new beam_las

Author: peter-yh-wu

tasks/Miami/beam_las/main.py

@@ -127,7 +127,7 @@ def forward(self, utterances, utterance_lengths):
         sorted_lengths, order = torch.sort(utterance_lengths, 0, descending=True)
         _, backorder = torch.sort(order, 0)
         h = h[:, order, :]
-        h = pack_padded_sequence(h, sorted_lengths) # .data.cpu().numpy())
+        h = pack_padded_sequence(h, sorted_lengths.data.cpu().numpy())
 
         # RNNs
         for rnn in self.rnns:
@@ -577,7 +577,7 @@ def main():
     print_log('%.2f Seconds' % (t1-t0), LOG_PATH)
 
     print("Running")
-    CKPT_PATH = os.path.join(args.save_directory, 'best_model.ckpt')
+    CKPT_PATH = os.path.join(args.save_directory, 'model.ckpt')
     if os.path.exists(CKPT_PATH):
         model.load_state_dict(torch.load(CKPT_PATH))
     if torch.cuda.is_available():
@@ -643,8 +643,6 @@ def main():
                 torch.save(model.state_dict(), CKPT_PATH)
             elif e - prev_best_epoch > args.patience:
                 break
-            torch.save(model.state_dict(), os.path.join(args.save_directory, f'model_{e}.ckpt'))
-            print(f'Saved model epoch {e}')
             print_log('Val Loss: %f' % val_loss, LOG_PATH)
             print_log('Avg Val Perplexity: %f' % (tot_perp/len(train_loader.dataset)), LOG_PATH)
             cer_val = cer(args, model, dev_loader, charset, dev_ys, device=args.cuda)

tasks/Miami/beam_las/model_utils.py

@@ -229,21 +229,9 @@ def __init__(self, ids, labels=None):
         '''
         parent_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
         self.mfcc_dir = os.path.join(parent_dir, 'data/mfcc')
-        mfcc_files = os.listdir(self.mfcc_dir)
-        mfcc_paths_set = set([os.path.join(self.mfcc_dir, f) for f in mfcc_files])
         self.ids = ids
         if labels:
             self.labels = [torch.from_numpy(y + 1).long() for y in labels]  # +1 for start/end token
-            new_ids = []
-            new_labels = []
-            for i, label in enumerate(self.labels):
-                curr_id = self.ids[i]
-                curr_mfcc_path = os.path.join(self.mfcc_dir, curr_id+'.mfcc')
-                if curr_mfcc_path in mfcc_paths_set:
-                    new_ids.append(curr_id)
-                    new_labels.append(label)
-            self.ids = new_ids
-            self.labels = new_labels
             assert len(self.ids) == len(self.labels)
         else:
             self.labels = None
@@ -304,7 +292,7 @@ def make_loader(ids, labels, args, shuffle=True, batch_size=64):
         labels: list of 1-dim int np arrays
     '''
     # Build the DataLoaders
-    kwargs = {'pin_memory': True, 'num_workers': args.num_workers} if args.cuda else {}
+    kwargs = {'pin_memory': True, 'num_workers': args.num_workers} if torch.cuda.is_available() else {}
     dataset = ASRDataset(ids, labels)
     loader = DataLoader(dataset, collate_fn=speech_collate_fn, shuffle=shuffle, batch_size=batch_size, **kwargs)
     return loader

tasks/Miami/beam_las/test_model.py

@@ -24,13 +24,12 @@
 from torch.nn.utils.rnn import PackedSequence
 from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
 
-from baseline import parse_args, Seq2SeqModel, write_transcripts
+from main import parse_args, Seq2SeqModel, write_transcripts
 from model_utils import *
 
 
 def main():
     args = parse_args()
-    args.cuda = not args.no_cuda and torch.cuda.is_available()
 
     t0 = time.time()
 
@@ -41,16 +40,9 @@ def main():
         pass
 
     print("Loading File Paths")
-    train_paths, dev_paths, test_paths = load_paths()
-    train_paths, dev_paths, test_paths = train_paths[:args.max_train], dev_paths[:args.max_dev], test_paths[:args.max_test]
-    t1 = time.time()
-    print_log('%.2f Seconds' % (t1-t0), LOG_PATH)
-
-    print("Loading Y Data")
-    test_paths = test_paths[:args.max_data]
-    train_ys = load_y_data('train') # 1-dim np array of strings
-    dev_ys = load_y_data('dev')
-    test_ys = load_y_data('test')
+    train_ids, train_ys = load_fid_and_y_data('train')
+    dev_ids, dev_ys = load_fid_and_y_data('dev')
+    test_ids, test_ys = load_fid_and_y_data('test')
     t1 = time.time()
     print_log('%.2f Seconds' % (t1-t0), LOG_PATH)
 
@@ -64,13 +56,13 @@ def main():
     print("Mapping Characters")
     testchars = map_characters(test_ys, charmap)
     print("Building Loader")
-    test_loader = make_loader(test_paths, testchars, args, shuffle=False, batch_size=1)
+    test_loader = make_loader(test_ids, testchars, args, shuffle=False, batch_size=1)
 
     print("Building Model")
     model = Seq2SeqModel(args, vocab_size=charcount, beam_width=args.beam_width)
 
     CKPT_PATH = os.path.join(args.save_directory, 'model.ckpt')
-    if args.cuda:
+    if torch.cuda.is_available():
         model.load_state_dict(torch.load(CKPT_PATH))
     else:
         gpu_dict = torch.load(CKPT_PATH, map_location=lambda storage, loc: storage)
@@ -80,7 +72,7 @@ def main():
         model.load_state_dict(cpu_model_dict)
     print("Loaded Checkpoint")
 
-    if args.cuda:
+    if torch.cuda.is_available():
         model = model.cuda()
 
     model.eval()