o pi7@sxddlmZddlZddlZddlmZddlmZmZddl m Z ddl m Z m Z ddlmZGd d d ee e ZdS) )OptionalN)nn) GPT2ConfigGPT2LMHeadModel)ModuleUtilsMixin) ConfigMixinregister_to_config) ModelMixinc(s>eZdZdZddgZe         dd6ed7ed8e d9eefd:d;ZZS)?UniDiffuserTextDecodera Text decoder model for a image-text [UniDiffuser](https://arxiv.org/pdf/2303.06555.pdf) model. This is used to generate text from the UniDiffuser image-text embedding. Parameters: prefix_length (`int`): Max number of prefix tokens that will be supplied to the model. prefix_inner_dim (`int`): The hidden size of the incoming prefix embeddings. For UniDiffuser, this would be the hidden dim of the CLIP text encoder. prefix_hidden_dim (`int`, *optional*): Hidden dim of the MLP if we encode the prefix. vocab_size (`int`, *optional*, defaults to 50257): Vocabulary size of the GPT-2 model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`GPT2Model`] or [`TFGPT2Model`]. n_positions (`int`, *optional*, defaults to 1024): The maximum sequence length that this model might ever be used with. Typically set this to something large just in case (e.g., 512 or 1024 or 2048). n_embd (`int`, *optional*, defaults to 768): Dimensionality of the embeddings and hidden states. n_layer (`int`, *optional*, defaults to 12): Number of hidden layers in the Transformer encoder. n_head (`int`, *optional*, defaults to 12): Number of attention heads for each attention layer in the Transformer encoder. n_inner (`int`, *optional*, defaults to None): Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd activation_function (`str`, *optional*, defaults to `"gelu"`): Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new"]`. resid_pdrop (`float`, *optional*, defaults to 0.1): The dropout probability for all fully connected layers in the embeddings, encoder, and pooler. embd_pdrop (`float`, *optional*, defaults to 0.1): The dropout ratio for the embeddings. attn_pdrop (`float`, *optional*, defaults to 0.1): The dropout ratio for the attention. layer_norm_epsilon (`float`, *optional*, defaults to 1e-5): The epsilon to use in the layer normalization layers. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. scale_attn_weights (`bool`, *optional*, defaults to `True`): Scale attention weights by dividing by sqrt(hidden_size).. use_cache (`bool`, *optional*, defaults to `True`): Whether or not the model should return the last key/values attentions (not used by all models). scale_attn_by_inverse_layer_idx (`bool`, *optional*, defaults to `False`): Whether to additionally scale attention weights by `1 / layer_idx + 1`. reorder_and_upcast_attn (`bool`, *optional*, defaults to `False`): Whether to scale keys (K) prior to computing attention (dot-product) and upcast attention dot-product/softmax to float() when training with mixed precision. zh\.\d+\.attn\.biaszh\.\d+\.attn\.masked_biasNQ gelu_new皙?h㈵>{Gz?TF prefix_lengthprefix_inner_dimprefix_hidden_dim vocab_size n_positionsn_embdn_layern_headn_inneractivation_function resid_pdrop embd_pdrop attn_pdroplayer_norm_epsiloninitializer_rangescale_attn_weights use_cachescale_attn_by_inverse_layer_idxreorder_and_upcast_attncst||_||kr|durtd|d|d||_||_|jdur.t|j|jnt|_ |jdur?t|j|nt|_ t did|d|d|d|d|d | d | d | d | d | d|d|d|d|d|d|}t ||_ dS)Nz>`prefix_hidden_dim` cannot be `None` when `prefix_inner_dim`: z and `n_embd`: z are not equal.rrrrrrrrrr r!r"r#r$r%r&)super__init__r ValueErrorrrrLinearIdentity encode_prefix decode_prefixrr transformer)selfrrrrrrrrrrrrr r!r"r#r$r%r& gpt_config __class__r's/home/ubuntu/SoloSpeech/.venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/modeling_text_decoder.pyr)Bsd       zUniDiffuserTextDecoder.__init__ input_ids prefix_embedsattention_masklabelsc Cs|jj|}||}||}tj||fdd}|dur1||jd|j}tj||fdd}|j|||d} |j durB| |fS| S)a* Args: input_ids (`torch.Tensor` of shape `(N, max_seq_len)`): Text tokens to use for inference. prefix_embeds (`torch.Tensor` of shape `(N, prefix_length, 768)`): Prefix embedding to preprend to the embedded tokens. attention_mask (`torch.Tensor` of shape `(N, prefix_length + max_seq_len, 768)`, *optional*): Attention mask for the prefix embedding. labels (`torch.Tensor`, *optional*): Labels to use for language modeling. dimNr) inputs_embedsr8r7) r/wter-r.torchcatget_dummy_tokenshapedevicer) r0r5r6r7r8embedding_texthidden embedding_cat dummy_tokenoutr'r'r4forwards   zUniDiffuserTextDecoder.forward batch_sizerBreturncCstj||jtj|dS)N)dtyperB)r>zerosrint64)r0rIrBr'r'r4r@sz&UniDiffuserTextDecoder.get_dummy_tokencCs ||S)N)r-)r0prefixr'r'r4encodes zUniDiffuserTextDecoder.encodec Cs~tj|ddd}g}g}|D]"}|||}|j|||d\}}||d||dqt|}t|}||fS)a Generate captions given text embedding features. Returns list[L]. Args: features (`torch.Tensor` of shape `(B, L, D)`): Text embedding features to generate captions from. eos_token_id (`int`): The token ID of the EOS token for the text decoder model. device: Device to perform text generation on. Returns: `List[str]`: A list of strings generated from the decoder model. r9rr:) input_embedsrB eos_token_id)r>splitr.to generate_beamappendstack) r0featuresrQrBgenerated_tokensgenerated_seq_lengthsfeature output_tokens seq_lengthsr'r'r4generate_captionss   z(UniDiffuserTextDecoder.generate_captionsC? beam_size entry_length temperaturerQcs|}dd} tj||tjdtj||tjd} |dur|} n|jj|} t|D]} |j| d} | j}|dddddf|dkrE|nd}| d }| dur| |d\} }| j |g| j ddR} |dd| d}} dur||n~j |gj ddRtj|fddnfttj || <d|| df<| dddf|}| d7<|dddf}|d |d\}}||j d}|||j d}|d}|tj|fdd| |} |} | |} |jj|| j ddd}tj| |fdd} | ||} | r$nq*| } | jd d }fd d |D}tj|dd}tjfd d |Djd|fS)a Generates text using the given tokenizer and text prompt or token embedding via beam search. This implementation is based on the beam search implementation from the [original UniDiffuser code](https://github.com/thu-ml/unidiffuser/blob/main/libs/caption_decoder.py#L89). Args: eos_token_id (`int`, *optional*): The token ID of the EOS token for the text decoder model. input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*): Tokenizer indices of input sequence tokens in the vocabulary. One of `input_ids` and `input_embeds` must be supplied. input_embeds (`torch.Tensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*): An embedded representation to directly pass to the transformer as a prefix for beam search. One of `input_ids` and `input_embeds` must be supplied. device: The device to perform beam search on. beam_size (`int`, *optional*, defaults to `5`): The number of best states to store during beam search. entry_length (`int`, *optional*, defaults to `67`): The number of iterations to run beam search. temperature (`float`, *optional*, defaults to 1.0): The temperature to use when performing the softmax over logits from the decoding model. Returns: `Tuple(torch.Tensor, torch.Tensor)`: A tuple of tensors where the first element is a tensor of generated token sequences sorted by score in descending order, and the second element is the sequence lengths corresponding to those sequences. N)rBrK)r<rr`r9r:T) descendingcg|]}|qSr'r'.0i)tokensr'r4 %z8UniDiffuserTextDecoder.generate_beam..crfr'r'rg)r\r'r4rk'rl)rK)r>onesintrLboolr/r=rangelogitssoftmaxlogtopkexpandrApermutesqueezer?floatnpinfview unsqueezeeqallargsortrVtensorrK)r0r5rPrBrarbrcrQstop_token_indexscores is_stopped generatedrioutputsrq next_tokens scores_sumscores_sum_averagenext_tokens_sourcenext_token_embedorder output_textsr')r\rjr4rTs`(  &  "  z$UniDiffuserTextDecoder.generate_beam)Nr r rrrNrrrrrrTTFF)NN)NNNr^r_r`N)__name__ __module__ __qualname____doc__"_keys_to_ignore_on_load_unexpectedr rnrstrrxror)r>TensorrHrBr@rOno_gradr]rT __classcell__r'r'r2r4r s1     D   r )typingrnumpyryr>r transformersrrtransformers.modeling_utilsrconfiguration_utilsrr modelsr r r'r'r'r4s