device = torch.device('cpu') from eval_utils import * warnings.simplefilter("ignore") from tqdm import tqdm config = configs.CONFIG_MAP['mlp_mfcc'] model = config['model'](linear_layer_size=config['linear_layer_size'], filter_sizes=config['filter_sizes']) model.set_device(device) model.to(device) checkpoint_dir = '../../checkpoints/comparisons/baseline_mfcc_trained_on_audioset' if os.path.exists(checkpoint_dir): torch_utils.load_checkpoint(checkpoint_dir + '/best.pth.tar', model) else: print("Checkpoint not found") model.eval() swb_val_distractor_df = pd.read_csv( '../../data/switchboard/annotations/clean_switchboard_val_distractor_annotations.csv' ) swb_test_distractor_df = pd.read_csv( '../../data/switchboard/annotations/clean_switchboard_test_distractor_annotations.csv' ) swb_val_df = pd.read_csv( '../../data/switchboard/annotations/clean_switchboard_val_laughter_annotations.csv' )
enc = Encoder(INPUT_DIM, HID_DIM, ENC_DROPOUT) dec = EmbeddingDecoder(OUTPUT_DIM, HID_DIM, DEC_EMB_DIM, output_embedding_matrix, word_output_vocab, DEC_DROPOUT, device) model = Seq2Seq(enc, dec, device, max_label_len).to(device) torch_utils.count_parameters(model) model.apply(torch_utils.init_weights) optimizer = optim.Adam(model.parameters()) criterion = nn.CrossEntropyLoss( ignore_index=word_output_vocab[text_utils.PAD_SYMBOL]) if os.path.exists(checkpoint_path): torch_utils.load_checkpoint(checkpoint_path, model, optimizer) else: print("Saving checkpoints to ", checkpoint_dir) print("Beginning training...") print("Batch Size: ", batch_size) def text_log(checkpoint_dir): text_path = os.path.join(checkpoint_dir, 'sample_preds.txt') i = np.random.randint(len(test_phoneme_files)) device = model.device model.set_device(torch.device('cpu')) pdr = torch_utils.Predictor(train_dataset, filepaths=test_tg_files[i:i + 1], model=model,
save_to_textgrid = bool(strtobool(args.save_to_textgrid)) output_dir = args.output_dir device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') print(f"Using device {device}") ##### Load the Model model = config['model'](dropout_rate=0.0, linear_layer_size=config['linear_layer_size'], filter_sizes=config['filter_sizes']) feature_fn = config['feature_fn'] model.set_device(device) if os.path.exists(model_path): torch_utils.load_checkpoint(model_path + '/best.pth.tar', model) model.eval() else: raise Exception(f"Model checkpoint not found at {model_path}") ##### Load the audio file and features inference_dataset = data_loaders.SwitchBoardLaughterInferenceDataset( audio_path=audio_path, feature_fn=feature_fn, sr=sample_rate) collate_fn = partial(audio_utils.pad_sequences_with_labels, expand_channel_dim=config['expand_channel_dim']) inference_generator = torch.utils.data.DataLoader(inference_dataset, num_workers=4, batch_size=8,
# hyperparameters # ============================================================================= json_path = os.path.join(args.exp_dir, 'params.json') assert os.path.isfile( json_path), "No json configuration file found at {}".format(json_path) params = utils.Params(json_path) # ============================================================================= # dataset # ============================================================================= transform = RandomTransform(10.0) data_path = osp.join(args.exp_dir, 'test_list') assert osp.isfile(data_path), "No test list file found at {}".format(data_path) test_dataset, path_list = load_dataset(data_path, transform) test_loader = DataLoader(test_dataset, batch_size=1, shuffle=False, num_workers=1) logging.info('{} testing data'.format(len(test_dataset))) # ============================================================================= # model # ============================================================================= device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') model = net.Net(params.nClassesTotal).to(device) load_checkpoint(osp.join(args.exp_dir, 'best.pth.tar'), model) acc, iou = test(test_loader) logging.info('Acc: {:.4f}, IoU: {:.4f}'.format(acc, iou))
return epoch_loss / len(iterator) print("Initializing model...") device = torch.device(torch_device if torch.cuda.is_available() else 'cpu') print("Using device", device) model = config['model'](dropout_rate=dropout_rate, linear_layer_size=config['linear_layer_size'], filter_sizes=config['filter_sizes']) model.set_device(device) torch_utils.count_parameters(model) model.apply(torch_utils.init_weights) optimizer = optim.Adam(model.parameters()) if os.path.exists(checkpoint_dir): torch_utils.load_checkpoint(checkpoint_dir + '/last.pth.tar', model, optimizer) else: print("Saving checkpoints to ", checkpoint_dir) print("Beginning training...") writer = SummaryWriter(checkpoint_dir) if augment_fn is not None: print("Loading background noise files...") noise_signals = load_noise_files() augment_fn = partial(augment_fn, noise_signals=noise_signals) print("Loading impulse responses...") impulse_responses = load_impulse_responses() augment_fn = partial(augment_fn, impulse_responses=impulse_responses) if supervised_augment: