CODEWITHSUNDEEP
amazon-sagemaker
dipalit
dipalit

Reputation: 21

How to save model.tar.gz file in sagemaker using Estimator

I am not able to save model artifacts in S3 bucket using below code. I am successfully able to save the result in output data path and training job is getting completed successfully. I am using the following piece of code.

Can anyone please confirm how do we save the model_artifacts in model-dir using below code.

# train.py code
#!/usr/bin/env python

from __future__ import print_function
import os
import sys
import pandas as pd

prefix = '/opt/ml/'

input_dir = prefix + 'input/data'
output_data_dir = os.path.join(prefix, 'output/data')
model_dir = os.path.join(prefix, 'model')

channel_name='training'
training_path = os.path.join(input_dir, channel_name)

# The function to execute the training.
def train():
    print('Starting the training.')
    # Take the set of files and read them all into a single pandas dataframe
    input_files = [ os.path.join(training_path, file) for file in os.listdir(training_path) ]
    raw_data = [ pd.read_csv(file, header=None) for file in input_files ]
    input_data = pd.concat(raw_data)
    
    print(pd.DataFrame(input_data))
    
    output_data = input_data.to_csv(os.path.join(output_data_dir, 'output.csv'), header=False, index=False)
    

if __name__ == '__main__':
    train()

# Below are the S3 input and output paths :
output_path = "s3://{}/{}".format(bucket, prefix_output)
S3_input = "s3://{}/{}".format(bucket, prefix)

#Estimator Code

test_estimator = sagemaker.estimator.Estimator(ecr_image,                                # ECR image arn,
                                          role=role,                                 # execution role
                                          instance_count=1,                     # no. of sagemaker instances
                                          instance_type='ml.m4.xlarge',         # instance type
                                          output_path=output_path,              # output path to store model outputs
                                          base_job_name='sagemaker-job1',       # job name prefix
                                          sagemaker_session=session             # session
                                         )

# Launch instance and start training
test_estimator.fit({'training':S3_input})

What is missing in this code?

Upvotes: 2

Views: 7758

Answers (1)

rok
rok

Reputation: 2765

Sagemaker save automatically to output_path everything that is inside your model directory, so everything that is in /opt/ml/model. If the training job complete successfully, at the end Sagemaker takes everything in that folder, create a model.tar.gz and upload to your output_path in a folder with the same name of your training job (sagemaker create this folder). You can also use the environmental variable SM_OUTPUT_DATA_DIR, which by default points to /opt/ml/output/data and put non-model training artifacts (e.g. evaluation results), Sagemaker will create an archive from that folder named output.tar.gz and will upload it in the same folder of the model.tar.gz on S3.

I don't understand exactly what you mean with "result", but whatever you want to put in that archive, it's up to you to save it in your model_dir. So for example this how I save my model in both json and H5, the first will be in the output.tar.gz archive, the latter in the model.tar.gz

output_artifacts = os.environ.get('SM_OUTPUT_DATA_DIR')

with open(os.path.join(output_artifacts,"model.json"), "w") as json_file:
        json_file.write(model_json) 

model_directory = os.environ.get('SM_MODEL_DIR')
model.save(os.path.join(model_directory, 'model.h5'))

Upvotes: 5

Related Questions