váthos or βάθος is in Greek which means Depth
A Monocular Depth Estimation and Segmentation Model, using ResNet-Multi-Tail-Encoder-Decoder or Vathos as i like to call it Architecture. See more in the Documentation below
ProjektDepth: A DNN for segmentation and depth estimation. The entire dataset was created from scraping images, augmenting them to finally create a dataset of 1.2B images. The DNN was custom made inspired from Resnet and Unet architecture, various loss functions and their combinations were tried on the model. LRFinder was used to find the best or for the model. The model was trained on Colab with a P100, experiments were done on TPU. The vathos package contains the runner, trainer, model, loss, and the notebooks. Everything is documented and added to readthedocs.
name: vathos_small_train
log_dir: logs
chkpt_dir: checkpoint
model_init: "models/model.pt"
device: GPU
model: ResUNet
augmentation: DepthLightAug
dataset:
name: DenseDepth
root: vathos_data
zip_dir: "/content/gdrive/My Drive/DepthProject/depth_dataset_zipped/"
div_factor: 1
loader_args:
batch_size: 128
num_workers: 4
shuffle: True
pin_memory: True
seg_loss: BCEDiceLoss
depth_loss: RMSELoss
lr_scheduler:
type: OneCycleLR
args:
max_lr: 0.6
optimizer:
type: AdamW
args:
lr: 0.01
training:
epochs: 3
___ ,---,
,---. ,--.'|_ ,--.' |
/__./| | | :,' | | : ,---.
,---.; ; | : : ' : : : : ' ,'\ .--.--.
/___/ \ | | ,--.--. .;__,' / : | |,--. / / | / / '
\ ; \ ' | / \ | | | | : ' |. ; ,. :| : /`./
\ \ \: | .--. .-. |:__,'| : | | /' :' | |: :| : ;_
; \ ' . \__\/: . . ' : |__ ' : | | |' | .; : \ \ `.
\ \ ' ," .--.; | | | '.'|| | ' | :| : | `----. \
\ ` ; / / ,. | ; : ;| : :_:,' \ \ / / /`--' /
: \ |; : .' \ | , / | | ,' `----' '--'. /
'---" | , .-./ ---`-' `--'' `--'---'
`--`---'
Author: Satyajit Ghana
Github: https://github.com/satyajitghana/ProjektDepth
[ 2020-05-23 20:55:28,682 - vathos.vathos.runner.runner ] INFO: Now simply setup_train and then start_train your model
import vathos
import vathos.runner as vrunner
import vathos.utils as vutils
config = vutils.load_config('small_data.yaml')
runner = vrunner.Runner(config)
runner.setup_train()
runner.start_train()
$ cd docs
$ pip install requirements.txt
$ make.bat html