How to train SageMaker BlazingText model using multiple channels - amazon-web-services

Question

I have two separate normalized text files that I want to train my BlazingText model on.
I am struggling to get this to work and the documentation is not helping.
Basically I need to figure out how to supply multiple files or S3 prefixes as "inputs" parameter to the sagemaker.estimator.Estimator.fit() method.
I first tried:
s3_train_data1 = 's3://{}/{}'.format(bucket, prefix1)
s3_train_data2 = 's3://{}/{}'.format(bucket, prefix2)
train_data1 = sagemaker.session.s3_input(s3_train_data1, distribution='FullyReplicated', content_type='text/plain', s3_data_type='S3Prefix')
train_data2 = sagemaker.session.s3_input(s3_train_data2, distribution='FullyReplicated', content_type='text/plain', s3_data_type='S3Prefix')
bt_model.fit(inputs={'train1': train_data1, 'train2': train_data2}, logs=True)
this doesn't work because SageMaker is looking for the key specifically to be "train" in the inputs parameter.
So then i tried:
bt_model.fit(inputs={'train': train_data1, 'train': train_data2}, logs=True)
This trains the model only on the second dataset and ignores the first one completely.
Now finally I tried using a Manifest file using the documentation here: https://docs.aws.amazon.com/sagemaker/latest/dg/API_S3DataSource.html
(see manifest file format under "S3Uri" section)
the documentation says the manifest file format is a JSON that looks like this example:
[
{"prefix": "s3://customer_bucket/some/prefix/"},
"relative/path/to/custdata-1",
"relative/path/custdata-2"
]
Well, I don't think this is valid JSON in the first place but what do I know, I still give it a try.
When I try this:
s3_train_data_manifest = 'https://s3.us-east-2.amazonaws.com/bucketpath/myfilename.manifest'
train_data_merged = sagemaker.session.s3_input(s3_train_data_manifest, distribution='FullyReplicated', content_type='text/plain', s3_data_type='ManifestFile')
data_channel_merged = {'train': train_data_merged}
bt_model.fit(inputs=data_channel_merged, logs=True)
I get an error saying:
ValueError: Error training blazingtext-2018-10-17-XX-XX-XX-XXX: Failed Reason: ClientError: Data download failed:Unable to parse manifest at s3://mybucketpath/myfilename.manifest - invalid format
I tried replacing square brackets in my manifest file with curly braces ...but still I feel the JSON file format seems to be missing something that documentation fails to describe correctly?

Answer 1

You can certainly match multiple files with the same prefix, so your first attempt could have worked as long as you organize your files in your S3 bucket to suit. For e.g. the prefix: s3://mybucket/foo/ will match the files s3://mybucket/foo/bar/data1.txt and s3://mybucket/foo/baz/data2.txt
However, if there is a third file in your bucket called s3://mybucket/foo/qux/data3.txt that you don't want matched (while still matching the first two) there is no way to do achieve that with a single prefix. In these cases a manifest would work. So, in the above example, the manifest would simply be:
[
{"prefix": "s3://mybucket/foo/"},
"bar/data1.txt",
"baz/data2.txt"
]
(and yes, this is valid json - it is an array whose first element is an object with an attribute called prefix and all subsequent elements are strings).
Please double check your manifest (you didn't actually post it so I can't do that for you) and make sure it conforms to the above syntax.
If you're still stuck please open up a thread on the AWS sagemaker forums - https://forums.aws.amazon.com/forum.jspa?forumID=285 and after you do that we can setup a PM to try and get to the bottom of this (never post your AWS account id in a public forum like StackOverflow or even in AWS forums).