Pytorch deployment failing with unexpected errors

[carlomazzaferro]

Please fill out the form below.

System Information

• Framework: Pytorch
• Framework Version: 1.0.0
• Python Version: 3
• CPU or GPU: CPU
• Python SDK Version: 1.18.4
• Are you using a custom image: No

Describe the problem

Model deployment fails with cryptic errors. See the logs below. The command issued to deploy the model is the following:

MODEL_PATH = 's3:///sagemaker-us-east-2-971148336196/improved-ner-training-0-25-0/output/model.tar.gz'

MODEL_NAME = 'improved-ner-model-model-' + os.environ['ENVIRONMENT']
ENDPOINT_NAME = 'improved-ner-model-sagemaker-endpoint-' + os.environ['ENVIRONMENT']

DEPLOY_INSTANCE = 'ml.m5.large'

model = PyTorchModel(model_data=MODEL_PATH, role=ROLE, entry_point='train_model.py',
 sagemaker_session=sm_session, py_version='py3', framework_version='1.0.0',
 name=ENDPOINT_NAME)

model.deploy(initial_instance_count=1, instance_type=DEPLOY_INSTANCE, endpoint_name=ENDPOINT_NAME)

The model is publicly available here:
https://s3.us-east-2.amazonaws.com/sagemaker-us-east-2-971148336196/improved-ner-training-0-25-0/output/model.tar.gz

It contains a directory called flair which contains the final_model.pt

The (relevant) part of the train_model.py script is the following:

def model_fn(model_dir):
 f_out = os.path.join(model_dir, 'flair')
 m = SequenceTagger.load_from_file(os.path.join(f_out, 'final-model.pt'))
 return m

def input_fn(request_body, request_content_type):
 if request_content_type.lower() != 'application/json':
 raise ValueError('Content type must be application/json')

 if 'sentence' not in request_body:
 raise ValueError('Request must be JSON formatted with key: sentence')
 return request_body['sentence']


def predict_fn(input_data, model):
 return model.predict(input_data)


if __name__ == "__main__":
 args, _ = parse_args()

 flair_out = os.path.join(args.model_dir, 'flair')
 trainer(flair_out) # This trains a model using flair.trainer.ModelTrainer

 model = SequenceTagger.load_from_file(os.path.join(flair_out, 'final-model.pt'))
 # create example sentence
 sentence = Sentence('I love Berlin')

 # predict tags and print
 model.predict(sentence)

Minimal repro / logs

The CloudWatch logs are very opaque. One of the errors is the following:

sagemaker_containers._errors.ClientError: [Errno 30] Read-only file system: '/opt/ml/model/flair/final-model.pt'

Then, much later, these errors pop up:

Processing /opt/ml/code
Could not install packages due to an EnvironmentError: [Errno 2] No such file or directory: '/tmp/pip-req-tracker-27gca9by/35241637574d11bf9bde50616c67372a334f94fa8356bc7164af8ca3'
You are using pip version 18.1, however version 19.0.3 is available.
[2019-04-12 03:49:26 +0000] [25] [ERROR] Error handling request /ping

Any ideas of what is actually causing the error, or some other steps to take to make it easier to debug?

[carlomazzaferro]

Got to the bottom of it, and fixed it with quite an unsavory hack:

def model_fn(model_dir):
 try:
 os.system('cp /opt/ml/model/final-model.pt /etc')
 except Exception as e:
 print("os.system('cp /opt/ml/model/final-model.pt /etc') failed because: ", e)

 return SequenceTagger.load_from_file('/etc/final-model.pt')

Apparently the /opt/ml/model directory is not in the same fs and hence is not writable. The reason why it needs to be writable is the specific to how flair loads models. See here.

I'll close this for now, but I'd really like to know if there is a more efficient way of testing the deployments locally. Granted, the docker containers are made available for Pytorch and Tensorflow, but there is no clear documentation on the steps required to build the images with the required entry_point script, model_data, etc. locally and run the inference container as it would on the SageMaker environment.

[ShaneRyan1977]

Hi Carlo,

I found your post: #752
I'm wondering if you could answer one thing. Did you have any issues getting flair installed? In the endpoint docker image? What version of flair are you using?

My issue is that I keep getting "No module named 'flair'". Your model.tar.gz file is no longer available. Did that have something else inside it other than the pt or pth file?

Thanks in advance! I am completely stumped.

Here is the output from my Cloudwatch
bash: cannot set terminal process group (-1): Inappropriate ioctl for device
bash: no job control in this shell
[2019-06-13 18:43:31 +0000] [21] [INFO] Starting gunicorn 19.9.0
[2019-06-13 18:43:31 +0000] [21] [INFO] Listening at: unix:/tmp/gunicorn.sock (21)
[2019-06-13 18:43:31 +0000] [21] [INFO] Using worker: gevent
[2019-06-13 18:43:31 +0000] [28] [INFO] Booting worker with pid: 28
[2019-06-13 18:43:31 +0000] [29] [INFO] Booting worker with pid: 29
[2019-06-13 18:43:31 +0000] [33] [INFO] Booting worker with pid: 33
[2019-06-13 18:43:31 +0000] [37] [INFO] Booting worker with pid: 37
Processing /opt/ml/code
Building wheels for collected packages: mnist Running setup.py bdist_wheel for mnist: started Running setup.py bdist_wheel for mnist: finished with status 'done' Stored in directory: /tmp/pip-ephem-wheel-cache-ufs3gl8c/wheels/35/24/16/37574d11bf9bde50616c67372a334f94fa8356bc7164af8ca3
Successfully built mnist
Installing collected packages: mnist
Successfully installed mnist-1.0.0
You are using pip version 18.1, however version 19.1.1 is available.
You should consider upgrading via the 'pip install --upgrade pip' command.
[2019-06-13 18:44:37 +0000] [28] [ERROR] Error handling request /ping
Traceback (most recent call last): File "/usr/local/lib/python3.5/dist-packages/sagemaker_containers/_functions.py", line 84, in wrapper return fn(*args, **kwargs) File "/usr/local/lib/python3.5/dist-packages/mnist.py", line 159, in model_fn model.load_state_dict(torch.load(f)) File "/usr/local/lib/python3.5/dist-packages/torch/serialization.py", line 303, in load return _load(f,
ImportError: No module named 'flair'

[carlomazzaferro]

Hey @ShaneRyan1977 AFAIK there are a couple of ways, one a bit hackier than the other. First one, is create a custom docker image:

# For more information on creating a Dockerfile
# https://docs.docker.com/compose/gettingstarted/#step-2-create-a-dockerfile
# https://github.com/awslabs/amazon-sagemaker-examples/master/advanced_functionality/pytorch_extending_our_containers/pytorch_extending_our_containers.ipynb
# SageMaker PyTorch image

ARG REGION

# These are the base images for pytorch-sagemaker 
FROM 520713654638.dkr.ecr.${REGION}.amazonaws.com/sagemaker-pytorch:1.0.0-cpu-py3 

RUN pip install flair # Install flair, and whatever else you need

ENV PATH="/opt/ml/code:${PATH}"

# /opt/ml and all subdirectories are utilized by SageMaker, we use the /code subdirectory to store our user code.
COPY /ner /opt/ml/code

# this environment variable is used by the SageMaker PyTorch container to determine our user code directory.
ENV SAGEMAKER_SUBMIT_DIRECTORY /opt/ml/code

You can then push that image to a private ECR repo with something like this:

algorithm_name=sagemaker-pytorch-ner
account=$(aws sts get-caller-identity --query Account --output text)

echo "ACCOUNT: ${account}"

region='us-east-1'

fullname="${account}.dkr.ecr.${region}.amazonaws.com/${algorithm_name}:latest"

# Get the login command from ECR and execute it directly
$(aws ecr get-login --region ${region} --no-include-email)

# Get the login command from ECR in order to pull down the SageMaker PyTorch image
$(aws ecr get-login --registry-ids 520713654638 --region ${region} --no-include-email)

# Build the docker image locally with the image name and then push it to ECR
# with the full name.

docker build -t ${algorithm_name} . --build-arg REGION=${region}
docker tag ${algorithm_name} ${fullname}

docker push ${fullname}

Then on training, you can specify your image as a base image to use: https://sagemaker.readthedocs.io/en/stable/sagemaker.pytorch.html#pytorch-estimator see the image_name parameter.

Alternatively, you can add a couple of lines at the very top of your training code:

import pip
pip._internal.main(['install', 'flair'])

import flair
import pytorch
etc....

Lastly, you can also add a bash file as your entry point. That bash script can run pip commands, and in the end it would call python train.py. There are some gotchas as far as using the bash command though, haven't had much success doing so.

[ShaneRyan1977]

Thanks for the incredibly quick reply. But I'm having a problem with deploying a model to an endpoint rather than training.

I actually trained the model using flair in a JupyterLab. Then I uploaded the model.tar.gz to s3.

I am trying to do something almost identical to your original code but deploying a trained model to an endpoint.

But I keep getting the errors I pasted.
I created a custom docker image with python 3.6 installed and python 3.5 not installed. But for some reason the code that runs in "/opt/ml/code" keeps installing python3.5 and then I get an error in Cloudwatch :
ImportError: No module named 'flair'

I am completely stumped. I am also amazed that you didn't have to build a custom image for deploying your endpoint!
Clearly something happens when the model is decompressed in:

[ShaneRyan1977]

The reason why python 3.5 is an issue is that flair 0.4.2 doesn't support it (it need python 3.6). Which is what I trained my model with.

[carlomazzaferro]

The same is valid, you can either define your own Docker image for deployment of add the pip install script in your main inference script. You can create an instance of a model from a pre-trained model just using this class right here: https://sagemaker.readthedocs.io/en/stable/sagemaker.pytorch.html#pytorch-model

Namely:

model = PytorchModel(model_data='s3://path/to/model.tar.gz', image_name='ecr_image_with_flair', ...)

Then you can model.deploy() as you normally would.

I haven't tested this in some time, I actually solved the issue by deploying models with terraform (which I'd advise you to look into, as it solves many problems as far as repeatable deployments go). See here: https://www.terraform.io/docs/providers/aws/r/sagemaker_model.html