Getting an IndexError from the tflite backend on Raspberry Pi

[aidenmitchell]

Traceback (most recent call last):
  File "/home/pi/Desktop/tester.py", line 6, in <module>
    result = model.predict_from_file('/home/pi/Desktop/image.png')
  File "/home/pi/.local/lib/python3.7/site-packages/lobe/ImageModel.py", line 43, in predict_from_file
    return self.predict(image_utils.get_image_from_file(path))
  File "/home/pi/.local/lib/python3.7/site-packages/lobe/ImageModel.py", line 47, in predict
    return self.__backend.predict(image_processed)
  File "/home/pi/.local/lib/python3.7/site-packages/lobe/backends/_backend_tflite.py", line 57, in predict
    output_details[1]["index"]
IndexError: list index out of range

Here is my code:

from lobe import ImageModel

model = ImageModel.load('/home/pi/Lobe/model')

result = model.predict_from_file('/home/pi/Desktop/image.png')
print(result.prediction)

[prettyflyforabeeguy]

I can reproduce this on my raspberry pi3 as far back as commit: 22c16f5
My old .tflite models work at this commit but my newly created models (using the naming convention saved_model.tflite) do not work and throw this same IndexError. I am not sure what versions of the UI aligns up with these commits, so I'm not sure what version of the UI I'd have to revert to in order to get a working tflite model.

[aidenmitchell]

@adiazulay Any idea what's going on? Hoping we can get someone from Lobe to take a look.

[mbeissinger]

I'm taking a look at this now and refactoring a bit how models are loaded and run. Should have a branch up later this evening that would be great to get some testing help.

[mbeissinger]

Sorry for the delay, turned into a larger refactor than anticipated! Here is the PR, would love to get some help testing in your cases: #15

[prettyflyforabeeguy]

Looking good!

[prettyflyforabeeguy]

Sorry for so many edits (too much coffee this afternoon!)
Some issues I've run into so far:

1. from lobe.Signature import Signature

 gives me an error:
Traceback (most recent call last):
File "motion.py", line 9, in <module>
  from tf_classify import TFClassify
File "/home/pi/src/BeeTracker/tf_classify.py", line 3, in <module>
 from lobe.Signature import Signature 

ModuleNotFoundError: No module named 'lobe.Signature'

I fixed this by changing my code to lower case: lobe.signature

2. it doesn't like something with the way i'm using this model.predict_from_file
   from lobe import ImageModel
   model: ImageModel = ImageModel.load_from_signature(sig)
   res = model.predict_from_file(item)

the item in this case is: img/PiZero_v1_2b762167-0c7b-42a4-9f6f-eab31ca35324.jpg

my main program doesn't give a helpful error, it just crashes at runtime, however, trying to duplicate this in a separate test
says:

Traceback (most recent call last):
File "imgmdltest.py", line 11, in
res = model.predict_from_file(item)
File "/home/pi/.local/lib/python3.7/site-packages/lobe/model/image_model.py", line 42, in predict_from_file
return self.predict(image_utils.get_image_from_file(path))
File "/home/pi/.local/lib/python3.7/site-packages/lobe/model/image_model.py", line 45, in predict
image_processed = image_utils.preprocess_image(image, self.signature.input_image_size)
AttributeError: 'Signature' object has no attribute 'input_image_size'

I don't think i've ever been supplying a size before.

[mbeissinger]

Ah gotcha! Yep I made it a subclass system -- you will want to use ImageClassificationSignature from lobe.signature (this one has the input_image_size property vs the generic Signature class).

It should also just work if you use ImageModel.load('path/to/your/signature.json') instead of using the load_from_signature function now.

[aidenmitchell]

(.venv) C:\Users\Aiden Mitchell>pip install git+https://github.com/lobe/lobe-python@v0.3-cleanup-and-match-lobe-export
Collecting git+https://github.com/lobe/lobe-python@v0.3-cleanup-and-match-lobe-export
  Cloning https://github.com/lobe/lobe-python (to revision v0.3-cleanup-and-match-lobe-export) to c:\users\aiden mitchell\appdata\local\temp\pip-req-build-gzt1j6n0

ERROR: Could not find a version that satisfies the requirement tensorflow<2,>=1.15 (from lobe)
ERROR: No matching distribution found for tensorflow<2,>=1.15

I got an error on Windows trying to install the new branch. I can try on a Linux machine instead.
Same error on Linux, Ubuntu 20.04.1.

[aidenmitchell]

Also of note, this step doesn't work (also noted in #13):

sudo apt install -y \
>     libatlas-base-dev \
>     libopenjp2-7 \
>     libtiff5 \
>     libjpeg62-turbo
[sudo] password for aiden:
Reading package lists... Done
Building dependency tree
Reading state information... Done
Package libjpeg62-turbo is not available, but is referred to by another package.
This may mean that the package is missing, has been obsoleted, or
is only available from another source

E: Package 'libjpeg62-turbo' has no installation candidate

libjpeg62-dev works instead, according to this

[prettyflyforabeeguy]

@mbeissinger thanks!
changing this:
from lobe.signature import Signature
to this:
from lobe.signature import ImageClassificationSignature

and this:
sig:Signature = Signature(sigpath)

to this:
sig: ImageClassificationSignature = ImageClassificationSignature(sigpath)

restored everything to working order.

[mbeissinger]

@aidenmitchell for the Tensorflow error, what version of Windows and Python are you running? I suspect it is a Python version issue -- have tested with 3.6 and 3.7, but 3.8 I believe only works in Tensorflow 2.x

[mbeissinger]

@aidenmitchell I just updated the branch to use tensorflow 2.4, could you please try now?

[aidenmitchell]

Works like a charm! Tested it on Ubuntu 20.10 with Python 3.8.6.