dirty fix to start exp

JianyuanZhong · JianyuanZhong · commit f06efbd7b753 · 2020-11-02T19:49:14.000-05:00
diff --git a/recipes/LibriSpeech/ASR/transformer/experiment_dual_schedulers.py b/recipes/LibriSpeech/ASR/transformer/experiment_dual_schedulers.py
@@ -284,61 +284,74 @@ def on_evaluate_start(self, max_key=None, min_key=None):
     sys.path.append(os.path.dirname(os.path.dirname(current_dir)))
     from librispeech_prepare import prepare_librispeech  # noqa E402
 
-    # Load hyperparameters file with command-line overrides
-    hparams_file, overrides = sb.parse_arguments(sys.argv[1:])
-    with open(hparams_file) as fin:
-        hparams = sb.load_extended_yaml(fin, overrides)
-
-    # Create experiment directory
-    sb.create_experiment_directory(
-        experiment_directory=hparams["output_folder"],
-        hyperparams_to_save=hparams_file,
-        overrides=overrides,
-    )
-
-    # Prepare data
-    prepare_librispeech(
-        data_folder=hparams["data_folder"],
-        splits=hparams["train_splits"]
-        + [hparams["dev_split"], "test-clean", "test-other"],
-        merge_lst=hparams["train_splits"],
-        merge_name=hparams["csv_train"],
-        save_folder=hparams["data_folder"],
-    )
-
-    # Creating tokenizer must be done after preparation
-    tokenizer = hparams["tokenizer"]()
-
-    # Load index2label dict for decoding
-    train_set = hparams["train_loader"]()
-    valid_set = hparams["valid_loader"]()
-    test_clean_set = hparams["test_clean_loader"]()
-    test_other_set = hparams["test_other_loader"]()
-    ind2lab = hparams["test_other_loader"].label_dict["wrd"]["index2lab"]
-    hparams["ind2lab"] = ind2lab
-    hparams["tokenizer"] = tokenizer
-
-    # Brain class initialization
-    asr_brain = ASR(
-        modules=hparams["model"],
-        opt_class=hparams["optimizer"],
-        hparams=hparams,
-        checkpointer=hparams["checkpointer"],
-    )
-
-    asr_brain.load_tokenizer()
-    if hasattr(asr_brain.hparams, "lm_ckpt_file"):
-        asr_brain.load_lm()
-
-    # Training
-    asr_brain.fit(asr_brain.hparams.epoch_counter, train_set, valid_set)
-
-    # Test
-    asr_brain.hparams.wer_file = (
-        hparams["output_folder"] + "/wer_test_clean.txt"
-    )
-    asr_brain.evaluate(test_clean_set, max_key="ACC")
-    asr_brain.hparams.wer_file = (
-        hparams["output_folder"] + "/wer_test_other.txt"
-    )
-    asr_brain.evaluate(test_other_set, max_key="ACC")
+    while True:
+        try:
+            # Load hyperparameters file with command-line overrides
+            hparams_file, overrides = sb.parse_arguments(sys.argv[1:])
+            with open(hparams_file) as fin:
+                hparams = sb.load_extended_yaml(fin, overrides)
+
+            # Create experiment directory
+            sb.create_experiment_directory(
+                experiment_directory=hparams["output_folder"],
+                hyperparams_to_save=hparams_file,
+                overrides=overrides,
+            )
+
+            # Prepare data
+            prepare_librispeech(
+                data_folder=hparams["data_folder"],
+                splits=hparams["train_splits"]
+                + [hparams["dev_split"], "test-clean", "test-other"],
+                merge_lst=hparams["train_splits"],
+                merge_name=hparams["csv_train"],
+                save_folder=hparams["data_folder"],
+            )
+
+            # Creating tokenizer must be done after preparation
+            tokenizer = hparams["tokenizer"]()
+
+            # Load index2label dict for decoding
+            train_set = hparams["train_loader"]()
+            valid_set = hparams["valid_loader"]()
+            test_clean_set = hparams["test_clean_loader"]()
+            test_other_set = hparams["test_other_loader"]()
+            ind2lab = hparams["test_other_loader"].label_dict["wrd"]["index2lab"]
+            hparams["ind2lab"] = ind2lab
+            hparams["tokenizer"] = tokenizer
+            
+
+            # Brain class initialization
+            asr_brain = ASR(
+                modules=hparams["model"],
+                opt_class=hparams["optimizer"],
+                hparams=hparams,
+                checkpointer=hparams["checkpointer"],
+            )
+
+            asr_brain.load_tokenizer()
+            if hasattr(asr_brain.hparams, "lm_ckpt_file"):
+                asr_brain.load_lm()
+
+            # Training
+            asr_brain.fit(asr_brain.hparams.epoch_counter, train_set, valid_set)
+
+            # Test
+            asr_brain.hparams.wer_file = (
+                hparams["output_folder"] + "/wer_test_clean.txt"
+            )
+            asr_brain.evaluate(test_clean_set, max_key="ACC")
+            asr_brain.hparams.wer_file = (
+                hparams["output_folder"] + "/wer_test_other.txt"
+            )
+            asr_brain.evaluate(test_other_set, max_key="ACC")
+
+        except KeyboardInterrupt:
+            sys.exit()
+
+        except:
+            msg = "rand {} failed with pkl I/O, retrying...."
+            print(msg)
+
+        
+            
diff --git a/speechbrain/data_io/data_io.py b/speechbrain/data_io/data_io.py
@@ -206,7 +206,7 @@ def get_dataloader(self, sampler=None):
             self.dataset,
             batch_size=self.batch_size,
             shuffle=self.shuffle if sampler is None else False,
-            pin_memory=(sampler is not None),
+            pin_memory=False,
             drop_last=self.drop_last,
             num_workers=self.num_workers,
             collate_fn=self.batch_creation,
diff --git a/speechbrain/decoders/ctc.py b/speechbrain/decoders/ctc.py
@@ -47,11 +47,11 @@ def __init__(self, x, xlens, blank, eos, margin=0):
         )
         # Pad the rest of posteriors in the batch
         # TODO(takaaki-hori): need a better way without for-loops
-        # for i, l in enumerate(xlens):
-        #     if l < self.input_length:
-        #         x[i, l:, :] = self.logzero
-        #         x[i, l:, blank] = 0
-        x[:, :, blank] = 0
+        for i, l in enumerate(xlens):
+            if l < self.input_length:
+                x[i, l:, :] = self.logzero
+                x[i, l:, blank] = 0
+
         # Reshape input x
         xn = x.transpose(0, 1)  # (B, T, O) -> (T, B, O)
         xb = xn[:, :, self.blank].unsqueeze(2).expand(-1, -1, self.odim)
diff --git a/speechbrain/decoders/seq2seq.py b/speechbrain/decoders/seq2seq.py
@@ -9,7 +9,7 @@
 import numpy as np
 
 import speechbrain as sb
-from speechbrain.decoders.ctc import CTCPrefixScorer
+from speechbrain.decoders.ctc import CTCPrefixScorer, CTCPrefixScoreTH
 
 
 class S2SBaseSearcher(torch.nn.Module):
@@ -538,11 +538,19 @@ def forward(self, enc_states, wav_len):  # noqa: C901
         if self.ctc_weight > 0:
             # (batch_size * beam_size, L, vocab_size)
             ctc_outputs = self.ctc_forward_step(enc_states)
-            ctc_scorer = CTCPrefixScorer(
+            # ctc_scorer = CTCPrefixScorer(
+            #     ctc_outputs,
+            #     enc_lens,
+            #     batch_size,
+            #     self.beam_size,
+            #     0,
+            #     self.eos_index,
+            # )
+            ctc_scorer = CTCPrefixScoreTH(
                 ctc_outputs,
                 enc_lens,
-                batch_size,
-                self.beam_size,
+                # batch_size,
+                # self.beam_size,
                 0,
                 self.eos_index,
             )
@@ -626,7 +634,8 @@ def forward(self, enc_states, wav_len):  # noqa: C901
 
             # adding CTC scores to log_prob if ctc_weight > 0
             if self.ctc_weight > 0:
-                g = alived_seq
+                # g = alived_seq
+                g = memory
                 # block blank token
                 log_probs[:, self.bos_index] = self.minus_inf
                 if self.ctc_weight != 1.0:
diff --git a/speechbrain/lobes/models/transformer/TransformerASR.py b/speechbrain/lobes/models/transformer/TransformerASR.py
@@ -140,6 +140,7 @@ def forward(
             memory=encoder_out,
             tgt_mask=tgt_mask,
             tgt_key_padding_mask=tgt_key_padding_mask,
+            # memory_key_padding_mask=src_key_padding_mask,
         )
 
         return encoder_out, decoder_out
diff --git a/speechbrain/processing/features.py b/speechbrain/processing/features.py
@@ -956,6 +956,7 @@ def __init__(
         norm_type="global",
         avg_factor=None,
         requires_grad=False,
+        update_until_epoch=3,
     ):
         super().__init__()
         self.mean_norm = mean_norm
@@ -972,8 +973,9 @@ def __init__(
         self.count = 0
         self.eps = 1e-10
         self.device_inp = torch.device("cpu")
+        self.update_until_epoch = update_until_epoch
 
-    def forward(self, x, lengths, spk_ids=torch.tensor([])):
+    def forward(self, x, lengths, spk_ids=torch.tensor([]), epoch=0):
         """Returns the tensor with the sourrounding context.
 
         Arguments
@@ -1059,7 +1061,7 @@ def forward(self, x, lengths, spk_ids=torch.tensor([])):
                     self.glob_mean = current_mean
                     self.glob_std = current_std
 
-                else:
+                elif epoch < self.update_until_epoch:
                     if self.avg_factor is None:
                         self.weight = 1 / (self.count + 1)
                     else:

Original file line number	Diff line number	Diff line change
`@@ -140,6 +140,7 @@ def forward(`
`140`	`140`	`memory=encoder_out,`
`141`	`141`	`tgt_mask=tgt_mask,`
`142`	`142`	`tgt_key_padding_mask=tgt_key_padding_mask,`
	`143`	`+ # memory_key_padding_mask=src_key_padding_mask,`
`143`	`144`	`)`
`144`	`145`
`145`	`146`	`return encoder_out, decoder_out`