o si  @sddlZddlZddlmZddlmZddZddZddd Zd d Zd d Z e ddej dej de dej fddZ ddZ ddZe ddZdS)N)nn) OrderedDictcCsjt|tjr |St|trdd|DSt|tr,|D] }t||||<q|Std t |)a{Transfer tensor, dict or list of tensors to GPU. Args: tensors (:class:`torch.Tensor`, list or dict): May be a single, a list or a dictionary of tensors. Returns: :class:`torch.Tensor`: Same as input but transferred to cuda. Goes through lists and dicts and transfers the torch.Tensor to cuda. Leaves the rest untouched. cSsg|]}t|qS)to_cuda.0tensrrN/home/ubuntu/.local/lib/python3.10/site-packages/asteroid/utils/torch_utils.py szto_cuda..zNtensors must be a tensor or a list or dict of tensors. Got tensors of type {}) isinstancetorchTensorcudalistdictkeysr TypeErrorformattype)tensorskeyrrr rs     rcsht|tjr |St|ttfrfdd|DSt|tr2|D] }t||||<q$|S|S)aTransfer tensor, dict or list of tensors to device. Args: tensors (:class:`torch.Tensor`): May be a single, a list or a dictionary of tensors. device (:class: `torch.device`): the device where to place the tensors. Returns: Union [:class:`torch.Tensor`, list, tuple, dict]: Same as input but transferred to device. Goes through lists and dicts and transfers the torch.Tensor to device. Leaves the rest untouched. csg|]}t|qSr)tensors_to_devicerdevicerr r 3sz%tensors_to_device..) r r r tortuplerrr)rrrrrr r"s    rcCsNt|dr|jSt|drt|jS|dur"tdt|dt|S)asGet the device of a tensor or a module. Args: tensor_or_module (Union[torch.Tensor, torch.nn.Module]): The object to get the device from. Can be a ``torch.Tensor``, a ``torch.nn.Module``, or anything else that has a ``device`` attribute or a ``parameters() -> Iterator[torch.Tensor]`` method. default (Optional[Union[str, torch.device]]): If the device can not be determined, return this device instead. If ``None`` (the default), raise a ``TypeError`` instead. Returns: torch.device: The device that ``tensor_or_module`` is on. r parametersNz Don't know how to get device of z object)hasattrrnextrrrr )tensor_or_moduledefaultrrr get_device<s   r!cCs tjS)z Returns ``True`` in tracing (if a function is called during the tracing of code with ``torch.jit.trace``) and ``False`` otherwise. )r _C _is_tracingrrrr is_tracingUs r$cs(tfdd_d_S)a Compiles ``fn`` when it is first called during tracing. ``torch.jit.script`` has a non-negligible start up time when it is first called due to lazy-initializations of many compiler builtins. Therefore you should not use it in library code. However, you may want to have parts of your library work in tracing even if they use control flow. In these cases, you should use ``@torch.jit.script_if_tracing`` to substitute for ``torch.jit.script``. Arguments: fn: A function to compile. Returns: If called during tracing, a :class:`ScriptFunction` created by ` `torch.jit.script`` is returned. Otherwise, the original function ``fn`` is returned. cs0ts |i|Stjj}||i|SN)r$r jitscript __original_fn)argskwargs compiled_fnfnwrapperrr r.qsz"script_if_tracing..wrapperT) functoolswrapsr(__script_if_tracing_wrapper)r-rr,r script_if_tracing^s r2xyaxisreturncCs6|dkrt|j|}|j|}tj|d||gS)a#Right-pad or right-trim first argument to have same size as second argument Args: x (torch.Tensor): Tensor to be padded. y (torch.Tensor): Tensor to pad `x` to. axis (int): Axis to pad on. Returns: torch.Tensor, `x` padded to match `y`'s shape. r3r)NotImplementedErrorshaper functionalpad)r4r5r6inp_len output_lenrrr pad_x_to_ys    r>cCsnz |j|ddW|Sty6t}|D]\}}||ddd}|||<q|j|ddY|Sw)aStrictly loads state_dict in model, or the next submodel. Useful to load standalone model after training it with System. Args: state_dict (OrderedDict): the state_dict to load. model (torch.nn.Module): the model to load it into Returns: torch.nn.Module: model with loaded weights. .. note:: Keys in a state_dict look like ``object1.object2.layer_name.weight.etc`` We first try to load the model in the classic way. If this fail we removes the first left part of the key to obtain ``object2.layer_name.weight.etc``. Blindly loading with ``strictly=False`` should be done with some logging of the missing keys in the state_dict and the model. T)strict.N)load_state_dict RuntimeErrorritemsfind) state_dictmodelnew_state_dictkvnew_krrr load_state_dict_ins  rLcCs<t||D]\}}|j|jdkrdSq dS)zCheck for weights equality between models. Args: model1 (nn.Module): model instance to be compared. model2 (nn.Module): second model instance to be compared. Returns: bool: Whether all model weights are equal. rFT)ziprdatanesum)model1model2p1p2rrr are_models_equals  rUcCs t|jS)aMGets shape of ``tensor`` as ``torch.Tensor`` type for jit compiler .. note:: Returning ``tensor.shape`` of ``tensor.size()`` directly is not torchscript compatible as return type would not be supported. Args: tensor (torch.Tensor): Tensor Returns: torch.Tensor: Shape of ``tensor`` )r tensorr9)rVrrr jitable_shapes rWr%)r3)r/r r collectionsrrrr!r$r2r intr>rLrUrWrrrr s    "$%