fix: stratified train/test split in notebook

[meilame-tayebjee]

avoid errors due to smaller num_classes than the max label

[micedre]

Still have this error when training on GPU (Nvidia T4 15G) :

2025-05-19 12:26:12 - torchFastText.torchFastText - Checking inputs...
2025-05-19 12:26:12 - torchFastText.torchFastText - Inputs successfully checked. Starting the training process..
2025-05-19 12:26:12 - torchFastText.torchFastText - Running on: cuda
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[31], line 1
----> 1 model.train(
      2     X_train,
      3     y_train,
      4     X_test,
      5     y_test,
      6     num_epochs=parameters_train['num_epochs'],
      7     batch_size=parameters_train['batch_size'],
      8     patience_scheduler=parameters_train['patience'],
      9     patience_train=parameters_train['patience'],
     10     lr=parameters_train['lr'],
     11     verbose = True
     12 )

File /usr/local/lib/python3.12/site-packages/torchFastText/torchFastText.py:589, in torchFastText.train(self, X_train, y_train, X_val, y_val, num_epochs, batch_size, cpu_run, num_workers, optimizer, optimizer_params, lr, scheduler, patience_scheduler, loss, patience_train, verbose, trainer_params)
    586         end = time.time()
    587         logger.info("Model successfully built in {:.2f} seconds.".format(end - start))
--> 589 self.pytorch_model = self.pytorch_model.to(self.device)
    591 # Dataloaders
    592 train_dataloader, val_dataloader = self.__build_data_loaders(
    593     train_categorical_variables=train_categorical_variables,
    594     training_text=training_text,
   (...)    600     num_workers=num_workers,
    601 )

File /usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py:1355, in Module.to(self, *args, **kwargs)
   1352         else:
   1353             raise
-> 1355 return self._apply(convert)

File /usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py:915, in Module._apply(self, fn, recurse)
    913 if recurse:
    914     for module in self.children():
--> 915         module._apply(fn)
    917 def compute_should_use_set_data(tensor, tensor_applied):
    918     if torch._has_compatible_shallow_copy_type(tensor, tensor_applied):
    919         # If the new tensor has compatible tensor type as the existing tensor,
    920         # the current behavior is to change the tensor in-place using `.data =`,
   (...)    925         # global flag to let the user control whether they want the future
    926         # behavior of overwriting the existing tensor or not.

File /usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py:942, in Module._apply(self, fn, recurse)
    938 # Tensors stored in modules are graph leaves, and we don't want to
    939 # track autograd history of `param_applied`, so we have to use
    940 # `with torch.no_grad():`
    941 with torch.no_grad():
--> 942     param_applied = fn(param)
    943 p_should_use_set_data = compute_should_use_set_data(param, param_applied)
    945 # subclasses may have multiple child tensors so we need to use swap_tensors

File /usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py:1341, in Module.to.<locals>.convert(t)
   1334     if convert_to_format is not None and t.dim() in (4, 5):
   1335         return t.to(
   1336             device,
   1337             dtype if t.is_floating_point() or t.is_complex() else None,
   1338             non_blocking,
   1339             memory_format=convert_to_format,
   1340         )
-> 1341     return t.to(
   1342         device,
   1343         dtype if t.is_floating_point() or t.is_complex() else None,
   1344         non_blocking,
   1345     )
   1346 except NotImplementedError as e:
   1347     if str(e) == "Cannot copy out of meta tensor; no data!":

RuntimeError: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

[micedre]

Suggested change

	"pip install torchFastText \n",
	"!pip install torchFastText \n",

[micedre]

This line cause an error when predicting (when training was done on GPU)

[micedre]

Still have this error when training on GPU (Nvidia T4 15G) :

2025-05-19 12:26:12 - torchFastText.torchFastText - Checking inputs...
2025-05-19 12:26:12 - torchFastText.torchFastText - Inputs successfully checked. Starting the training process..
2025-05-19 12:26:12 - torchFastText.torchFastText - Running on: cuda
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[31], line 1
----> 1 model.train(
      2     X_train,
      3     y_train,
      4     X_test,
      5     y_test,
      6     num_epochs=parameters_train['num_epochs'],
      7     batch_size=parameters_train['batch_size'],
      8     patience_scheduler=parameters_train['patience'],
      9     patience_train=parameters_train['patience'],
     10     lr=parameters_train['lr'],
     11     verbose = True
     12 )

File /usr/local/lib/python3.12/site-packages/torchFastText/torchFastText.py:589, in torchFastText.train(self, X_train, y_train, X_val, y_val, num_epochs, batch_size, cpu_run, num_workers, optimizer, optimizer_params, lr, scheduler, patience_scheduler, loss, patience_train, verbose, trainer_params)
    586         end = time.time()
    587         logger.info("Model successfully built in {:.2f} seconds.".format(end - start))
--> 589 self.pytorch_model = self.pytorch_model.to(self.device)
    591 # Dataloaders
    592 train_dataloader, val_dataloader = self.__build_data_loaders(
    593     train_categorical_variables=train_categorical_variables,
    594     training_text=training_text,
   (...)    600     num_workers=num_workers,
    601 )

File /usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py:1355, in Module.to(self, *args, **kwargs)
   1352         else:
   1353             raise
-> 1355 return self._apply(convert)

File /usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py:915, in Module._apply(self, fn, recurse)
    913 if recurse:
    914     for module in self.children():
--> 915         module._apply(fn)
    917 def compute_should_use_set_data(tensor, tensor_applied):
    918     if torch._has_compatible_shallow_copy_type(tensor, tensor_applied):
    919         # If the new tensor has compatible tensor type as the existing tensor,
    920         # the current behavior is to change the tensor in-place using `.data =`,
   (...)    925         # global flag to let the user control whether they want the future
    926         # behavior of overwriting the existing tensor or not.

File /usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py:942, in Module._apply(self, fn, recurse)
    938 # Tensors stored in modules are graph leaves, and we don't want to
    939 # track autograd history of `param_applied`, so we have to use
    940 # `with torch.no_grad():`
    941 with torch.no_grad():
--> 942     param_applied = fn(param)
    943 p_should_use_set_data = compute_should_use_set_data(param, param_applied)
    945 # subclasses may have multiple child tensors so we need to use swap_tensors

File /usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py:1341, in Module.to.<locals>.convert(t)
   1334     if convert_to_format is not None and t.dim() in (4, 5):
   1335         return t.to(
   1336             device,
   1337             dtype if t.is_floating_point() or t.is_complex() else None,
   1338             non_blocking,
   1339             memory_format=convert_to_format,
   1340         )
-> 1341     return t.to(
   1342         device,
   1343         dtype if t.is_floating_point() or t.is_complex() else None,
   1344         non_blocking,
   1345     )
   1346 except NotImplementedError as e:
   1347     if str(e) == "Cannot copy out of meta tensor; no data!":

RuntimeError: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

Could not reproduce it today... Either way, it works with cpu.