brew install pyenv pyenv exec python -V pyenv exec python -m venv .venv source .venv/bin/activate Create requirements.txt pip install -r requirements.txt

Install Library install conda

create conda environment conda create -n learn python=3.9 activate the conda environment conda activate learn install pytorch (https://pytorch.org/get-started/locally/)

without gpu:

conda install pytorch torchvision torchaudio cpuonly -c pytorch or with gpu:

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch install pandas

conda install pandas install tensorboard

pip install tensorboard (segmentation) install albumentation

conda install -c conda-forge imgaug conda install -c conda-forge albumentations (seq2seq) install spacy

pip install -U spacy python -m spacy download en_core_web_sm python -m spacy download de_core_news_sm (seq2seq) install torchtext and torch-data

conda install -c pytorch torchtext torchdata Download Dataset https://drive.google.com/drive/folders/1feL5X6epYQiGaT-dHpgV7yWU7_ZthVVI?usp=sharing

1.Get dataset loader 1.1 define preprocessing transform steps 1.2 create dataset object (define where to load data) 1.3 create dataloader object (define batchsize and how to load data)

2.Define model components 2.1 network 2.2 loss function 2.3 optimizer

3.Define logger object

4.Training loop 4.0 set device 4.1 make prediction 4.2 compute loss 4.3 compute gradients 4.4 adjust the weights 4.5 collect result into the logger

#git git remote add origin [email protected]:angpao/classifier.git git push -u origin main git remote add origin [email protected]:angpao/segmentation.git git branch -M main git push -u origin main

#tensorboard

tensorboard --logdir=runs/ http://localhost:6006/?darkMode=true#timeseries

#recall case 1 : if recall is low in tensorboard because predict is little in image (ignore some class So should weight increase ) predict mistake , loss is high in loss_fn = nn.crossentropyloss ([2.0,1.0,1.0]) case 2 : no label some data maybe class 1 ,class 2 or class 3 So ignore it by adding unknown class (0.0) loss_fn = nn.crossentropyloss ([2.0,1.0,1.0,0.0])

#camera python io stream snapshot -> forward part in model

#loss no reduce check prediction to make sense if not make sense is bug image input uint8 to float32 change model to big and deep create big dataset

#auto label image use GAN to generate segmentation