问题
I would like to run a program on my laptop (Gazebo simulator) and send a stream of image data to a GCE instance, where it will be run through an object-detection network and sent back to my laptop in near real-time. Is such a set-up possible?
My best idea right now is, for each image:
- Save the image as a JPEG on my personal machine
- Stream the JPEG to a Cloud Storage bucket
- Access the storage bucket from my GCE instance and transfer the file to the instance
- In my python script, convert the JPEG image to numpy array and run through the object detection network
- Save the detection results in a text file and transfer to the Cloud Storage bucket
- Access the storage bucket from my laptop and download the detection results file
- Convert the detection results file to a numpy array for further processing
This seems like a lot of steps, and I am curious if there are ways to speed it up, such as reducing the number of save and load operations or transporting the image in a better format.
回答1:
If your question is "is it possible to set up such a system and do those actions in real time?" then I think the answer is yes I think so. If your question is "how can I reduce the number of steps in doing the above" then I am not sure I can help and will defer to one of the experts on here and can't wait to hear the answer!
I have implemented a system that I think is similar to what you describe for research of Forex trading algorithms (e.g. upload data to storage from my laptop, compute engine workers pull the data and work on it, post results back to storage and I download the compiled results from my laptop).
I used the Google PubSub architecture - apologies if you have already read up on this. It allows near-realtime messaging between programs. For example you can have code looping on your laptop that scans a folder that looks out for new images. When they appear it automatically uploads the files to a bucket and once theyre in the bucket it can send a message to the instance(s) telling them that there are new files there to process, or you can use the "change notification" feature of Google Storage buckets. The instances can do the work, send the results back to the storage and send a notification to the code running on your laptop that work is done and results are available for pick-up.
Note that I set this up for my project above and encountered problems to the point that I gave up with PubSub. The reason was that the Python Client Library for PubSub only supports 'asynchronous' message pulls, which seems to mean that the subscribers will pull multiple messages from the queue and process them in parallel. There are some features to help manage 'flow control' of messages built into the API, but even with them implemented I couldn't get it to work the way I wanted. For my particular application I wanted to process everything in order, one file at a time because it was important to me that I'm clear what the instance is doing and the order its doing it in. There are several threads on google search, StackOverflow and Google groups that discuss workarounds for this using queues, classes, allocating specific tasks for specific instances, etc which I tried, but even these presented problems for me. Some of these links are:
Run synchronous pull in PubSub using Python client API and pubsub problems pulling one message at a time and there are plenty more if you would like them!
You may find that if the processing of an image is relatively quick, order isn't too important and you don't mind an instance working on multiple things in parallel that my problems don't really apply to your case.
FYI, I ended up just making a simple loop on my 'worker instances' that scans the 'task list' bucket every 30 seconds or whatever to look for new files to process, but obviously this isn't quite the real-time approach that you were originally looking for. Good luck!
来源：https://stackoverflow.com/questions/50053968/is-it-possible-to-perform-real-time-communication-with-a-google-compute-engine-i