How do I get the filename if I'm using bentoml.io.File API descriptor for the input?

[mgmalana]

I understand that the input/output becomes a Filelike object. I want to get the filename because I will use it for post processing.

I tried calling input.name but it returns None. I also tried input.path but I get an "TypeError: Expected str, got bytes" on "raw_path = urllib.request.url2pathname(urllib.parse.unquote(parsed.path))" line.

It seems very obvious but I cant seem to find the way to get the filename. Thank you in advance for your guidance!

[sauyon]

Hi! There currently isn't a way to get a filename, as there is no actual file on disk. If you need to pass a filename to something unfortunately you'll have to write it manually first.

[mgmalana]

@sauyon thank you for reply! I tried to use the Multipart API descriptor so that I can write the file name and also upload the data. I used the following code

input_spec = Multipart(vid=File(), filename=Text())
@svc.api(input=input_spec, output=JSON())

However, I got a BentoMLException: File should have Content-Type b'application/octet-stream' or b'multipart/form-data', got b'video/mp4' instead. I tried using multipart with other descriptor like NumpyNdarray, it did not raise an error. What do you think?

[aarnphm]

@aarnphm, does that mean I have to change the request format? When I use @svc.api(input=File(), output=JSON()), I don't encounter any error with the video. When I changed it to MultiPart, I get the error that I mentioned above.

This means in your curl command you can do sth like

curl -X POST -H "Content-Type: multipart/form-data" -F [email protected] -F arr='[5,4,3,2]'

[mgmalana]

@aarnphm I'm sorry I am not able to follow. First, I need to convert the video to byte stream so that I can use it with multipart. How do I convert the video to byte stream?

[aarnphm]

sorry for the confusion, what I meant is you can just set the header to form-data. You can also stream your video as binary this way

curl --data-binary ...

@mgmalana

[mgmalana]

@aarnphm thank you for your explanation. I was able to fix it by removing "type=video/mp4" from the curl command. I was using "[email protected];type=video/mp4". Now it works as expected!

[sauyon]

I think we should introduce a Video IO type instead. I prefer File to be accepted as a stream. But the argument is what is really the benefits of introducing a Video IO type? Unless we want to do sth special I think for most cases file can converted to byte stream @parano?

I think it would be valuable, because users that wish to only accept video content might have something use the Video descriptor to
a) describe that what they're looking for is an API endpoint that takes video, maybe giving us hints at serving-time what kinds of inputs we might expect,
b) document to other developers that the API endpoint takes video, and
c) restrict the API endpoint to content-types related to video, maybe with options like Video(kind='mp4') to limit to specific video formats.

IMO the File descriptor should probably just be a catch-all which accepts any content type, and then we'd have subclasses like Video which specifically restrict the content type.

I've opened #2387 in case we want to address situations like this in the future.