IsADirectoryError: [Errno 21] Is a directory: '../saved_models/conv1d_transformer_decoder_tts_joint'

[GO0108]

Hello, I attempted to test the model by following the steps provided in etc/inference.ipynb. I downloaded the YoloStutter checkpoint from Google Drive, but it seems the code expects a .pt file or a similar format. Am I missing something here? Should I be using the pretrained VITS weights at this step?

---

IsADirectoryError Traceback (most recent call last)
Cell In[14], line 20
16 device = torch.device("cuda:0")
18 # decoder = Decoder(in_channels, out_channels, kernel_size, kernel_stride, num_blocks, num_classes, downsample_factor)
---> 20 decoder = torch.load("../saved_models/conv1d_transformer_decoder_tts_joint", map_location=device)
21 dataset = Dataset(
22 "../stutter_VCTK_joint/", hps, downsample_factor, skip_raw_label_comparison=True
23 )
24 train_set, val_set = torch.utils.data.random_split(dataset, [0, 1])

File /opt/conda/envs/yolostutter/lib/python3.10/site-packages/torch/serialization.py:986, in load(f, map_location, pickle_module, weights_only, mmap, **pickle_load_args)
983 if 'encoding' not in pickle_load_args.keys():
984 pickle_load_args['encoding'] = 'utf-8'
--> 986 with _open_file_like(f, 'rb') as opened_file:
987 if _is_zipfile(opened_file):
988 # The zipfile reader is going to advance the current file position.
989 # If we want to actually tail call to torch.jit.load, we need to
990 # reset back to the original position.
991 orig_position = opened_file.tell()

File /opt/conda/envs/yolostutter/lib/python3.10/site-packages/torch/serialization.py:435, in _open_file_like(name_or_buffer, mode)
433 def _open_file_like(name_or_buffer, mode):
434 if _is_path(name_or_buffer):
--> 435 return _open_file(name_or_buffer, mode)
436 else:
437 if 'w' in mode:

File /opt/conda/envs/yolostutter/lib/python3.10/site-packages/torch/serialization.py:416, in _open_file.init(self, name, mode)
415 def init(self, name, mode):
--> 416 super().init(open(name, mode))

IsADirectoryError: [Errno 21] Is a directory: '../saved_models/conv1d_transformer_decoder_tts_joint'

[rorizzz]

hey, I feel the issue is not about the checkpoint, but the path you put into the torch.load should be decoder = torch.load("../saved_models/decoder_tts_new_l", map_location=device).

Besides, pretrained VITS weights are not applicable here, since they are used for obtaining the soft_alignment, but here we are uploading the weights for decoder.

[GO0108]

Got it, but I still have the same issue with this path. Is this error related to a specific torch version?