combilatent / src /xp1 /4_run.py
younadi's picture
Add files using upload-large-folder tool
4417138 verified
Raw History Blame Contribute Delete
5.78 kB
from tqdm import tqdm
from src.recover_schedules import recover_schedules
params = [
{
"testing": False,
"seed": 97,
"init_mode": "random",
"n_optimization_steps": 2000,
"epsilon": 0.01,
"nb_sinkhorn_iters": 40,
"decay_ls": False,
"ls_partitions_ratios": [0.66],
"init_ls": 2.0,
"max_ls": 10,
"min_ls": 0.1,
"ls_warmup_iters_ratio": 0.2,
"ls_decay_iters_ratio": 0.90,
"decay_lr": True,
"lr_partitions_ratios": [0.66],
"init_lr": 1e-2,
"max_lr": 1e-1,
"min_lr": 1e-4,
"lr_warmup_iters_ratio": 0.1,
"lr_decay_iters_ratio": 0.90,
"checkpoint_interval": 1,
"data_dir": f"outputs/exhaustive_{nb_jobs}_{nb_machines}/top_{top_k}",
"model_dir": f"outputs/exhaustive_{nb_jobs}_{nb_machines}/top_{top_k}/train_nEmbd{n_embd}_nHead{n_head}_nLayer{n_layer}",
"output_dir": f"outputs/exhaustive_{nb_jobs}_{nb_machines}/top_{top_k}/train_nEmbd{n_embd}_nHead{n_head}_nLayer{n_layer}/recover_{yes_or_no_mf}_mf",
"apply_metropolized_filling": apply_mf,
"mf_patience": 10,
"mf_fifo_size": 50,
"mf_penalty_strength": 1.0,
"mf_min_delta": 0.05,
}
for yes_or_no_mf, apply_mf in [("no", False)]
for nb_jobs in [9]
for nb_machines in [6]
for top_k in [0]
for n_embd in [6, 8, 12, 24, 36, 48, 64]
for n_head in [1, 2]
for n_layer in [1, 2, 3, 4]
]
for param in (pbar:=tqdm(params)):
config = (
f"n_embd={param['model_dir'].split('nEmbd')[1].split('_')[0]}, "
f"n_head={param['model_dir'].split('nHead')[1].split('_')[0]}, "
f"n_layer={param['model_dir'].split('nLayer')[1].split('/')[0]}, "
f"apply_metropolized_filling={param['apply_metropolized_filling']}"
)
pbar.set_description(f"Section 1. Recovering schedules for {config}")
recover_schedules(**param)
params = [
{
"testing": False,
"seed": 97,
"init_mode": "random",
"n_optimization_steps": 2000,
"epsilon": 0.01,
"nb_sinkhorn_iters": 40,
"decay_ls": False,
"ls_partitions_ratios": [0.66],
"init_ls": 2.0,
"max_ls": 10,
"min_ls": 0.1,
"ls_warmup_iters_ratio": 0.2,
"ls_decay_iters_ratio": 0.90,
"decay_lr": True,
"lr_partitions_ratios": [0.66],
"init_lr": 1e-2,
"max_lr": 1e-1,
"min_lr": 1e-4,
"lr_warmup_iters_ratio": 0.1,
"lr_decay_iters_ratio": 0.90,
"checkpoint_interval": 1,
"data_dir": f"outputs/exhaustive_{nb_jobs}_{nb_machines}/top_{top_k}",
"model_dir": f"outputs/exhaustive_{nb_jobs}_{nb_machines}/top_{top_k}/train_nEmbd{n_embd}_nHead{n_head}_nLayer{n_layer}",
"output_dir": f"outputs/exhaustive_{nb_jobs}_{nb_machines}/top_{top_k}/train_nEmbd{n_embd}_nHead{n_head}_nLayer{n_layer}/recover_{yes_or_no_mf}_mf",
"apply_metropolized_filling": apply_mf,
"mf_patience": 10,
"mf_fifo_size": 50,
"mf_penalty_strength": 1.0,
"mf_min_delta": 0.05,
}
for yes_or_no_mf, apply_mf in [("no", False)]
for nb_jobs in [9]
for nb_machines in [6]
for top_k in [0]
for n_embd in [6 + 3*i for i in range(0, 21)]
for n_head in [3]
for n_layer in [1, 2, 3, 4]
]
for param in (pbar:=tqdm(params)):
config = (
f"n_embd={param['model_dir'].split('nEmbd')[1].split('_')[0]}, "
f"n_head={param['model_dir'].split('nHead')[1].split('_')[0]}, "
f"n_layer={param['model_dir'].split('nLayer')[1].split('/')[0]}, "
f"apply_metropolized_filling={param['apply_metropolized_filling']}"
)
pbar.set_description(f"Section 2. Recovering schedules for {config}")
recover_schedules(**param)
params = [
{
"testing": False,
"seed": 97,
"init_mode": "random",
"n_optimization_steps": 2000,
"epsilon": 0.01,
"nb_sinkhorn_iters": 40,
"decay_ls": False,
"ls_partitions_ratios": [0.66],
"init_ls": 2.0,
"max_ls": 10,
"min_ls": 0.1,
"ls_warmup_iters_ratio": 0.2,
"ls_decay_iters_ratio": 0.90,
"decay_lr": True,
"lr_partitions_ratios": [0.66],
"init_lr": 1e-2,
"max_lr": 1e-1,
"min_lr": 1e-4,
"lr_warmup_iters_ratio": 0.1,
"lr_decay_iters_ratio": 0.90,
"checkpoint_interval": 1,
"data_dir": f"outputs/exhaustive_{nb_jobs}_{nb_machines}/top_{top_k}",
"model_dir": f"outputs/exhaustive_{nb_jobs}_{nb_machines}/top_{top_k}/train_nEmbd{n_embd}_nHead{n_head}_nLayer{n_layer}",
"output_dir": f"outputs/exhaustive_{nb_jobs}_{nb_machines}/top_{top_k}/train_nEmbd{n_embd}_nHead{n_head}_nLayer{n_layer}/recover_{yes_or_no_mf}_mf",
"apply_metropolized_filling": apply_mf,
"mf_patience": 10,
"mf_fifo_size": 50,
"mf_penalty_strength": 1.0,
"mf_min_delta": 0.05,
}
for yes_or_no_mf, apply_mf in [("no", False)]
for nb_jobs in [9]
for nb_machines in [6]
for top_k in [0]
for n_embd in [8 + 4*i for i in range(0, 15)]
for n_head in [4]
for n_layer in [1, 2, 3, 4]
]
for param in (pbar:=tqdm(params)):
config = (
f"n_embd={param['model_dir'].split('nEmbd')[1].split('_')[0]}, "
f"n_head={param['model_dir'].split('nHead')[1].split('_')[0]}, "
f"n_layer={param['model_dir'].split('nLayer')[1].split('/')[0]}, "
f"apply_metropolized_filling={param['apply_metropolized_filling']}"
)
pbar.set_description(f"Section 3. Recovering schedules for {config}")
recover_schedules(**param)