shared/experiment/showcase.yaml

--- # ---------------------------
# `sample` can be
# - `uniform` (np.random.uniform(*from))
# - `range` (np.choice(np.arange(*from)))
# - `list` (np.choice(from))
# - `cartesian` special case where a cartesian product of all keys with the `cartesian` sampling scheme
#               is created and iterated over in order. `from` MUST be a list
#               As we iterate over the cartesian product of all
#               such keys, others are sampled as usual. If n_search is larger than the size of the cartesian
#               product, it will cycle again through the product in the same order
#               example with A being `cartesian` from [1, 2] and B from [y, z] and 5 searches:
#                  => {A:1, B: y}, {A:1, B: z}, {A:2, B: y}, {A:2, B: z}, {A:1, B: y}
# - `sequential` samples will loop through the values in `from`. `from` MUST be a list

# ---------------------------
# -----  SBATCH config  -----
cpus: 8
partition: long
mem: 32G
gres: "gpu:rtx8000:1"
codeloc: $HOME/ccai/climategan

modules: "module load anaconda/3 && module load pytorch"
conda: "conda activate climatenv && conda deactivate && conda activate climatenv"

n_search: -1

# ------------------------
# -----  Train Args  -----
# ------------------------

"args.note": "Hyper Parameter search #1"
"args.comet_tags": ["masker_search", "v1"]
"args.config": "config/trainer/my_config.yaml"

# --------------------------
# -----  Model config  -----
# --------------------------
"gen.opt.lr":
  sample: list
  from: [0.01, 0.001, 0.0001, 0.00001]

"dis.opt.lr":
  sample: uniform
  from: [0.01, 0.001]

"dis.opt.optimizer":
  sample: cartesian
  from:
    - ExtraAdam
    - Adam

"gen.opt.optimizer":
  sample: cartesian
  from:
    - ExtraAdam
    - Adam

"gen.lambdas.C":
  sample: cartesian
  from:
    - 0.1
    - 0.5
    - 1

"data.loaders.batch_size":
  sample: sequential
  from:
    - 2
    - 4
    - 6