An AutoEncoder model to reconstruct and generate new images of edge-on Proto Planetary disks using physical parameters as input.
The model architecture is the following:
The edge-on disk images used for training were generated using the MCFOST Radiative Transfer code. The training set looks like this:
with a wide variety of shapes and sizes.
Each image is associated with the physical parameters used to simulate the source. We used 8 physical parameters to as inputs to MCFOST:
m_dust = 'mass of the dust'
Rc = 'critical radius when exp drops(size)'
f_exp = 'flare exponent'
H0 = 'scale hight'
Rin = 'inner raidus'
sd_exp = 'surface density exponent'
alpha = 'dust stettling'
inc = 'inclination'All physical parameters where sampled from a evenly spaced grid and a random sampling between the range of possible values in order to fill up the gaps.
Use ae_main.py to train a AE model with the following parameters:
-h, --help show this help message and exit
--dry-run Load data and initialize models [False]
--machine MACHINE were to is running (local, [colab], exalearn)
--data DATA data used for training (MNIST, [PPD])
--img-norm IMG_NORM type of normalization for images (global, [image])
--par-norm PAR_NORM physical parameters are 0-1 scaled ([T],F)
--subset SUBSET data subset ([25052021], fexp1)
--optim OPTIM Optimizer ([Adam], SGD)
--lr LR learning rate [1e-4]
--lr-sch LR_SCH learning rate shceduler ([None], step, exp, cosine, plateau)
--batch-size BATCH_SIZE
batch size [128]
--num-epochs NUM_EPOCHS
total number of training epochs [100]
--early-stop Early stoping
--cond COND physics conditioned AE (F,[T])
--feed-phy FEED_PHY feed physics to decoder ([F],T)
--latent-dim LATENT_DIM
dimension of latent space [8]
--dropout DROPOUT dropout for all layers [0.2]
--kernel-size KERNEL_SIZE
2D conv kernel size, encoder [3]
--conv-blocks CONV_BLOCKS
conv+actfx+pool blocks [5]
--model-name MODEL_NAME
name of model [ConvLinTrans_AE]
--comment COMMENT extra commentsEdge-on images (upper row), AE reconstruction (middle), and residuals (lower row).
Training logs and models are stored at W&B here: https://wandb.ai/deep_ppd/PPD-AE
We use poetry to manage dependencies and environment. First install poetry:
pip install poetry
Then install the necessary libraries listed in poetry.toml using the install command from
poetry inside the repo directory:
poetry install
More info on how to use poetry -> https://python-poetry.org/docs/