U &c|K@sVdZddlZddlZddlZddlZddlZddlZddl m Z ddl m Z ddl mZmZmZmZmZmZddlmZeeZGdd d Zd d ZGd d d ejjeZddZddZddZ dde!d dfddZ"ddZ#ddZ$Gddde%Z&Gdddejj'j(Z)Gd d!d!ejj'j(Z*Gd"d#d#ejj'j(Z+d$d%Z,d)d'd(Z-dS)*zTF general model utils.N) hdf5_format)PretrainedConfig) DUMMY_INPUTSTF2_WEIGHTS_NAME WEIGHTS_NAME cached_path hf_bucket_url is_remote_url)$load_pytorch_checkpoint_in_tf2_modelc@s"eZdZdZdeedddZdS)TFModelUtilsMixinzG A few utilities for `tf.keras.Model`s, to be used as a mixin. F)only_trainablereturncCs(|rttdd|jDS|SdS)zP Get number of (optionally, trainable) parameters in the model. css|]}t|jVqdSN)npprodshapeas_list).0wrB/tmp/pip-unpacked-wheel-ymerj3tt/transformers/modeling_tf_utils.py ,sz3TFModelUtilsMixin.num_parameters..N)intsumZtrainable_variablesZ count_params)selfr rrrnum_parameters'sz TFModelUtilsMixin.num_parametersN)F)__name__ __module__ __qualname____doc__boolrrrrrrr "sr csjtdddkr"tdtfdd}|_tdsRtdtjdrpfd d }|_d _tt j j d rt j j S) al Decorate a Keras Layer class to support Keras serialization. This is done by: 1. adding a `transformers_config` dict to the Keras config dictionary in `get_config` (called by Keras at serialization time 2. wrapping `__init__` to accept that `transformers_config` dict (passed by Keras at deserialization time) and convert it to a config object for the actual layer initializer 3. registering the class as a custom object in Keras (if the Tensorflow version supports this), so that it does not need to be supplied in `custom_objects` in the call to `tf.keras.models.load_model` :param cls: a tf.keras.layers.Layers subclass that accepts a `config` argument to its initializer (typically a `TF*MainLayer` class in this project) :return: the same class object, with modifications for Keras deserialization. config_classNz2Must set `config_class` to use @keras_serializablecs|dd}|r&t|dtr&|dn |dd}|dk rL|dk rLtdnH|dk rf|f||n.|dk r|}||f||ntd||_dS)Ntransformers_configrconfigz.wrapped_init get_configz@Only use @keras_serializable on tf.keras.layers.Layer subclassesZ _is_defaultcs t|}|j|d<|S)Nr#)superr/r*Zto_dict)rcfg)clsrrr/^sz&keras_serializable..get_configTregister_keras_serializable) __init__getattrAttributeError functoolswrapshasattr TypeErrorr/Z_keras_serializabletfkerasutilsr3)r2r.r/r)r2r"r-rkeras_serializable1s     r>cseZdZdZdZiZdZeddZfddZ dd Z d d Z d"d d Z d#ddZ ddZddZeddZddZddZd$ddZddZddZed d!ZZS)%TFPreTrainedModela4 Base class for all TF models. :class:`~transformers.TFPreTrainedModel` takes care of storing the configuration of the models and handles methods for loading/downloading/saving models as well as a few methods common to all models to (i) resize the input embeddings and (ii) prune heads in the self-attention heads. Class attributes (overridden by derived classes): - ``config_class``: a class derived from :class:`~transformers.PretrainedConfig` to use as configuration class for this model architecture. - ``pretrained_model_archive_map``: a python ``dict`` of with `short-cut-names` (string) as keys and `url` (string) of associated pretrained weights as values. - ``load_tf_weights``: a python ``method`` for loading a TensorFlow checkpoint in a PyTorch model, taking as arguments: - ``model``: an instance of the relevant subclass of :class:`~transformers.PreTrainedModel`, - ``config``: an instance of the relevant subclass of :class:`~transformers.PretrainedConfig`, - ``path``: a path (string) to the TensorFlow checkpoint. - ``base_model_prefix``: a string indicating the attribute associated to the base model in derived classes of the same architecture adding modules on top of the base model. NcCsdttiS)zf Dummy inputs to build the network. Returns: tf.Tensor with dummy inputs input_ids)r;Zconstantrrrrr dummy_inputsszTFPreTrainedModel.dummy_inputscs:tj||t|ts0td|jj|jj||_dS)NzParameter config in `{}(config)` should be an instance of class `PretrainedConfig`. To create a model from a pretrained model use `model = {}.from_pretrained(PRETRAINED_MODEL_NAME)`) r0r4r&rr(format __class__rr$)rr$inputsr,rErrr4s zTFPreTrainedModel.__init__cCs&t||j|}||k r|StdS)z Returns the model's input embeddings. Returns: :obj:`tf.keras.layers.Layer`: A torch module mapping vocabulary to hidden states. N)r5base_model_prefixget_input_embeddingsNotImplementedError)rZ base_modelrrrrIsz&TFPreTrainedModel.get_input_embeddingscCsdS)z Returns the model's output embeddings. Returns: :obj:`tf.keras.layers.Layer`: A torch module mapping hidden states to vocabulary. NrrBrrrget_output_embeddingssz'TFPreTrainedModel.get_output_embeddingscCsdS)a Build a resized Embedding Variable from a provided token Embedding Module. Increasing the size will add newly initialized vectors at the end Reducing the size will remove vectors from the end Args: new_num_tokens: (`optional`) int New number of tokens in the embedding matrix. Increasing the size will add newly initialized vectors at the end Reducing the size will remove vectors from the end If not provided or None: return the provided token Embedding Module. Return: ``tf.Variable`` Pointer to the resized Embedding Module or the old Embedding Module if new_num_tokens is None Nr)rZold_embeddingsnew_num_tokensrrr_get_resized_embeddingssz)TFPreTrainedModel._get_resized_embeddingscCstdS)a Resize input token embeddings matrix of the model if new_num_tokens != config.vocab_size. Take care of tying weights embeddings afterwards if the model class has a `tie_weights()` method. Arguments: new_num_tokens: (`optional`) int: New number of tokens in the embedding matrix. Increasing the size will add newly initialized vectors at the end. Reducing the size will remove vectors from the end. If not provided or None: does nothing and just returns a pointer to the input tokens ``tf.Variable`` Module of the model. Return: ``tf.Variable`` Pointer to the input tokens Embeddings Module of the model NrJ)rrLrrrresize_token_embeddingss z)TFPreTrainedModel.resize_token_embeddingscCstdS)z Prunes heads of the base model. Arguments: heads_to_prune: dict with keys being selected layer indices (`int`) and associated values being the list of heads to prune in said layer (list of `int`). NrN)rZheads_to_prunerrr prune_headsszTFPreTrainedModel.prune_headscCsLtj|std|j|tj|t}||t d |dS)z Save a model and its configuration file to a directory, so that it can be re-loaded using the :func:`~transformers.PreTrainedModel.from_pretrained` class method. zPSaving path should be a directory where the model and configuration can be savedzModel weights saved in {}N) ospathisdirAssertionErrorr$save_pretrainedjoinrZ save_weightsloggerinforD)rZsave_directoryZoutput_model_filerrrrUs  z!TFPreTrainedModel.save_pretrainedc Os|dd}|dd}|dd}|dd}|dd}|dd} |d d} |d d } t|ts|dk rv|n|} |jj| f||d ||d |\}} n|} |dk r>||jkr|j|}ntj|r>tj tj |t rtj |t }nB|r(tj tj |t r(tj |t }nt d t t g|nRtj |sVt|r\|}n4tj |drx|d}nt||rt nt | d}zt||||| d}Wnft k r }zF||jkrtd|n td|d |j||W5d}~XYnX||kr*td|ntd||nd}||f|| }|rft||d dS||jddtj |std|z|j|d dWntk rtdYnX||jddt|d4}d|jkrd|kr|d}tt |d}W5QRXtdd|j!D}t"||}t"||}g}t#|d krjtd!|j$j%|t#|d krtd"|j$j%|t#|d krt&d#|j$j%d$ || r|||d%}||fS|S)&a[Instantiate a pretrained TF 2.0 model from a pre-trained model configuration. The warning ``Weights from XXX not initialized from pretrained model`` means that the weights of XXX do not come pre-trained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning task. The warning ``Weights from XXX not used in YYY`` means that the layer XXX is not used by YYY, therefore those weights are discarded. Parameters: pretrained_model_name_or_path: either: - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``. - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``. - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``. - a path or url to a `PyTorch state_dict save file` (e.g. `./pt_model/pytorch_model.bin`). In this case, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the PyTorch checkpoint in a TensorFlow model using the provided conversion scripts and loading the TensorFlow model afterwards. model_args: (`optional`) Sequence of positional arguments: All remaning positional arguments will be passed to the underlying model's ``__init__`` method config: (`optional`) one of: - an instance of a class derived from :class:`~transformers.PretrainedConfig`, or - a string valid as input to :func:`~transformers.PretrainedConfig.from_pretrained()` Configuration for the model to use instead of an automatically loaded configuation. Configuration can be automatically loaded when: - the model is a model provided by the library (loaded with the ``shortcut-name`` string of a pretrained model), or - the model was saved using :func:`~transformers.PreTrainedModel.save_pretrained` and is reloaded by suppling the save directory. - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory. from_pt: (`optional`) boolean, default False: Load the model weights from a PyTorch state_dict save file (see docstring of pretrained_model_name_or_path argument). cache_dir: (`optional`) string: Path to a directory in which a downloaded pre-trained model configuration should be cached if the standard cache should not be used. force_download: (`optional`) boolean, default False: Force to (re-)download the model weights and configuration files and override the cached versions if they exists. resume_download: (`optional`) boolean, default False: Do not delete incompletely recieved file. Attempt to resume the download if such a file exists. proxies: (`optional`) dict, default None: A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}. The proxies are used on each request. output_loading_info: (`optional`) boolean: Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages. kwargs: (`optional`) Remaining dictionary of keyword arguments: Can be used to update the configuration object (after it being loaded) and initiate the model. (e.g. ``output_attention=True``). Behave differently depending on whether a `config` is provided or automatically loaded: - If a configuration is provided with ``config``, ``**kwargs`` will be directly passed to the underlying model's ``__init__`` method (we assume all relevant updates to the configuration have already been done) - If a configuration is not provided, ``kwargs`` will be first passed to the configuration class initialization function (:func:`~transformers.PretrainedConfig.from_pretrained`). Each key of ``kwargs`` that corresponds to a configuration attribute will be used to override said attribute with the supplied ``kwargs`` value. Remaining keys that do not correspond to any configuration attribute will be passed to the underlying model's ``__init__`` function. Examples:: # For example purposes. Not runnable. model = BertModel.from_pretrained('bert-base-uncased') # Download model and configuration from S3 and cache. model = BertModel.from_pretrained('./test/saved_model/') # E.g. model was saved using `save_pretrained('./test/saved_model/')` model = BertModel.from_pretrained('bert-base-uncased', output_attention=True) # Update configuration during loading assert model.config.output_attention == True # Loading from a TF checkpoint file instead of a PyTorch model (slower) config = BertConfig.from_json_file('./tf_model/my_tf_model_config.json') model = BertModel.from_pretrained('./tf_model/my_tf_checkpoint.ckpt.index', from_pt=True, config=config) r$N cache_dirfrom_ptFforce_downloadresume_downloadproxiesoutput_loading_infouse_cdnT)rYZreturn_unused_kwargsr[r\zFError no file named {} found in directory {} or `from_pt` set to Falsez.index)filenamer_)rYr[r\r]z=Couldn't reach server at '{}' to download pretrained weights.zModel name '{}' was not found in model name list ({}). We assumed '{}' was a path or url but couldn't find any file associated to this path or url.z, zloading weights file {}z(loading weights file {} from cache at {})Zallow_missing_keystrainingzError retrieving file {})Zby_namez}Unable to load weights from h5 file. If you tried to load a TF 2.0 model from a PyTorch checkpoint, please set from_pt=True. rZ layer_namesZ model_weightscss|] }|jVqdSr)name)rZlayerrrrrsz4TFPreTrainedModel.from_pretrained..rz6Layers of {} not initialized from pretrained model: {}z/Layers from pretrained model not used in {}: {}z'Error(s) in loading weights for {}: {}z ) missing_keysunexpected_keys error_msgs)'r%r&rr"from_pretrainedpretrained_model_archive_maprQrRrSisfilerVrrEnvironmentErrorrDr r rrWerrorkeysrXr rCrTZ load_weightsOSErrorh5pyZFileattrssetrZload_attributes_from_hdf5_grouplayerslistlenrEr RuntimeError)r2Zpretrained_model_name_or_pathZ model_argsr,r$rYrZr[r\r]r^r_Z config_pathZ model_kwargsZ archive_fileZresolved_archive_fileemodelfZhdf5_layer_namesZmodel_layer_namesrerfrgZ loading_inforrrrhsC                       z!TFPreTrainedModel.from_pretrainedcKsd|iS)NrFr)rrFr,rrrprepare_inputs_for_generationsz/TFPreTrainedModel.prepare_inputs_for_generationcCs8t|dks|dkrdSt|jdr4|jjdkr4dSdS)zWDuring generation, decide whether to pass the `past` variable to the next forward pass.rFmem_lenrT)rtr9r$rz)routputs use_cacherrr _use_caches zTFPreTrainedModel._use_cachecCs|dkrtd|dk r |n|jj}|dk r4|n|jj}|dk rH|n|jj}|dk r\|n|jj}|dk rp|n|jj}|dk r|n|jj}|dk r|n|jj }|dk r|n|jj }| dk r| n|jj } | dk r| n|jj } | dk r| n|jj } | dk r| n|jj} |dk r|n|jj}|dk r(|n|jj}|dk r>|n|jj}| dk rT| n|jj} |dk rj|n|jj}|dk r|n|jj}|dk rt|d}nd}t|tr|dkstdt|tr|dkstdt|tstdt|tstdt|tstd t|tr2|dks:td |dksLtd t|trb|dksjtd d| krdksntd | dkstd|dk st| tr| dkstd| dkst| tr| dkstd|dkst|tr|dkstd|dks(tdt|tr>|dksFtd| dkstt| trlt| dtsttd|dkrt| tr| dkstdt|df| }ntt|dkstd|dkr|dkr|dkstdn||kstd|dkr@| dk r@| |kr@tjtj !|| tj"d}n|dkrTt#|}| dkr||dk r|t$%d&||} t|d}|jj'}|r||}|}n|}d}|jj(r |dkr| }|dk stdt)|dstd &|t*|j+s td!&|j+|+}|||d"}|dks4|dkrt|d#}t,t-|d||||f}t,t-|d||||f}t.||||f}t.||||f}|jj(rRtj/||dftj"d|}d}||dj0dkstd$|d%|dj0dd&tj.tj1t-t2|d#||dd'd(d)}tj3|d|dd*f|dd}nd}t|d#}|dkr|j4||||||||| | || | | ||||||||||d+}n2|j5|||||||| | || | | |||||||d,}|S)-ar Generates sequences for models with a LM head. The method currently supports greedy or penalized greedy decoding, sampling with top-k or nucleus sampling and beam-search. Adapted in part from `Facebook's XLM beam search code`_. .. _`Facebook's XLM beam search code`: https://github.com/facebookresearch/XLM/blob/9e6f6814d17be4fe5b15f2e6c43eb2b2d76daeb4/src/model/transformer.py#L529 Parameters: input_ids: (`optional`) `tf.Tensor` of `dtype=tf.int32` of shape `(batch_size, sequence_length)` The sequence used as a prompt for the generation. If `None` the method initializes it as an empty `tf.Tensor` of shape `(1,)`. max_length: (`optional`) int The max length of the sequence to be generated. Between 1 and infinity. Default to 20. min_length: (`optional`) int The min length of the sequence to be generated. Between 0 and infinity. Default to 0. do_sample: (`optional`) bool If set to `False` greedy decoding is used. Otherwise sampling is used. Defaults to `False` as defined in `configuration_utils.PretrainedConfig`. early_stopping: (`optional`) bool if set to `True` beam search is stopped when at least `num_beams` sentences finished per batch. Defaults to `False` as defined in `configuration_utils.PretrainedConfig`. num_beams: (`optional`) int Number of beams for beam search. Must be between 1 and infinity. 1 means no beam search. Default to 1. temperature: (`optional`) float The value used to module the next token probabilities. Must be strictely positive. Default to 1.0. top_k: (`optional`) int The number of highest probability vocabulary tokens to keep for top-k-filtering. Between 1 and infinity. Default to 50. top_p: (`optional`) float The cumulative probability of parameter highest probability vocabulary tokens to keep for nucleus sampling. Must be between 0 and 1. Default to 1. repetition_penalty: (`optional`) float The parameter for repetition penalty. Between 1.0 and infinity. 1.0 means no penalty. Default to 1.0. bos_token_id: (`optional`) int Beginning of sentence token if no prompt is provided. Default to specicic model bos_token_id or None if it does not exist. pad_token_id: (`optional`) int Pad token. Defaults to pad_token_id as defined in the models config. eos_token_id: (`optional`) int EOS token. Defaults to eos_token_id as defined in the models config. length_penalty: (`optional`) float Exponential penalty to the length. Default to 1. no_repeat_ngram_size: (`optional`) int If set to int > 0, all ngrams of size `no_repeat_ngram_size` can only occur once. bad_words_ids: (`optional`) list of lists of int `bad_words_ids` contains tokens that are not allowed to be generated. In order to get the tokens of the words that should not appear in the generated text, use `tokenizer.encode(bad_word, add_prefix_space=True)`. num_return_sequences: (`optional`) int The number of independently computed returned sequences for each element in the batch. Default to 1. attention_mask (`optional`) obj: `tf.Tensor` with `dtype=tf.int32` of same shape as `input_ids` Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``: ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens. Defaults to `None`. `What are attention masks? <../glossary.html#attention-mask>`__ decoder_start_token_id=None: (`optional`) int If an encoder-decoder model starts decoding with a different token than BOS. Defaults to `None` and is changed to `BOS` later. use_cache: (`optional`) bool If `use_cache` is True, past key values are used to speed up decoding if applicable to model. Defaults to `True`. Return: output: `tf.Tensor` of `dtype=tf.int32` shape `(batch_size * num_return_sequences, sequence_length)` sequence_length is either equal to max_length or shorter if all batches finished early due to the `eos_token_id` Examples:: tokenizer = AutoTokenizer.from_pretrained('distilgpt2') # Initialize tokenizer model = TFAutoModelWithLMHead.from_pretrained('distilgpt2') # Download model and configuration from S3 and cache. outputs = model.generate(max_length=40) # do greedy decoding print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True))) tokenizer = AutoTokenizer.from_pretrained('openai-gpt') # Initialize tokenizer model = TFAutoModelWithLMHead.from_pretrained('openai-gpt') # Download model and configuration from S3 and cache. input_context = 'The dog' input_ids = tokenizer.encode(input_context, return_tensors='tf') # encode input context outputs = model.generate(input_ids=input_ids, num_beams=5, num_return_sequences=3, temperature=1.5) # generate 3 independent sequences using beam search decoding (5 beams) with sampling from initial context 'The dog' for i in range(3): # 3 output sequences were generated print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True))) tokenizer = AutoTokenizer.from_pretrained('distilgpt2') # Initialize tokenizer model = TFAutoModelWithLMHead.from_pretrained('distilgpt2') # Download model and configuration from S3 and cache. input_context = 'The dog' input_ids = tokenizer.encode(input_context, return_tensors='tf') # encode input context outputs = model.generate(input_ids=input_ids, max_length=40, temperature=0.7, num_return_sequences=3) # 3 generate sequences using by sampling for i in range(3): # 3 output sequences were generated print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True))) tokenizer = AutoTokenizer.from_pretrained('ctrl') # Initialize tokenizer model = TFAutoModelWithLMHead.from_pretrained('ctrl') # Download model and configuration from S3 and cache. input_context = 'Legal My neighbor is' # "Legal" is one of the control codes for ctrl input_ids = tokenizer.encode(input_context, return_tensors='tf') # encode input context outputs = model.generate(input_ids=input_ids, max_length=50, temperature=0.7, repetition_penalty=1.2) # generate sequences print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True))) tokenizer = AutoTokenizer.from_pretrained('gpt2') # Initialize tokenizer model = TFAutoModelWithLMHead.from_pretrained('gpt2') # Download model and configuration from S3 and cache. input_context = 'My cute dog' # "Legal" is one of the control codes for ctrl bad_words_ids = [tokenizer.encode(bad_word, add_prefix_space=True) for bad_word in ['idiot', 'stupid', 'shut up']] input_ids = tokenizer.encode(input_context, return_tensors='tf') # encode input context outputs = model.generate(input_ids=input_ids, max_length=100, do_sample=True, bad_words_ids=bad_words_ids) # generate sequences without allowing bad_words to be generated NaYou tried to generate sequences with a model that does not have a LM Head.Please use another model class (e.g. `TFOpenAIGPTLMHeadModel`, `TFXLNetLMHeadModel`, `TFGPT2LMHeadModel`, `TFCTRLLMHeadModel`, `TFT5ForConditionalGeneration`, `TFTransfoXLLMHeadModel`)rrz4`max_length` should be a strictely positive integer.z*`min_length` should be a positive integer.z `do_sample` should be a boolean.z%`early_stopping` should be a boolean.z `use_cache` should be a boolean.z3`num_beams` should be a strictely positive integer.z+`temperature` should be strictely positive.z%`top_k` should be a positive integer.z"`top_p` should be between 0 and 1.?z$`repetition_penalty` should be >= 1.zIIf input_ids is not defined, `bos_token_id` should be a positive integer.z,`pad_token_id` should be a positive integer.z,`eos_token_id` should be a positive integer.z.`length_penalty` should be strictely positive.z>`num_return_sequences` should be a strictely positive integer.zZ`bad_words_ids` is either `None` or a list of lists of tokens that should not be generatedzyou should either supply a context to complete as `input_ids` input or a `bos_token_id` (integer >= 0) as a first token to start the generation.z>Input prompt should be of shape (batch_size, sequence length).FzGreedy decoding will always produce the same output for num_beams == 1 and num_return_sequences > 1. Please set num_return_sequences = 1zxGreedy beam search decoding cannot return more sequences than it has beams. Please set num_beams >= num_return_sequencesdtypezHSetting `pad_token_id` to {} (first `eos_token_id`) to generate sequencezWdecoder_start_token_id or bos_token_id has to be defined for encoder-decoder generation get_encoderz/{} should have a 'get_encoder' function definedz{} should be a method)attention_maskz5expected encoder_outputs[0] to have 1st dimension bs=z, got  )Zrepeatsaxis)r)rr)cur_len max_length min_length do_sampleearly_stopping temperaturetop_ktop_prepetition_penaltyno_repeat_ngram_size bad_words_ids bos_token_id pad_token_id eos_token_iddecoder_start_token_id batch_sizenum_return_sequenceslength_penalty num_beams vocab_sizeencoder_outputsrr|)rrrrrrrrrrrrrrrrrrr|)6rKr6r$rrrrr|rrrrrrrrrrrrr shape_listr&rrTr!rsr;fillrtnumpycastmath not_equalint32 ones_likerWwarningrDris_encoder_decoderr9callabler broadcast_to expand_dimsreshapeonesrrepeatrangegather_generate_beam_search_generate_no_beam_search)rrArrrrrrrrrrrrrrrrrrr|rrrZeffective_batch_sizeZeffective_batch_multencoderrZ input_ids_lenZexpanded_batch_idxsoutputrrrgenerates  "       "             $  zTFPreTrainedModel.generatec* st|dddf}t|dddf|}|}||kr|j||||d}|f|}|ddddddf}|||r|d}| dkrt||| }tj||}| dkr t||| |}g}|D] |fddt |Dqt |tj |tj d t d  }| dk rht|| }g}|D]"|fd dt |Dq&t |tj |tj d t d  }dk r||krtj fd dt |Dtj d }t|||g}t ||t d  }|r|dkr||}t|||d }tjtjj|tjdddd} ntjj|dtjd} dk r:| || d|}!n| }!t|t|!dgd}dk r|!k}"tj|t|"tj}#|d|#||#}||#8}tj|dkrq|jjdkrtj|tjt|ddftjd gdd}|d}q4tj|}$tj|}%|$|%kr| dk s(tdtj||% gtjd | }&tt|d||%g}'t!ttt |d||g}(t"|(|'k||&})n|})|)S)z Generate sequences for each example without beam search (num_beams == 1). All returned sequence are generated independantly. Nrpastrr|rrr~csg|]}|krdndqSTFrrtokenbanned_tokens_slicerr nsz>TFPreTrainedModel._generate_no_beam_search..rinfcsg|]}|krdndqSrrrrrrr|scsg|]}|krdndqSrrrrrrrs)rrrZ num_samplesr)r output_typeFzB`Pad_token_id` has to be defined if batches have different lengths)#r;rryr}#_create_next_token_logits_penaltiesrmultiplycalc_banned_ngram_tokensappendrset_tensor_by_indices_to_valueconvert_to_tensorr!floatcalc_banned_bad_words_idsrtf_top_k_top_p_filteringsqueezerandom categoricalrZargmaxconcatrr reduce_maxr$rrr reduce_minrTr transposewhere)*rrArrrrrrrrrrrrrrrrrrr|Zunfinished_sents sent_lengthsr model_inputsr{next_token_logitsnext_token_logits_penalties banned_tokensbanned_tokens_indices_maskis_token_logit_eos_tokeneos_token_indices_maskZ next_tokenZ tokens_to_addZ eos_in_sentsZ+is_sents_unfinished_and_token_to_add_is_eosZmin_sent_lengthZmax_sent_lengthpaddingZbroad_casted_sent_lengthsZbroad_casted_rangedecodedr)rrrr3s                z*TFPreTrainedModel._generate_no_beam_searchcL s fddt|D}|dkrftj|dftjd}tj|dftjdd}t||gd}ntj|ftjd}t||f}|}ddt|D}|kr4|j|||d }|f|} | d d d dd d f}!|| |r| d}| d krt |!| }"tj |!|"}!|d kr,|!|}!tj j |!dd }#d k r||kr|}$tjfddtDtjd}%t|%|$g}&t|#|&td }#| d kr|}$t|$| |}'g}(|'D]"|(fddtDqt|#tj|(tjdtd }#| d k rdt| }'g}(|'D]"|(fddtDq"t|#tj|(tjdtd }#t|#|gks~t|r&|#t|d d d f|f})t|)|| dd})t|)|f})tjj|)tjdd}*tj|)|*dd}+tj|+ddd},tj|+|,dd}+tj|*|,dd}*nR|#t|d d d f|f}+t|+|f}+tj j|+ddd\}+}*t|+t|*kr|dgksntg}-t|D]}.||.rt||.kstd d k r|d k std|-!d |d fgqg}/t"t#|*|.|+|.D]\}0\}1}2|1}3|1}4|.|3}5d k r|4$kr|0k}6|6rq2||.%t&|5|2$n|/|2|4|5ft|/kr2q̐q2||.p||.j't(|+|.$|d||.<t|/kstd|-!|/t|-|.dkstqt)|rFq4t|-|ks\ttjdd|-Dtjd}tjdd|-Dtjd}7tjd d|-Dtjd}8t*fd!d|8Dtjt+|7dgdd |d k r|,||8}|j-j.dkr*tj|tjt|d dftjdgdd }|d}qt|D]}.||.rNq<d k rt)fd"d#|*|.Drt/|+|.d ft||f|.kstd$ |+d d d f|.t||f|.tD]:}3|.|3}5||5$0}9|5}:||.%|:|9qސq<|r(|n||};|r:dn|}t"|D]P\}?}@t1|@j2d%d&d'}At|<D]*}B|A3d}C|=t|C|>|CqpqN|;t|>kstd( |;t|>tj|=tjd}Dt4|D$t(|D$kr|d k std)t5t(|D$d}Eg}Ft"|>D]\}?}G|D|?t|Gd ksHt|D|?|Ekr\|G}Hnv|E|D|?}I|tj|Iftjd}Jtj|G|Jgdd }H|D|?krt6tj|Etjd|D|?ktj|Eftjd|H}H|F|Hq&t*|F}Kn"fd*d#|>D stt*|>}K|KS)+z? Generate sequences for each example with beam search. csg|]}tdqS))r)BeamHypothesesr_)rrrrrrrsz;TFPreTrainedModel._generate_beam_search..FrrgercSsg|]}dqS)FrrrrrrsrrNr~rcsg|]}|krdndqSrrrrrrr)srcsg|]}|krdndqSrrrrrrr:scsg|]}|krdndqSrrrrrrrHsr)rrmin_tokens_to_keepr batch_dims DESCENDING) directionrT)ksortedz?Batch can only be done if at least {} beams have been generatedzMgenerated beams >= num_beams -> eos_token_id and pad_token have to be defined)rzBeam should always be fullcSsg|] }|dqS)rrrxrrrrscSsg|] }|dqS)rrrrrrrscSsg|] }|dqS)rrrrrrrscs"g|]}t|ddfqSr)r;identityr)rArrrsc3s"|]}|k VqdSr)ritem)rtoken_id)rrrrrsz:TFPreTrainedModel._generate_beam_search..z\If batch_idx is not done, final next scores: {} have to equal to accumulated beam_scores: {}cSs|dS)Nrr)rrrrz9TFPreTrainedModel._generate_beam_search..)keyz9Output batch size {} must match output beam hypotheses {}z `Pad_token_id` has to be definedc3s|]}t|kVqdSrrt)rhypo)rrrrs)7rr;zerosfloat32rrrryr}rrrnnZ log_softmaxrr!rrrrrrrrTrrrrrargsortrrtrDextend enumeratezipraddris_donerallstackr_reorder_cacher$rZ reduce_allrrbeamsr%rminr)LrrArrrrrrrrrrrrrrrrrrrrrrr|Zgenerated_hypsZbeam_scores_beginZbeam_scores_endZ beam_scoresrdonerr{rrZscoresZnum_batch_hypothesesrrrrZ_scoresZ next_tokensZ next_scoresZnext_scores_indicesZnext_batch_beamZ batch_idxZnext_sent_beamZbeam_token_rankZ beam_token_idZbeam_token_scoreZbeam_idrZeffective_beam_idZ&is_beam_token_worse_than_top_num_beamsZ beam_tokensbeam_idxZ final_scoreZ final_tokensZoutput_batch_sizeZ%output_num_return_sequences_per_batchZsent_lengths_listbestiZ hypothesesZ sorted_hypsjZbest_hyprZ sent_max_lenZ decoded_listrZ decoded_sliceZnum_pad_tokensrrr)rrrrArrrrrrs           .             "       z'TFPreTrainedModel._generate_beam_searchcstfdd|DS)Nc3s|]}tj|ddVqdS)rrN)r;r)rZ layer_pastrrrrsz3TFPreTrainedModel._reorder_cache..)tuple)rrrrrrsz TFPreTrainedModel._reorder_cache)N)N)NNNNNNNNNNNNNNNNNNNN)rrrr r"rirHpropertyrCr4rIrKrMrOrPrU classmethodrhryr}rrr staticmethodr __classcell__rrrGrr?ks`   "   G  k$Cr?c Cstt|}dd|D}t|D]R\}}|||}t|j}|||dk<d|||dk<t||||q(tj |tj dS)NcSsg|]}t|qSr)runique)rZinput_idrrrrsz7_create_next_token_logits_penalties..rrr) rrrrrrrputr;rr) rAlogitsrZtoken_penaltiesprev_input_idsrZ prev_input_idZlogit_penalizedZlogit_penaltiesrrrrs  rc sdkrddt|DSddt|Dt|D]f}||}tfddtDD].}t|dd}||g|dg||<qnq8fddfd dt|D}|S) NrcSsg|]}gqSrrrrrrr.sz,calc_banned_ngram_tokens..cSsg|]}iqSrrrrrrr/scsg|]}|dqSrr)rr) gen_tokensrrr3srcs8d}t||f}||gS)Nr)rrtolistr')hypo_idxZ start_idxZ ngram_idx)rgenerated_ngramsrr rr_get_generated_ngrams7s z7calc_banned_ngram_tokens.._get_generated_ngramscsg|] }|qSrr)rr)rrrr=s)rrr rrr') r Z num_hyposrridxZgenerated_ngramZngramZprev_ngram_tuplerr)rrr rrr rr*s  rcs~g}fdd}D]d}g}|D]L}t|dks>td||||dddkr^q ||dq ||q|S)NcsFt|dkrdSt|tkr$dS|t| d|kr>dSdSdS)NrTFr)Z prev_tokenstokensr rr _tokens_matchDs z0calc_banned_bad_words_ids.._tokens_matchrz9Banned words token sequences {} cannot have an empty listrF)rtrTrDrr r)r rrrZprev_input_ids_slicerZbanned_token_seqrrrrAs  rr~ZInfc Cs8t|}|dkrLtt|||d}|tjj||dddk}t|||}|dkr4tj|dd}tj||ddd }tjj tj j |dd dd } | |k} |dkrt t | d d d |f| d d |d fgd} tj| ddd } t t | d d d df| d d dd fgd} t| |}t|||}|S) a Filter a distribution of logits using top-k and/or nucleus (top-p) filtering Args: logits: logits distribution shape (batch size, vocabulary size) if top_k > 0: keep only top k tokens with highest probability (top-k filtering). if top_p < 1.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering). Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751) Make sure we keep at least min_tokens_to_keep per batch example in the output From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317 rr)r).rNr~r)rr)rrrN)rrmaxr;rrrrrZcumsumrZsoftmaxr zeros_likeZrollscatter_values_on_batch_indices) r rrZ filter_valuerZ logits_shapeZindices_to_removeZsorted_indicesZ sorted_logitsZcumulative_probsZsorted_indices_to_removerrrres<    ,  rc Csnt|}tttjt|ddd|ddg}tt|t|ddggd}t|t|dg|S)Nrrrr) rr;rrrrrrZ scatter_nd)valuesZ batch_indicesrZbroad_casted_batch_dimsZ pair_indicesrrrrs,"rcCst||}t|||Sr)r;rr)ZtensorindicesvalueZ value_tensorrrrrsrc@s.eZdZddZddZddZd dd ZdS) rcCs,|d|_||_||_||_g|_d|_dS)z7 Initialize n-best list of hypotheses. rgeAN)rrrrr worst_score)rrrrrrrrr4s  zBeamHypotheses.__init__cCs t|jS)z3 Number of hypotheses in the list. )rtrrBrrr__len__szBeamHypotheses.__len__cCs|t||j}t||jks*||jkr|j||ft||jkrtddt|jD}|j|dd=|dd|_nt||j|_dS)z3 Add a new hypothesis to the list. cSsg|]\}\}}||fqSrr)rrsrrrrrs z&BeamHypotheses.add..rrN) rtrrrrrrrr)rZhypZ sum_logprobsZscoreZ sorted_scoresrrrrszBeamHypotheses.addNcCsJt||jkrdS|jrdS|dkr*|j}|||j}|j|k}|SdS)z If there are enough hypotheses and that none of the hypotheses being generated can become better than the worst one in the heap, then we are done with this sentence. FTN)rtrrrrr)rZbest_sum_logprobsrZ cur_scoreretrrrrs zBeamHypotheses.is_done)N)rrrr4rrrrrrrrs rcs.eZdZdfdd ZddZddZZS) TFConv1D{Gz?c s$tjf|||_||_||_dS)z TFConv1D layer as defined by Radford et al. for OpenAI GPT (and also used in GPT-2) Basically works like a Linear layer but the weights are transposed N)r0r4nfnxinitializer_range)rr!r"r#r,rGrrr4szTFConv1D.__init__cCs@|jd|j|jgt|jd|_|jdd|jgtd|_dS)Nweightrr-biasr) add_weightr"r!get_initializerr#r$r;Zzeros_initializerr&rZ input_shaperrrbuilds  zTFConv1D.buildcCsRt|dd\}}t|d|jg}t||j|j}t||||jg}|S)Nrr)rr;rr"matmulr$r&r!)rrZbzslrrrcalls z TFConv1D.call)r )rrrr4r*r-rrrrGrrs rcsHeZdZdZdfdd ZfddZddd Zd d Zd d ZZ S)TFSharedEmbeddingsz'Construct shared token embeddings. Nc s4tjf|||_||_|dkr*|dn||_dS)Ng)r0r4r hidden_sizer#)rrr/r#r,rGrrr4szTFSharedEmbeddings.__init__cs0|jd|j|jgt|jd|_t|dS)zBuild shared token embedding layer Shared weights logic adapted from https://github.com/tensorflow/models/blob/a009f4fb9d2fc4949e32192a944688925ef78659/official/transformer/v2/embedding_layer.py#L24 r$r%N)r'rr/r(r#r$r0r*r)rGrrr*s  zTFSharedEmbeddings.build embeddingcCs6|dkr||S|dkr$||Std|dS)a Get token embeddings of inputs. Args: inputs: list of three int64 tensors with shape [batch_size, length]: (input_ids, position_ids, token_type_ids) mode: string, a valid value is one of "embedding" and "linear". Returns: outputs: (1) If mode == "embedding", output embedding tensor, float32 with shape [batch_size, length, embedding_size]; (2) mode == "linear", output linear tensor, float32 with shape [batch_size, length, vocab_size]. Raises: ValueError: if mode is not valid. Shared weights logic adapted from https://github.com/tensorflow/models/blob/a009f4fb9d2fc4949e32192a944688925ef78659/official/transformer/v2/embedding_layer.py#L24 r0Zlinearzmode {} is not valid.N) _embedding_linearr(rD)rrFmoderrrr-s   zTFSharedEmbeddings.callcCst|j|S)z)Applies embedding based on inputs tensor.)r;rr$)rrArrrr1szTFSharedEmbeddings._embeddingcCsHt|dd}t|d|jg}tj||jdd}t|||jgS)zComputes logits by running inputs through a linear layer. Args: inputs: A float32 tensor with shape [..., hidden_size] Returns: float32 tensor with shape [..., vocab_size]. NrT)Z transpose_b)rr;rr/r+r$r)rrFZ first_dimsrr rrrr2"szTFSharedEmbeddings._linear)N)r0) rrrr r4r*r-r1r2rrrrGrr.s  r.cs,eZdZdZdfdd Zd ddZZS) TFSequenceSummarya Compute a single vector summary of a sequence hidden states according to various possibilities: Args of the config class: summary_type: - 'last' => [default] take the last token hidden state (like XLNet) - 'first' => take the first token hidden state (like Bert) - 'mean' => take the mean of all tokens hidden states - 'cls_index' => supply a Tensor of classification token position (GPT/GPT-2) - 'attn' => Not implemented now, use multi-head attention summary_use_proj: Add a projection after the vector extraction summary_proj_to_labels: If True, the projection outputs to config.num_labels classes (otherwise to hidden_size). Default: False. summary_activation: 'tanh' => add a tanh activation to the output, Other => no activation. Default summary_first_dropout: Add a dropout before the projection and activation summary_last_dropout: Add a dropout after the projection and activation r c stjf|t|dr|jnd|_|jdkr2tt|do@|j|_|jrt|drl|jrl|jdkrl|j}n|j }t j j j |t|dd|_t|do|jd k|_|jrt j jj|_t|d o|jdk|_|jrt j j |j|_t|d o|jdk|_|jrt j j |j|_dS) Nsummary_use_projlastattnsummary_proj_to_labelsrsummary)Zkernel_initializerrdsummary_activationtanhsummary_first_dropoutsummary_last_dropout)r0r4r9 summary_typerJr5 has_summaryr8Z num_labelsr/r;r<rrZDenser(r9r:has_activationZ activationsr; activationr<has_first_dropoutZDropout first_dropoutr=has_last_dropout last_dropout)rr$r#r,Z num_classesrGrrr4As.  zTFSequenceSummary.__init__FcCst|tttfs|}d}nZt|ttfr^|d}t|dkrD|dnd}t|dksttdn|d}|dd}|jdkr|ddd f}n|jd kr|dddf}n|jd krtj |dd }n|jdkrZt |}|dkrt |dd |d d}t |}t|t|dkr*|dtj f}tj ||t|dd}tj|t|dd }n|jdkrjt|jr|j||d}|jr||}|jr||}|jr|j||d}|S)a hidden_states: float Tensor in shape [bsz, seq_len, hidden_size], the hidden-states of the last layer. cls_index: [optional] position of the classification token if summary_type == 'cls_index', shape (bsz,) or more generally (bsz, ...) where ... are optional leading dimensions of hidden_states. if summary_type == 'cls_index' and cls_index is None: we take the last token of the sequence as classification token NrrrzToo many inputs. hidden_states cls_indexr6rfirstZmeanr.rr7ra)r&dictrrsrtrTr'r>r;Z reduce_meanrrZnewaxisrrrJrBrCr?r9r@rArDrE)rrFrbrFrGrZ hidden_shapeZ cls_shaperrrr-asR           zTFSequenceSummary.call)r )F)rrrr r4r-rrrrGrr41s r4cs*|j}t|fddt|DS)z.Deal with dynamic shape in tensorflow cleanly.cs$g|]\}}|dkr|n|qSrr)rrrZdynamicrrrszshape_list..)rrr;r)rZstaticrrKrrs  rr cCstjjj|dS)zCreates a `tf.initializers.truncated_normal` with the given range. Args: initializer_range: float, initializer range for stddev. Returns: TruncatedNormal initializer with stddev = `initializer_range`. )stddev)r;r<Z initializersZTruncatedNormal)r#rrrr(sr()r ).r r7loggingrQrorrZ tensorflowr;Ztensorflow.python.keras.savingrZconfiguration_utilsrZ file_utilsrrrrr r Zmodeling_tf_pytorch_utilsr getLoggerrrWr r>r<ZModelr?rrrrrrrobjectrrrZLayerrr.r4rr(rrrrsF     ::$2 2=k