Closed mattroos closed 3 years ago
For now, I resolved this by importing the SegDetectorModel
class for building the model rather than using the BasicModel
class directly. I still want to hone this down to minimum code footprint, but don't quite understand what's going on in SegDetectorModel.__init__()
, particularly in the parallelize()
call, without which the model doesn't perform as expected.
For now, I resolved this by importing the
SegDetectorModel
class for building the model rather than using theBasicModel
class directly. I still want to hone this down to minimum code footprint, but don't quite understand what's going on inSegDetectorModel.__init__()
, particularly in theparallelize()
call, without which the model doesn't perform as expected.
Hello! Did you understand, what is the strange behaviour with parallelize()? I've encountered similar problem...
This is a general question, not an issue with the code.
I'm trying to write code that straightforwardly builds a model (one with a MobileNetv3 backbone that I trained on ICDAR 2015), gives the same results as
demo.py
, and can (hopefully) be used to convert the model to ONNX. But the code in this repo is so deeply abstracted into a multitude of classes that I can't figure out how to do it. I'm very close, but...When I run
demo.py
on a training image from icdar, I get the prediction below. Looks good. But when I write code that does nothing more that build the MobileNetv3 and decoder models, and load the weights, I get the results below. I think the features from the backbone are fine. But the decoder output is a checkered pattern (at single-pixel resolution... the larger-scale checkered features in the image below are not "real" but are due to image aliasing). It seems like it should be straightforward, but I'm obviously missing something. Any guesses? @MhLiao? The output looks like what one might get if the decoder weights were not loaded properly (or not loaded at all), but I believe I'm doing it correctly. See the## Build the model
section in my code.Here is my code: