bnel1201 / hog_finder Goto Github PK

View Code? Open in Web Editor NEW

0.0 0.0 0.0 127.1 MB

Python 0.17% Batchfile 0.01% Jupyter Notebook 63.70% HTML 36.12% Dockerfile 0.01%

hog_finder's People

Contributors

Watchers

hog_finder's Issues

add test set and dice score to get baseline

then start exploring other generalization strategies to avoid overfitting or other models / libraries for practice and comparison

sd image size is 640x360, that should be the default size for inference

this would make inference run fast, lower memory constraints, etc.
one issue with progressize resizing could be that the largest size is 448x448 which is larger than the smallest dimension of 640x360, maybe that could be one reason why progressize resizing isn't working as expected? #4

batch inference

single cpu inference is slow

use skimage.measure.regionsprops orientation attribue to have overlays rotate with Xiaomi

https://scikit-image.org/docs/dev/api/skimage.measure.html#skimage.measure.regionprops

something went wrong with progressive resizing model

need to add model names in addition to date when saving models to allow for easier comparisons
then retrain normalization and progressive resizing models to see what is going on

issue reading windows Path string on docker

This is likely because I save a fastai model using a windows Path() oject so when unpickling it doesnt know what to do. I should try resaving the model setting the path with a naked string, rather than a pathlib Path object.

File "/usr/local/lib/python3.9/site-packages/streamlit/script_runner.py", line 333, in _run_script
    exec(code, module.__dict__)
File "/code/streamlit_app.py", line 50, in <module>
    main()
File "/code/streamlit_app.py", line 29, in main
    show_image(imlist[selected_frame_index], show_mask, predict)
File "/code/streamlit_app.py", line 44, in show_image
    hf = hedgiefinder.HedgieFinder(fname, cleanup=False).predict([fname])
File "/code/hedgiefinder/inference.py", line 31, in __init__
    self.model = load_learner(model_dir/model_name)
File "/usr/local/lib/python3.9/site-packages/fastai/learner.py", line 374, in load_learner
    res = torch.load(fname, map_location='cpu' if cpu else None, pickle_module=pickle_module)
File "/usr/local/lib/python3.9/site-packages/torch/serialization.py", line 594, in load
    return _load(opened_zipfile, map_location, pickle_module, **pickle_load_args)
File "/usr/local/lib/python3.9/site-packages/torch/serialization.py", line 853, in _load
    result = unpickler.load()
File "/usr/local/lib/python3.9/pathlib.py", line 1073, in __new__
    raise NotImplementedError("cannot instantiate %r on your system"

Explore LSTM and Transformer models to improve performance

These architectures are good for time series data which our hog finder can be classified as. This is because the position at any given time is related to the position in the past frames

bnel1201 / hog_finder Goto Github PK

hog_finder's People

Contributors

Watchers

hog_finder's Issues

add test set and dice score to get baseline

sd image size is 640x360, that should be the default size for inference

batch inference

use skimage.measure.regionsprops orientation attribue to have overlays rotate with Xiaomi

something went wrong with progressive resizing model

issue reading windows Path string on docker

Explore LSTM and Transformer models to improve performance

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent