# coding=utf-8 # Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved. # Copyright 2022 The Salesforce Team Authors and The HuggingFace Team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Paddle BLIP model.""" from dataclasses import dataclass from functools import partial from typing import Any, Optional, Tuple, Union import paddle import paddle.nn as nn import paddle.nn.functional as F from paddle.distributed.fleet.utils import recompute from ...utils.initializer import normal_, ones_, zeros_ from ..activations import ACT2FN from ..model_outputs import BaseModelOutput, BaseModelOutputWithPooling, ModelOutput from ..model_utils import PretrainedModel from .configuration import BlipConfig, BlipTextConfig, BlipVisionConfig from .modeling_text import BlipTextLMHeadModel, BlipTextModel BLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [ "Salesforce/blip-vqa-base", "Salesforce/blip-vqa-capfilt-large", "Salesforce/blip-image-captioning-base", "Salesforce/blip-image-captioning-large", "Salesforce/blip-itm-base-coco", "Salesforce/blip-itm-large-coco", "Salesforce/blip-itm-base-flickr", "Salesforce/blip-itm-large-flickr", ] __all__ = [ "BlipPretrainedModel", "BlipVisionModel", "BlipModel", "BlipForConditionalGeneration", "BlipForQuestionAnswering", "BlipForImageTextRetrieval", ] # Copied from transformers.models.clip.modeling_clip.contrastive_loss def contrastive_loss(logits: paddle.Tensor) -> paddle.Tensor: return F.cross_entropy(logits, paddle.arange(len(logits))) # Copied from transformers.models.clip.modeling_clip.clip_loss with clip->blip def blip_loss(similarity: paddle.Tensor) -> paddle.Tensor: caption_loss = contrastive_loss(similarity) image_loss = contrastive_loss(similarity.t()) return (caption_loss + image_loss) / 2.0 def Parameter(tensor): return paddle.create_parameter(tensor.shape, dtype=tensor.dtype, default_initializer=nn.initializer.Assign(tensor)) @dataclass class BlipForConditionalGenerationModelOutput(ModelOutput): """ Adapted from the base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states. This class also adds the loss term from the text decoder. Args: loss (`paddle.Tensor`, *optional*, returned when `labels` is provided, `paddle.Tensor` of shape `(1,)`): Languge modeling loss from the text decoder. decoder_logits (`paddle.Tensor` of shape `(batch_size, sequence_length, config.vocab_size)`, *optional*): Prediction scores of the language modeling head of the text decoder model. image_embeds (`paddle.Tensor` of shape `(batch_size, output_dim)`, *optional*): The image embeddings obtained after applying the Vision Transformer model to the input image. last_hidden_state (`paddle.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Sequence of hidden-states at the output of the last layer of the model. hidden_states (`tuple(paddle.Tensor)`, *optional*, returned when `output_hidden_states=True`): Tuple of `paddle.Tensor` (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the optional initial embedding outputs. attentions (`tuple(paddle.Tensor)`, *optional*, returned when `output_attentions=True` is passed): Tuple of `paddle.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. """ loss: Optional[Tuple[paddle.Tensor]] = None decoder_logits: Optional[Tuple[paddle.Tensor]] = None image_embeds: Optional[paddle.Tensor] = None last_hidden_state: paddle.Tensor = None hidden_states: Optional[Tuple[paddle.Tensor]] = None attentions: Optional[Tuple[paddle.Tensor]] = None @dataclass class BlipTextVisionModelOutput(ModelOutput): """ Adapted from the base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states. This class also adds the loss term from the text decoder. Args: loss (`paddle.Tensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Languge modeling loss from the text decoder. image_embeds (`paddle.Tensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`): The image embeddings obtained by applying the projection layer to the pooler_output. last_hidden_state (`paddle.Tensor` of shape `(batch_size, sequence_length, hidden_size)`): Sequence of hidden-states at the output of the last layer of the model. hidden_states (`tuple(paddle.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `paddle.Tensor` (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the optional initial embedding outputs. attentions (`tuple(paddle.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `paddle.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. """ loss: Optional[paddle.Tensor] = None image_embeds: Optional[paddle.Tensor] = None last_hidden_state: paddle.Tensor = None hidden_states: Optional[Tuple[paddle.Tensor]] = None attentions: Optional[Tuple[paddle.Tensor]] = None @dataclass class BlipImageTextMatchingModelOutput(ModelOutput): """ Adapted from the base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states. This class also adds the loss term from the text decoder as well as the image-text similarity scores. Args: itm_score (`paddle.Tensor`): The image-text similarity scores. loss (`paddle.Tensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Languge modeling loss from the text decoder. image_embeds (`paddle.Tensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`): The image embeddings obtained by applying the projection layer to the pooler_output. last_hidden_state (`paddle.Tensor` of shape `(batch_size, sequence_length, hidden_size)`): Sequence of hidden-states at the output of the last layer of the model. hidden_states (`tuple(paddle.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `paddle.Tensor` (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the optional initial embedding outputs. vision_pooler_output (`paddle.Tensor` of shape `(batch_size, hidden_size)`, *optional*): Last layer hidden-state of the vision of the vision-only branch of the model. attentions (`tuple(paddle.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `paddle.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. question_embeds (`paddle.Tensor`): The question embeddings obtained by the text projection layer. """ itm_score: Optional[paddle.Tensor] = None loss: Optional[paddle.Tensor] = None image_embeds: Optional[paddle.Tensor] = None last_hidden_state: paddle.Tensor = None hidden_states: Optional[Tuple[paddle.Tensor]] = None vision_pooler_output: Optional[paddle.Tensor] = None attentions: Optional[Tuple[paddle.Tensor]] = None question_embeds: Optional[Tuple[paddle.Tensor]] = None @dataclass class BlipOutput(ModelOutput): """ Args: loss (`paddle.Tensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`): Contrastive loss for image-text similarity. logits_per_image:(`paddle.Tensor` of shape `(image_batch_size, text_batch_size)`): The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text similarity scores. logits_per_text:(`paddle.Tensor` of shape `(text_batch_size, image_batch_size)`): The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image similarity scores. text_embeds(`paddle.Tensor` of shape `(batch_size, output_dim`): The text embeddings obtained by applying the projection layer to the pooled output of [`BlipTextModel`]. image_embeds(`paddle.Tensor` of shape `(batch_size, output_dim`): The image embeddings obtained by applying the projection layer to the pooled output of [`BlipVisionModel`]. text_model_output(`BaseModelOutputWithPooling`): The output of the [`BlipTextModel`]. vision_model_output(`BaseModelOutputWithPooling`): The output of the [`BlipVisionModel`]. """ loss: Optional[paddle.Tensor] = None logits_per_image: paddle.Tensor = None logits_per_text: paddle.Tensor = None text_embeds: paddle.Tensor = None image_embeds: paddle.Tensor = None text_model_output: BaseModelOutputWithPooling = None vision_model_output: BaseModelOutputWithPooling = None def to_tuple(self) -> Tuple[Any]: return tuple( self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple() for k in self.keys() ) class BlipVisionEmbeddings(nn.Layer): def __init__(self, config: BlipVisionConfig): super().__init__() self.config = config self.embed_dim = config.hidden_size self.image_size = config.image_size self.patch_size = config.patch_size self.class_embedding = Parameter(paddle.randn([1, 1, self.embed_dim])) self.patch_embedding = nn.Conv2D( in_channels=3, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size ) self.num_patches = (self.image_size // self.patch_size) ** 2 self.num_positions = self.num_patches + 1 self.position_embedding = Parameter(paddle.randn([1, self.num_positions, self.embed_dim])) def forward(self, pixel_values: paddle.Tensor) -> paddle.Tensor: batch_size = pixel_values.shape[0] target_dtype = self.patch_embedding.weight.dtype patch_embeds = self.patch_embedding(pixel_values) # shape = [*, width, grid, grid] patch_embeds = patch_embeds.flatten(2).transpose([0, 2, 1]) class_embeds = self.class_embedding.expand([batch_size, 1, -1]).cast(target_dtype) embeddings = paddle.concat([class_embeds, patch_embeds], axis=1) embeddings = embeddings + self.position_embedding[:, : embeddings.shape[1], :].cast(target_dtype) return embeddings # Copied from transformers.models.clip.modeling_clip.CLIPTextEmbeddings with CLIP->Blip class BlipTextEmbeddings(nn.Layer): def __init__(self, config: BlipTextConfig): super().__init__() embed_dim = config.hidden_size self.token_embedding = nn.Embedding(config.vocab_size, embed_dim) self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim) # position_ids (1, len position emb) is contiguous in memory and exported when serialized self.register_buffer("position_ids", paddle.arange(config.max_position_embeddings).reshape((1, -1))) def forward( self, input_ids: Optional[paddle.Tensor] = None, position_ids: Optional[paddle.Tensor] = None, inputs_embeds: Optional[paddle.Tensor] = None, ) -> paddle.Tensor: seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2] if position_ids is None: position_ids = self.position_ids[:, :seq_length] if inputs_embeds is None: inputs_embeds = self.token_embedding(input_ids) position_embeddings = self.position_embedding(position_ids) embeddings = inputs_embeds + position_embeddings return embeddings class BlipAttention(nn.Layer): """Multi-headed attention from 'Attention Is All You Need' paper""" def __init__(self, config): super().__init__() self.config = config self.embed_dim = config.hidden_size self.num_heads = config.num_attention_heads self.head_dim = self.embed_dim // self.num_heads if self.head_dim * self.num_heads != self.embed_dim: raise ValueError( f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:" f" {self.num_heads})." ) self.scale = self.head_dim**-0.5 self.dropout = nn.Dropout(config.attention_dropout) self.qkv = nn.Linear(self.embed_dim, 3 * self.embed_dim) self.projection = nn.Linear(self.embed_dim, self.embed_dim) def _shape(self, tensor: paddle.Tensor, seq_len: int, bsz: int): return tensor.reshape([bsz, seq_len, self.num_heads, self.head_dim]).transpose([0, 2, 1, 3]) def forward( self, hidden_states: paddle.Tensor, attention_mask: Optional[paddle.Tensor] = None, output_attentions: Optional[bool] = False, ) -> Tuple[paddle.Tensor, Optional[paddle.Tensor], Optional[Tuple[paddle.Tensor]]]: """Input shape: Batch x Time x Channel""" bsz, tgt_len, embed_dim = hidden_states.shape mixed_qkv = self.qkv(hidden_states) mixed_qkv = ( self.qkv(hidden_states) .reshape([bsz, tgt_len, 3, self.num_heads, embed_dim // self.num_heads]) .transpose([2, 0, 3, 1, 4]) ) query_states, key_states, value_states = ( mixed_qkv[0], mixed_qkv[1], mixed_qkv[2], ) # Take the dot product between "query" and "key" to get the raw attention scores. attention_scores = paddle.matmul(query_states, key_states, transpose_y=True) attention_scores = attention_scores * self.scale # Normalize the attention scores to probabilities. attention_probs = F.softmax(attention_scores, axis=-1) # This is actually dropping out entire tokens to attend to, which might # seem a bit unusual, but is taken from the original Transformer paper. attention_probs = self.dropout(attention_probs) context_layer = paddle.matmul(attention_probs, value_states).transpose([0, 2, 1, 3]) new_context_layer_shape = context_layer.shape[:-2] + [ self.embed_dim, ] context_layer = context_layer.reshape(new_context_layer_shape) output = self.projection(context_layer) outputs = (output, attention_probs) if output_attentions else (output, None) return outputs # Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Blip class BlipMLP(nn.Layer): def __init__(self, config: BlipVisionConfig): super().__init__() self.config = config self.activation_fn = ACT2FN[config.hidden_act] self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size) self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size) def forward(self, hidden_states: paddle.Tensor) -> paddle.Tensor: hidden_states = self.fc1(hidden_states) hidden_states = self.activation_fn(hidden_states) hidden_states = self.fc2(hidden_states) return hidden_states class BlipEncoderLayer(nn.Layer): def __init__(self, config: BlipVisionConfig): super().__init__() self.embed_dim = config.hidden_size self.self_attn = BlipAttention(config) self.layer_norm1 = nn.LayerNorm(self.embed_dim, epsilon=config.layer_norm_eps) self.mlp = BlipMLP(config) self.layer_norm2 = nn.LayerNorm(self.embed_dim, epsilon=config.layer_norm_eps) def forward( self, hidden_states: paddle.Tensor, attention_mask: paddle.Tensor, output_attentions: Optional[bool] = False, ) -> Tuple[paddle.Tensor]: """ Args: hidden_states (`paddle.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`paddle.Tensor`): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. `(config.encoder_attention_heads,)`. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. """ residual = hidden_states hidden_states = self.layer_norm1(hidden_states) hidden_states, attn_weights = self.self_attn( hidden_states=hidden_states, attention_mask=attention_mask, output_attentions=output_attentions, ) hidden_states = hidden_states + residual residual = hidden_states hidden_states = self.layer_norm2(hidden_states) hidden_states = self.mlp(hidden_states) hidden_states = hidden_states + residual outputs = (hidden_states,) if output_attentions: outputs += (attn_weights,) return outputs class BlipPretrainedModel(PretrainedModel): """ An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. """ config_class = BlipConfig base_model_prefix = "blip" supports_gradient_checkpointing = True _keys_to_ignore_on_load_missing = [r"position_ids"] def init_weights(self): """ A method executed at the end of each Transformer model initialization, to execute code that needs the model's modules properly initialized (such as weight initialization). """ self.apply(self._init_weights) def gradient_checkpointing_enable(self): """ Activates gradient checkpointing for the current model. Note that in other frameworks this feature can be referred to as "activation checkpointing" or "checkpoint activations". """ if not self.supports_gradient_checkpointing: raise ValueError(f"{self.__class__.__name__} does not support gradient checkpointing.") self.apply(partial(self._set_gradient_checkpointing, value=True)) def gradient_checkpointing_disable(self): """ Deactivates gradient checkpointing for the current model. Note that in other frameworks this feature can be referred to as "activation checkpointing" or "checkpoint activations". """ if self.supports_gradient_checkpointing: self.apply(partial(self._set_gradient_checkpointing, value=False)) def _init_weights(self, module): """Initialize the weights""" factor = self.config.initializer_range if isinstance(module, nn.Conv2D) or isinstance(module, nn.Embedding) or isinstance(module, nn.Linear): normal_(module.weight, mean=0.0, std=factor) if hasattr(module, "bias") and module.bias is not None: zeros_(module.bias) if isinstance(module, BlipVisionEmbeddings): if hasattr(self.config, "vision_config"): factor = self.config.vision_config.initializer_range trunc_normal_ = nn.initializer.TruncatedNormal(mean=0.0, std=factor) trunc_normal_(module.position_embedding) trunc_normal_( module.class_embedding, ) elif isinstance(module, nn.LayerNorm): zeros_(module.bias) ones_(module.weight) elif isinstance(module, nn.Linear) and module.bias is not None: zeros_(module.bias) def _set_gradient_checkpointing(self, module, value=False): if isinstance(module, BlipEncoder): module.gradient_checkpointing = value class BlipEncoder(nn.Layer): """ Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a [`BlipEncoderLayer`]. Args: config (`BlipVisionConfig`): The corresponding vision configuration for the `BlipEncoder`. """ def __init__(self, config: BlipVisionConfig): super().__init__() self.config = config self.layers = nn.LayerList([BlipEncoderLayer(config) for _ in range(config.num_hidden_layers)]) self.gradient_checkpointing = False def forward( self, inputs_embeds, attention_mask: Optional[paddle.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, BaseModelOutput]: r""" Args: inputs_embeds (`paddle.Tensor` of shape `(batch_size, sequence_length, hidden_size)`): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. attention_mask (`paddle.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`BaseModelOutput`] instead of a plain tuple. """ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict encoder_states = () if output_hidden_states else None all_attentions = () if output_attentions else None hidden_states = inputs_embeds for encoder_layer in self.layers: if output_hidden_states: encoder_states = encoder_states + (hidden_states,) if self.gradient_checkpointing and self.training: def create_custom_forward(module): def custom_forward(*inputs): return module(*inputs, output_attentions) return custom_forward layer_outputs = recompute( create_custom_forward(encoder_layer), hidden_states, attention_mask, ) else: layer_outputs = encoder_layer( hidden_states, attention_mask, output_attentions=output_attentions, ) hidden_states = layer_outputs[0] if output_attentions: all_attentions = all_attentions + (layer_outputs[1],) if output_hidden_states: encoder_states = encoder_states + (hidden_states,) if not return_dict: return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None) return BaseModelOutput( last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions ) class BlipVisionModel(BlipPretrainedModel): r""" The vision model from BLIP without any head or projection on top. This model inherits from :class:`~paddlenlp.transformers.model_utils.PretrainedModel`. Refer to the superclass documentation for the generic methods. This model is also a Paddle `paddle.nn.Layer `__ subclass. Use it as a regular Paddle Layer and refer to the Paddle documentation for all matter related to general usage and behavior. Args: config (:class:`BlipVisionConfig`): An instance of BlipVisionConfig used to construct BlipVisionModel. """ main_input_name = "pixel_values" config_class = BlipVisionConfig def __init__(self, config: BlipVisionConfig): super().__init__(config) self.config = config embed_dim = config.hidden_size self.embeddings = BlipVisionEmbeddings(config) self.encoder = BlipEncoder(config) self.post_layernorm = nn.LayerNorm(embed_dim, epsilon=config.layer_norm_eps) self.init_weights() def get_input_embeddings(self) -> nn.Layer: return self.embeddings def forward( self, pixel_values: Optional[paddle.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, BaseModelOutputWithPooling]: r""" Args: pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using [`BlipImageProcessor`]. See [`BlipImageProcessor.__call__`] for details. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~BaseModelOutputWithPooling`] instead of a plain tuple. Returns: An instance of :class:`BaseModelOutputWithPooling` if `return_dict=True`. Otherwise it returns a tuple of tensors corresponding to ordered and not None (depending on the input arguments) fields of :class:`BaseModelOutputWithPooling`. Examples: ```python >>> from PIL import Image >>> import requests >>> from paddlenlp.transformers import BLIPProcessor, BLIPVisionModel >>> model = BLIPVisionModel.from_pretrained("Salesforce/blip-image-captioning-base") >>> model.eval() >>> processor = BLIPProcessor.from_pretrained("Salesforce/blip-image-captioning-base") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor(images=image, return_tensors="pd") >>> outputs = model(**inputs) >>> last_hidden_state = outputs.last_hidden_state >>> pooled_output = outputs.pooler_output # pooled CLS states ```""" output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict if pixel_values is None: raise ValueError("You have to specify pixel_values") hidden_states = self.embeddings(pixel_values) encoder_outputs = self.encoder( inputs_embeds=hidden_states, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) last_hidden_state = encoder_outputs[0] last_hidden_state = self.post_layernorm(last_hidden_state) pooled_output = last_hidden_state[:, 0, :] if not return_dict: return (last_hidden_state, pooled_output) + encoder_outputs[1:] return BaseModelOutputWithPooling( last_hidden_state=last_hidden_state, pooler_output=pooled_output, hidden_states=encoder_outputs.hidden_states, attentions=encoder_outputs.attentions, ) class BlipModel(BlipPretrainedModel): r""" The bare BLIP Model outputting logits_per_image and logits_per_text. This model inherits from :class:`~paddlenlp.transformers.model_utils.PretrainedModel`. Refer to the superclass documentation for the generic methods. This model is also a Paddle `paddle.nn.Layer `__ subclass. Use it as a regular Paddle Layer and refer to the Paddle documentation for all matter related to general usage and behavior. Args: config (:class:`BlipConfig`): An instance of BlipConfig used to construct BlipModel. """ config_class = BlipConfig def __init__(self, config: BlipConfig): super().__init__(config) if not isinstance(config.text_config, BlipTextConfig): raise ValueError( "config.text_config is expected to be of type BlipTextConfig but is of type" f" {type(config.text_config)}." ) if not isinstance(config.vision_config, BlipVisionConfig): raise ValueError( "config.vision_config is expected to be of type BlipVisionConfig but is of type" f" {type(config.vision_config)}." ) text_config = config.text_config vision_config = config.vision_config self.projection_dim = config.projection_dim self.text_embed_dim = text_config.hidden_size self.vision_embed_dim = vision_config.hidden_size self.text_model = BlipTextModel(text_config) self.vision_model = BlipVisionModel(vision_config) self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias_attr=False) self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias_attr=False) self.logit_scale = Parameter( paddle.ones( [ 1, ] ) * config.logit_scale_init_value ) # Initialize weights and apply final processing self.init_weights() def get_text_features( self, input_ids: Optional[paddle.Tensor] = None, attention_mask: Optional[paddle.Tensor] = None, position_ids: Optional[paddle.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> paddle.Tensor: r""" Args: input_ids (`paddle.Tensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [`BertTokenizer`]. attention_mask (`paddle.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. position_ids (`paddle.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.max_position_embeddings - 1]`. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`BaseModelOutputWithPooling`] instead of a plain tuple. Returns: text_features (`paddle.Tensor` of shape `(batch_size, output_dim`): The text embeddings obtained by applying the projection layer to the pooled output of [`BlipTextModel`]. Examples: ```python >>> from paddlenlp.transformers import BlipProcessor, BlipModel >>> model = BlipModel.from_pretrained("Salesforce/blip-image-captioning-base") >>> model.eval() >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base") >>> inputs = processor(text=["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pd") >>> text_features = model.get_text_features(**inputs) ```""" output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions return_dict = return_dict if return_dict is not None else self.config.use_return_dict text_outputs = self.text_model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) pooled_output = text_outputs[1] text_features = self.text_projection(pooled_output) return text_features def get_image_features( self, pixel_values: Optional[paddle.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> paddle.Tensor: r""" Args: pixel_values (`paddle.Tensor` of shape `(batch_size, num_channels, height, width)`): Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using [`BLIPImageProcessor`]. See [`BLIPImageProcessor.__call__`] for details. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`BaseModelOutputWithPooling`] instead of a plain tuple. Returns: image_features (`paddle.Tensor` of shape `(batch_size, output_dim`): The image embeddings obtained by applying the projection layer to the pooled output of [`BlipVisionModel`]. Examples: ```python >>> from PIL import Image >>> import requests >>> from paddlenlp.transformers import BlipProcessor, BlipModel >>> model = BlipModel.from_pretrained("Salesforce/blip-image-captioning-base") >>> model.eval() >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor(images=image, return_tensors="pd") >>> image_features = model.get_image_features(**inputs) ```""" output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict vision_outputs = self.vision_model( pixel_values=pixel_values, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) pooled_output = vision_outputs[1] # pooled_output image_features = self.visual_projection(pooled_output) return image_features def forward( self, input_ids: Optional[paddle.Tensor] = None, pixel_values: Optional[paddle.Tensor] = None, attention_mask: Optional[paddle.Tensor] = None, position_ids: Optional[paddle.Tensor] = None, return_loss: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, BlipOutput]: r""" The BLIPPModel forward method, overrides the `__call__()` special method. Args: input_ids (Tensor): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Its data type should be `int64` and it has a shape of [text_batch_size, sequence_length]. pixel_values (Tensor): Pixel values. Padding will be ignored by default should you provide it. Its data type should be `float32` and it has a shape of [image_batch_size, num_channels, height, width]. position_ids(Tensor, optional): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0, max_text_length - 1]``. Shape as `(batch_size, num_tokens)` and dtype as int64. Defaults to `None`. attention_mask (Tensor, optional): Mask used in multi-head attention to avoid performing attention on to some unwanted positions, usually the paddings or the subsequent positions. Its data type can be int, float and bool. When the data type is bool, the `masked` tokens have `False` values and the others have `True` values. When the data type is int, the `masked` tokens have `0` values and the others have `1` values. When the data type is float, the `masked` tokens have `0.0` values and the others have `1.0` values. It is a tensor with shape `[batch_size, sequence_length`. Defaults to `None`, which means nothing needed to be prevented attention to. return_loss (`bool`, *optional*): Whether or not to return the contrastive loss. output_hidden_states (bool, optional): Whether to return the hidden states of all layers. Defaults to `False`. output_attentions (bool, optional): Whether to return the attentions tensors of all attention layers. Defaults to `False`. return_dict (bool, optional): Whether to return a :class:`BlipOutput` object. If `False`, the output will be a tuple of tensors. Defaults to `True`. Returns: An instance of :class:`BlipOutput` if `return_dict=True`. Otherwise it returns a tuple of tensors corresponding to ordered and not None (depending on the input arguments) fields of :class:`BlipOutput`. Examples: ```python >>> from PIL import Image >>> import requests >>> from paddlenlp.transformers import BlipProcessor, BlipModel >>> model = BlipModel.from_pretrained("Salesforce/blip-image-captioning-base") >>> model.eval() >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor( ... text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pd", padding=True ... ) >>> outputs = model(**inputs) >>> logits_per_image = outputs.logits_per_image # this is the image-text similarity score >>> probs = F.softmax(logits_per_image, axis=1) # we can take the softmax to get the label probabilities ```""" # Use BLIP model's config for some fields (if specified) instead of those of vision & text components. output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict vision_outputs = self.vision_model( pixel_values=pixel_values, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) text_outputs = self.text_model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) image_embeds = vision_outputs[1] image_embeds = self.visual_projection(image_embeds) text_embeds = text_outputs[1] text_embeds = self.text_projection(text_embeds) # normalized features image_embeds = F.normalize(image_embeds, axis=-1) text_embeds = F.normalize(text_embeds, axis=-1) # cosine similarity as logits logit_scale = self.logit_scale.exp() logits_per_text = paddle.matmul(text_embeds, image_embeds, transpose_y=True) * logit_scale logits_per_image = logits_per_text.t() loss = None if return_loss: loss = blip_loss(logits_per_text) if not return_dict: output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs) return ((loss,) + output) if loss is not None else output return BlipOutput( loss=loss, logits_per_image=logits_per_image, logits_per_text=logits_per_text, text_embeds=text_embeds, image_embeds=image_embeds, text_model_output=text_outputs, vision_model_output=vision_outputs, ) class BlipForConditionalGeneration(BlipPretrainedModel): r""" BLIP Model for image captioning. The model consists of a vision encoder and a text decoder. One can optionally pass `input_ids` to the model, which serve as a text prompt, to make the text decoder continue the prompt. Otherwise, the decoder starts generating text from the [BOS] (beginning-of-sequence) token. will start generating the caption from the text input. If no text input is provided, the decoder will start with the [BOS] token only. This model inherits from :class:`~paddlenlp.transformers.model_utils.PretrainedModel`. Refer to the superclass documentation for the generic methods. This model is also a Paddle `paddle.nn.Layer `__ subclass. Use it as a regular Paddle Layer and refer to the Paddle documentation for all matter related to general usage and behavior. Args: config (:class:`BlipConfig`): An instance of BlipConfig used to construct BlipForConditionalGeneration. """ config_class = BlipConfig _keys_to_ignore_on_load_missing = [r"text_decoder.cls.predictions.decoder.bias"] main_input_name = "pixel_values" def __init__(self, config: BlipConfig): super().__init__(config) self.vision_model = BlipVisionModel(config.vision_config) self.text_decoder = BlipTextLMHeadModel(config.text_config) self.decoder_input_ids = config.text_config.bos_token_id self.decoder_pad_token_id = config.text_config.pad_token_id # Initialize weights and apply final processing self.init_weights() def get_input_embeddings(self) -> nn.Layer: return self.vision_model.embeddings.patch_embedding def forward( self, pixel_values: paddle.Tensor, input_ids: Optional[paddle.Tensor] = None, position_ids: Optional[paddle.Tensor] = None, attention_mask: Optional[paddle.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, labels: Optional[paddle.Tensor] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, BlipForConditionalGenerationModelOutput]: r""" Args: pixel_values (Tensor): Pixel values. Padding will be ignored by default should you provide it. Its data type should be `float32` and it has a shape of [image_batch_size, num_channels, height, width]. input_ids (Tensor): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Its data type should be `int64` and it has a shape of [text_batch_size, sequence_length]. position_ids (Tensor, optional): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0, max_text_length - 1]``. Shape as `(batch_size, num_tokens)` and dtype as int64. Defaults to `None`. attention_mask (Tensor, optional): Mask used in multi-head attention to avoid performing attention on to some unwanted positions, usually the paddings or the subsequent positions. Its data type can be int, float and bool. When the data type is bool, the `masked` tokens have `False` values and the others have `True` values. When the data type is int, the `masked` tokens have `0` values and the others have `1` values. When the data type is float, the `masked` tokens have `0.0` values and the others have `1.0` values. It is a tensor with shape `[batch_size, sequence_length`. Defaults to `None`, which means nothing needed to be prevented attention to. output_attentions (bool, optional): Whether to return the attentions tensors of all attention layers. Defaults to `False`. output_hidden_states (bool, optional): Whether to return the hidden states of all layers. Defaults to `False`. return_dict (`bool`, *optional*): Whether or not to return a [`BlipForConditionalGenerationModelOutput`] instead of a plain tuple. Examples: ```python >>> from PIL import Image >>> import requests >>> from paddlenlp.transformers import BlipProcessor, BlipForConditionalGeneration >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base") >>> model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base") >>> model.eval() >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor(images=image, return_tensors="pd") >>> outputs = model(**inputs) ```""" batch_size = pixel_values.shape[0] return_dict = return_dict if return_dict is not None else self.config.use_return_dict vision_outputs = self.vision_model( pixel_values=pixel_values, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) image_embeds = vision_outputs[0] if input_ids is None: input_ids = paddle.to_tensor([[self.decoder_input_ids] * batch_size]) if labels is None: labels = paddle.where(input_ids == self.decoder_pad_token_id, paddle.to_tensor(-100), input_ids) outputs = self.text_decoder( input_ids=input_ids, position_ids=position_ids, attention_mask=attention_mask, encoder_hidden_states=image_embeds, labels=labels, return_dict=return_dict, ) if not return_dict: outputs = (outputs[0], outputs[1], image_embeds, vision_outputs[0]) + vision_outputs[2:] return tuple(output for output in outputs if output is not None) return BlipForConditionalGenerationModelOutput( loss=outputs.loss, decoder_logits=outputs.logits, image_embeds=image_embeds, last_hidden_state=vision_outputs.last_hidden_state, hidden_states=vision_outputs.hidden_states, attentions=vision_outputs.attentions, ) @paddle.no_grad() def generate( self, pixel_values: paddle.Tensor, input_ids: Optional[paddle.Tensor] = None, attention_mask: Optional[paddle.Tensor] = None, **generate_kwargs ) -> paddle.Tensor: r""" Overrides *generate* function to be able to use the model as a conditional generator Args: pixel_values (*paddle.Tensor* of shape *(batch_size, image_width, image_height)*: Input image to be processed input_ids (*paddle.Tensor* of shape *(batch_size, sequence_length)*, *optional*): The sequence used as a prompt for the generation. attention_mask (*paddle.Tensor* of shape *(batch_size, sequence_length)*, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: Examples: ```python >>> from PIL import Image >>> import requests >>> from paddlenlp.transformers import BlipProcessor, BlipForConditionalGeneration >>> model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base") >>> model.eval() >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor(images=image, return_tensors="pd") >>> outputs = model.generate(**inputs)[0] >>> print(processor.decode(outputs[0], skip_special_tokens=True)) two cats are laying on a couch ``` """ batch_size = pixel_values.shape[0] vision_outputs = self.vision_model( pixel_values=pixel_values, ) image_embeds = vision_outputs[0] image_attention_mask = paddle.ones(image_embeds.shape[:-1], dtype=paddle.int64) if isinstance(input_ids, list): input_ids = paddle.to_tensor(input_ids) elif input_ids is None: input_ids = paddle.to_tensor([[self.decoder_input_ids, self.config.text_config.eos_token_id]]).tile( [batch_size, 1] ) input_ids[:, 0] = self.config.text_config.bos_token_id attention_mask = attention_mask[:, :-1] if attention_mask is not None else None outputs = self.text_decoder.generate( input_ids=input_ids[:, :-1], eos_token_id=self.config.text_config.sep_token_id, pad_token_id=self.config.text_config.pad_token_id, attention_mask=attention_mask, encoder_hidden_states=image_embeds, encoder_attention_mask=image_attention_mask, **generate_kwargs, ) return outputs class BlipForQuestionAnswering(BlipPretrainedModel): r""" BLIP Model for visual question answering. The model consists of a vision encoder, a text encoder as well as a text decoder. The vision encoder will encode the input image, the text encoder will encode the input question together with the encoding of the image, and the text decoder will output the answer to the question. This model inherits from :class:`~paddlenlp.transformers.model_utils.PretrainedModel`. Refer to the superclass documentation for the generic methods. This model is also a Paddle `paddle.nn.Layer `__ subclass. Use it as a regular Paddle Layer and refer to the Paddle documentation for all matter related to general usage and behavior. Args: config (:class:`BlipConfig`): An instance of BlipConfig used to construct BlipForQuestionAnswering. """ config_class = BlipConfig _keys_to_ignore_on_load_missing = [r"text_decoder.cls.predictions.decoder.bias"] def __init__(self, config: BlipConfig): super().__init__(config) self.vision_model = BlipVisionModel(config.vision_config) self.text_encoder = BlipTextModel(config.text_config, add_pooling_layer=False) self.text_decoder = BlipTextLMHeadModel(config.text_config) self.decoder_pad_token_id = config.text_config.pad_token_id self.decoder_bos_token_id = config.text_config.bos_token_id # Initialize weights and apply final processing self.init_weights() def get_input_embeddings(self) -> nn.Layer: return self.vision_model.embeddings.patch_embedding def forward( self, input_ids: paddle.Tensor, pixel_values: paddle.Tensor, decoder_input_ids: Optional[paddle.Tensor] = None, decoder_attention_mask: Optional[paddle.Tensor] = None, attention_mask: Optional[paddle.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, labels: Optional[paddle.Tensor] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, BlipTextVisionModelOutput]: r""" Args: input_ids (Tensor): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Its data type should be `int64` and it has a shape of [text_batch_size, sequence_length]. pixel_values (Tensor): Pixel values. Padding will be ignored by default should you provide it. Its data type should be `float32` and it has a shape of [image_batch_size, num_channels, height, width]. decoder_input_ids (Tensor, optional): Indices of decoder input sequence tokens in the vocabulary. Its data type should be `int64` and it has a shape of [batch_size, sequence_length]. Defaults to `None`, which means no `decoder_input_ids` is provided, the model will create the tensor by shifting the `input_ids` to the right. decoder_attention_mask (Tensor, optional): Mask used in multi-head attention to avoid performing attention to some unwanted positions in `decoder_input_ids`. Its data type and shape is the same as `attention_mask`. Defaults to `None`. attention_mask (Tensor, optional): Mask used in multi-head attention to avoid performing attention on to some unwanted positions, usually the paddings or the subsequent positions. Its data type can be int, float and bool. When the data type is bool, the `masked` tokens have `False` values and the others have `True` values. When the data type is int, the `masked` tokens have `0` values and the others have `1` values. When the data type is float, the `masked` tokens have `0.0` values and the others have `1.0` values. It is a tensor with shape `[batch_size, sequence_length`. Defaults to `None`, which means nothing needed to be prevented attention to. output_attentions (bool, optional): Whether to return the attentions tensors of all attention layers. Defaults to `False`. output_hidden_states (bool, optional): Whether to return the hidden states of all layers. Defaults to `False`. return_dict (`bool`, *optional*): Whether or not to return a [`BlipTextVisionModelOutput`] instead of a plain tuple. Examples: ```python >>> from PIL import Image >>> import requests >>> from paddlenlp.transformers import BlipProcessor, BlipForQuestionAnswering >>> model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base") >>> model.eval() >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> text = "How many cats are in the picture?" >>> inputs = processor(images=image, text=text, return_tensors="pd") >>> outputs = model(**inputs) ```""" return_dict = return_dict if return_dict is not None else self.config.use_return_dict batch_size = input_ids.shape[0] vision_outputs = self.vision_model( pixel_values=pixel_values, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) image_embeds = vision_outputs[0] image_attention_mask = paddle.ones(image_embeds.shape[:-1], dtype=paddle.int64) question_embeds = self.text_encoder( input_ids=input_ids, attention_mask=attention_mask, encoder_hidden_states=image_embeds, encoder_attention_mask=image_attention_mask, return_dict=return_dict, ) question_embeds = question_embeds[0] if not return_dict else question_embeds.last_hidden_state if decoder_input_ids is None: # (TODO, junnyu) [batch_size, 2] decoder_input_ids = paddle.to_tensor([self.decoder_bos_token_id]).tile((batch_size, 2)) if labels is None: labels = paddle.where( decoder_input_ids == self.decoder_pad_token_id, paddle.to_tensor(-100), decoder_input_ids ) answer_output = self.text_decoder( input_ids=decoder_input_ids, attention_mask=decoder_attention_mask, encoder_hidden_states=question_embeds, encoder_attention_mask=attention_mask, labels=labels, return_dict=return_dict, reduction="none", ) decoder_loss = answer_output.loss.mean() if return_dict else answer_output[0].mean() if not return_dict: outputs = (decoder_loss, image_embeds, vision_outputs[0]) + vision_outputs[2:] return tuple(output for output in outputs if output is not None) return BlipTextVisionModelOutput( loss=decoder_loss, image_embeds=image_embeds, last_hidden_state=vision_outputs.last_hidden_state, hidden_states=vision_outputs.hidden_states, attentions=vision_outputs.attentions, ) @paddle.no_grad() def generate( self, input_ids: paddle.Tensor, pixel_values: paddle.Tensor, attention_mask: Optional[paddle.Tensor] = None, **generate_kwargs ) -> paddle.Tensor: r""" Overrides *generate* function to be able to use the model as a conditional generator Args: input_ids (*paddle.Tensor* of shape *(batch_size, sequence_length)*): The sequence used as a prompt for the generation. pixel_values (*paddle.Tensor* of shape *(batch_size, image_width, image_height)*: Input image to be processed attention_mask (*paddle.Tensor* of shape *(batch_size, sequence_length)*, *optional*): 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. **generate_kwargs: Additional arguments passed to the *generate* function of the decoder Examples: ```python >>> from PIL import Image >>> import requests >>> from paddlenlp.transformers import BlipProcessor, BlipForQuestionAnswering >>> model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base") >>> model.eval() >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> text = "How many cats are in the picture?" >>> inputs = processor(images=image, text=text, return_tensors="pd") >>> outputs = model.generate(**inputs) >>> print(processor.decode(outputs[0], skip_special_tokens=True)) 2 ``` """ vision_outputs = self.vision_model( pixel_values=pixel_values, ) image_embeds = vision_outputs[0] image_attention_mask = paddle.ones(image_embeds.shape[:-1], dtype=paddle.int64) if isinstance(input_ids, list): input_ids = paddle.to_tensor(input_ids) question_outputs = self.text_encoder( input_ids=input_ids, attention_mask=attention_mask, encoder_hidden_states=image_embeds, encoder_attention_mask=image_attention_mask, ) question_embeds = question_outputs[0] question_attention_mask = paddle.ones(question_embeds.shape[:-1], dtype=paddle.int64) bos_ids = paddle.full((question_embeds.shape[0], 1), fill_value=self.decoder_bos_token_id, dtype=paddle.int64) outputs = self.text_decoder.generate( input_ids=bos_ids, eos_token_id=self.config.text_config.sep_token_id, pad_token_id=self.config.text_config.pad_token_id, encoder_hidden_states=question_embeds, encoder_attention_mask=question_attention_mask, **generate_kwargs, ) return outputs class BlipForImageTextRetrieval(BlipPretrainedModel): r""" BLIP Model with a vision and text projector, and a classification head on top. The model is used in the context of image-text retrieval. Given an image and a text, the model returns the probability of the text being relevant to the image. This model inherits from :class:`~paddlenlp.transformers.model_utils.PretrainedModel`. Refer to the superclass documentation for the generic methods. This model is also a Paddle `paddle.nn.Layer `__ subclass. Use it as a regular Paddle Layer and refer to the Paddle documentation for all matter related to general usage and behavior. Args: config (:class:`BlipConfig`): An instance of BlipConfig used to construct BlipForImageTextRetrieval. """ config_class = BlipConfig def __init__(self, config: BlipConfig): super().__init__(config) self.vision_model = BlipVisionModel(config.vision_config) self.text_encoder = BlipTextModel(config.text_config, add_pooling_layer=False) # vision projection layer self.vision_proj = nn.Linear(config.vision_config.hidden_size, config.image_text_hidden_size) # text projection layer self.text_proj = nn.Linear(config.text_config.hidden_size, config.image_text_hidden_size) # image text matching head self.itm_head = nn.Linear(config.text_config.hidden_size, 2) self.decoder_pad_token_id = config.text_config.pad_token_id self.decoder_bos_token_id = config.text_config.bos_token_id # Initialize weights and apply final processing self.init_weights() def get_input_embeddings(self) -> nn.Layer: return self.vision_model.embeddings.patch_embedding def forward( self, input_ids: paddle.Tensor, pixel_values: paddle.Tensor, use_itm_head: Optional[bool] = True, attention_mask: Optional[paddle.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, BlipTextVisionModelOutput]: r""" Args: input_ids (Tensor): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Its data type should be `int64` and it has a shape of [text_batch_size, sequence_length]. pixel_values (Tensor): Pixel values. Padding will be ignored by default should you provide it. Its data type should be `float32` and it has a shape of [image_batch_size, num_channels, height, width]. use_itm_head (bool, optional): Whether to use itm head. Defaults to `True`. attention_mask (Tensor, optional): Mask used in multi-head attention to avoid performing attention on to some unwanted positions, usually the paddings or the subsequent positions. Its data type can be int, float and bool. When the data type is bool, the `masked` tokens have `False` values and the others have `True` values. When the data type is int, the `masked` tokens have `0` values and the others have `1` values. When the data type is float, the `masked` tokens have `0.0` values and the others have `1.0` values. It is a tensor with shape `[batch_size, sequence_length`. Defaults to `None`, which means nothing needed to be prevented attention to. output_attentions (bool, optional): Whether to return the attentions tensors of all attention layers. Defaults to `False`. output_hidden_states (bool, optional): Whether to return the hidden states of all layers. Defaults to `False`. return_dict (`bool`, *optional*): Whether or not to return a [`BlipTextVisionModelOutput`] instead of a plain tuple. Examples: ```python >>> from PIL import Image >>> import requests >>> from paddlenlp.transformers import BlipProcessor, BlipForImageTextRetrieval >>> model = BlipForImageTextRetrieval.from_pretrained("Salesforce/blip-itm-base") >>> model.eval() >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-itm-base") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> text = "an image of a cat" >>> inputs = processor(images=image, text=text, return_tensors="pd") >>> outputs = model(**inputs) ``` """ return_dict = return_dict if return_dict is not None else self.config.use_return_dict vision_outputs = self.vision_model( pixel_values=pixel_values, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) image_embeds = vision_outputs[0] image_atts = paddle.ones(image_embeds.shape[:-1], dtype=paddle.int64) if use_itm_head: question_embeds = self.text_encoder( input_ids=input_ids, attention_mask=attention_mask, encoder_hidden_states=image_embeds, encoder_attention_mask=image_atts, return_dict=return_dict, ) question_embeds = question_embeds[0] if not return_dict else question_embeds.last_hidden_state output = self.itm_head(question_embeds[:, 0, :]) else: question_embeds = self.text_encoder( input_ids=input_ids, attention_mask=attention_mask, return_dict=return_dict, ) question_embeds = question_embeds[0] if not return_dict else question_embeds.last_hidden_state image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), axis=-1) text_feat = F.normalize(self.text_proj(question_embeds[:, 0, :]), axis=-1) output = paddle.matmul(image_feat, text_feat, transpose_y=True) if not return_dict: outputs = (output, vision_outputs[0]) + vision_outputs[2:] + (question_embeds,) return tuple(output for output in outputs if output is not None) return BlipImageTextMatchingModelOutput( itm_score=output, last_hidden_state=vision_outputs.last_hidden_state, hidden_states=vision_outputs.hidden_states, attentions=vision_outputs.attentions, question_embeds=question_embeds, )