args = parser.parse_args()
srun = TorchDatasetSearchRun(args)
params = srun.default_params(args.activation_type)
params['verbose'] = True
base_model_name = (f'{params["net"]}_{params["activation_type"]}_'
f'lr_{params["lr"]}')
# change gridsearch values as desired
if params["net"].startswith('nin'):
weight_decay_list = [1e-4]
else:
weight_decay_list = [5e-4]
search_len = len(weight_decay_list)
start_idx, end_idx = srun.init_indexes(params['log_dir'], search_len)
for idx in range(start_idx, end_idx):
srun.update_index_json(idx)
weight_decay = weight_decay_list[idx]
params['weight_decay'] = weight_decay
params['model_name'] = base_model_name + '_weight_decay_{:.1E}'.format(
weight_decay)
combination_str = (f'\nsearch idx {idx}/{end_idx-1}, '
'weight decay {:.1E}.'.format(weight_decay))
params['combination_str'] = combination_str
main_prog(copy.deepcopy(params))
if (n_rings > 0):
print("\nThis is a cyclic species !!!\nProgram exiting...")
print("\n==========================================\n")
sys.exit(1)
m_blk = Chem.MolToMolBlock(m)
#print (m_blk)
lines = m_blk.split('\n')
if (len(sys.argv) > (i + 2)):
if (sys.argv[3][0] != '-'):
name = sys.argv[3][:9]
else:
name = "Molecule"
else:
name = "Molecule"
#print(lines)
main.main_prog(lines, name, out_file_path, doc_file_path)
print("\n==========================================\n")
elif (sys.argv[1] == "-sf"): #SMILES NAME
s_file = open(sys.argv[sys.argv.index("-sf") + 1], 'r')
m_lines = s_file.readlines()
s_file.close()
i = 0
for l in m_lines:
m_splt = l.split()
if (len(m_splt) == 0):
continue
elif (m_splt[0][0] == '#'):
continue
i = i + 1
m = Chem.MolFromSmiles(m_splt[0])
base_model_name += f'_S_apl_{params["S_apl"]}'
else:
base_model_name += f'_size{params["spline_size"]}'
start_idx, end_idx = srun.init_indexes(params['log_dir'], args.num_runs)
for idx in range(start_idx, end_idx):
srun.update_index_json(idx)
params['model_name'] = base_model_name + f'_run{idx}'
combination_str = (f'\nrun {idx}/{end_idx-1}')
params['combination_str'] = combination_str
params['verbose'] = True
start_time = time.time()
results = main_prog(copy.deepcopy(params))
end_time = time.time()
max_memory = torch.cuda.max_memory_allocated(device=params['device'])
torch.cuda.reset_max_memory_cached(params['device'])
# Log time/memory in train_results json file
results_dict = Project.load_results_dict(params['log_dir'])
results_dict[params['model_name']]['time'] = end_time - start_time
results_dict[params['model_name']]['max_memory'] = max_memory
Project.dump_results_dict(results_dict, params['log_dir'])
threshold_list = np.concatenate(
(np.zeros(1), np.arange(0.0002, 0.004,
0.0002), np.arange(0.004, 1, 0.05),
np.arange(1, 3, 0.2), np.arange(3, 10,
0.5), np.arange(10, 100, 2)))
for k in range(threshold_list.shape[0]):
threshold = threshold_list[k]
params[
'model_name'] = base_model_name + '_slope_diff_threshold_{:.4f}'.format(
threshold)
params['slope_diff_threshold'] = threshold
sys.stdout = open(os.devnull, "w")
main_prog(copy.deepcopy(params), isloaded_params=True)
sys.stdout = sys.__stdout__
results_dict = Project.load_results_dict(args.out_log_dir)
model_dict = results_dict[params['model_name']]
if k == 0:
assert np.allclose(threshold, 0)
base_train_acc = model_dict['latest_train_acc']
acc_drop = np.clip(
(model_dict['latest_train_acc'] - base_train_acc),
a_max=100,
a_min=-100)
print('\nThreshold: {:.4f}'.format(threshold))