# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # This file was automatically generated from src/transformers/models/metaclip_2/modular_metaclip_2.py. # Do NOT edit this file manually as any edits will be overwritten by the generation of # the file from the modular. If any change should be done, please apply the change to the # modular_metaclip_2.py file directly. One of our CI enforces this. # 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 from dataclasses import dataclass from typing import Any, Callable, Optional, Union import torch from torch import nn from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss from ...activations import ACT2FN from ...modeling_attn_mask_utils import _create_4d_causal_attention_mask, _prepare_4d_attention_mask from ...modeling_layers import GradientCheckpointingLayer from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, ImageClassifierOutput from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel from ...processing_utils import Unpack from ...utils import ModelOutput, TransformersKwargs, auto_docstring, can_return_tuple, logging, torch_int from ...utils.generic import check_model_inputs from .configuration_metaclip_2 import MetaClip2Config, MetaClip2TextConfig, MetaClip2VisionConfig logger = logging.get_logger(__name__) class MetaClip2TextEmbeddings(nn.Module): def __init__(self, config: MetaClip2TextConfig): 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", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False ) def forward( self, input_ids: Optional[torch.LongTensor] = None, position_ids: Optional[torch.LongTensor] = None, inputs_embeds: Optional[torch.FloatTensor] = None, ) -> torch.Tensor: seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2] max_position_embedding = self.position_embedding.weight.shape[0] if seq_length > max_position_embedding: raise ValueError( f"Sequence length must be less than max_position_embeddings (got `sequence length`: " f"{seq_length} and max_position_embeddings: {max_position_embedding}" ) 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 MetaClip2VisionEmbeddings(nn.Module): def __init__(self, config: MetaClip2VisionConfig): 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 = nn.Parameter(torch.randn(self.embed_dim)) self.patch_embedding = nn.Conv2d( in_channels=config.num_channels, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size, bias=False, ) self.num_patches = (self.image_size // self.patch_size) ** 2 self.num_positions = self.num_patches + 1 self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim) self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False) def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int) -> torch.Tensor: """ This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution images. This method is also adapted to support torch.jit tracing. Adapted from: - https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174-L194, and - https://github.com/facebookresearch/dinov2/blob/e1277af2ba9496fbadf7aec6eba56e8d882d1e35/dinov2/models/vision_transformer.py#L179-L211 """ num_patches = embeddings.shape[1] - 1 position_embedding = self.position_embedding.weight.unsqueeze(0) num_positions = position_embedding.shape[1] - 1 # always interpolate when tracing to ensure the exported model works for dynamic input shapes if not torch.jit.is_tracing() and num_patches == num_positions and height == width: return self.position_embedding(self.position_ids) class_pos_embed = position_embedding[:, :1] patch_pos_embed = position_embedding[:, 1:] dim = embeddings.shape[-1] new_height = height // self.patch_size new_width = width // self.patch_size sqrt_num_positions = torch_int(num_positions**0.5) patch_pos_embed = patch_pos_embed.reshape(1, sqrt_num_positions, sqrt_num_positions, dim) patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2) patch_pos_embed = nn.functional.interpolate( patch_pos_embed, size=(new_height, new_width), mode="bicubic", align_corners=False, ) patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim) return torch.cat((class_pos_embed, patch_pos_embed), dim=1) def forward(self, pixel_values: torch.FloatTensor, interpolate_pos_encoding=False) -> torch.Tensor: batch_size, _, height, width = pixel_values.shape if not interpolate_pos_encoding and (height != self.image_size or width != self.image_size): raise ValueError( f"Input image size ({height}*{width}) doesn't match model ({self.image_size}*{self.image_size})." ) target_dtype = self.patch_embedding.weight.dtype patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype)) # shape = [*, width, grid, grid] patch_embeds = patch_embeds.flatten(2).transpose(1, 2) class_embeds = self.class_embedding.expand(batch_size, 1, -1) embeddings = torch.cat([class_embeds, patch_embeds], dim=1) if interpolate_pos_encoding: embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width) else: embeddings = embeddings + self.position_embedding(self.position_ids) return embeddings def eager_attention_forward( module: nn.Module, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, attention_mask: Optional[torch.Tensor], scaling: float, dropout: float = 0.0, output_attentions: bool = True, **kwargs, ): attn_weights = torch.matmul(query, key.transpose(-1, -2)) * scaling if attention_mask is not None: attn_weights = attn_weights + attention_mask attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) attn_output = torch.matmul(attn_weights, value) attn_output = attn_output.transpose(1, 2).contiguous() if not output_attentions: attn_weights = None return attn_output, attn_weights class MetaClip2Attention(nn.Module): """Multi-headed attention from 'Attention Is All You Need' paper""" def __init__(self, config: Union[MetaClip2VisionConfig, MetaClip2TextConfig]): 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 = config.attention_dropout self.is_causal = False self.k_proj = nn.Linear(self.embed_dim, self.embed_dim) self.v_proj = nn.Linear(self.embed_dim, self.embed_dim) self.q_proj = nn.Linear(self.embed_dim, self.embed_dim) self.out_proj = nn.Linear(self.embed_dim, self.embed_dim) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, causal_attention_mask: Optional[torch.Tensor] = None, output_attentions: Optional[bool] = False, ) -> tuple[torch.Tensor, Optional[torch.Tensor]]: """Input shape: Batch x Time x Channel""" batch_size, seq_length, embed_dim = hidden_states.shape queries = self.q_proj(hidden_states) keys = self.k_proj(hidden_states) values = self.v_proj(hidden_states) queries = queries.view(batch_size, seq_length, -1, self.head_dim).transpose(1, 2) keys = keys.view(batch_size, seq_length, -1, self.head_dim).transpose(1, 2) values = values.view(batch_size, seq_length, -1, self.head_dim).transpose(1, 2) # METACLIP_2 text model uses both `causal_attention_mask` and `attention_mask` # in case FA2 kernel is called, `is_causal` should be inferred from `causal_attention_mask` if self.config._attn_implementation == "flash_attention_2": self.is_causal = causal_attention_mask is not None else: if attention_mask is not None and causal_attention_mask is not None: attention_mask = attention_mask + causal_attention_mask elif causal_attention_mask is not None: attention_mask = causal_attention_mask attention_interface: Callable = eager_attention_forward if self.config._attn_implementation != "eager": if self.config._attn_implementation == "sdpa" and output_attentions: logger.warning_once( "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to " 'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' ) else: attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation] attn_output, attn_weights = attention_interface( self, queries, keys, values, attention_mask, is_causal=self.is_causal, scaling=self.scale, dropout=0.0 if not self.training else self.dropout, output_attentions=output_attentions, ) attn_output = attn_output.reshape(batch_size, seq_length, embed_dim).contiguous() attn_output = self.out_proj(attn_output) if not output_attentions: attn_weights = None return attn_output, attn_weights class MetaClip2MLP(nn.Module): def __init__(self, config): 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: torch.Tensor) -> torch.Tensor: hidden_states = self.fc1(hidden_states) hidden_states = self.activation_fn(hidden_states) hidden_states = self.fc2(hidden_states) return hidden_states @auto_docstring class MetaClip2PreTrainedModel(PreTrainedModel): config: MetaClip2Config base_model_prefix = "metaclip_2" supports_gradient_checkpointing = True _supports_sdpa = True _supports_flash_attn = True _supports_flex_attn = True _supports_attention_backend = True def _init_weights(self, module): """Initialize the weights""" factor = self.config.initializer_factor if isinstance(module, MetaClip2TextEmbeddings): module.token_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02) module.position_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02) elif isinstance(module, MetaClip2VisionEmbeddings): factor = self.config.initializer_factor nn.init.normal_(module.class_embedding, mean=0.0, std=module.embed_dim**-0.5 * factor) nn.init.normal_(module.patch_embedding.weight, std=module.config.initializer_range * factor) nn.init.normal_(module.position_embedding.weight, std=module.config.initializer_range * factor) elif isinstance(module, MetaClip2Attention): factor = self.config.initializer_factor in_proj_std = (module.embed_dim**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor out_proj_std = (module.embed_dim**-0.5) * factor nn.init.normal_(module.q_proj.weight, std=in_proj_std) nn.init.normal_(module.k_proj.weight, std=in_proj_std) nn.init.normal_(module.v_proj.weight, std=in_proj_std) nn.init.normal_(module.out_proj.weight, std=out_proj_std) elif isinstance(module, MetaClip2MLP): factor = self.config.initializer_factor in_proj_std = (module.config.hidden_size**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor fc_std = (2 * module.config.hidden_size) ** -0.5 * factor nn.init.normal_(module.fc1.weight, std=fc_std) nn.init.normal_(module.fc2.weight, std=in_proj_std) elif isinstance(module, MetaClip2Model): nn.init.normal_( module.text_projection.weight, std=module.text_embed_dim**-0.5 * self.config.initializer_factor, ) nn.init.normal_( module.visual_projection.weight, std=module.vision_embed_dim**-0.5 * self.config.initializer_factor, ) elif isinstance(module, MetaClip2VisionModelWithProjection): nn.init.normal_( module.visual_projection.weight, std=self.config.hidden_size**-0.5 * self.config.initializer_factor, ) elif isinstance(module, MetaClip2TextModelWithProjection): nn.init.normal_( module.text_projection.weight, std=self.config.hidden_size**-0.5 * self.config.initializer_factor, ) elif isinstance(module, MetaClip2ForImageClassification): nn.init.normal_( module.classifier.weight, std=self.config.vision_config.hidden_size**-0.5 * self.config.initializer_factor, ) if isinstance(module, nn.LayerNorm): module.bias.data.zero_() module.weight.data.fill_(1.0) if isinstance(module, nn.Linear) and module.bias is not None: module.bias.data.zero_() class MetaClip2EncoderLayer(GradientCheckpointingLayer): def __init__(self, config: Union[MetaClip2VisionConfig, MetaClip2TextConfig]): super().__init__() self.embed_dim = config.hidden_size self.self_attn = MetaClip2Attention(config) self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps) self.mlp = MetaClip2MLP(config) self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps) def forward( self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, causal_attention_mask: torch.Tensor, output_attentions: Optional[bool] = False, ) -> tuple[torch.FloatTensor]: """ Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`): 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, causal_attention_mask=causal_attention_mask, output_attentions=output_attentions, ) hidden_states = residual + hidden_states residual = hidden_states hidden_states = self.layer_norm2(hidden_states) hidden_states = self.mlp(hidden_states) hidden_states = residual + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (attn_weights,) return outputs class MetaClip2Encoder(nn.Module): """ Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a [`MetaClip2EncoderLayer`]. Args: config: MetaClip2Config """ def __init__(self, config: MetaClip2Config): super().__init__() self.config = config self.layers = nn.ModuleList([MetaClip2EncoderLayer(config) for _ in range(config.num_hidden_layers)]) self.gradient_checkpointing = False def forward( self, inputs_embeds, attention_mask: Optional[torch.Tensor] = None, causal_attention_mask: Optional[torch.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, ) -> BaseModelOutput: r""" Args: inputs_embeds (`torch.FloatTensor` 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 (`torch.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**. [What are attention masks?](../glossary#attention-mask) causal_attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Causal mask for the text model. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) 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 [`~utils.ModelOutput`] 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 ) encoder_states = () if output_hidden_states else None all_attentions = () if output_attentions else None hidden_states = inputs_embeds for idx, encoder_layer in enumerate(self.layers): if output_hidden_states: encoder_states = encoder_states + (hidden_states,) layer_outputs = encoder_layer( hidden_states, attention_mask, causal_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,) return BaseModelOutput( last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions, ) class MetaClip2TextTransformer(nn.Module): def __init__(self, config: MetaClip2TextConfig): super().__init__() self.config = config embed_dim = config.hidden_size self.embeddings = MetaClip2TextEmbeddings(config) self.encoder = MetaClip2Encoder(config) self.final_layer_norm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) # For `pooled_output` computation self.eos_token_id = config.eos_token_id # For attention mask, it differs between `flash_attention_2` and other attention implementations self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2" @check_model_inputs @auto_docstring def forward( self, input_ids, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.Tensor] = None, use_cache: Optional[bool] = None, **kwargs: Unpack[TransformersKwargs], ) -> BaseModelOutputWithPooling: input_shape = input_ids.size() input_ids = input_ids.view(-1, input_shape[-1]) hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids) # CLIP's text model uses causal mask, prepare it here. # https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324 causal_attention_mask = _create_4d_causal_attention_mask( input_shape, hidden_states.dtype, device=hidden_states.device ) # expand attention_mask if attention_mask is not None and not self._use_flash_attention_2: # [batch_size, seq_len] -> [batch_size, 1, tgt_seq_len, src_seq_len] attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_states.dtype) encoder_outputs: BaseModelOutput = self.encoder( inputs_embeds=hidden_states, attention_mask=attention_mask, causal_attention_mask=causal_attention_mask, **kwargs, ) last_hidden_state = encoder_outputs.last_hidden_state last_hidden_state = self.final_layer_norm(last_hidden_state) # Use robust pooling like CLIP - finds the first EOS token position per sequence pooled_output = last_hidden_state[ torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device), (input_ids.to(dtype=torch.int, device=last_hidden_state.device) == self.eos_token_id).int().argmax(dim=-1), ] return BaseModelOutputWithPooling( last_hidden_state=last_hidden_state, pooler_output=pooled_output, hidden_states=encoder_outputs.hidden_states, attentions=encoder_outputs.attentions, ) @auto_docstring( custom_intro=""" The text model from METACLIP_2 without any head or projection on top. """ ) class MetaClip2TextModel(MetaClip2PreTrainedModel): config: MetaClip2TextConfig _no_split_modules = ["MetaClip2TextEmbeddings", "MetaClip2EncoderLayer"] _supports_flash_attn = False # mask creation only accounts for sdpa/eager def __init__(self, config: MetaClip2TextConfig): super().__init__(config) self.text_model = MetaClip2TextTransformer(config) # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self) -> nn.Module: return self.text_model.embeddings.token_embedding def set_input_embeddings(self, value): self.text_model.embeddings.token_embedding = value @can_return_tuple @auto_docstring def forward( self, input_ids: Optional[torch.Tensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, ) -> BaseModelOutputWithPooling: r""" Examples: ```python >>> from transformers import AutoTokenizer, MetaClip2TextModel >>> model = MetaClip2TextModel.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> tokenizer = AutoTokenizer.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt") >>> outputs = model(**inputs) >>> last_hidden_state = outputs.last_hidden_state >>> pooled_output = outputs.pooler_output # pooled (EOS token) states ```""" return 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, ) @dataclass @auto_docstring( custom_intro=""" Base class for text model's outputs that also contains a pooling of the last hidden states. """ ) class MetaClip2TextModelOutput(ModelOutput): r""" text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`): The text embeddings obtained by applying the projection layer to the pooler_output. """ text_embeds: Optional[torch.FloatTensor] = None last_hidden_state: Optional[torch.FloatTensor] = None hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None attentions: Optional[tuple[torch.FloatTensor, ...]] = None @auto_docstring class MetaClip2TextModelWithProjection(MetaClip2PreTrainedModel): config: MetaClip2TextConfig _no_split_modules = ["MetaClip2TextEmbeddings", "MetaClip2EncoderLayer"] def __init__(self, config: MetaClip2TextConfig): super().__init__(config) text_model = MetaClip2TextModel._from_config(config) self.text_model = text_model.text_model self.text_projection = nn.Linear(config.hidden_size, config.projection_dim, bias=False) # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self) -> nn.Module: return self.text_model.embeddings.token_embedding def set_input_embeddings(self, value): self.text_model.embeddings.token_embedding = value @can_return_tuple @auto_docstring def forward( self, input_ids: Optional[torch.Tensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, ) -> MetaClip2TextModelOutput: r""" Examples: ```python >>> from transformers import AutoTokenizer, MetaClip2TextModelWithProjection >>> model = MetaClip2TextModelWithProjection.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> tokenizer = AutoTokenizer.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt") >>> outputs = model(**inputs) >>> text_embeds = outputs.text_embeds ```""" text_outputs: BaseModelOutputWithPooling = 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, ) pooled_output = text_outputs.pooler_output text_embeds = self.text_projection(pooled_output) return MetaClip2TextModelOutput( text_embeds=text_embeds, last_hidden_state=text_outputs.last_hidden_state, hidden_states=text_outputs.hidden_states, attentions=text_outputs.attentions, ) @dataclass @auto_docstring class MetaClip2Output(ModelOutput): r""" loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`): Contrastive loss for image-text similarity. logits_per_image (`torch.FloatTensor` 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 (`torch.FloatTensor` 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 (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by applying the projection layer to the pooled output of [`MetaClip2TextModel`]. image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by applying the projection layer to the pooled output of [`MetaClip2VisionModel`]. text_model_output (`BaseModelOutputWithPooling`): The output of the [`MetaClip2TextModel`]. vision_model_output (`BaseModelOutputWithPooling`): The output of the [`MetaClip2VisionModel`]. """ loss: Optional[torch.FloatTensor] = None logits_per_image: Optional[torch.FloatTensor] = None logits_per_text: Optional[torch.FloatTensor] = None text_embeds: Optional[torch.FloatTensor] = None image_embeds: Optional[torch.FloatTensor] = 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() ) # contrastive loss function, adapted from # https://sachinruk.github.io/blog/2021-03-07-metaclip_2.html def contrastive_loss(logits: torch.Tensor) -> torch.Tensor: return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device)) def metaclip_2_loss(similarity: torch.Tensor) -> torch.Tensor: caption_loss = contrastive_loss(similarity) image_loss = contrastive_loss(similarity.t()) return (caption_loss + image_loss) / 2.0 def _get_vector_norm(tensor: torch.Tensor) -> torch.Tensor: """ This method is equivalent to tensor.norm(p=2, dim=-1, keepdim=True) and used to make model `executorch` exportable. See issue https://github.com/pytorch/executorch/issues/3566 """ square_tensor = torch.pow(tensor, 2) sum_tensor = torch.sum(square_tensor, dim=-1, keepdim=True) normed_tensor = torch.pow(sum_tensor, 0.5) return normed_tensor @auto_docstring class MetaClip2Model(MetaClip2PreTrainedModel): config: MetaClip2Config _no_split_modules = ["MetaClip2TextEmbeddings", "MetaClip2EncoderLayer", "MetaClip2VisionEmbeddings"] _supports_flash_attn = False # mask creation only accounts for sdpa/eager def __init__(self, config: MetaClip2Config): super().__init__(config) if not isinstance(config.text_config, MetaClip2TextConfig): raise TypeError( "config.text_config is expected to be of type MetaClip2TextConfig but is of type" f" {type(config.text_config)}." ) if not isinstance(config.vision_config, MetaClip2VisionConfig): raise TypeError( "config.vision_config is expected to be of type MetaClip2VisionConfig 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 text_model = MetaClip2TextModel._from_config(text_config) self.text_model = text_model.text_model vision_model = MetaClip2VisionModel._from_config(vision_config) self.vision_model = vision_model.vision_model self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias=False) self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias=False) self.logit_scale = nn.Parameter(torch.tensor(self.config.logit_scale_init_value)) # Initialize weights and apply final processing self.post_init() @auto_docstring def get_text_features( self, input_ids: Optional[torch.Tensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, ) -> torch.FloatTensor: r""" Returns: text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by applying the projection layer to the pooled output of [`MetaClip2TextModel`]. Examples: ```python >>> from transformers import AutoTokenizer, MetaClip2Model >>> model = MetaClip2Model.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> tokenizer = AutoTokenizer.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt") >>> text_features = model.get_text_features(**inputs) ```""" # Use METACLIP_2 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 ) text_outputs: BaseModelOutputWithPooling = 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, ) pooled_output = text_outputs.pooler_output text_features = self.text_projection(pooled_output) return text_features @auto_docstring def get_image_features( self, pixel_values: Optional[torch.FloatTensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, interpolate_pos_encoding: bool = False, ) -> torch.FloatTensor: r""" Returns: image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by applying the projection layer to the pooled output of [`MetaClip2VisionModel`]. Examples: ```python >>> from PIL import Image >>> import requests >>> from transformers import AutoProcessor, MetaClip2Model >>> model = MetaClip2Model.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> processor = AutoProcessor.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor(images=image, return_tensors="pt") >>> image_features = model.get_image_features(**inputs) ```""" # Use METACLIP_2 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 ) vision_outputs: BaseModelOutputWithPooling = self.vision_model( pixel_values=pixel_values, output_attentions=output_attentions, output_hidden_states=output_hidden_states, interpolate_pos_encoding=interpolate_pos_encoding, ) pooled_output = vision_outputs.pooler_output image_features = self.visual_projection(pooled_output) return image_features @can_return_tuple @auto_docstring def forward( self, input_ids: Optional[torch.LongTensor] = None, pixel_values: Optional[torch.FloatTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, return_loss: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, interpolate_pos_encoding: bool = False, ) -> MetaClip2Output: r""" return_loss (`bool`, *optional*): Whether or not to return the contrastive loss. Examples: ```python >>> from PIL import Image >>> import requests >>> from transformers import AutoProcessor, MetaClip2Model >>> model = MetaClip2Model.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> processor = AutoProcessor.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> 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="pt", padding=True ... ) >>> outputs = model(**inputs) >>> logits_per_image = outputs.logits_per_image # this is the image-text similarity score >>> probs = logits_per_image.softmax(dim=1) # we can take the softmax to get the label probabilities ```""" # Use METACLIP_2 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 ) vision_outputs: BaseModelOutputWithPooling = self.vision_model( pixel_values=pixel_values, output_attentions=output_attentions, output_hidden_states=output_hidden_states, interpolate_pos_encoding=interpolate_pos_encoding, ) text_outputs: BaseModelOutputWithPooling = 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, ) image_embeds = vision_outputs.pooler_output image_embeds = self.visual_projection(image_embeds) text_embeds = text_outputs.pooler_output text_embeds = self.text_projection(text_embeds) # normalized features image_embeds = image_embeds / _get_vector_norm(image_embeds) text_embeds = text_embeds / _get_vector_norm(text_embeds) # cosine similarity as logits logits_per_text = torch.matmul(text_embeds, image_embeds.t().to(text_embeds.device)) logits_per_text = logits_per_text * self.logit_scale.exp().to(text_embeds.device) logits_per_image = logits_per_text.t() loss = None if return_loss: loss = metaclip_2_loss(logits_per_text) return MetaClip2Output( 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 MetaClip2VisionTransformer(nn.Module): def __init__(self, config: MetaClip2VisionConfig): super().__init__() self.config = config embed_dim = config.hidden_size self.embeddings = MetaClip2VisionEmbeddings(config) self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) self.encoder = MetaClip2Encoder(config) self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) @auto_docstring def forward( self, pixel_values: Optional[torch.FloatTensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, interpolate_pos_encoding: Optional[bool] = False, ) -> BaseModelOutputWithPooling: 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 ) if pixel_values is None: raise ValueError("You have to specify pixel_values") hidden_states = self.embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding) hidden_states = self.pre_layrnorm(hidden_states) encoder_outputs: BaseModelOutput = self.encoder( inputs_embeds=hidden_states, output_attentions=output_attentions, output_hidden_states=output_hidden_states, ) last_hidden_state = encoder_outputs.last_hidden_state pooled_output = last_hidden_state[:, 0, :] pooled_output = self.post_layernorm(pooled_output) return BaseModelOutputWithPooling( last_hidden_state=last_hidden_state, pooler_output=pooled_output, hidden_states=encoder_outputs.hidden_states, attentions=encoder_outputs.attentions, ) @auto_docstring( custom_intro=""" The vision model from METACLIP_2 without any head or projection on top. """ ) class MetaClip2VisionModel(MetaClip2PreTrainedModel): config: MetaClip2VisionConfig main_input_name = "pixel_values" _no_split_modules = ["MetaClip2EncoderLayer"] def __init__(self, config: MetaClip2VisionConfig): super().__init__(config) self.vision_model = MetaClip2VisionTransformer(config) # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self) -> nn.Module: return self.vision_model.embeddings.patch_embedding @can_return_tuple @auto_docstring def forward( self, pixel_values: Optional[torch.FloatTensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, interpolate_pos_encoding: bool = False, ) -> BaseModelOutputWithPooling: r""" Example: ```python >>> from PIL import Image >>> import requests >>> from transformers import AutoProcessor, MetaClip2VisionModel >>> model = MetaClip2VisionModel.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> processor = AutoProcessor.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor(images=image, return_tensors="pt") >>> outputs = model(**inputs) >>> last_hidden_state = outputs.last_hidden_state >>> pooled_output = outputs.pooler_output # pooled CLS states ```""" return self.vision_model( pixel_values=pixel_values, output_attentions=output_attentions, output_hidden_states=output_hidden_states, interpolate_pos_encoding=interpolate_pos_encoding, ) @dataclass @auto_docstring( custom_intro=""" Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states. """ ) class MetaClip2VisionModelOutput(ModelOutput): r""" image_embeds (`torch.FloatTensor` 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. """ image_embeds: Optional[torch.FloatTensor] = None last_hidden_state: Optional[torch.FloatTensor] = None hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None attentions: Optional[tuple[torch.FloatTensor, ...]] = None @auto_docstring class MetaClip2VisionModelWithProjection(MetaClip2PreTrainedModel): config: MetaClip2VisionConfig main_input_name = "pixel_values" def __init__(self, config: MetaClip2VisionConfig): super().__init__(config) vision_model = MetaClip2VisionModel._from_config(config) self.vision_model = vision_model.vision_model self.visual_projection = nn.Linear(config.hidden_size, config.projection_dim, bias=False) # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self) -> nn.Module: return self.vision_model.embeddings.patch_embedding @can_return_tuple @auto_docstring def forward( self, pixel_values: Optional[torch.FloatTensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, interpolate_pos_encoding: bool = False, ) -> MetaClip2VisionModelOutput: r""" Examples: ```python >>> from PIL import Image >>> import requests >>> from transformers import AutoProcessor, MetaClip2VisionModelWithProjection >>> model = MetaClip2VisionModelWithProjection.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> processor = AutoProcessor.from_pretrained("openai/metaclip_2-vit-base-patch32") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor(images=image, return_tensors="pt") >>> outputs = model(**inputs) >>> image_embeds = outputs.image_embeds ```""" vision_outputs: BaseModelOutputWithPooling = self.vision_model( pixel_values=pixel_values, output_attentions=output_attentions, output_hidden_states=output_hidden_states, interpolate_pos_encoding=interpolate_pos_encoding, ) pooled_output = vision_outputs.pooler_output image_embeds = self.visual_projection(pooled_output) return MetaClip2VisionModelOutput( image_embeds=image_embeds, last_hidden_state=vision_outputs.last_hidden_state, hidden_states=vision_outputs.hidden_states, attentions=vision_outputs.attentions, ) @auto_docstring( custom_intro=""" METACLIP_2 vision encoder with an image classification head on top (a linear layer on top of the pooled final hidden states of the patch tokens) e.g. for ImageNet. """ ) class MetaClip2ForImageClassification(MetaClip2PreTrainedModel): main_input_name = "pixel_values" def __init__(self, config: MetaClip2Config) -> None: super().__init__(config) self.num_labels = config.num_labels vision_model = MetaClip2VisionModel._from_config(config.vision_config) self.vision_model = vision_model.vision_model # Classifier head self.classifier = ( nn.Linear(config.vision_config.hidden_size, config.num_labels) if config.num_labels > 0 else nn.Identity() ) # Initialize weights and apply final processing self.post_init() @can_return_tuple @auto_docstring def forward( self, pixel_values: Optional[torch.Tensor] = None, labels: Optional[torch.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, ) -> ImageClassifierOutput: r""" labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the image classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). """ 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 ) outputs: BaseModelOutputWithPooling = self.vision_model( pixel_values, output_attentions=output_attentions, output_hidden_states=output_hidden_states, ) sequence_output = outputs.last_hidden_state # average pool the patch tokens sequence_output = torch.mean(sequence_output[:, 1:, :], dim=1) # apply classifier logits = self.classifier(sequence_output) loss = None if labels is not None: # move labels to correct device to enable model parallelism labels = labels.to(logits.device) if self.config.problem_type is None: if self.num_labels == 1: self.config.problem_type = "regression" elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int): self.config.problem_type = "single_label_classification" else: self.config.problem_type = "multi_label_classification" if self.config.problem_type == "regression": loss_fct = MSELoss() if self.num_labels == 1: loss = loss_fct(logits.squeeze(), labels.squeeze()) else: loss = loss_fct(logits, labels) elif self.config.problem_type == "single_label_classification": loss_fct = CrossEntropyLoss() loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1)) elif self.config.problem_type == "multi_label_classification": loss_fct = BCEWithLogitsLoss() loss = loss_fct(logits, labels) return ImageClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) __all__ = [ "MetaClip2Model", "MetaClip2PreTrainedModel", "MetaClip2TextModel", "MetaClip2TextModelWithProjection", "MetaClip2VisionModel", "MetaClip2VisionModelWithProjection", "MetaClip2ForImageClassification", ]