git push -fCustom checkpoint loading hook

speechbrain/decoders/seq2seq.py

@@ -7,7 +7,6 @@
  * Mirco Ravanelli 2020
  * Sung-Lin Yeh 2020
 """
-
 import torch
 from speechbrain.decoders.utils import (
     inflate_tensor,
@@ -18,7 +17,7 @@
 
 
 class AlivedHypotheses(torch.nn.Module):
-    """This class handle the data for the hypotheses during the decoding.
+    """ This class handle the data for the hypotheses during the decoding.
 
     Arguments
     ---------
@@ -721,7 +720,7 @@ def _max_attn_shift_permute_memory_step(self, prev_attn_peak, predecessors):
         return prev_attn_peak
 
     def _update_reset_memory(self, enc_states, enc_lens):
-        """Call reset memory for each module.
+        """ Call reset memory for each module.
 
         Arguments
         ---------
@@ -1813,7 +1812,7 @@ def set_n_out(self):
     def forward_step(self, inp_tokens, memory, enc_states, enc_lens):
         """Performs a step in the implemented beamsearcher."""
         memory = _update_mem(inp_tokens, memory)
-        (dec_out, attn,) = self.model.forward_decoder(enc_states, memory)
+        dec_out, attn, = self.model.forward_decoder(enc_states, memory)
         log_probs = self.softmax(dec_out[:, -1] / self.temperature)
         return log_probs, memory, attn
 

speechbrain/lobes/models/huggingface_transformers/speecht5.py

@@ -8,6 +8,7 @@
  * Haroun Elleuch 2024
 """
 
+import pathlib
 import torch
 import logging
 
@@ -16,12 +17,17 @@
 )
 
 from transformers import SpeechT5ForSpeechToText, SpeechT5Config
+from speechbrain.utils.checkpoints import (
+    mark_as_loader,
+    register_checkpoint_hooks,
+)
 from speechbrain.utils.fetching import fetch
 
 
 logger = logging.getLogger(__name__)
 
 
+@register_checkpoint_hooks
 class SpeechT5ForASR(HFTransformersInterface):
     """This lobe enables the integration of HuggingFace and SpeechBrain
     pretrained SpeechT5 models for Automatic Speech Recognition.
@@ -286,16 +292,19 @@ def _from_pretrained(self, source: str, save_path: str, cache_dir: str):
         is_sb, ckpt_file, _ = self._check_model_source(source, save_path)
 
         if is_sb or self.for_pretraining:
-            self.model = SpeechT5ForSpeechToText.from_config(self.config)
+            self.model = SpeechT5ForSpeechToText._from_config(self.config)
 
         if is_sb:
             self.model.gradient_checkpointing_disable()  # Required by DDP
             # fetch the checkpoint file
             ckpt_full_path = fetch(
-                filename=ckpt_file, source=source, savedir=save_path,
+                filename=ckpt_file,
+                source=source,
+                savedir=save_path,
+                huggingface_cache_dir=cache_dir,
             )
             # We transfer the parameters from the checkpoint.
-            self._load_sb_pretrained_parameters(ckpt_full_path)
+            self._load_sb_pretrained_parameters(path=ckpt_full_path,)
         elif not self.for_pretraining:
             self.model = SpeechT5ForSpeechToText.from_pretrained(
                 source,
@@ -305,6 +314,30 @@ def _from_pretrained(self, source: str, save_path: str, cache_dir: str):
                 ignore_mismatched_sizes=True,
             )
 
+    @mark_as_loader
+    def _on_load_checkpoint(
+        self, path: pathlib.Path | str, end_of_epoch: bool
+    ) -> None:
+        loaded_state_dict = torch.load(path)
+        model_state_dict = self.state_dict()
+        is_changed = False
+        for k in loaded_state_dict:
+            if k in model_state_dict:
+                if loaded_state_dict[k].shape != model_state_dict[k].shape:
+                    logger.warning(
+                        f"Skip loading parameter: {k}, "
+                        f"required shape: {model_state_dict[k].shape}, "
+                        f"loaded shape: {loaded_state_dict[k].shape}"
+                    )
+                    loaded_state_dict[k] = model_state_dict[k]
+                    is_changed = True
+            else:
+                logger.warning(f"Dropping parameter {k}")
+                is_changed = True
+
+        if is_changed:
+            loaded_state_dict.pop("optimizer_states", None)
+
 
 def custom_padding(x, org_pad, custom_pad):
     """