o ei7@sdZddlZddlZddlZddlZddlmZddlmZddl m Z e e Z Gdddej jZGdd d eZGd d d ej jZGd d d ej jZdS)zkLibrary for implementing cascade (sequences) of different neural modules. Authors * Peter Plantinga 2020 N)Linear)lengths_arg_exists) get_loggercsDeZdZdZddfdd ZddddZd d Zd d ZZS) SequentialaA sequence of modules with potentially inferring shape on construction. If layers are passed with names, these can be referenced with dot notation. Arguments --------- *layers : tuple Layers to be applied in sequence. input_shape : iterable A list or tuple of ints or None, representing the expected shape of an input tensor. None represents a variable-length dimension. If no ``input_shape`` is passed, no shape inference will be performed. **named_layers : dict The inputs are treated as a list of layers to be applied in sequence. The output shape of each layer is used to infer the shape of the following layer. If a tuple is returned, only the shape of the first element is used to determine input shape of the next layer (e.g. RNN returns output, hidden). Example ------- >>> inputs = torch.rand(10, 40, 50) >>> model = Sequential(input_shape=inputs.shape) >>> model.append(Linear, n_neurons=100, layer_name="layer1") >>> model.append(Linear, n_neurons=200, layer_name="layer2") >>> outputs = model(inputs) >>> outputs.shape torch.Size([10, 40, 200]) >>> outputs = model.layer1(inputs) >>> outputs.shape torch.Size([10, 40, 100]) N input_shapecst|s|dur|stdg|_||_|r=d|vr=t||_t|jD]\}}|dkr5|dur5d}|p8d|j|<q'|D]}||q?|D] \}}|j||dqKdS)Nz&Must pass either layers or input shaper layer_name) super__init__ ValueError length_layersrlist enumerateappenditems)selfrlayers named_layersidimlayername __class__Y/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/speechbrain/nnet/containers.pyr 6s    zSequential.__init__r cOs|dur tt|}n#||vr.d}|d||vr'|d7}|d||vs|d|}|jrKt|}d|j|jvrK|}||d|i|}z |||WdSt y_t dw)aAdd a layer to the list of layers, inferring shape if necessary. Arguments --------- layer : A torch.nn.Module class or object If the layer is a class, it should accept an argument called ``input_shape`` which will be inferred and passed. If the layer is a module object, it is added as-is. *args : tuple These are passed to the layer if it is constructed. layer_name : str The name of the layer, for reference. If the name is in use, ``_{count}`` will be appended. **kwargs : dict These are passed to the layer if it is constructed. Nr_rrzwMust pass `input_shape` at initialization and use modules that take `input_shape` to infer shape when using `append()`.) strlenrinspectgetfullargspecargs kwonlyargsget_output_shape add_module TypeErrorr)rrr r$kwargsindexargspecrrrrrUs(  zSequential.appendcCsFtt|j}||}Wd|jS1swY|jS)zReturns expected shape of the output. Computed by passing dummy input constructed with the ``self.input_shape`` attribute. Returns ------- Expected shape of the output after all layers applied. N)torchno_gradzerosrshape)r dummy_input dummy_outputrrrr&s   zSequential.get_output_shapecCs,|D]}||}t|tr|d}q|S)a0Applies layers in sequence, passing only the first element of tuples. Arguments --------- x : torch.Tensor The input tensor to run through the network. Returns ------- x : torch.Tensor Output after all layers are applied. r)values isinstancetuplerxrrrrforwards  zSequential.forward) __name__ __module__ __qualname____doc__r rr&r7 __classcell__rrrrrs !,rcs6eZdZdZfddZfddZd ddZZS) LengthsCapableSequentialaSequential model that can take ``lengths`` in the forward method. This is useful for Sequential models that include RNNs where it is important to avoid padding, or for some feature normalization layers. Unfortunately, this module is not jit-able because the compiler doesn't know ahead of time if the length will be passed, and some layers don't accept the length parameter. csg|_tj|i|dSN) takes_lengthsr r )rr$r)rrrr sz!LengthsCapableSequential.__init__cs8tj|i|t|dj}|jt|dS)z@Add a layer to the list of layers, inferring shape if necessary.N)r rrr2r7r?r)rr$r)latest_forward_methodrrrrszLengthsCapableSequential.appendNcCsJt||jD]\}}|r|||d}n||}t|tr"|d}q|S)a.Applies layers in sequence, passing only the first element of tuples. In addition, forward the ``lengths`` argument to all layers that accept a ``lengths`` argument in their ``forward()`` method (e.g. RNNs). Arguments --------- x : torch.Tensor The input tensor to run through the network. lengths : torch.Tensor The relative lengths of each signal in the tensor. Returns ------- x : torch.Tensor The outputs after all layers are applied. )lengthsr)zipr2r?r3r4)rr6rBr give_lengthsrrrr7s z LengthsCapableSequential.forwardr>)r8r9r:r;r rr7r<rrrrr=s  r=cs@eZdZdZfddZddZddZdd Zd d ZZ S) ModuleListaThis class implements a wrapper to torch.nn.ModuleList with a forward() method to forward all the layers sequentially. For some pretrained model with the SpeechBrain older implementation of Sequential class, user can use this class to load those pretrained models Arguments --------- *layers : torch class Torch objects to be put in a ModuleList. csttj||_dSr>)r r r,nnrEr)rrrrrr s zModuleList.__init__cCs*|jD]}||}t|tr|d}q|S)z!Applies the computation pipeline.r)rr3r4r5rrrr7s  zModuleList.forwardcC|j|dSz"Appends module to the layers list.N)rr)rmodulerrrrzModuleList.appendcCrGrH)rextend)rmodulesrrrrKrJzModuleList.extendcCs|j|dS)z"Inserts module to the layers list.N)rinsert)rr*rIrrrrMrJzModuleList.insert) r8r9r:r;r r7rrKrMr<rrrrrEs rEcsDeZdZdZddejffdd ZddZdd Zd d d Z Z S) ConnectBlocksaConnect a sequence of blocks with shortcut connections. Note: all shortcuts start from the output of the first block, since the first block may change the shape significantly. Arguments --------- input_shape : tuple The shape of the shortcut_type : str One of: * "residual" - first block output passed to final output, * "dense" - input of each block is from all previous blocks, * "skip" - output of each block is passed to final output. shortcut_projection : bool Only has an effect if `shortcut_type` is passed. Whether to add a linear projection layer to the shortcut connection before combining with the output, to handle different sizes. shortcut_combine_fn : str or function Either a pre-defined function (one of "add", "sub", "mul", "div", "avg", "cat") or a user-defined function that takes the shortcut and next input, and combines them, as well as `init_params` in case parameters need to be initialized inside of the function. Example ------- >>> inputs = torch.rand(10, 100, 20) >>> model = ConnectBlocks( ... input_shape=inputs.shape, shortcut_projection=True ... ) >>> model.append(Linear, n_neurons=10) >>> model.append(Linear, n_neurons=10, end_of_block=True) >>> model.append(Linear, n_neurons=10) >>> model.append(Linear, n_neurons=10, end_of_block=True) >>> outputs = model(inputs) >>> outputs.shape torch.Size([10, 100, 10]) residualFcs^t||_||_d|_tj|_|dvrt d||_ ||_ |r*tj|_ ||_ dS)NT)rOdenseskipz9'shortcuts' must be one of 'residual', 'dense', or 'skip')r r first_input_shapeblock_input_shape new_blockr,rFrEblocksr shortcut_typeshortcut_projection projectionsshortcut_combine_fn)rrrVrWrYrrrr $s    zConnectBlocks.__init__c Os|jr|jt|jdd|_d}d|vr|d}|d=|jdj|g|Ri||ryt|j}|jd|}|jrit t j |j ddd}|j dkrZ|j}t|j}n|j}|jt||dd d d |_|||d}|j |_dSdS) aAppends the specified module to the shortcut model. Arguments --------- layer : torch.nn.Module class This layer will get initialized with *args and **kwargs. Also, the argument ``input_shape`` will be passed if the layer takes it. *args : tuple **kwargs : dict Passed unchanged to the layer **EXCEPT** the kwarg ``end_of_block`` which is used to indicate that the shortcut should be added in. rF end_of_blockr@NrrOT) n_neuronsrbias combine_dims)rTrUrrrSr,r.rW functoolsreduceoperatormulr/rVrRrXr_combine) rrr$r)rZr0r1projection_sizer/rrrr;s>     zConnectBlocks.appendcCs~|}t|jD].\}}||}|jdkr||||}|jdkr)||||}}|jdkr5||||}q|jdkr=|S|S)z Arguments --------- x : torch.Tensor The inputs to the replicated modules. Returns ------- x : torch.Tensor The output processed by all blocks. rQrPrO)rrUrVrc)rr6shortcutrblockrrrr7ss     zConnectBlocks.forwardrcCs,|jr|j||}||j}|||S)z'Handle combining shortcut with outputs.)rWrXreshaper/rY)rrer6 block_indexrrrrcs  zConnectBlocks._combine)r) r8r9r:r;r,addr rr7rcr<rrrrrNs*8rN)r;r_r"rar,speechbrain.nnet.linearrspeechbrain.utils.callchainsrspeechbrain.utils.loggerrr8loggerrF ModuleDictrr=ModulerErNrrrrs   2%