BAAI
/

Feature Extraction
Transformers
PyTorch
clip
custom_code
EVA-CLIP-18B / modeling_evaclip.py
QuanSun's picture
init
15d3eab verified
# coding=utf-8
""" EvaCLIP model configuration"""
# Code mainly copied here: https://github.com/huggingface/transformers/blob/main/src/transformers/models/clip/configuration_clip.py
# and adjusted for evaclip
from dataclasses import dataclass
from typing import Any, Optional, Tuple, Union
import torch
import torch.utils.checkpoint
from torch import nn
from transformers.activations import ACT2FN
from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import (
ModelOutput,
logging,
)
from .configuration_evaclip import EvaCLIPConfig, EvaCLIPTextConfig, EvaCLIPVisionConfig
# try:
# from xformers import ops as xops
# except ImportError:
# xops = None
logger = logging.get_logger(__name__)
class RMSNorm(nn.Module):
"""
adepted from transformers T5LayerNorm
"""
def __init__(self, hidden_size, eps=1e-6):
"""
Construct a layernorm module in the T5 style. No bias and no subtraction of mean.
"""
super().__init__()
self.weight = nn.Parameter(torch.ones(hidden_size))
self.variance_epsilon = eps
def forward(self, hidden_states):
# T5 uses a layer_norm which only scales and doesn't shift, which is also known as Root Mean
# Square Layer Normalization https://arxiv.org/abs/1910.07467 thus varience is calculated
# w/o mean and there is no bias. Additionally we want to make sure that the accumulation for
# half-precision inputs is done in fp32
variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
# convert into half-precision if necessary
if self.weight.dtype in [torch.float16, torch.bfloat16]:
hidden_states = hidden_states.to(self.weight.dtype)
return self.weight * hidden_states
# Copied from transformers.models.bart.modeling_bart._expand_mask
def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
"""
Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
"""
bsz, src_len = mask.size()
tgt_len = tgt_len if tgt_len is not None else src_len
expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
inverted_mask = 1.0 - expanded_mask
return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
# 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
class EvaCLIPVisionModelOutput(ModelOutput):
image_embeds: Optional[torch.FloatTensor] = None
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
attentions: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class EvaCLIPTextModelOutput(ModelOutput):
text_embeds: Optional[torch.FloatTensor] = None
last_hidden_state: torch.FloatTensor = None
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
attentions: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class EvaCLIPOutput(ModelOutput):
loss: Optional[torch.FloatTensor] = None
logits_per_image: torch.FloatTensor = None
logits_per_text: torch.FloatTensor = None
text_embeds: torch.FloatTensor = None
image_embeds: 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()
)
class EvaCLIPVisionEmbeddings(nn.Module):
def __init__(self, config: EvaCLIPVisionConfig):
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=True,
)
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 forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
batch_size = pixel_values.shape[0]
patch_embeds = self.patch_embedding(pixel_values) # 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)
embeddings = embeddings + self.position_embedding(self.position_ids)
return embeddings
class EvaCLIPTextEmbeddings(nn.Module):
def __init__(self, config: EvaCLIPTextConfig):
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]
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 EvaCLIPAttention(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.k_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.k_bias)
self.v_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.v_bias)
self.q_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.q_bias)
self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=True)
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
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], Optional[Tuple[torch.Tensor]]]:
"""Input shape: Batch x Time x Channel"""
bsz, tgt_len, embed_dim = hidden_states.size()
query_states = self.q_proj(hidden_states) * self.scale
key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
proj_shape = (bsz * self.num_heads, -1, self.head_dim)
query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
key_states = key_states.view(*proj_shape)
value_states = value_states.view(*proj_shape)
src_len = key_states.size(1)
attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
raise ValueError(
f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
f" {attn_weights.size()}"
)
# apply the causal_attention_mask first
if causal_attention_mask is not None:
if causal_attention_mask.size() != (bsz, 1, tgt_len, src_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
f" {causal_attention_mask.size()}"
)
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + causal_attention_mask
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, tgt_len, src_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
if output_attentions:
# this operation is a bit akward, but it's required to
# make sure that attn_weights keeps its gradient.
# In order to do so, attn_weights have to reshaped
# twice and have to be reused in the following
attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
else:
attn_weights_reshaped = None
attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
attn_output = torch.bmm(attn_probs, value_states)
if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
attn_output = attn_output.transpose(1, 2)
attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
attn_output = self.out_proj(attn_output)
return attn_output, attn_weights_reshaped
class EvaCLIPTextAttention(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.k_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.k_bias)
self.v_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.v_bias)
self.q_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.q_bias)
self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=True)
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
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], Optional[Tuple[torch.Tensor]]]:
"""Input shape: Batch x Time x Channel"""
bsz, tgt_len, embed_dim = hidden_states.size()
# get query proj
query_states = self.q_proj(hidden_states)
key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
proj_shape = (bsz * self.num_heads, -1, self.head_dim)
query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
key_states = key_states.view(*proj_shape)
value_states = value_states.view(*proj_shape)
src_len = key_states.size(1)
attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
raise ValueError(
f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
f" {attn_weights.size()}"
)
# apply the causal_attention_mask first
if causal_attention_mask is not None:
if causal_attention_mask.size() != (bsz, 1, tgt_len, src_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
f" {causal_attention_mask.size()}"
)
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + causal_attention_mask
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, tgt_len, src_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
if output_attentions:
# this operation is a bit akward, but it's required to
# make sure that attn_weights keeps its gradient.
# In order to do so, attn_weights have to reshaped
# twice and have to be reused in the following
attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
else:
attn_weights_reshaped = None
attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
attn_output = torch.bmm(attn_probs, value_states)
if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
attn_output = attn_output.transpose(1, 2)
attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
attn_output = self.out_proj(attn_output)
return attn_output, attn_weights_reshaped
class EvaCLIPMLP(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 EvaCLIPEncoderLayer(nn.Module):
def __init__(self, config: EvaCLIPConfig):
super().__init__()
self.config = config
norm_layer = RMSNorm if config.use_rms_norm else nn.LayerNorm
self.embed_dim = config.hidden_size
self.post_layernorm = config.post_layernorm if config.post_layernorm is not None else False
self.self_attn = EvaCLIPAttention(config)
self.layer_norm1 = norm_layer(self.embed_dim, eps=config.layer_norm_eps)
self.mlp = EvaCLIPMLP(config)
self.layer_norm2 = norm_layer(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]:
residual = hidden_states
if not self.post_layernorm:
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,
)
if self.post_layernorm:
hidden_states = self.layer_norm1(hidden_states)
hidden_states = residual + hidden_states
residual = hidden_states
if not self.post_layernorm:
hidden_states = self.layer_norm2(hidden_states)
hidden_states = self.mlp(hidden_states)
if self.post_layernorm:
hidden_states = self.layer_norm2(hidden_states)
hidden_states = residual + hidden_states
outputs = (hidden_states,)
if output_attentions:
outputs += (attn_weights,)
return outputs
class EvaCLIPPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = EvaCLIPConfig
base_model_prefix = "clip"
supports_gradient_checkpointing = True
_keys_to_ignore_on_load_missing = [r"position_ids"]
def _init_weights(self, module):
"""Initialize the weights"""
factor = self.config.initializer_factor
if isinstance(module, EvaCLIPTextEmbeddings):
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, EvaCLIPVisionEmbeddings):
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, EvaCLIPAttention):
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, EvaCLIPMLP):
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, EvaCLIPModel):
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, EvaCLIPVisionModelWithProjection):
nn.init.normal_(
module.visual_projection.weight,
std=self.config.hidden_size**-0.5 * self.config.initializer_factor,
)
elif isinstance(module, EvaCLIPTextModelWithProjection):
nn.init.normal_(
module.text_projection.weight,
std=self.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_()
def _set_gradient_checkpointing(self, module, value=False):
if isinstance(module, EvaCLIPEncoder):
module.gradient_checkpointing = value
class EvaCLIPEncoder(nn.Module):
"""
Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
[`CLIPEncoderLayer`].
Args:
config: CLIPConfig
"""
def __init__(self, config: EvaCLIPConfig):
super().__init__()
self.config = config
self.layers = nn.ModuleList([EvaCLIPEncoderLayer(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,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutput]:
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 idx, encoder_layer in enumerate(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 = torch.utils.checkpoint.checkpoint(
create_custom_forward(encoder_layer),
hidden_states,
attention_mask,
causal_attention_mask,
)
else:
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,)
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 EvaCLIPTextTransformer(EvaCLIPPreTrainedModel):
def __init__(self, config: EvaCLIPTextConfig):
super().__init__(config)
self.config = config
embed_dim = config.hidden_size
norm_layer = RMSNorm if config.use_rms_norm else nn.LayerNorm
self.embeddings = EvaCLIPTextEmbeddings(config)
self.encoder = EvaCLIPEncoder(config)
self.final_layer_norm = norm_layer(embed_dim, eps=config.layer_norm_eps)
def gradient_checkpointing_enable(self):
self.encoder.gradient_checkpointing = True
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,
return_dict: Optional[bool] = None,
) -> Union[Tuple, 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
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if input_ids is None:
raise ValueError("You have to specify input_ids")
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)
bsz, seq_len = input_shape
# CLIP's text model uses causal mask, prepare it here.
# https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324
causal_attention_mask = self._build_causal_attention_mask(bsz, seq_len, hidden_states.dtype).to(
hidden_states.device
)
# expand attention_mask
if attention_mask is not None:
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
attention_mask = _expand_mask(attention_mask, hidden_states.dtype)
encoder_outputs = self.encoder(
inputs_embeds=hidden_states,
attention_mask=attention_mask,
causal_attention_mask=causal_attention_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
last_hidden_state = encoder_outputs[0]
last_hidden_state = self.final_layer_norm(last_hidden_state)
# text_embeds.shape = [batch_size, sequence_length, transformer.width]
# take features from the eot embedding (eot_token is the highest number in each sequence)
# casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
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).argmax(dim=-1),
]
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,
)
def _build_causal_attention_mask(self, bsz, seq_len, dtype):
# lazily create causal attention mask, with full attention between the vision tokens
# pytorch uses additive attention mask; fill with -inf
mask = torch.empty(bsz, seq_len, seq_len, dtype=dtype)
mask.fill_(torch.tensor(torch.finfo(dtype).min))
mask.triu_(1) # zero out the lower diagonal
mask = mask.unsqueeze(1) # expand mask
return mask
class EvaCLIPTextModel(EvaCLIPPreTrainedModel):
config_class = EvaCLIPTextConfig
_no_split_modules = ["EvaCLIPEncoderLayer"]
def __init__(self, config: EvaCLIPTextConfig):
super().__init__(config)
self.text_model = EvaCLIPTextTransformer(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
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,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutputWithPooling]:
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
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,
return_dict=return_dict,
)
class EvaCLIPVisionTransformer(EvaCLIPPreTrainedModel):
def __init__(self, config: EvaCLIPVisionConfig):
super().__init__(config)
self.config = config
embed_dim = config.hidden_size
norm_layer = RMSNorm if config.use_rms_norm else nn.LayerNorm
self.embeddings = EvaCLIPVisionEmbeddings(config)
self.encoder = EvaCLIPEncoder(config)
self.post_layernorm = norm_layer(embed_dim, eps=config.layer_norm_eps)
def gradient_checkpointing_enable(self):
self.encoder.gradient_checkpointing = True
def forward(
self,
pixel_values: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, 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
)
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]
pooled_output = last_hidden_state[:, 0, :]
pooled_output = self.post_layernorm(pooled_output)
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 EvaCLIPVisionModel(nn.Module):
config_class = EvaCLIPVisionConfig
main_input_name = "pixel_values"
def __init__(self, config: EvaCLIPVisionConfig):
super().__init__(config)
# super().__init__()
self.vision_model = EvaCLIPVisionTransformer(config)
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self) -> nn.Module:
return self.vision_model.embeddings.patch_embedding
def forward(
self,
pixel_values: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutputWithPooling]:
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
return self.vision_model(
pixel_values=pixel_values,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
class EvaCLIPModel(EvaCLIPPreTrainedModel):
config_class = EvaCLIPConfig
def __init__(self, config: EvaCLIPConfig):
super().__init__(config)
if not (type(config.text_config).__name__ == "EvaCLIPTextConfig"):
raise ValueError(
"config.text_config is expected to be of type EvaCLIPTextConfig but is of type"
f" {type(config.text_config)}."
)
if not (type(config.vision_config).__name__ == "EvaCLIPVisionConfig"):
raise ValueError(
"config.vision_config is expected to be of type EvaCLIPVisionConfig 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 = EvaCLIPTextTransformer(text_config)
self.vision_model = EvaCLIPVisionTransformer(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 = torch.tensor(100., requires_grad=False)
# Initialize weights and apply final processing
self.post_init()
def encode_text(
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,
return_dict: Optional[bool] = None,
) -> torch.FloatTensor:
# Use CLIP 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
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 encode_image(
self,
pixel_values: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> torch.FloatTensor:
# Use EvaCLIP 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,
)
pooled_output = vision_outputs[1] # pooled_output
image_features = self.visual_projection(pooled_output)
return image_features
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,
return_dict: Optional[bool] = None,
) -> Union[Tuple, EvaCLIPOutput]:
# Use CLIP 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 = 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)
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 EvaCLIPOutput(
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 EvaCLIPTextModelWithProjection(EvaCLIPPreTrainedModel):
config_class = EvaCLIPTextConfig
_no_split_modules = ["EvaCLIPEncoderLayer"]
def __init__(self, config: EvaCLIPTextConfig):
super().__init__(config)
self.text_model = EvaCLIPTextTransformer(config)
self.text_projection = nn.Linear(config.hidden_size, config.projection_dim, bias=False)
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
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,
return_dict: Optional[bool] = None,
) -> Union[Tuple, EvaCLIPTextModelOutput]:
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_embeds = self.text_projection(pooled_output)
if not return_dict:
outputs = (text_embeds, text_outputs[0]) + text_outputs[2:]
return tuple(output for output in outputs if output is not None)
return EvaCLIPTextModelOutput(
text_embeds=text_embeds,
last_hidden_state=text_outputs.last_hidden_state,
hidden_states=text_outputs.hidden_states,
attentions=text_outputs.attentions,
)
class EvaCLIPVisionModelWithProjection(EvaCLIPPreTrainedModel):
config_class = EvaCLIPVisionConfig
main_input_name = "pixel_values"
def __init__(self, config: EvaCLIPVisionConfig):
super().__init__(config)
self.vision_model = EvaCLIPVisionTransformer(config)
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
def forward(
self,
pixel_values: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, EvaCLIPVisionModelOutput]:
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_embeds = self.visual_projection(pooled_output)
if not return_dict:
outputs = (image_embeds, vision_outputs[0]) + vision_outputs[2:]
return tuple(output for output in outputs if output is not None)
return EvaCLIPVisionModelOutput(
image_embeds=image_embeds,
last_hidden_state=vision_outputs.last_hidden_state,
hidden_states=vision_outputs.hidden_states,
attentions=vision_outputs.attentions,
)