can be any bash command. Ideally, the command would help you validate if the process is healthy.
By using exit 0, you assume that the container is always healthy.
This is not ideal, but I haven't found an alternative approach yet.
Takeaways
GFPGAN:
- is a ML project able to restore damaged or low-quality photos
- is written in Python but the interface is fairly easy to modify, even if you're not a Python Developer
- can be containerized in Docker
- can be deployed to ECS Fargate, I am using 2 GB o RAM and 1 CPU
- you can easily integrate it with a queue system and integrate it with your app via an API
Conclusions
Hopefully, I've provided you with a real-world scenario in which GFPGAN is used and deployed to AWS.
I didn't include the entire Terraform code to keep the blog post readable, but I am happy to share more.
Just reach out to me if you need any help with this; I am more than happy to give you a hand deploying GFPGAN to AWS.
Top comments (8)
What if we want to use the GPU? What would be the best option to deploy?
Hey! Sorry I haven't seen this comment before. I moved my deployment to a GPU one because it's much faster. I am using Vast AI which works very well and it's very affordable. Let me know if you need any help with that ))
@andreafalzetti How's the scaling like on VastAI? Do we have to use Kubernetes or anything? or do we just deploy the docker image and it'll auto-scale?
even if we deploy on gpu will the docker image will be same ?
very similar, if you need help, reach out to me and I can assist further!
I would gladly to see a multi part article where you step by step go through on each part. How to set up amazon, how to create each service and connects them together. I'm planning to create a similar service but to use real-esrgan but I'm not home in amazon services.
how long take time to compute in cpu and gpu ?
CPU can take up to a couple of minutes. With GPU is mostly seconds.