def test_model_generator_config_copy(self): class MyArchFactory(tpmaf.ArchitectureFactory): def new_architecture(self): return tpma.ModdedLeNet5Net(channels=1) arch = MyArchFactory() # setup the xforms to ensure we can test the callables def data_xform(x): return x * x def label_xform(y): return y * y * y data = tpmd.DataManager(self.experiment_path, self.train_file, self.clean_test_file, triggered_test_file=self.triggered_file, data_transform=data_xform, label_transform=label_xform, file_loader='image', shuffle_train=True, shuffle_clean_test=False, shuffle_triggered_test=False) num_models = 1 mgc1 = tpmc.ModelGeneratorConfig(arch, data, self.model_save_dir, self.stats_save_dir, num_models) mgc2 = copy.deepcopy(mgc1) self.assertEqual(mgc1, mgc2)
optim='adam', objective='cross_entropy_loss', early_stopping=early_stopping_argin, train_val_split=a.train_val_split) reporting_params = tpmc.ReportingConfig( num_batches_per_logmsg=500, num_epochs_per_metric=1, num_batches_per_metrics=default_nbpvdm, tensorboard_output_dir=a.tensorboard_dir, experiment_name=experiment_cfg['name']) optimizer_cfg = tpmc.DefaultOptimizerConfig(training_params, reporting_params) cfg = tpmc.ModelGeneratorConfig(arch, data_obj, model_save_dir, stats_save_dir, num_models, optimizer=optimizer_cfg, experiment_cfg=experiment_cfg, parallel=True) # may also provide lists of run_ids or filenames are arguments to ModelGeneratorConfig to have more control # of saved model file names; see RunnerConfig and ModelGeneratorConfig for more information modelgen_cfgs.append(cfg) model_generator = mg.ModelGenerator(modelgen_cfgs) start = time.time() model_generator.run() print("\nTime to run: ", (time.time() - start) / 60 / 60, 'hours')
def train_models(top_dir, data_folder, experiment_folder, experiment_list, model_save_folder, stats_save_folder, early_stopping, train_val_split, tensorboard_dir, gpu, uge, uge_dir): """ Given paths to the experiments and specifications to where models and model statistics should be saved, create triggered models for each experiment in the experiment directory. :param top_dir: (str) path to top level directory for text classification data and models are to be stored :param data_folder: (str) name of folder containing the experiments folder :param experiment_folder: (str) name of folder containing the experiments used to generate models :param model_save_folder: (str) name of folder under which models are to be saved :param stats_save_folder: (str) name of folder under which model training information is to be saved :param tensorboard_dir: (str) name of folder under which tensorboard information is to be saved :param gpu: (bool) use a gpu in training :param uge: (bool) use a Univa Grid Engine (UGE) to generate models :param uge_dir: (str) working directory for UGE models :return: None """ class MyArchFactory(tpm_af.ArchitectureFactory): def new_architecture(self, input_dim=25000, embedding_dim=100, hidden_dim=256, output_dim=1, n_layers=2, bidirectional=True, dropout=0.5, pad_idx=-999): return tpta.EmbeddingLSTM(input_dim, embedding_dim, hidden_dim, output_dim, n_layers, bidirectional, dropout, pad_idx) def arch_factory_kwargs_generator(train_dataset_desc, clean_test_dataset_desc, triggered_test_dataset_desc): # Note: the arch_factory_kwargs_generator returns a dictionary, which is used as kwargs input into an # architecture factory. Here, we allow the input-dimension and the pad-idx to be set when the model gets # instantiated. This is useful because these indices and the vocabulary size are not known until the # vocabulary is built. output_dict = dict(input_dim=train_dataset_desc.vocab_size, pad_idx=train_dataset_desc.pad_idx) return output_dict # get all available experiments from the experiment root directory experiment_path = os.path.join(top_dir, data_folder, experiment_folder) modelgen_cfgs = [] arch_factory_kwargs = dict( input_dim=25000, embedding_dim=100, hidden_dim=256, output_dim=1, n_layers=2, bidirectional=True, dropout=0.5 ) for i in range(len(experiment_list)): experiment_cfg = experiment_list[i] data_obj = dm.DataManager(experiment_path, experiment_cfg['train_file'], experiment_cfg['clean_test_file'], data_type='text', triggered_test_file=experiment_cfg['triggered_test_file'], shuffle_train=True, data_configuration=dc.TextDataConfiguration( max_vocab_size=arch_factory_kwargs['input_dim'], embedding_dim=arch_factory_kwargs['embedding_dim'])) num_models = 5 if uge: if gpu: device = torch.device('cuda') else: device = torch.device('cpu') else: device = torch.device('cuda' if torch.cuda.is_available() and gpu else 'cpu') default_nbpvdm = None if device.type == 'cpu' else 500 early_stopping_argin = tpmc.EarlyStoppingConfig() if early_stopping else None training_params = tpmc.TrainingConfig(device=device, epochs=10, batch_size=64, lr=1e-3, optim='adam', objective='BCEWithLogitsLoss', early_stopping=early_stopping_argin, train_val_split=train_val_split) reporting_params = tpmc.ReportingConfig(num_batches_per_logmsg=100, num_epochs_per_metric=1, num_batches_per_metrics=default_nbpvdm, tensorboard_output_dir=tensorboard_dir, experiment_name=experiment_cfg['name']) lstm_optimizer_config = tpmc.TorchTextOptimizerConfig(training_cfg=training_params, reporting_cfg=reporting_params, copy_pretrained_embeddings=True) optimizer = tptto.TorchTextOptimizer(lstm_optimizer_config) # There seem to be some issues w/ using the DataParallel w/ RNN's (hence, parallel=False). # See here: # - https://discuss.pytorch.org/t/pack-padded-sequence-with-multiple-gpus/33458 # - https://pytorch.org/docs/master/notes/faq.html#pack-rnn-unpack-with-data-parallelism # - https://github.com/pytorch/pytorch/issues/10537 # Although these issues are "old," the solutions provided in these forums haven't yet worked # for me to try to resolve the data batching error. For now, we suffice to using the single # GPU version. cfg = tpmc.ModelGeneratorConfig(MyArchFactory(), data_obj, model_save_folder, stats_save_folder, num_models, arch_factory_kwargs=arch_factory_kwargs, arch_factory_kwargs_generator=arch_factory_kwargs_generator, optimizer=optimizer, experiment_cfg=experiment_cfg, parallel=False, save_with_hash=True) # may also provide lists of run_ids or filenames as arguments to ModelGeneratorConfig to have more control # of saved model file names; see RunnerConfig and ModelGeneratorConfig for more information modelgen_cfgs.append(cfg) if uge: if gpu: q1 = tpmc.UGEQueueConfig("gpu-k40.q", True) q2 = tpmc.UGEQueueConfig("gpu-v100.q", True) q_cfg = tpmc.UGEConfig([q1, q2], queue_distribution=None) else: q1 = tpmc.UGEQueueConfig("htc.q", False) q_cfg = tpmc.UGEConfig(q1, queue_distribution=None) working_dir = uge_dir try: shutil.rmtree(working_dir) except IOError: pass model_generator = ugemg.UGEModelGenerator(modelgen_cfgs, q_cfg, working_directory=working_dir) else: model_generator = mg.ModelGenerator(modelgen_cfgs) start = time.time() model_generator.run() logger.debug("Time to run: ", (time.time() - start) / 60 / 60, 'hours')