# 设定并存储了一个best.th 的模型
# When training a model, many sets of weights are saved. By default we want to
# save/load this set of weights.
_DEFAULT_WEIGHTS = "best.th"class Model(torch.nn.Module, Registrable):"""This abstract class represents a model to be trained. Rather than relying completelyon the Pytorch Module, we modify the output spec of ``forward`` to be a dictionary.
Models built using this API are still compatible with other pytorch models and canbe used naturally as modules within other models - outputs are dictionaries, whichcan be unpacked and passed into other layers. One caveat to this is that if youwish to use an AllenNLP model inside a Container (such as nn.Sequential), you mustinterleave the models with a wrapper module which unpacks the dictionary intoa list of tensors.
In order for your model to be trained using the :class:`~allennlp.training.Trainer`api, the output dictionary of your Model must include a "loss" key, which will beoptimised during the training process.
Finally, you can optionally implement :func:`Model.get_metrics` in order to make useof early stopping and best-model serialization based on a validation metric in:class:`~allennlp.training.Trainer`. Metrics that begin with "_" will not be loggedto the progress bar by :class:`~allennlp.training.Trainer`."""_warn_for_unseparable_batches: Set[str] = set()# 必然是要传入一个词表的~def __init__(self,vocab: Vocabulary,regularizer: RegularizerApplicator = None) -> None:super().__init__()self.vocab = vocabself._regularizer = regularizer# 得到对某些层的参数的惩罚def get_regularization_penalty(self) -> Union[float, torch.Tensor]:"""Computes the regularization penalty for the model.Returns 0 if the model was not configured to use regularization."""if self._regularizer is None:return 0.0else:return self._regularizer(self)def get_parameters_for_histogram_tensorboard_logging( # pylint: disable=invalid-nameself) -> List[str]:"""Returns the name of model parameters used for logging histograms to tensorboard."""return [name for name, _ in self.named_parameters()]# 关键的forward函数，看英文注释便可def forward(self, *inputs) -> Dict[str, torch.Tensor]:  # pylint: disable=arguments-differ"""Defines the forward pass of the model. In addition, to facilitate easy training,this method is designed to compute a loss function defined by a user.
The input is comprised of everything required to perform atraining update, `including` labels - you define the signature here!It is down to the user to ensure that inference can be performedwithout the presence of these labels. Hence, any inputs not available atinference time should only be used inside a conditional block.
The intended sketch of this method is as follows::
def forward(self, input1, input2, targets=None):........output1 = self.layer1(input1)output2 = self.layer2(input2)output_dict = {"output1": output1, "output2": output2}if targets is not None:# Function returning a scalar torch.Tensor, defined by the user.loss = self._compute_loss(output1, output2, targets)output_dict["loss"] = lossreturn output_dict
Parameters----------inputs:Tensors comprising everything needed to perform a training update, `including` labels,which should be optional (i.e have a default value of ``None``).  At inference time,simply pass the relevant inputs, not including the labels.
Returns-------output_dict: ``Dict[str, torch.Tensor]``The outputs from the model. In order to train a model using the:class:`~allennlp.training.Trainer` api, you must provide a "loss" key pointing to ascalar ``torch.Tensor`` representing the loss to be optimized."""raise NotImplementedError# 很机制，干脆直接调用处理多个输入的方法def forward_on_instance(self, instance: Instance) -> Dict[str, numpy.ndarray]:"""Takes an :class:`~allennlp.data.instance.Instance`, which typically has raw text in it,converts that text into arrays using this model's :class:`Vocabulary`, passes those arraysthrough :func:`self.forward()` and :func:`self.decode()` (which by default does nothing)and returns the result.  Before returning the result, we convert any``torch.Tensors`` into numpy arrays and remove the batch dimension."""return self.forward_on_instances([instance])[0]# 这里非常重要！！！！！！！！！！# 我们自行做DIY完全可以按照这种思路来做！# 比如Batch的操作~！def forward_on_instances(self,instances: List[Instance]) -> List[Dict[str, numpy.ndarray]]:"""Takes a list of  :class:`~allennlp.data.instance.Instance`s, converts that text intoarrays using this model's :class:`Vocabulary`, passes those arrays through:func:`self.forward()` and :func:`self.decode()` (which by default does nothing)and returns the result.  Before returning the result, we convert any``torch.Tensors`` into numpy arrays and separate thebatched output into a list of individual dicts per instance. Note that typicallythis will be faster on a GPU (and conditionally, on a CPU) than repeated calls to:func:`forward_on_instance`.
Parameters----------instances : List[Instance], requiredThe instances to run the model on.
Returns-------A list of the models output for each instance."""batch_size = len(instances)with torch.no_grad():cuda_device = self._get_prediction_device()# 将我们的instances原本不等长# 变成一个batch，这样 之后 就可以使用padding等操作dataset = Batch(instances)# 使用index_instances，词表，将我们的instance变成数值化dataset.index_instances(self.vocab)# 使用as_tensor_dict方法，将数据变成tensormodel_input = util.move_to_device(dataset.as_tensor_dict(), cuda_device)# 调用模型decode方法~，将数值化的结果再反解回 文本outputs = self.decode(self(**model_input))# 将batch拆解成 一个个单独的实例，并且返回numpy的格式instance_separated_output: List[Dict[str, numpy.ndarray]] = [{} for _ in dataset.instances]for name, output in list(outputs.items()):if isinstance(output, torch.Tensor):# NOTE(markn): This is a hack because 0-dim pytorch tensors are not iterable.# This occurs with batch size 1, because we still want to include the loss in that case.if output.dim() == 0:output = output.unsqueeze(0)if output.size(0) != batch_size:self._maybe_warn_for_unseparable_batches(name)continueoutput = output.detach().cpu().numpy()elif len(output) != batch_size:self._maybe_warn_for_unseparable_batches(name)continuefor instance_output, batch_element in zip(instance_separated_output, output):instance_output[name] = batch_elementreturn instance_separated_output# 解码部分def decode(self, output_dict: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:"""Takes the result of :func:`forward` and runs inference / decoding / whateverpost-processing you need to do your model.  The intent is that ``model.forward()`` shouldproduce potentials or probabilities, and then ``model.decode()`` can take those results andrun some kind of beam search or constrained inference or whatever is necessary.  This doesnot handle all possible decoding use cases, but it at least handles simple kinds ofdecoding.
This method `modifies` the input dictionary, and also `returns` the same dictionary.
By default in the base class we do nothing.  If your model has some special decoding step,override this method."""# pylint: disable=no-self-usereturn output_dict# 返回模型的评估结果def get_metrics(self, reset: bool = False) -> Dict[str, float]:"""Returns a dictionary of metrics. This method will be called by:class:`allennlp.training.Trainer` in order to compute and use model metrics for earlystopping and model serialization.  We return an empty dictionary here rather than raisingas it is not required to implement metrics for a new model.  A boolean `reset` parameter ispassed, as frequently a metric accumulator will have some state which should be resetbetween epochs. This is also compatible with :class:`~allennlp.training.Metric`s. Metricsshould be populated during the call to ``forward``, with the:class:`~allennlp.training.Metric` handling the accumulation of the metric until thismethod is called."""# pylint: disable=unused-argument,no-self-usereturn {}def _get_prediction_device(self) -> int:"""This method checks the device of the model parameters to determine the cuda_devicethis model should be run on for predictions.  If there are no parameters, it returns -1.
Returns-------The cuda device this model should run on for predictions."""devices = {util.get_device_of(param) for param in self.parameters()}if len(devices) > 1:devices_string = ", ".join(str(x) for x in devices)raise ConfigurationError(f"Parameters have mismatching cuda_devices: {devices_string}")elif len(devices) == 1:return devices.pop()else:return -1def _maybe_warn_for_unseparable_batches(self, output_key: str):"""This method warns once if a user implements a model which returns a dictionary withvalues which we are unable to split back up into elements of the batch. This is controlledby a class attribute ``_warn_for_unseperable_batches`` because it would be extremely verboseotherwise."""if  output_key not in self._warn_for_unseparable_batches:logger.warning(f"Encountered the {output_key} key in the model's return dictionary which ""couldn't be split by the batch size. Key will be ignored.")# We only want to warn once for this key,# so we set this to false so we don't warn again.self._warn_for_unseparable_batches.add(output_key)@classmethoddef _load(cls,config: Params,serialization_dir: str,weights_file: str = None,cuda_device: int = -1) -> 'Model':"""Instantiates an already-trained model, based on the experimentconfiguration and some optional overrides."""weights_file = weights_file or os.path.join(serialization_dir, _DEFAULT_WEIGHTS)# Load vocabulary from filevocab_dir = os.path.join(serialization_dir, 'vocabulary')# If the config specifies a vocabulary subclass, we need to use it.vocab_params = config.get("vocabulary", Params({}))vocab_choice = vocab_params.pop_choice("type", Vocabulary.list_available(), True)vocab = Vocabulary.by_name(vocab_choice).from_files(vocab_dir)model_params = config.get('model')# The experiment config tells us how to _train_ a model, including where to get pre-trained# embeddings from. We're now _loading_ the model, so those embeddings will already be# stored in our weights. We don't need any pretrained weight file anymore, and we don't# want the code to look for it, so we remove it from the parameters here.remove_pretrained_embedding_params(model_params)model = Model.from_params(vocab=vocab, params=model_params)# If vocab+embedding extension was done, the model initialized from from_params# and one defined by state dict in weights_file might not have same embedding shapes.# Eg. when model embedder module was transferred along with vocab extension, the# initialized embedding weight shape would be smaller than one in the state_dict.# So calling model embedding extension is required before load_state_dict.# If vocab and model embeddings are in sync, following would be just a no-op.# 拓展embedding的vocabularymodel.extend_embedder_vocab()model_state = torch.load(weights_file, map_location=util.device_mapping(cuda_device))model.load_state_dict(model_state)# Force model to cpu or gpu, as appropriate, to make sure that the embeddings are# in sync with the weightsif cuda_device >= 0:model.cuda(cuda_device)else:model.cpu()return model@classmethoddef load(cls,config: Params,serialization_dir: str,weights_file: str = None,cuda_device: int = -1) -> 'Model':"""Instantiates an already-trained model, based on the experimentconfiguration and some optional overrides.
Parameters----------config: ParamsThe configuration that was used to train the model. It should definitelyhave a `model` section, and should probably have a `trainer` sectionas well.serialization_dir: str = NoneThe directory containing the serialized weights, parameters, and vocabularyof the model.weights_file: str = NoneBy default we load the weights from `best.th` in the serializationdirectory, but you can override that value here.cuda_device: int = -1By default we load the model on the CPU, but if you want to load itfor GPU usage you can specify the id of your GPU hereReturns-------model: ModelThe model specified in the configuration, loaded with the serializedvocabulary and the trained weights."""# Peak at the class of the model.model_type = config["model"] if isinstance(config["model"], str) else config["model"]["type"]# Load using an overridable _load method.# This allows subclasses of Model to override _load.# pylint: disable=protected-accessreturn cls.by_name(model_type)._load(config, serialization_dir, weights_file, cuda_device)# 扩充词表def extend_embedder_vocab(self, embedding_sources_mapping: Dict[str, str] = None) -> None:"""Iterates through all embedding modules in the model and assures it can embedwith the extended vocab. This is required in fine-tuning or transfer learningscenarios where model was trained with original vocabulary but duringfine-tuning/transfer-learning, it will have it work with extended vocabulary(original + new-data vocabulary).
Parameters----------embedding_sources_mapping : Dict[str, str], (optional, default=None)Mapping from model_path to pretrained-file path of the embeddingmodules. If pretrained-file used at time of embedding initializationisn't available now, user should pass this mapping. Model path ispath traversing the model attributes upto this embedding module.Eg. "_text_field_embedder.token_embedder_tokens"."""# self.named_modules() gives all sub-modules (including nested children)# The path nesting is already separated by ".": eg. parent_module_name.child_module_nameembedding_sources_mapping = embedding_sources_mapping or {}for model_path, module in self.named_modules():if hasattr(module, 'extend_vocab'):pretrained_file = embedding_sources_mapping.get(model_path, None)module.extend_vocab(self.vocab,extension_pretrained_file=pretrained_file,model_path=model_path)def remove_pretrained_embedding_params(params: Params):if isinstance(params, Params):  # The model could possible be a string, for example.keys = params.keys()if 'pretrained_file' in keys:del params['pretrained_file']for value in params.values():if isinstance(value, Params):remove_pretrained_embedding_params(value)