(self, opt)
| 22 | return parser |
| 23 | |
| 24 | def modify_options(self, opt): |
| 25 | opt = self.override_options(opt) |
| 26 | |
| 27 | iterations = opt.lr_decay_epochs.split(',') |
| 28 | opt.lr_decay_epochs = list([]) |
| 29 | for it in iterations: |
| 30 | opt.lr_decay_epochs.append(int(it)) |
| 31 | |
| 32 | # set up saving name |
| 33 | opt.model_name = '{}_{}_{}_Jig_{}_{}_aug_{}_{}_{}'.format( |
| 34 | opt.method, opt.arch, opt.modal, opt.jigsaw, opt.mem, |
| 35 | opt.aug, opt.head, opt.nce_t |
| 36 | ) |
| 37 | if opt.amp: |
| 38 | opt.model_name = '{}_amp_{}'.format(opt.model_name, opt.opt_level) |
| 39 | if opt.cosine: |
| 40 | opt.model_name = '{}_cosine'.format(opt.model_name) |
| 41 | |
| 42 | # warm-up for large-batch training, e.g. 1024 with multiple nodes |
| 43 | if opt.batch_size > 256: |
| 44 | opt.warm = True |
| 45 | if opt.warm: |
| 46 | opt.model_name = '{}_warm'.format(opt.model_name) |
| 47 | opt.warmup_from = 0.01 |
| 48 | if opt.epochs > 500: |
| 49 | opt.warm_epochs = 10 |
| 50 | else: |
| 51 | opt.warm_epochs = 5 |
| 52 | if opt.cosine: |
| 53 | eta_min = opt.learning_rate * (opt.lr_decay_rate ** 3) |
| 54 | opt.warmup_to = eta_min + (opt.learning_rate - eta_min) * ( |
| 55 | 1 + math.cos(math.pi * opt.warm_epochs / opt.epochs)) / 2 |
| 56 | else: |
| 57 | opt.warmup_to = opt.learning_rate |
| 58 | |
| 59 | # create folders |
| 60 | opt.model_folder = os.path.join(opt.model_path, opt.model_name) |
| 61 | if not os.path.isdir(opt.model_folder): |
| 62 | os.makedirs(opt.model_folder) |
| 63 | opt.tb_folder = os.path.join(opt.tb_path, opt.model_name) |
| 64 | if not os.path.isdir(opt.tb_folder): |
| 65 | os.makedirs(opt.tb_folder) |
| 66 | |
| 67 | return opt |
nothing calls this directly
no test coverage detected