def launch_grid(grid):
grid = copy.deepcopy(grid)
grid['model'] = 'InnerProduct'
grid['lr_decay'] = 'linear'
grid['emb_size'] = 9
for cfg in jobs.param_grid(grid):
cfg['train_dir'] = jobs.make_train_dir(cfg, keys_for_dir_name)
jobs.submit(commands, cfg, get_slurm_script_gpu)
grid = copy.deepcopy(grid)
grid['model'] = ['Deep', 'ResidualInnerProduct']
grid['emb_size'] = 8
grid['eval_batch_size'] = 1024
grid['lr_decay'] = 'plateau'
grid['hidden_size'] = 128
grid['learning_rate'] = 0.01
for cfg in jobs.param_grid(grid):
cfg['train_dir'] = jobs.make_train_dir(cfg, keys_for_dir_name)
jobs.submit(commands, cfg, get_slurm_script_gpu)
base_grid.map_items = False
base_grid.eval_recall_max = 100
base_grid.test_recall_max = 1000
base_grid.tokenize = False
base_grid.target_publication = 0
base_grid.batch_size = 32
base_grid.learning_rate = [2e-5, 3e-5, 4e-5]
base_grid.use_gpu = True
base_grid.frequency = 200
base_grid.eval_batch_size = 500
base_grid.dict_dir = pathlib.Path(
"/scratch/network/altosaar/dat/longform-data/main/dictionaries")
base_grid.tokenizer_file = (
"/scratch/network/altosaar/dat/longform-data/main/bert-base-uncased.txt"
)
base_grid.model_path = "/scratch/network/altosaar/dat/longform-data/BERT/model"
base_grid.index_file_path = (
"/scratch/network/altosaar/dat/longform-data/BERT/eval_indices_list.txt"
)
# 100 warmup steps
grid = copy.deepcopy(base_grid)
grid.warmup_steps = 1000
grid.training_steps = [5000, 50000, 100000]
keys_for_dir_name = jobs.get_keys_for_dir_name(grid)
keys_for_dir_name.insert(0, "warmup_steps")
for cfg in jobs.param_grid(grid):
cfg["output_dir"] = jobs.make_output_dir(log_dir, experiment_name, cfg,
keys_for_dir_name)
jobs.submit(commands, cfg, get_slurm_script_gpu)