/Image_Classification_Trainer_Base

Image Classification Boilerplate Code. Using EfficientNet-B2

Primary LanguagePython

Image_Classification_Trainer_Base

Intro

Boilerplate code for Image Classification task. (using EfficientNet-B2)

Model Train

using model

how to train?

python codes/train.py \
    --learning_rate=5e-05 \
    --train_dataset_path=TRAIN_DATASET_PATH \
    --valid_dataset_path=VALID_DATASET_PATH \
    --device=DEVICE \
    --train_batch_size=64 \
    --valid_batch_size=32 \
    --epoch=5

parameters

parameter type description default
dropout_percent float value for dropout percent 0.25
use_loss_weight bool decising apply loss_weight False
use_weight_sampler bool decising apply weight_sampler False
train_dataset_path str train dataset path -
valid_dataset_path str valid dataset path -
use_custom_normalize bool decising apply temp data-only normalize False
train_batch_size int batch size for model train 64
valid_batch_size int batch size for model valid 32
learning_rate float decise learning rate for train 5e-05
lr_warmup_rate float percent for learning-rate warmup 0.1
lr_warmup_steps int steps for learning-rate warmup None
weight_decay_rate float percent for weight_decay 0.0
epoch int epoch use in training 5
device str device use in training "cuda"
base_checkpoint_path str EfficientNet-B2 base model checkpoint path None
wandb_name str A name using for Wandb project "Image_Classification_Project" 
NOTE) 
1. Only one of "use_loss_weight" and "use_weight_sampler" can be True.
2. if base_checkpoint_path is None, code will loading EfficientNet-B2 from "timm" library.

Contact

Reference