Aluode/PerceptionLabPortable
0
1o
2 .�YiQF � @ s` d dl Z d dlmZ d dlmZ d dlmZmZmZm Z m3Z4 d dl Z d dl m Z d dl
mZ d dlmZ dd lmZ dd5lmZ dd lmZ dd lmZ dd
lmZmZ ddl m Z e� rad dl Z er�d dlm Z ddl m!Z! ddl"m#Z# ddl$m%Z% ddl&m'Z' ddl(m)Z) ddl*m+Z+ ddl,m-Z- ddl.m/Z/m0Z0 ddl1m2Z2 e�3e4�Z5G dd� de�Z6dS )� N)�deepcopy)�Path)�
TYPE_CHECKING�Any�Callable�Optional�Union)�nn)�FullyShardedDataParallel)�Dataset� )�GenerationConfig)�is_deepspeed_zero3_enabled)�is_fsdp_managed_module)�Trainer)�is_datasets_available�logging)�deprecate_kwarg)�IterableDataset)� DataCollator)�FeatureExtractionMixin)�BaseImageProcessor)�PreTrainedModel)�ProcessorMixin)�PreTrainedTokenizerBase)�TrainerCallback)�EvalPrediction�PredictionOutput)�TrainingArgumentsc s� e Zd Zeddddd� d1deed ejf d6ed d ed
deeeddf deeee e7ef f deed dee g d f dee dee dge f dee d de
eejj eejjj f dee ejejgejf f� fdd �
�Zed ee8ef d efd d ��Z !d2dee d"ee e9 d#e10d e e11ef f� fd$d%�
Z &d3d'ed"ee e12 d#e13d d(f� fd)d*�
Z d4dejd+e e14eejef f d,e d"ee e15 d e
ee eej eej f f16� fd-d.�
Z d/d0� Z � Z S )5�Seq2SeqTrainer� tokenizer�processing_classz5.0.0T)�new_name�versionZraise_if_both_namesN�NN�modelr �argsr �
data_collatorr �
train_datasetr zdatasets.Dataset� eval_dataset)r r r r �17model_init�compute_loss_func�compute_metricsr � callbacksr �18optimizers� preprocess_logits_for_metricsc
sN t � j||||||||| |19| | d� | jjd ur%| �| jj�}
|
| j_d S d S )N) r% r& r' r( r) r! r* r+ r, r- r. r/ )�super�__init__r& �generation_config�load_generation_configr% )�selfr% r& r' r( r) r! r* r+ r, r- r. r/ �20gen_config�� __class__� ��E:\DocsHouse\542 percep lab latest\PerceptionLab\PerceptionLab_Portable\python_embed\Lib\site-packages\transformers/trainer_seq2seq.pyr1 6 s$ � �zSeq2SeqTrainer.__init__�gen_config_arg�returnc 21 C s� t | t�r22t| �}n%t | t�rt| �n| }d}|�� r"|j}|j}n|�� r'n| }t� ||�}z |j23dd� W |S t yL } zt t|�d ��d}~ww )aW 24 Loads a `~generation.GenerationConfig` from the `Seq2SeqTrainingArguments.generation_config` arguments.25 26 Args:27 gen_config_arg (`str` or [`~generation.GenerationConfig]`):28 `Seq2SeqTrainingArguments.generation_config` argument.29 30 Returns:31 A `~generation.GenerationConfig`.32 NT)�strictz'33 34Fix these issues to train your model.) �35isinstancer
r �strr �is_file�name�parent�is_dirZfrom_pretrained�validate�36ValueError)r: r5 Zpretrained_model_nameZconfig_file_name�excr8 r8 r9 r3 ] s$ 3738 ���z%Seq2SeqTrainer.load_generation_config�eval� ignore_keys�metric_key_prefixc � |� � }|�d�du r |�d�du r | jjdur | jj|d< |�d�du r1| jjdur1| jj|d< | jj| _|| _t � j39|||d�S )a� 40 Run evaluation and returns metrics.41 42 The calling script will be responsible for providing a method to compute metrics, as they are task-dependent43 (pass it to the init `compute_metrics` argument).44 45 You can also subclass and override this method to inject custom behavior.46 47 Args:48 eval_dataset (`Dataset`, *optional*):49 Pass a dataset if you wish to override `self.eval_dataset`. If it is an [`~datasets.Dataset`], columns50 not accepted by the `model.forward()` method are automatically removed. It must implement the `__len__`51 method.52 ignore_keys (`list[str]`, *optional*):53 A list of keys in the output of your model (if it is a dictionary) that should be ignored when54 gathering predictions.55 metric_key_prefix (`str`, *optional*, defaults to `"eval"`):56 An optional prefix to be used as the metrics key prefix. For example the metrics "bleu" will be named57 "eval_bleu" if the prefix is `"eval"` (default)58 max_length (`int`, *optional*):59 The maximum target length to use when predicting with the generate method.60 num_beams (`int`, *optional*):61 Number of beams for beam search that will be used when predicting with the generate method. 1 means no62 beam search.63 gen_kwargs:64 Additional `generate` specific kwargs.65 66 Returns:67 A dictionary containing the evaluation loss and the potential metrics computed from the predictions. The68 dictionary also contains the epoch number which comes from the training state.69 �70max_lengthN�max_new_tokens� num_beams�rG rH ) �copy�getr& �generation_max_length�generation_num_beams� accelerator�gather�gather_function� _gen_kwargsr0 �evaluate)r4 r) rG rH �71gen_kwargsr6 r8 r9 rV � s ' 72zSeq2SeqTrainer.evaluate�test� test_datasetr c rI )a& 73 Run prediction and returns predictions and potential metrics.74 75 Depending on the dataset and your use case, your test dataset may contain labels. In that case, this method76 will also return metrics, like in `evaluate()`.77 78 Args:79 test_dataset (`Dataset`):80 Dataset to run the predictions on. If it is a [`~datasets.Dataset`], columns not accepted by the81 `model.forward()` method are automatically removed. Has to implement the method `__len__`82 ignore_keys (`list[str]`, *optional*):83 A list of keys in the output of your model (if it is a dictionary) that should be ignored when84 gathering predictions.85 metric_key_prefix (`str`, *optional*, defaults to `"eval"`):86 An optional prefix to be used as the metrics key prefix. For example the metrics "bleu" will be named87 "eval_bleu" if the prefix is `"eval"` (default)88 max_length (`int`, *optional*):89 The maximum target length to use when predicting with the generate method.90 num_beams (`int`, *optional*):91 Number of beams for beam search that will be used when predicting with the generate method. 1 means no92 beam search.93 gen_kwargs:94 Additional `generate` specific kwargs.95 96 <Tip>97 98 If your predictions or labels have different sequence lengths (for instance because you're doing dynamic99 padding in a token classification task) the predictions will be padded (on the right) to allow for100 concatenation into one array. The padding index is -100.101 102 </Tip>103 104 Returns: *NamedTuple* A namedtuple with the following keys:105 106 - predictions (`np.ndarray`): The predictions on `test_dataset`.107 - label_ids (`np.ndarray`, *optional*): The labels (if the dataset contained some).108 - metrics (`dict[str, float]`, *optional*): The potential dictionary of metrics (if the dataset contained109 labels).110 rJ NrK rL rM ) rN rO r& rP rQ rR rS rT rU r0 �predict)r4 rY rG rH rW r6 r8 r9 rZ � s / 111zSeq2SeqTrainer.predict�inputs�prediction_loss_onlyc s� | j jr|rt� j||||d�S d|v }| �|�}t|�dkr)t| d�r)| j�� }d|v r8|d du r8|� d� d|v rG|d du rG|� d� t112� pNt | j �}|�
d|�|d< |�� }d|v rvd |v rv|d j|d jkrvd113d � |�� D �}t| j t�r�t�| j �nt�� } | � | j jdi |�|��}114W d � n1 s�w Y | j jjr�d | j j_| j j} |115jd
| jk r�| �|116| j�}117n| jdur�|118jd
| jd k r�| �|119| jd �}120t� � �N |�r"| � � � |di |��} W d � n1 s�w Y | j du�r| � | |d �� � � � }
nt| t!��r| d n| d � � � � }
nd}
W d � n 1 �s/w Y | j j"�r>|
ddfS |�rq|d }|jd
| jk �rV| �|| j�}n | jdu�rp|jd
| jd k �rp| �|| jd �}nd}|
|121|fS )a� 122 Perform an evaluation step on `model` using `inputs`.123 124 Subclass and override to inject custom behavior.125 126 Args:127 model (`nn.Module`):128 The model to evaluate.129 inputs (`dict[str, Union[torch.Tensor, Any]]`):130 The inputs and targets of the model.131 132 The dictionary will be unpacked before being fed to the model. Most models expect the targets under the133 argument `labels`. Check your model's documentation for all accepted arguments.134 prediction_loss_only (`bool`):135 Whether or not to return the loss only.136 gen_kwargs:137 Additional `generate` specific kwargs.138 139 Return:140 tuple[Optional[float], Optional[torch.Tensor], Optional[torch.Tensor]]: A tuple with the loss, logits and141 labels (each being optional).142 )r\ rG �labelsr rU rL NrJ Z synced_gpus�decoder_input_idsc S s i | ] \}}|d vr||�qS ))r^ Zdecoder_attention_maskr8 )�.0�k�vr8 r8 r9 �143<dictcomp>< s z2Seq2SeqTrainer.prediction_step.<locals>.<dictcomp>F�����r �lossr8 )#r& Zpredict_with_generater0 �prediction_stepZ_prepare_inputs�len�hasattrrU rN �popr r r% rO �shape�itemsr= r144 Zsummon_full_params�145contextlib� nullcontext�generater2 Z_from_model_configrJ �_pad_tensors_to_max_lenrK �torchZno_gradZ compute_loss_context_managerZlabel_smoother�detach�mean�dictr\ )r4 r% r[ r\ rG rW Z146has_labelsZdefault_synced_gpusZgeneration_inputsZsummon_full_params_contextZgenerated_tokensr5 Zoutputsrd r] r6 r8 r9 re sr �147148149150�151�� �152153 154155 � &� �156 157"�158z Seq2SeqTrainer.prediction_stepc C s� | j d urt| j d�r| j jd ur| j jn| j j}n| jjjd ur'| jjj}ntd��|tj|j d |f|j159|j d� }||d d �d |j d �f< |S )N� pad_token_idzSPad_token_id must be set in the configuration of the model, in order to pad tensorsr )�dtype�devicerc ) r! rg rs Z eos_token_idr% �configrD ro Zonesri rt ru )r4 ZtensorrJ rs Z
padded_tensorr8 r8 r9 rn p s �� �z&Seq2SeqTrainer._pad_tensors_to_max_len) NNNNNNNNNNr$ N)NNrF )NrX )N) �__name__�160__module__� __qualname__r r r r �Moduler rr r> r �list�tuplero ZoptimZ OptimizerZ lr_schedulerZLambdaLRZTensorr1 � staticmethodr
r3 �floatrV rZ r �boolre rn �
__classcell__r8 r8 r6 r9 r 5 s� ��������161� � �162
� ��& -��163��164�;��165���E����166� �or )7rk rN r �pathlibr �typingr r r r r ro r Ztorch.distributed.fsdpr167 Ztorch.utils.datar Z generation.configuration_utilsr
Zintegrations.deepspeedr Zintegrations.fsdpr Ztrainerr �utilsr r Zutils.deprecationr Zdatasetsr Zdata.data_collatorr Zfeature_extraction_utilsr Zimage_processing_utilsr Zmodeling_utilsr Zprocessing_utilsr Ztokenization_utils_baser Ztrainer_callbackr Z
trainer_utilsr r Z
training_argsr Z168get_loggerrw �loggerr r8 r8 r8 r9 �<module> s: 169