How can I make this process more efficient - Importing data file into numpy array

[VirajVaitha123]

Hi all,

I am new to FastAPI but I love this library, thank you so much!

So here is my problem statement.

My code works but it is taking very long to return the response via the browser. I have many steps I want to do to investigate this, but first I wanted to share the problem here to see what your thoughts are. My bottleneck is most likely occurring at the transformation/preprocessing using machine learning, but I can optimise this part myself. The part I can't optimise yet is the method of reading and returning data using FastAPI.

I have only included the relevant code to make things easier to follow.

Problem statement

I have this endpoint where users can upload an image (UploadFile = File(...)).

Once the image is uploaded into UploadFile = File(...), I use data = await file.read() to convert it into bytes. Am I correct in doing this, and is it in fact bytes?

Secondly, I have this function that converts it into an numpy array, np.array(Image.open(BytesIO(data_obj)))/255. Is there a better way for me to convert the uploaded image into an numpy array?

Thirdly, I use matplotlib to plot an image from the pixels and save it locally. I will look for a better way to save an image from numpy array. If anyone has recommendations let me know (will look into opencv, PIL ect...)

Finally, Ideally I don't save this image, but the only way I can think of returning it to the end user is by using return FileResponse(path=filepath, filename=filename, media_type='image/jpg'). This means I need an extra step to remove the file after this is all complete. Ideally I don't save the file locally and just return the image to the user. Does anyone know how to do this?

@router.post('/submit', response_class=HTMLResponse)
async def post_form(request: Request,k_cluster: int = Form(...), file: UploadFile = File(...)):
    # Read uploaded file as bytes
    data = await file.read()
   
    # Convert Bytes to numpy (great to work with images and processing for machine learning)
    img_array = bytes_to_numpy_array(data, scale = True) # where img_array function is basically: image = np.array(Image.open(BytesIO(data_obj)))/255

   # I do some transformations (not included) then I save the numpy array as an image
     fig = plt.figure(figsize=fig_size)
    segmented_img = cluster_image(k_cluster,img_array)

    plt.subplots_adjust(0,0,1,1)
    plt.axis('off')
    plt.imshow(segmented_img,interpolation='nearest')
    # plt.title("new image")

   return FileResponse(path=filepath, filename=filename, media_type='image/jpg')

[asaah18]

You can shift some of the functions that don't necessarily need to be run before returning a response using background tasks.
A good candidate for that is "removing the stored image after returning it as a response".

[VirajVaitha123]

Thank you for your response 💯

I'll definitely ensure I add background tasks for these types of scenarios.

[jgould22]

You are using matplotlib just to convert the array to image?

Use Pillow to convert it to an image https://pillow.readthedocs.io/en/stable/reference/Image.html#PIL.Image.fromarray

Then use Pillow to Save the image to a BytesIO. https://pillow.readthedocs.io/en/stable/reference/Image.html#PIL.Image.Image.save

then return just a normal response, something like

return Response(the_bytesio, media_type='img/jpg')