# Copyright 2022 The BAAI Teams Authors and The HuggingFace Inc. 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. """PyTorch AltCLIP model.""" import math from collections.abc import Callable from dataclasses import dataclass from typing import Any import torch import torch.nn as nn from ... import initialization as init from ...activations import ACT2FN from ...modeling_layers import GradientCheckpointingLayer from ...modeling_outputs import ( BaseModelOutput, BaseModelOutputWithPooling, BaseModelOutputWithPoolingAndProjection, ) from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel from ...processing_utils import Unpack from ...pytorch_utils import apply_chunking_to_forward from ...utils import ModelOutput, TransformersKwargs, auto_docstring, can_return_tuple, logging, torch_int from ...utils.generic import merge_with_config_defaults from ...utils.output_capturing import capture_outputs from .configuration_altclip import AltCLIPConfig, AltCLIPTextConfig, AltCLIPVisionConfig logger = logging.get_logger(__name__) # contrastive loss function, adapted from # https://sachinruk.github.io/blog/pytorch/pytorch%20lightning/loss%20function/gpu/2021/03/07/CLIP.html def contrastive_loss(logits: torch.Tensor) -> torch.Tensor: return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device)) def clip_loss(similarity: torch.Tensor) -> torch.Tensor: caption_loss = contrastive_loss(similarity) image_loss = contrastive_loss(similarity.t()) return (caption_loss + image_loss) / 2.0 @dataclass @auto_docstring # Copied from transformers.models.clip.modeling_clip.CLIPOutput with CLIP->AltCLIP class AltCLIPOutput(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 [`AltCLIPTextModel`]. image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by applying the projection layer to the pooled output of [`AltCLIPVisionModel`]. text_model_output (`BaseModelOutputWithPooling`): The output of the [`AltCLIPTextModel`]. vision_model_output (`BaseModelOutputWithPooling`): The output of the [`AltCLIPVisionModel`]. """ loss: torch.FloatTensor | None = None logits_per_image: torch.FloatTensor | None = None logits_per_text: torch.FloatTensor | None = None text_embeds: torch.FloatTensor | None = None image_embeds: torch.FloatTensor | None = None text_model_output: BaseModelOutputWithPooling = None vision_model_output: BaseModelOutputWithPooling = None def to_tuple(self) -> tuple[Any]: return tuple(v.to_tuple() if isinstance(v, ModelOutput) else v for v in self.values()) # Copied from transformers.models.roberta.modeling_roberta.RobertaEmbeddings with Roberta->AltRoberta class AltRobertaEmbeddings(nn.Module): """Construct the embeddings from word, position and token_type embeddings.""" def __init__(self, config): super().__init__() self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id) self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size) self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) self.dropout = nn.Dropout(config.hidden_dropout_prob) # 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 ) self.register_buffer( "token_type_ids", torch.zeros(self.position_ids.size(), dtype=torch.long), persistent=False ) self.padding_idx = config.pad_token_id self.position_embeddings = nn.Embedding( config.max_position_embeddings, config.hidden_size, padding_idx=self.padding_idx ) def forward( self, input_ids: torch.LongTensor | None = None, token_type_ids: torch.LongTensor | None = None, position_ids: torch.LongTensor | None = None, inputs_embeds: torch.FloatTensor | None = None, past_key_values_length: int = 0, ) -> torch.Tensor: if position_ids is None: if input_ids is not None: # Create the position ids from the input token ids. Any padded tokens remain padded. position_ids = self.create_position_ids_from_input_ids( input_ids, self.padding_idx, past_key_values_length ) else: position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds, self.padding_idx) if input_ids is not None: input_shape = input_ids.size() else: input_shape = inputs_embeds.size()[:-1] batch_size, seq_length = input_shape # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves # issue #5664 if token_type_ids is None: if hasattr(self, "token_type_ids"): # NOTE: We assume either pos ids to have bsz == 1 (broadcastable) or bsz == effective bsz (input_shape[0]) buffered_token_type_ids = self.token_type_ids.expand(position_ids.shape[0], -1) buffered_token_type_ids = torch.gather(buffered_token_type_ids, dim=1, index=position_ids) token_type_ids = buffered_token_type_ids.expand(batch_size, seq_length) else: token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device) if inputs_embeds is None: inputs_embeds = self.word_embeddings(input_ids) token_type_embeddings = self.token_type_embeddings(token_type_ids) embeddings = inputs_embeds + token_type_embeddings position_embeddings = self.position_embeddings(position_ids) embeddings = embeddings + position_embeddings embeddings = self.LayerNorm(embeddings) embeddings = self.dropout(embeddings) return embeddings @staticmethod def create_position_ids_from_inputs_embeds(inputs_embeds, padding_idx): """ We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids. Args: inputs_embeds: torch.Tensor Returns: torch.Tensor """ input_shape = inputs_embeds.size()[:-1] sequence_length = input_shape[1] position_ids = torch.arange( padding_idx + 1, sequence_length + padding_idx + 1, dtype=torch.long, device=inputs_embeds.device ) return position_ids.unsqueeze(0).expand(input_shape) @staticmethod def create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_length=0): """ Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols are ignored. This is modified from fairseq's `utils.make_positions`. Args: x: torch.Tensor x: Returns: torch.Tensor """ # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA. mask = input_ids.ne(padding_idx).int() incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask return incremental_indices.long() + padding_idx class AltRobertaSelfAttention(nn.Module): def __init__(self, config): super().__init__() self.config = config if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"): raise ValueError( f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention " f"heads ({config.num_attention_heads})" ) self.num_attention_heads = config.num_attention_heads self.attention_head_size = int(config.hidden_size / config.num_attention_heads) self.all_head_size = self.num_attention_heads * self.attention_head_size self.query = nn.Linear(config.hidden_size, self.all_head_size) self.key = nn.Linear(config.hidden_size, self.all_head_size) self.value = nn.Linear(config.hidden_size, self.all_head_size) self.dropout = nn.Dropout(config.attention_probs_dropout_prob) def forward( self, hidden_states: torch.Tensor, attention_mask: torch.FloatTensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> tuple[torch.Tensor]: input_shape = hidden_states.shape[:-1] hidden_shape = (*input_shape, -1, self.attention_head_size) query_layer = self.query(hidden_states).view(hidden_shape).transpose(1, 2) key_layer = self.key(hidden_states).view(hidden_shape).transpose(1, 2) value_layer = self.value(hidden_states).view(hidden_shape).transpose(1, 2) # Take the dot product between "query" and "key" to get the raw attention scores. attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2)) attention_scores = attention_scores / math.sqrt(self.attention_head_size) if attention_mask is not None: # Apply the attention mask is (precomputed for all layers in AltRobertaModel forward() function) attention_scores = attention_scores + attention_mask # Normalize the attention scores to probabilities. attention_probs = nn.functional.softmax(attention_scores, dim=-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 = torch.matmul(attention_probs, value_layer) context_layer = context_layer.permute(0, 2, 1, 3).contiguous() new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,) context_layer = context_layer.view(new_context_layer_shape) return context_layer, attention_probs # Copied from transformers.models.roberta.modeling_roberta.RobertaSelfOutput class AltRobertaSelfOutput(nn.Module): def __init__(self, config): super().__init__() self.dense = nn.Linear(config.hidden_size, config.hidden_size) self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) self.dropout = nn.Dropout(config.hidden_dropout_prob) def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor: hidden_states = self.dense(hidden_states) hidden_states = self.dropout(hidden_states) hidden_states = self.LayerNorm(hidden_states + input_tensor) return hidden_states ALT_ROBERTA_SELF_ATTENTION_CLASSES = { "eager": AltRobertaSelfAttention, } class AltRobertaAttention(nn.Module): def __init__(self, config): super().__init__() self.self = ALT_ROBERTA_SELF_ATTENTION_CLASSES[config._attn_implementation](config) self.output = AltRobertaSelfOutput(config) def forward( self, hidden_states: torch.Tensor, attention_mask: torch.FloatTensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> torch.Tensor: attention_output, _ = self.self( hidden_states, attention_mask=attention_mask, **kwargs, ) attention_output = self.output(attention_output, hidden_states) return attention_output # Copied from transformers.models.roberta.modeling_roberta.RobertaIntermediate with Roberta->AltRoberta class AltRobertaIntermediate(nn.Module): def __init__(self, config): super().__init__() self.dense = nn.Linear(config.hidden_size, config.intermediate_size) if isinstance(config.hidden_act, str): self.intermediate_act_fn = ACT2FN[config.hidden_act] else: self.intermediate_act_fn = config.hidden_act def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: hidden_states = self.dense(hidden_states) hidden_states = self.intermediate_act_fn(hidden_states) return hidden_states # Copied from transformers.models.roberta.modeling_roberta.RobertaOutput class AltRobertaOutput(nn.Module): def __init__(self, config): super().__init__() self.dense = nn.Linear(config.intermediate_size, config.hidden_size) self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) self.dropout = nn.Dropout(config.hidden_dropout_prob) def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor: hidden_states = self.dense(hidden_states) hidden_states = self.dropout(hidden_states) hidden_states = self.LayerNorm(hidden_states + input_tensor) return hidden_states # Copied from transformers.models.align.modeling_align.AlignTextLayer with AlignText->AltRoberta class AltRobertaLayer(GradientCheckpointingLayer): def __init__(self, config): super().__init__() self.chunk_size_feed_forward = config.chunk_size_feed_forward self.seq_len_dim = 1 self.attention = AltRobertaAttention(config) self.intermediate = AltRobertaIntermediate(config) self.output = AltRobertaOutput(config) def forward( self, hidden_states: torch.Tensor, attention_mask: torch.FloatTensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> torch.Tensor: hidden_states = self.attention( hidden_states, attention_mask=attention_mask, **kwargs, ) hidden_states = apply_chunking_to_forward( self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, hidden_states ) return hidden_states def feed_forward_chunk(self, attention_output): intermediate_output = self.intermediate(attention_output) layer_output = self.output(intermediate_output, attention_output) return layer_output # Copied from transformers.models.align.modeling_align.AlignTextEncoder with AlignText->AltRoberta class AltRobertaEncoder(nn.Module): def __init__(self, config): super().__init__() self.config = config self.layer = nn.ModuleList([AltRobertaLayer(config) for i in range(config.num_hidden_layers)]) self.gradient_checkpointing = False def forward( self, hidden_states: torch.Tensor, attention_mask: torch.FloatTensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> BaseModelOutput: for layer_module in self.layer: hidden_states = layer_module( hidden_states, attention_mask, **kwargs, ) return BaseModelOutput( last_hidden_state=hidden_states, ) # Copied from transformers.models.roberta.modeling_roberta.RobertaPooler class AltRobertaPooler(nn.Module): def __init__(self, config): super().__init__() self.dense = nn.Linear(config.hidden_size, config.hidden_size) self.activation = nn.Tanh() def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: # We "pool" the model by simply taking the hidden state corresponding # to the first token. first_token_tensor = hidden_states[:, 0] pooled_output = self.dense(first_token_tensor) pooled_output = self.activation(pooled_output) return pooled_output # Copied from transformers.models.siglip.modeling_siglip.eager_attention_forward def eager_attention_forward( module: nn.Module, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, attention_mask: torch.Tensor | None, scaling: float, dropout: float = 0.0, **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() return attn_output, attn_weights class AltCLIPAttention(nn.Module): """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 = 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: torch.Tensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> tuple[torch.Tensor, torch.Tensor | None]: """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, self.num_heads, self.head_dim).transpose(1, 2) keys = keys.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2) values = values.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2) attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface( self.config._attn_implementation, eager_attention_forward ) attn_output, attn_weights = attention_interface( self, queries, keys, values, attention_mask, scaling=self.scale, dropout=0.0 if not self.training else self.dropout, **kwargs, ) attn_output = attn_output.reshape(batch_size, seq_length, embed_dim).contiguous() attn_output = self.out_proj(attn_output) return attn_output, attn_weights # Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->AltCLIP class AltCLIPMLP(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 class AltCLIPEncoderLayer(GradientCheckpointingLayer): def __init__(self, config: AltCLIPConfig): super().__init__() self.embed_dim = config.hidden_size self.self_attn = AltCLIPAttention(config) self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps) self.mlp = AltCLIPMLP(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, **kwargs: Unpack[TransformersKwargs], ) -> tuple[torch.FloatTensor, torch.Tensor | None]: residual = hidden_states hidden_states = self.layer_norm1(hidden_states) hidden_states, _ = self.self_attn( hidden_states=hidden_states, attention_mask=attention_mask, **kwargs, ) 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 return hidden_states class AltCLIPEncoder(nn.Module): """ Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a [`AltCLIPEncoderLayer`]. Args: config: AltCLIPConfig """ def __init__(self, config: AltCLIPConfig): super().__init__() self.config = config self.layers = nn.ModuleList([AltCLIPEncoderLayer(config) for _ in range(config.num_hidden_layers)]) self.gradient_checkpointing = False def forward( self, inputs_embeds, attention_mask: torch.Tensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> tuple | 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) """ hidden_states = inputs_embeds for encoder_layer in self.layers: hidden_states = encoder_layer( hidden_states, attention_mask, **kwargs, ) return BaseModelOutput( last_hidden_state=hidden_states, ) # Copied from transformers.models.clip.modeling_clip.CLIPVisionEmbeddings with CLIP->AltCLIP class AltCLIPVisionEmbeddings(nn.Module): def __init__(self, config: AltCLIPVisionConfig): 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 @auto_docstring class AltCLIPPreTrainedModel(PreTrainedModel): config: AltCLIPConfig base_model_prefix = "altclip" input_modalities = ("image", "text") supports_gradient_checkpointing = True _no_split_module = [] @torch.no_grad() def _init_weights(self, module): """Initialize the weights""" factor = self.config.initializer_factor if isinstance(module, AltCLIPVisionEmbeddings): factor = self.config.initializer_factor init.normal_(module.class_embedding, mean=0.0, std=module.embed_dim**-0.5 * factor) init.normal_(module.patch_embedding.weight, std=module.config.initializer_range * factor) init.normal_(module.position_embedding.weight, std=module.config.initializer_range * factor) init.copy_(module.position_ids, torch.arange(module.num_positions).expand((1, -1))) elif isinstance(module, AltCLIPAttention): 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 init.normal_(module.q_proj.weight, std=in_proj_std) init.normal_(module.k_proj.weight, std=in_proj_std) init.normal_(module.v_proj.weight, std=in_proj_std) init.normal_(module.out_proj.weight, std=out_proj_std) elif isinstance(module, AltCLIPMLP): 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 init.normal_(module.fc1.weight, std=fc_std) init.normal_(module.fc2.weight, std=in_proj_std) elif isinstance(module, AltCLIPModel): init.normal_( module.text_projection.weight, std=module.text_embed_dim**-0.5 * self.config.initializer_factor, ) init.normal_( module.visual_projection.weight, std=module.vision_embed_dim**-0.5 * self.config.initializer_factor, ) elif isinstance(module, nn.LayerNorm): init.zeros_(module.bias) init.ones_(module.weight) elif isinstance(module, nn.Linear): init.normal_(module.weight, mean=0.0, std=self.config.initializer_factor) if module.bias is not None: init.zeros_(module.bias) elif isinstance(module, nn.Embedding): init.normal_(module.weight, mean=0.0, std=self.config.initializer_factor) # Here we need the check explicitly, as we slice the weight in the `zeros_` call, so it looses the flag if module.padding_idx is not None and not getattr(module.weight, "_is_hf_initialized", False): init.zeros_(module.weight[module.padding_idx]) elif isinstance(module, AltRobertaEmbeddings): init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1))) init.zeros_(module.token_type_ids) class AltCLIPVisionTransformer(nn.Module): def __init__(self, config: AltCLIPVisionConfig): super().__init__() self.config = config embed_dim = config.hidden_size self.embeddings = AltCLIPVisionEmbeddings(config) self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) self.encoder = AltCLIPEncoder(config) self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) @auto_docstring def forward( self, pixel_values: torch.FloatTensor | None = None, interpolate_pos_encoding: bool | None = False, **kwargs: Unpack[TransformersKwargs], ) -> tuple | BaseModelOutputWithPooling: 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 = self.encoder( inputs_embeds=hidden_states, **kwargs, ) last_hidden_state = encoder_outputs[0] 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, ) class AltCLIPVisionModel(AltCLIPPreTrainedModel): config: AltCLIPVisionConfig main_input_name = "pixel_values" input_modalities = ("image",) _can_record_outputs = { "hidden_states": AltCLIPEncoderLayer, "attentions": AltCLIPAttention, } def __init__(self, config: AltCLIPVisionConfig): super().__init__(config) self.vision_model = AltCLIPVisionTransformer(config) # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self) -> nn.Module: return self.vision_model.embeddings.patch_embedding @merge_with_config_defaults @capture_outputs(tie_last_hidden_states=False) @auto_docstring def forward( self, pixel_values: torch.FloatTensor | None = None, interpolate_pos_encoding: bool = False, **kwargs: Unpack[TransformersKwargs], ) -> BaseModelOutputWithPooling: r""" Examples: ```python >>> from PIL import Image >>> import httpx >>> from io import BytesIO >>> from transformers import AutoProcessor, AltCLIPVisionModel >>> model = AltCLIPVisionModel.from_pretrained("BAAI/AltCLIP") >>> processor = AutoProcessor.from_pretrained("BAAI/AltCLIP") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> with httpx.stream("GET", url) as response: ... image = Image.open(BytesIO(response.read())) >>> 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, interpolate_pos_encoding=interpolate_pos_encoding, **kwargs, ) @auto_docstring( custom_intro=""" The model behaves as an encoder following the architecture described in *Attention is all you need*_ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin. .. _*Attention is all you need*: https://huggingface.co/papers/1706.03762 """ ) class AltRobertaModel(AltCLIPPreTrainedModel): config: AltCLIPTextConfig _can_record_outputs = { "hidden_states": AltRobertaLayer, "attentions": AltRobertaSelfAttention, } # Copied from transformers.models.clap.modeling_clap.ClapTextModel.__init__ with ClapText->AltRoberta def __init__(self, config, add_pooling_layer=True): r""" add_pooling_layer (bool, *optional*, defaults to `True`): Whether to add a pooling layer """ super().__init__(config) self.config = config self.embeddings = AltRobertaEmbeddings(config) self.encoder = AltRobertaEncoder(config) self.pooler = AltRobertaPooler(config) if add_pooling_layer else None # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self): return self.embeddings.word_embeddings def set_input_embeddings(self, value): self.embeddings.word_embeddings = value @merge_with_config_defaults @capture_outputs @auto_docstring # Copied from transformers.models.clap.modeling_clap.ClapTextModel.forward def forward( self, input_ids: torch.Tensor | None = None, attention_mask: torch.Tensor | None = None, token_type_ids: torch.Tensor | None = None, position_ids: torch.Tensor | None = None, inputs_embeds: torch.Tensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> BaseModelOutputWithPooling: if input_ids is not None and inputs_embeds is not None: raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") elif input_ids is not None: self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask) input_shape = input_ids.size() elif inputs_embeds is not None: input_shape = inputs_embeds.size()[:-1] else: raise ValueError("You have to specify either input_ids or inputs_embeds") batch_size, seq_length = input_shape device = input_ids.device if input_ids is not None else inputs_embeds.device if attention_mask is None: attention_mask = torch.ones(((batch_size, seq_length)), device=device) # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length] # ourselves in which case we just need to make it broadcastable to all heads. extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape) embedding_output = self.embeddings( input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds, ) encoder_outputs = self.encoder( embedding_output, attention_mask=extended_attention_mask, **kwargs, ) sequence_output = encoder_outputs[0] pooled_output = self.pooler(sequence_output) if self.pooler is not None else None return BaseModelOutputWithPooling( last_hidden_state=sequence_output, pooler_output=pooled_output, ) class AltCLIPTextModel(AltCLIPPreTrainedModel): config: AltCLIPTextConfig input_modalities = ("text",) def __init__(self, config): super().__init__(config) self.roberta = AltRobertaModel(config, add_pooling_layer=False) self.transformation = nn.Linear(config.hidden_size, config.project_dim) self.pre_LN = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) self.post_init() def get_input_embeddings(self) -> nn.Module: return self.roberta.embeddings.word_embeddings def set_input_embeddings(self, value: nn.Embedding) -> None: self.roberta.embeddings.word_embeddings = value def resize_token_embeddings(self, new_num_tokens: int | None = None) -> nn.Embedding: return super().resize_token_embeddings(new_num_tokens) @can_return_tuple @auto_docstring def forward( self, input_ids: torch.Tensor | None = None, attention_mask: torch.Tensor | None = None, token_type_ids: torch.Tensor | None = None, position_ids: torch.Tensor | None = None, inputs_embeds: torch.Tensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> tuple | BaseModelOutputWithPoolingAndProjection: r""" Examples: ```python >>> from transformers import AutoProcessor, AltCLIPTextModel >>> model = AltCLIPTextModel.from_pretrained("BAAI/AltCLIP") >>> processor = AutoProcessor.from_pretrained("BAAI/AltCLIP") >>> texts = ["it's a cat", "it's a dog"] >>> inputs = processor(text=texts, padding=True, return_tensors="pt") >>> outputs = model(**inputs) >>> last_hidden_state = outputs.last_hidden_state >>> pooled_output = outputs.pooler_output # pooled CLS states ```""" outputs = self.roberta( input_ids=input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids, position_ids=position_ids, inputs_embeds=inputs_embeds, **kwargs, ) # last module outputs sequence_output = outputs[0] # project every module sequence_output = self.pre_LN(sequence_output) # pooler projection_state = self.transformation(sequence_output) pooler_output = projection_state[:, 0] return BaseModelOutputWithPoolingAndProjection( last_hidden_state=projection_state, pooler_output=pooler_output, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) class AltCLIPModel(AltCLIPPreTrainedModel): config: AltCLIPConfig _can_record_outputs = { "hidden_states": AltCLIPEncoderLayer, "attentions": AltCLIPAttention, } def __init__(self, config: AltCLIPConfig): super().__init__(config) if not isinstance(config.vision_config, AltCLIPVisionConfig): raise TypeError( "config.vision_config is expected to be of type AltCLIPVisionConfig but is of type" f" {type(config.vision_config)}." ) if not isinstance(config.text_config, AltCLIPTextConfig): raise TypeError( "config.text_config is expected to be of type AltCLIPTextConfig but is of type" f" {type(config.text_config)}." ) text_config = config.text_config vision_config = config.vision_config # The module using it is not a PreTrainedModel subclass so we need this vision_config._attn_implementation = config._attn_implementation self.projection_dim = config.projection_dim self.text_embed_dim = text_config.project_dim self.vision_embed_dim = vision_config.hidden_size self.text_model = AltCLIPTextModel(text_config) self.vision_model = AltCLIPVisionTransformer(vision_config) 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() @can_return_tuple @auto_docstring def get_text_features( self, input_ids: torch.Tensor, attention_mask: torch.Tensor | None = None, position_ids: torch.Tensor | None = None, token_type_ids: torch.Tensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> tuple | BaseModelOutputWithPooling: r""" Examples: ```python >>> import torch >>> from transformers import AutoProcessor, AltCLIPModel >>> model = AltCLIPModel.from_pretrained("BAAI/AltCLIP") >>> processor = AutoProcessor.from_pretrained("BAAI/AltCLIP") >>> inputs = processor(text=["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt") >>> with torch.inference_mode(): ... text_features = model.get_text_features(**inputs) ```""" text_outputs: BaseModelOutputWithPoolingAndProjection = self.text_model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, token_type_ids=token_type_ids, **kwargs, ) pooled_output = text_outputs.pooler_output text_outputs.pooler_output = self.text_projection(pooled_output) return text_outputs @merge_with_config_defaults @capture_outputs(tie_last_hidden_states=False) @auto_docstring def get_image_features( self, pixel_values: torch.FloatTensor, interpolate_pos_encoding: bool = False, **kwargs: Unpack[TransformersKwargs], ) -> tuple | BaseModelOutputWithPooling: r""" Examples: ```python >>> import torch >>> from transformers import AutoProcessor, AltCLIPModel >>> from transformers.image_utils import load_image >>> model = AltCLIPModel.from_pretrained("BAAI/AltCLIP") >>> processor = AutoProcessor.from_pretrained("BAAI/AltCLIP") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = load_image(url) >>> inputs = processor(images=image, return_tensors="pt") >>> with torch.inference_mode(): ... image_features = model.get_image_features(**inputs) ```""" vision_outputs = self.vision_model( pixel_values=pixel_values, interpolate_pos_encoding=interpolate_pos_encoding, **kwargs, ) pooled_output = vision_outputs.pooler_output vision_outputs.pooler_output = self.visual_projection(pooled_output) return vision_outputs @can_return_tuple @auto_docstring def forward( self, input_ids: torch.LongTensor | None = None, pixel_values: torch.FloatTensor | None = None, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, token_type_ids: torch.Tensor | None = None, return_loss: bool | None = None, interpolate_pos_encoding: bool = False, **kwargs: Unpack[TransformersKwargs], ) -> tuple | AltCLIPOutput: r""" return_loss (`bool`, *optional*): Whether or not to return the contrastive loss. Examples: ```python >>> from PIL import Image >>> import httpx >>> from io import BytesIO >>> from transformers import AutoProcessor, AltCLIPModel >>> model = AltCLIPModel.from_pretrained("BAAI/AltCLIP") >>> processor = AutoProcessor.from_pretrained("BAAI/AltCLIP") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> with httpx.stream("GET", url) as response: ... image = Image.open(BytesIO(response.read())) >>> 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 ```""" text_outputs = self.text_model( input_ids=input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids, position_ids=position_ids, **kwargs, ) vision_outputs = self.vision_model( pixel_values=pixel_values, interpolate_pos_encoding=interpolate_pos_encoding, **kwargs, ) 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 = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True) text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True) # cosine similarity as logits logit_scale = self.logit_scale.exp() logits_per_text = torch.matmul(text_embeds, image_embeds.t()) * logit_scale logits_per_image = logits_per_text.T loss = None if return_loss: loss = clip_loss(logits_per_text) return AltCLIPOutput( 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, ) __all__ = ["AltCLIPPreTrainedModel", "AltCLIPVisionModel", "AltCLIPTextModel", "AltCLIPModel"]