Updates model outputs to make ctcdecode stop at the right timestep

decoder.py

@@ -105,6 +105,8 @@ def convert_to_strings(self, out, seq_len):
             for p, utt in enumerate(batch):
                 size = seq_len[b][p]
                 if size > 0:
+#                    print(utt.size())
+#                    print(utt[0:size])
                     transcript = ''.join(map(lambda x: self.int_to_char[x], utt[0:size]))
                 else:
                     transcript = ''
@@ -137,6 +139,7 @@ def decode(self, probs, sizes=None):
         """
         probs = probs.cpu().transpose(0, 1).contiguous()
         out, scores, offsets, seq_lens = self._decoder.decode(probs)
+        #print(seq_lens)
 
         strings = self.convert_to_strings(out, seq_lens)
         offsets = self.convert_tensor(offsets, seq_lens)

test.py

@@ -1,6 +1,7 @@
 import argparse
 
 import numpy as np
+import torch
 from torch.autograd import Variable
 from tqdm import tqdm
 
@@ -63,8 +64,10 @@
     output_data = []
     for i, (data) in tqdm(enumerate(test_loader), total=len(test_loader)):
         inputs, targets, input_percentages, target_sizes = data
+        #print(input_percentages)
 
-        inputs = Variable(inputs, volatile=True)
+        with torch.no_grad():
+            inputs = Variable(inputs)
 
         # unflatten targets
         split_targets = []
@@ -73,14 +76,23 @@
             split_targets.append(targets[offset:offset + size])
             offset += size
 
+        #print(inputs.size())
         if args.cuda:
             inputs = inputs.cuda()
 
+
         out = model(inputs)
         out = out.transpose(0, 1)  # TxNxH
+        #print(out.size())
         seq_length = out.size(0)
         sizes = input_percentages.mul_(int(seq_length)).int()
 
+        for i in range(out.size(1)):
+            start_idx = sizes[i] - 1
+            out.data[start_idx:,i, :] = torch.zeros(seq_length-start_idx, out.size(2))-1 
+
+        #print((out.transpose(0,1)).data)
+
         if decoder is None:
             # add output to data array, and continue
             output_data.append((out.data.cpu().numpy(), sizes.numpy()))
@@ -90,7 +102,8 @@
         target_strings = target_decoder.convert_to_strings(split_targets)
         wer, cer = 0, 0
         for x in range(len(target_strings)):
-            transcript, reference = decoded_output[x][0], target_strings[x][0]
+            transcript, reference = decoded_output[x][0], target_strings[x][0].replace(" '", "'")
+            print("Prediction: {}\nReference: {}\n-------------------------".format(transcript, reference))
             wer_inst = decoder.wer(transcript, reference) / float(len(reference.split()))
             cer_inst = decoder.cer(transcript, reference) / float(len(reference))
             wer += wer_inst

tune_decoder.py

@@ -60,11 +60,18 @@ def decode_dataset(logits, test_dataset, batch_size, lm_alpha, lm_beta, mesh_x,
         out = torch.from_numpy(logits[i][0])
         sizes = torch.from_numpy(logits[i][1])
 
+        #print(logits[i][0])
+#        seq_length = out.size(0)
+#        for i in range(out.size(1)):
+#            start_idx = sizes[i] - 1
+#            out.data[start_idx:,i, :] = torch.zeros(seq_length-start_idx, out.size(2))-1 
+
         decoded_output, _ = decoder.decode(out, sizes)
         target_strings = target_decoder.convert_to_strings(split_targets)
         wer, cer = 0, 0
         for x in range(len(target_strings)):
             transcript, reference = decoded_output[x][0], target_strings[x][0]
+            print("Prediction: {}\nReference: {}\n-------------------------".format(transcript, reference))
             wer_inst = decoder.wer(transcript, reference) / float(len(reference.split()))
             cer_inst = decoder.cer(transcript, reference) / float(len(reference))
             wer += wer_inst