wehub-resource-sync
1365 lines
58 KiB
Python
1365 lines
58 KiB
Python
# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
|
|
# Copyright 2022 The OpenAI Team Authors and The HuggingFace Team. All rights reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
""" PaddlePaddle CLIPSeg model."""
|
|
|
|
import copy
|
|
import math
|
|
from dataclasses import dataclass
|
|
from typing import Any, Optional, Tuple, Union
|
|
|
|
import paddle
|
|
import paddle.nn.functional as F
|
|
from paddle import nn
|
|
from paddle.distributed.fleet.utils import recompute
|
|
|
|
from ...utils.initializer import normal_, ones_, zeros_
|
|
from ..activations import ACT2FN
|
|
from ..model_outputs import BaseModelOutput, BaseModelOutputWithPooling, ModelOutput
|
|
from ..model_utils import PretrainedModel
|
|
from .configuration import CLIPSegConfig, CLIPSegTextConfig, CLIPSegVisionConfig
|
|
|
|
_CHECKPOINT_FOR_DOC = "CIDAS/clipseg-rd64-refined"
|
|
|
|
CLIPSEG_PRETRAINED_MODEL_ARCHIVE_LIST = [
|
|
"CIDAS/clipseg-rd64-refined",
|
|
]
|
|
|
|
__all__ = [
|
|
"CLIPSegPreTrainedModel",
|
|
"CLIPSegTextModel",
|
|
"CLIPSegVisionModel",
|
|
"CLIPSegModel",
|
|
"CLIPSegForImageSegmentation",
|
|
]
|
|
|
|
|
|
# Copied from transformers.models.bart.modeling_bart._expand_mask
|
|
def _expand_mask(mask: paddle.Tensor, 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.shape
|
|
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])
|
|
|
|
inverted_mask = 1.0 - expanded_mask
|
|
|
|
def masked_fill(x, mask, value):
|
|
y = paddle.full(x.shape, value, x.dtype)
|
|
return paddle.where(mask, y, x)
|
|
|
|
return masked_fill(inverted_mask, inverted_mask.cast("bool"), -1e4)
|
|
|
|
|
|
# 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: paddle.Tensor) -> paddle.Tensor:
|
|
return F.cross_entropy(logits, paddle.arange(len(logits)))
|
|
|
|
|
|
# Copied from paddlenlp.transformers.clip.modeling.clip_loss with clip->clipseg
|
|
def clipseg_loss(similarity: paddle.Tensor) -> paddle.Tensor:
|
|
caption_loss = contrastive_loss(similarity)
|
|
image_loss = contrastive_loss(similarity.t())
|
|
return (caption_loss + image_loss) / 2.0
|
|
|
|
|
|
@dataclass
|
|
# Copied from paddlenlp.transformers.clip.modeling.CLIPOutput with CLIP->CLIPSeg
|
|
class CLIPSegOutput(ModelOutput):
|
|
"""
|
|
Args:
|
|
loss (`paddle.Tensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
|
|
Contrastive loss for image-text similarity.
|
|
logits_per_image:(`paddle.Tensor` of shape `(image_batch_size, text_batch_size)`):
|
|
The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
|
|
similarity scores.
|
|
logits_per_text:(`paddle.Tensor` of shape `(text_batch_size, image_batch_size)`):
|
|
The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
|
|
similarity scores.
|
|
text_embeds(`paddle.Tensor` of shape `(batch_size, output_dim`):
|
|
The text embeddings obtained by applying the projection layer to the pooled output of [`CLIPSegTextModel`].
|
|
image_embeds(`paddle.Tensor` of shape `(batch_size, output_dim`):
|
|
The image embeddings obtained by applying the projection layer to the pooled output of
|
|
[`CLIPSegVisionModel`].
|
|
text_model_output(`BaseModelOutputWithPooling`):
|
|
The output of the [`CLIPSegTextModel`].
|
|
vision_model_output(`BaseModelOutputWithPooling`):
|
|
The output of the [`CLIPSegVisionModel`].
|
|
"""
|
|
|
|
loss: Optional[paddle.Tensor] = None
|
|
logits_per_image: paddle.Tensor = None
|
|
logits_per_text: paddle.Tensor = None
|
|
text_embeds: paddle.Tensor = None
|
|
image_embeds: paddle.Tensor = None
|
|
text_model_output: BaseModelOutputWithPooling = None
|
|
vision_model_output: BaseModelOutputWithPooling = None
|
|
|
|
def to_tuple(self) -> Tuple[Any]:
|
|
return tuple(
|
|
self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
|
|
for k in self.keys()
|
|
)
|
|
|
|
|
|
@dataclass
|
|
class CLIPSegDecoderOutput(ModelOutput):
|
|
"""
|
|
Args:
|
|
logits (`paddle.Tensor` of shape `(batch_size, height, width)`):
|
|
Classification scores for each pixel.
|
|
hidden_states (`tuple(paddle.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
|
Tuple of `paddle.Tensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
|
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
|
|
attentions (`tuple(paddle.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
|
|
Tuple of `paddle.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
|
sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in
|
|
the self-attention heads.
|
|
"""
|
|
|
|
logits: paddle.Tensor = None
|
|
hidden_states: Optional[Tuple[paddle.Tensor]] = None
|
|
attentions: Optional[Tuple[paddle.Tensor]] = None
|
|
|
|
|
|
@dataclass
|
|
class CLIPSegImageSegmentationOutput(ModelOutput):
|
|
"""
|
|
Args:
|
|
loss (`paddle.Tensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
|
|
Contrastive loss for image-text similarity.
|
|
...
|
|
vision_model_output (`BaseModelOutputWithPooling`):
|
|
The output of the [`CLIPSegVisionModel`].
|
|
"""
|
|
|
|
loss: Optional[paddle.Tensor] = None
|
|
logits: paddle.Tensor = None
|
|
conditional_embeddings: paddle.Tensor = None
|
|
pooled_output: paddle.Tensor = None
|
|
vision_model_output: BaseModelOutputWithPooling = None
|
|
decoder_output: CLIPSegDecoderOutput = None
|
|
|
|
def to_tuple(self) -> Tuple[Any]:
|
|
return tuple(
|
|
self[k] if k not in ["vision_model_output", "decoder_output"] else getattr(self, k).to_tuple()
|
|
for k in self.keys()
|
|
)
|
|
|
|
|
|
class CLIPSegVisionEmbeddings(nn.Layer):
|
|
# Copied from paddlenlp.transformers.clip.modeling.CLIPVisionEmbeddings.__init__
|
|
def __init__(self, config: CLIPSegVisionConfig):
|
|
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 = paddle.create_parameter(
|
|
(self.embed_dim,),
|
|
dtype=paddle.get_default_dtype(),
|
|
default_initializer=nn.initializer.Assign(paddle.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_attr=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", paddle.arange(self.num_positions).expand((1, -1)), persistable=False)
|
|
|
|
def interpolate_position_embeddings(self, new_size):
|
|
if len(new_size) != 2:
|
|
raise ValueError("new_size should consist of 2 values")
|
|
|
|
num_patches_one_direction = int(self.num_patches**0.5)
|
|
# we interpolate the position embeddings in 2D
|
|
a = self.position_embedding.weight[1:].T.reshape(
|
|
[1, self.config.hidden_size, num_patches_one_direction, num_patches_one_direction]
|
|
)
|
|
b = (
|
|
nn.functional.interpolate(a, new_size, mode="bicubic", align_corners=False)
|
|
.squeeze(0)
|
|
.reshape([self.config.hidden_size, new_size[0] * new_size[1]])
|
|
.T
|
|
)
|
|
result = paddle.concat([self.position_embedding.weight[:1], b])
|
|
|
|
return result
|
|
|
|
def forward(self, pixel_values: paddle.Tensor) -> paddle.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([0, 2, 1])
|
|
|
|
class_embeds = self.class_embedding.expand([batch_size, 1, -1])
|
|
embeddings = paddle.concat([class_embeds, patch_embeds], axis=1)
|
|
|
|
if embeddings.shape[1] != self.num_positions:
|
|
new_shape = int(math.sqrt(embeddings.shape[1] - 1))
|
|
embeddings = embeddings + self.interpolate_position_embeddings((new_shape, new_shape))
|
|
embeddings = embeddings
|
|
else:
|
|
embeddings = embeddings + self.position_embedding(self.position_ids)
|
|
|
|
return embeddings
|
|
|
|
|
|
# Copied from paddlenlp.transformers.clip.modeling.CLIPTextEmbeddings with CLIP->CLIPSeg
|
|
class CLIPSegTextEmbeddings(nn.Layer):
|
|
def __init__(self, config: CLIPSegTextConfig):
|
|
super().__init__()
|
|
embed_dim = config.hidden_size
|
|
|
|
self.token_embedding = nn.Embedding(config.vocab_size, embed_dim)
|
|
self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)
|
|
|
|
# position_ids (1, len position emb) is contiguous in memory and exported when serialized
|
|
self.register_buffer(
|
|
"position_ids",
|
|
paddle.arange(config.max_position_embeddings, dtype="int64").expand((1, -1)),
|
|
persistable=False,
|
|
)
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: Optional[paddle.Tensor] = None,
|
|
position_ids: Optional[paddle.Tensor] = None,
|
|
inputs_embeds: Optional[paddle.Tensor] = None,
|
|
) -> paddle.Tensor:
|
|
seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]
|
|
|
|
if position_ids is None:
|
|
position_ids = self.position_ids[:, :seq_length]
|
|
|
|
if inputs_embeds is None:
|
|
inputs_embeds = self.token_embedding(input_ids)
|
|
|
|
position_embeddings = self.position_embedding(position_ids)
|
|
embeddings = inputs_embeds + position_embeddings
|
|
|
|
return embeddings
|
|
|
|
|
|
# Copied from paddlenlp.transformers.clip.modeling.CLIPAttention with CLIP->CLIPSeg
|
|
class CLIPSegAttention(nn.Layer):
|
|
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
|
|
|
def __init__(self, config):
|
|
super().__init__()
|
|
self.config = config
|
|
self.embed_dim = config.hidden_size
|
|
self.num_heads = config.num_attention_heads
|
|
self.head_dim = self.embed_dim // self.num_heads
|
|
if self.head_dim * self.num_heads != self.embed_dim:
|
|
raise ValueError(
|
|
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
|
|
f" {self.num_heads})."
|
|
)
|
|
self.scale = self.head_dim**-0.5
|
|
self.dropout = config.attention_dropout
|
|
|
|
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 _shape(self, tensor: paddle.Tensor, seq_len: int, bsz: int):
|
|
return tensor.reshape([bsz, seq_len, self.num_heads, self.head_dim]).transpose([0, 2, 1, 3])
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: paddle.Tensor,
|
|
attention_mask: Optional[paddle.Tensor] = None,
|
|
causal_attention_mask: Optional[paddle.Tensor] = None,
|
|
output_attentions: Optional[bool] = False,
|
|
) -> Tuple[paddle.Tensor, Optional[paddle.Tensor], Optional[Tuple[paddle.Tensor]]]:
|
|
"""Input shape: Batch x Time x Channel"""
|
|
|
|
bsz, tgt_len, embed_dim = hidden_states.shape
|
|
|
|
# get query proj
|
|
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).reshape(proj_shape)
|
|
key_states = key_states.reshape(proj_shape)
|
|
value_states = value_states.reshape(proj_shape)
|
|
|
|
src_len = key_states.shape[1]
|
|
attn_weights = paddle.bmm(query_states, key_states.transpose([0, 2, 1]))
|
|
|
|
if attn_weights.shape != [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.shape}"
|
|
)
|
|
|
|
# apply the causal_attention_mask first
|
|
if causal_attention_mask is not None:
|
|
if causal_attention_mask.shape != [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.shape}"
|
|
)
|
|
attn_weights = attn_weights.reshape([bsz, self.num_heads, tgt_len, src_len]) + causal_attention_mask
|
|
attn_weights = attn_weights.reshape([bsz * self.num_heads, tgt_len, src_len])
|
|
|
|
if attention_mask is not None:
|
|
if attention_mask.shape != [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.shape}"
|
|
)
|
|
attn_weights = attn_weights.reshape([bsz, self.num_heads, tgt_len, src_len]) + attention_mask
|
|
attn_weights = attn_weights.reshape([bsz * self.num_heads, tgt_len, src_len])
|
|
|
|
attn_weights = nn.functional.softmax(attn_weights, axis=-1)
|
|
|
|
if output_attentions:
|
|
# this operation is a bit awkward, 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.reshape([bsz, self.num_heads, tgt_len, src_len])
|
|
attn_weights = attn_weights_reshaped.reshape([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 = paddle.bmm(attn_probs, value_states)
|
|
|
|
if attn_output.shape != [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.shape}"
|
|
)
|
|
|
|
attn_output = attn_output.reshape([bsz, self.num_heads, tgt_len, self.head_dim])
|
|
attn_output = attn_output.transpose([0, 2, 1, 3])
|
|
attn_output = attn_output.reshape([bsz, tgt_len, embed_dim])
|
|
|
|
attn_output = self.out_proj(attn_output)
|
|
|
|
return attn_output, attn_weights_reshaped
|
|
|
|
|
|
# Copied from paddlenlp.transformers.clip.modeling.CLIPMLP with CLIP->CLIPSeg
|
|
class CLIPSegMLP(nn.Layer):
|
|
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: paddle.Tensor) -> paddle.Tensor:
|
|
hidden_states = self.fc1(hidden_states)
|
|
hidden_states = self.activation_fn(hidden_states)
|
|
hidden_states = self.fc2(hidden_states)
|
|
return hidden_states
|
|
|
|
|
|
# Copied from paddlenlp.transformers.clip.modeling.CLIPEncoderLayer with CLIP->CLIPSeg
|
|
class CLIPSegEncoderLayer(nn.Layer):
|
|
def __init__(self, config: CLIPSegConfig):
|
|
super().__init__()
|
|
self.embed_dim = config.hidden_size
|
|
self.self_attn = CLIPSegAttention(config)
|
|
self.layer_norm1 = nn.LayerNorm(self.embed_dim, epsilon=config.layer_norm_eps)
|
|
self.mlp = CLIPSegMLP(config)
|
|
self.layer_norm2 = nn.LayerNorm(self.embed_dim, epsilon=config.layer_norm_eps)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: paddle.Tensor,
|
|
attention_mask: paddle.Tensor,
|
|
causal_attention_mask: paddle.Tensor,
|
|
output_attentions: Optional[bool] = False,
|
|
) -> Tuple[paddle.Tensor]:
|
|
"""
|
|
Args:
|
|
hidden_states (`paddle.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
|
attention_mask (`paddle.Tensor`): attention mask of size
|
|
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
|
|
`(config.encoder_attention_heads,)`.
|
|
output_attentions (`bool`, *optional*):
|
|
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
|
returned tensors for more detail.
|
|
"""
|
|
residual = hidden_states
|
|
|
|
hidden_states = self.layer_norm1(hidden_states)
|
|
hidden_states, attn_weights = self.self_attn(
|
|
hidden_states=hidden_states,
|
|
attention_mask=attention_mask,
|
|
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 CLIPSegPreTrainedModel(PretrainedModel):
|
|
"""
|
|
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
|
models.
|
|
"""
|
|
|
|
config_class = CLIPSegConfig
|
|
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, CLIPSegTextEmbeddings):
|
|
normal_(module.token_embedding.weight, mean=0.0, std=factor * 0.02)
|
|
normal_(module.position_embedding.weight, mean=0.0, std=factor * 0.02)
|
|
elif isinstance(module, CLIPSegVisionEmbeddings):
|
|
factor = self.config.initializer_factor
|
|
normal_(module.class_embedding, mean=0.0, std=module.embed_dim**-0.5 * factor)
|
|
normal_(module.patch_embedding.weight, std=module.config.initializer_range * factor)
|
|
normal_(module.position_embedding.weight, std=module.config.initializer_range * factor)
|
|
elif isinstance(module, CLIPSegAttention):
|
|
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
|
|
normal_(module.q_proj.weight, std=in_proj_std)
|
|
normal_(module.k_proj.weight, std=in_proj_std)
|
|
normal_(module.v_proj.weight, std=in_proj_std)
|
|
normal_(module.out_proj.weight, std=out_proj_std)
|
|
elif isinstance(module, CLIPSegMLP):
|
|
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
|
|
normal_(module.fc1.weight, std=fc_std)
|
|
normal_(module.fc2.weight, std=in_proj_std)
|
|
elif isinstance(module, CLIPSegModel):
|
|
normal_(
|
|
module.text_projection.weight,
|
|
std=module.text_embed_dim**-0.5 * self.config.initializer_factor,
|
|
)
|
|
normal_(
|
|
module.visual_projection.weight,
|
|
std=module.vision_embed_dim**-0.5 * self.config.initializer_factor,
|
|
)
|
|
|
|
if isinstance(module, nn.LayerNorm):
|
|
zeros_(module.bias)
|
|
ones_(module.weight)
|
|
if isinstance(module, nn.Linear) and module.bias is not None:
|
|
zeros_(module.bias)
|
|
|
|
def _set_gradient_checkpointing(self, module, value=False):
|
|
if isinstance(module, CLIPSegEncoder):
|
|
module.enable_recompute = value
|
|
|
|
|
|
# Copied from paddlenlp.transformers.clip.modeling.CLIPEncoder with CLIP->CLIPSeg
|
|
class CLIPSegEncoder(nn.Layer):
|
|
"""
|
|
Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
|
|
[`CLIPSegEncoderLayer`].
|
|
Args:
|
|
config: CLIPSegConfig
|
|
"""
|
|
|
|
def __init__(self, config: CLIPSegConfig):
|
|
super().__init__()
|
|
self.config = config
|
|
self.layers = nn.LayerList([CLIPSegEncoderLayer(config) for _ in range(config.num_hidden_layers)])
|
|
self.gradient_checkpointing = False
|
|
|
|
def forward(
|
|
self,
|
|
inputs_embeds,
|
|
attention_mask: Optional[paddle.Tensor] = None,
|
|
causal_attention_mask: Optional[paddle.Tensor] = None,
|
|
output_attentions: Optional[bool] = None,
|
|
output_hidden_states: Optional[bool] = None,
|
|
return_dict: Optional[bool] = None,
|
|
) -> Union[Tuple, BaseModelOutput]:
|
|
r"""
|
|
Args:
|
|
inputs_embeds (`paddle.Tensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
|
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
|
|
This is useful if you want more control over how to convert `input_ids` indices into associated vectors
|
|
than the model's internal embedding lookup matrix.
|
|
attention_mask (`paddle.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
|
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
|
- 1 for tokens that are **not masked**,
|
|
- 0 for tokens that are **masked**.
|
|
[What are attention masks?](../glossary#attention-mask)
|
|
causal_attention_mask (`paddle.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
|
|
)
|
|
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 = recompute(
|
|
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 CLIPSegTextTransformer(nn.Layer):
|
|
# Copied from paddlenlp.transformers.clip.modeling.CLIPTextTransformer.__init__ with CLIP->CLIPSeg
|
|
def __init__(self, config: CLIPSegTextConfig):
|
|
super().__init__()
|
|
self.config = config
|
|
embed_dim = config.hidden_size
|
|
self.embeddings = CLIPSegTextEmbeddings(config)
|
|
self.encoder = CLIPSegEncoder(config)
|
|
self.final_layer_norm = nn.LayerNorm(embed_dim, epsilon=config.layer_norm_eps)
|
|
|
|
# For `pooled_output` computation
|
|
self.eos_token_id = config.eos_token_id
|
|
|
|
# Copied from paddlenlp.transformers.clip.modeling.CLIPTextTransformer.forward with clip->clipseg, CLIP->CLIPSeg
|
|
def forward(
|
|
self,
|
|
input_ids: Optional[paddle.Tensor] = None,
|
|
attention_mask: Optional[paddle.Tensor] = None,
|
|
position_ids: Optional[paddle.Tensor] = None,
|
|
output_attentions: Optional[bool] = None,
|
|
output_hidden_states: Optional[bool] = None,
|
|
return_dict: Optional[bool] = None,
|
|
) -> Union[Tuple, BaseModelOutputWithPooling]:
|
|
r"""
|
|
Returns:
|
|
"""
|
|
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.shape
|
|
input_ids = input_ids.reshape([-1, input_shape[-1]])
|
|
|
|
hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids)
|
|
|
|
bsz, seq_len = input_shape
|
|
# CLIPSeg's text model uses causal mask, prepare it here.
|
|
# https://github.com/openai/CLIPSeg/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clipseg/model.py#L324
|
|
causal_attention_mask = self._build_causal_attention_mask(bsz, seq_len, hidden_states.dtype)
|
|
# 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)
|
|
|
|
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)
|
|
|
|
if self.eos_token_id == 2:
|
|
# The `eos_token_id` was incorrect before PR #24773: Let's keep what have been done here.
|
|
# A CLIPSeg model with such `eos_token_id` in the config can't work correctly with extra new tokens added
|
|
# ------------------------------------------------------------
|
|
# 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 paddle.int32 for onnx compatibility: argmax doesn't support int64 inputs with opset 14
|
|
pooled_output = last_hidden_state.gather_nd(
|
|
paddle.stack(
|
|
[paddle.arange(last_hidden_state.shape[0], dtype="int32"), input_ids.argmax(-1, dtype="int32")],
|
|
axis=-1,
|
|
)
|
|
)
|
|
else:
|
|
# The config gets updated `eos_token_id` from PR #24773 (so the use of exta new tokens is possible)
|
|
# We need to get the first position of `eos_token_id` value (`pad_token_ids` might equal to `eos_token_id`)
|
|
pooled_output = last_hidden_state.gather_nd(
|
|
paddle.stack(
|
|
[
|
|
paddle.arange(last_hidden_state.shape[0], dtype="int32"),
|
|
(input_ids == self.eos_token_id).cast("int32").argmax(axis=-1, dtype="int32"),
|
|
],
|
|
axis=-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 = paddle.full([bsz, seq_len, seq_len], fill_value=-1e9, dtype=dtype)
|
|
mask = paddle.triu(mask, diagonal=1) # zero out the upper diagonal
|
|
mask = mask.unsqueeze(1) # expand mask
|
|
return mask
|
|
|
|
|
|
class CLIPSegTextModel(CLIPSegPreTrainedModel):
|
|
config_class = CLIPSegTextConfig
|
|
|
|
_no_split_modules = ["CLIPSegEncoderLayer"]
|
|
|
|
def __init__(self, config: CLIPSegTextConfig):
|
|
super().__init__(config)
|
|
self.text_model = CLIPSegTextTransformer(config)
|
|
|
|
def get_input_embeddings(self) -> nn.Layer:
|
|
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[paddle.Tensor] = None,
|
|
attention_mask: Optional[paddle.Tensor] = None,
|
|
position_ids: Optional[paddle.Tensor] = None,
|
|
output_attentions: Optional[bool] = None,
|
|
output_hidden_states: Optional[bool] = None,
|
|
return_dict: Optional[bool] = None,
|
|
) -> Union[Tuple, BaseModelOutputWithPooling]:
|
|
r"""
|
|
Returns:
|
|
Examples:
|
|
```python
|
|
>>> from paddlenlp.transformers import AutoTokenizer, CLIPSegTextModel
|
|
>>> tokenizer = AutoTokenizer.from_pretrained("CIDAS/clipseg-rd64-refined")
|
|
>>> model = CLIPSegTextModel.from_pretrained("CIDAS/clipseg-rd64-refined")
|
|
>>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pd")
|
|
>>> 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,
|
|
return_dict=return_dict,
|
|
)
|
|
|
|
|
|
class CLIPSegVisionTransformer(nn.Layer):
|
|
# Copied from paddlenlp.transformers.clip.modeling.CLIPVisionTransformer.__init__ with CLIP->CLIPSeg
|
|
def __init__(self, config: CLIPSegVisionConfig):
|
|
super().__init__()
|
|
self.config = config
|
|
embed_dim = config.hidden_size
|
|
|
|
self.embeddings = CLIPSegVisionEmbeddings(config)
|
|
self.pre_layrnorm = nn.LayerNorm(embed_dim, epsilon=config.layer_norm_eps)
|
|
self.encoder = CLIPSegEncoder(config)
|
|
self.post_layernorm = nn.LayerNorm(embed_dim, epsilon=config.layer_norm_eps)
|
|
|
|
# Copied from paddlenlp.transformers.clip.modeling.CLIPVisionTransformer.forward
|
|
def forward(
|
|
self,
|
|
pixel_values: Optional[paddle.Tensor] = None,
|
|
output_attentions: Optional[bool] = None,
|
|
output_hidden_states: Optional[bool] = None,
|
|
return_dict: Optional[bool] = None,
|
|
) -> Union[Tuple, BaseModelOutputWithPooling]:
|
|
r"""
|
|
Returns:
|
|
"""
|
|
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)
|
|
hidden_states = self.pre_layrnorm(hidden_states)
|
|
|
|
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 CLIPSegVisionModel(CLIPSegPreTrainedModel):
|
|
config_class = CLIPSegVisionConfig
|
|
main_input_name = "pixel_values"
|
|
|
|
def __init__(self, config: CLIPSegVisionConfig):
|
|
super().__init__(config)
|
|
self.vision_model = CLIPSegVisionTransformer(config)
|
|
|
|
def get_input_embeddings(self) -> nn.Layer:
|
|
return self.vision_model.embeddings.patch_embedding
|
|
|
|
def forward(
|
|
self,
|
|
pixel_values: Optional[paddle.Tensor] = None,
|
|
output_attentions: Optional[bool] = None,
|
|
output_hidden_states: Optional[bool] = None,
|
|
return_dict: Optional[bool] = None,
|
|
) -> Union[Tuple, BaseModelOutputWithPooling]:
|
|
r"""
|
|
Returns:
|
|
Examples:
|
|
```python
|
|
>>> from PIL import Image
|
|
>>> import requests
|
|
>>> from paddlenlp.transformers import AutoProcessor, CLIPSegVisionModel
|
|
>>> processor = AutoProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
|
|
>>> model = CLIPSegVisionModel.from_pretrained("CIDAS/clipseg-rd64-refined")
|
|
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
|
|
>>> image = Image.open(requests.get(url, stream=True).raw)
|
|
>>> inputs = processor(images=image, return_tensors="pd")
|
|
>>> outputs = model(**inputs)
|
|
>>> last_hidden_state = outputs.last_hidden_state
|
|
>>> pooled_output = outputs.pooler_output # pooled CLS states
|
|
```"""
|
|
return self.vision_model(
|
|
pixel_values=pixel_values,
|
|
output_attentions=output_attentions,
|
|
output_hidden_states=output_hidden_states,
|
|
return_dict=return_dict,
|
|
)
|
|
|
|
|
|
class CLIPSegModel(CLIPSegPreTrainedModel):
|
|
config_class = CLIPSegConfig
|
|
|
|
def __init__(self, config: CLIPSegConfig):
|
|
super().__init__(config)
|
|
|
|
if not isinstance(config.text_config, CLIPSegTextConfig):
|
|
raise ValueError(
|
|
"config.text_config is expected to be of type CLIPSegTextConfig but is of type"
|
|
f" {type(config.text_config)}."
|
|
)
|
|
|
|
if not isinstance(config.vision_config, CLIPSegVisionConfig):
|
|
raise ValueError(
|
|
"config.vision_config is expected to be of type CLIPSegVisionConfig 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 = CLIPSegTextTransformer(text_config)
|
|
self.vision_model = CLIPSegVisionTransformer(vision_config)
|
|
|
|
self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias_attr=False)
|
|
self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias_attr=False)
|
|
self.logit_scale = paddle.create_parameter(
|
|
(1,),
|
|
dtype=paddle.get_default_dtype(),
|
|
default_initializer=nn.initializer.Constant(self.config.logit_scale_init_value),
|
|
)
|
|
|
|
def get_text_features(
|
|
self,
|
|
input_ids: Optional[paddle.Tensor] = None,
|
|
attention_mask: Optional[paddle.Tensor] = None,
|
|
position_ids: Optional[paddle.Tensor] = None,
|
|
output_attentions: Optional[bool] = None,
|
|
output_hidden_states: Optional[bool] = None,
|
|
return_dict: Optional[bool] = None,
|
|
) -> paddle.Tensor:
|
|
r"""
|
|
Returns:
|
|
text_features (`paddle.Tensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
|
|
applying the projection layer to the pooled output of [`CLIPSegTextModel`].
|
|
Examples:
|
|
```python
|
|
>>> from paddlenlp.transformers import AutoTokenizer, CLIPSegModel
|
|
>>> tokenizer = AutoTokenizer.from_pretrained("CIDAS/clipseg-rd64-refined")
|
|
>>> model = CLIPSegModel.from_pretrained("CIDAS/clipseg-rd64-refined")
|
|
>>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pd")
|
|
>>> text_features = model.get_text_features(**inputs)
|
|
```"""
|
|
# Use CLIPSEG 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 get_image_features(
|
|
self,
|
|
pixel_values: Optional[paddle.Tensor] = None,
|
|
output_attentions: Optional[bool] = None,
|
|
output_hidden_states: Optional[bool] = None,
|
|
return_dict: Optional[bool] = None,
|
|
) -> paddle.Tensor:
|
|
r"""
|
|
Returns:
|
|
image_features (`paddle.Tensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
|
|
applying the projection layer to the pooled output of [`CLIPSegVisionModel`].
|
|
Examples:
|
|
```python
|
|
>>> from PIL import Image
|
|
>>> import requests
|
|
>>> from paddlenlp.transformers import AutoProcessor, CLIPSegModel
|
|
>>> processor = AutoProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
|
|
>>> model = CLIPSegModel.from_pretrained("CIDAS/clipseg-rd64-refined")
|
|
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
|
|
>>> image = Image.open(requests.get(url, stream=True).raw)
|
|
>>> inputs = processor(images=image, return_tensors="pd")
|
|
>>> image_features = model.get_image_features(**inputs)
|
|
```"""
|
|
# Use CLIPSEG 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[paddle.Tensor] = None,
|
|
pixel_values: Optional[paddle.Tensor] = None,
|
|
attention_mask: Optional[paddle.Tensor] = None,
|
|
position_ids: Optional[paddle.Tensor] = None,
|
|
return_loss: Optional[bool] = None,
|
|
output_attentions: Optional[bool] = None,
|
|
output_hidden_states: Optional[bool] = None,
|
|
return_dict: Optional[bool] = None,
|
|
) -> Union[Tuple, CLIPSegOutput]:
|
|
r"""
|
|
Returns:
|
|
Examples:
|
|
```python
|
|
>>> from PIL import Image
|
|
>>> import requests
|
|
>>> from paddlenlp.transformers import AutoProcessor, CLIPSegModel
|
|
>>> processor = AutoProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
|
|
>>> model = CLIPSegModel.from_pretrained("CIDAS/clipseg-rd64-refined")
|
|
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
|
|
>>> image = Image.open(requests.get(url, stream=True).raw)
|
|
>>> inputs = processor(
|
|
... text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pd", padding=True
|
|
... )
|
|
>>> outputs = model(**inputs)
|
|
>>> logits_per_image = outputs.logits_per_image # this is the image-text similarity score
|
|
>>> probs = logits_per_image.softmax(dim=1) # we can take the softmax to get the label probabilities
|
|
```"""
|
|
# Use CLIPSEG 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, axis=-1, keepdim=True)
|
|
text_embeds = text_embeds / text_embeds.norm(p=2, axis=-1, keepdim=True)
|
|
|
|
# cosine similarity as logits
|
|
logit_scale = self.logit_scale.exp()
|
|
logits_per_text = paddle.matmul(text_embeds, image_embeds, transpose_y=True) * logit_scale
|
|
logits_per_image = logits_per_text.t()
|
|
|
|
loss = None
|
|
if return_loss:
|
|
loss = clipseg_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 CLIPSegOutput(
|
|
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 CLIPSegDecoderLayer(nn.Layer):
|
|
"""
|
|
CLIPSeg decoder layer, which is identical to `CLIPSegEncoderLayer`, except that normalization is applied after
|
|
self-attention/MLP, rather than before.
|
|
"""
|
|
|
|
# Copied from paddlenlp.transformers.clip.modeling.CLIPEncoderLayer.__init__ with CLIP->CLIPSeg
|
|
def __init__(self, config: CLIPSegConfig):
|
|
super().__init__()
|
|
self.embed_dim = config.hidden_size
|
|
self.self_attn = CLIPSegAttention(config)
|
|
self.layer_norm1 = nn.LayerNorm(self.embed_dim, epsilon=config.layer_norm_eps)
|
|
self.mlp = CLIPSegMLP(config)
|
|
self.layer_norm2 = nn.LayerNorm(self.embed_dim, epsilon=config.layer_norm_eps)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: paddle.Tensor,
|
|
attention_mask: paddle.Tensor,
|
|
causal_attention_mask: paddle.Tensor,
|
|
output_attentions: Optional[bool] = False,
|
|
) -> Tuple[paddle.Tensor]:
|
|
"""
|
|
Args:
|
|
hidden_states (`paddle.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
|
attention_mask (`paddle.Tensor`): attention mask of size
|
|
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
|
|
`(config.encoder_attention_heads,)`.
|
|
output_attentions (`bool`, *optional*):
|
|
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
|
returned tensors for more detail.
|
|
"""
|
|
residual = hidden_states
|
|
|
|
hidden_states, 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
|
|
hidden_states = self.layer_norm1(hidden_states)
|
|
|
|
residual = hidden_states
|
|
hidden_states = self.mlp(hidden_states)
|
|
hidden_states = residual + hidden_states
|
|
hidden_states = self.layer_norm2(hidden_states)
|
|
|
|
outputs = (hidden_states,)
|
|
|
|
if output_attentions:
|
|
outputs += (attn_weights,)
|
|
|
|
return outputs
|
|
|
|
|
|
class CLIPSegDecoder(CLIPSegPreTrainedModel):
|
|
def __init__(self, config: CLIPSegConfig):
|
|
super().__init__(config)
|
|
|
|
self.conditional_layer = config.conditional_layer
|
|
|
|
self.film_mul = nn.Linear(config.projection_dim, config.reduce_dim)
|
|
self.film_add = nn.Linear(config.projection_dim, config.reduce_dim)
|
|
|
|
if config.use_complex_transposed_convolution:
|
|
transposed_kernels = (config.vision_config.patch_size // 4, config.vision_config.patch_size // 4)
|
|
|
|
self.transposed_convolution = nn.Sequential(
|
|
nn.Conv2D(config.reduce_dim, config.reduce_dim, kernel_size=3, padding=1),
|
|
nn.ReLU(),
|
|
nn.Conv2DTranspose(
|
|
config.reduce_dim,
|
|
config.reduce_dim // 2,
|
|
kernel_size=transposed_kernels[0],
|
|
stride=transposed_kernels[0],
|
|
),
|
|
nn.ReLU(),
|
|
nn.Conv2DTranspose(
|
|
config.reduce_dim // 2, 1, kernel_size=transposed_kernels[1], stride=transposed_kernels[1]
|
|
),
|
|
)
|
|
else:
|
|
self.transposed_convolution = nn.Conv2DTranspose(
|
|
config.reduce_dim, 1, config.vision_config.patch_size, stride=config.vision_config.patch_size
|
|
)
|
|
|
|
depth = len(config.extract_layers)
|
|
self.reduces = nn.LayerList(
|
|
[nn.Linear(config.vision_config.hidden_size, config.reduce_dim) for _ in range(depth)]
|
|
)
|
|
|
|
decoder_config = copy.deepcopy(config.vision_config)
|
|
decoder_config.hidden_size = config.reduce_dim
|
|
decoder_config.num_attention_heads = config.decoder_num_attention_heads
|
|
decoder_config.intermediate_size = config.decoder_intermediate_size
|
|
decoder_config.hidden_act = "relu"
|
|
self.layers = nn.LayerList([CLIPSegDecoderLayer(decoder_config) for _ in range(len(config.extract_layers))])
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: Tuple[paddle.Tensor],
|
|
conditional_embeddings: paddle.Tensor,
|
|
output_attentions: Optional[bool] = None,
|
|
output_hidden_states: Optional[bool] = None,
|
|
return_dict: Optional[bool] = True,
|
|
):
|
|
all_hidden_states = () if output_hidden_states else None
|
|
all_attentions = () if output_attentions else None
|
|
|
|
activations = hidden_states[::-1]
|
|
|
|
output = None
|
|
for i, (activation, layer, reduce) in enumerate(zip(activations, self.layers, self.reduces)):
|
|
if output is not None:
|
|
output = reduce(activation) + output
|
|
else:
|
|
output = reduce(activation)
|
|
|
|
if i == self.conditional_layer:
|
|
output = self.film_mul(conditional_embeddings) * output.transpose([1, 0, 2]) + self.film_add(
|
|
conditional_embeddings
|
|
)
|
|
output = output.transpose([1, 0, 2])
|
|
|
|
layer_outputs = layer(
|
|
output, attention_mask=None, causal_attention_mask=None, output_attentions=output_attentions
|
|
)
|
|
|
|
output = layer_outputs[0]
|
|
|
|
if output_hidden_states:
|
|
all_hidden_states += (output,)
|
|
|
|
if output_attentions:
|
|
all_attentions += (layer_outputs[1],)
|
|
|
|
output = output[:, 1:, :].transpose(
|
|
[0, 2, 1]
|
|
) # remove cls token and reshape to [batch_size, reduce_dim, seq_len]
|
|
|
|
size = int(math.sqrt(output.shape[2]))
|
|
|
|
batch_size = conditional_embeddings.shape[0]
|
|
output = output.reshape([batch_size, output.shape[1], size, size])
|
|
|
|
logits = self.transposed_convolution(output).squeeze()
|
|
|
|
if not return_dict:
|
|
return tuple(v for v in [logits, all_hidden_states, all_attentions] if v is not None)
|
|
|
|
return CLIPSegDecoderOutput(
|
|
logits=logits,
|
|
hidden_states=all_hidden_states,
|
|
attentions=all_attentions,
|
|
)
|
|
|
|
|
|
class CLIPSegForImageSegmentation(CLIPSegPreTrainedModel):
|
|
config_class = CLIPSegConfig
|
|
|
|
def __init__(self, config: CLIPSegConfig):
|
|
super().__init__(config)
|
|
|
|
self.config = config
|
|
|
|
self.clip = CLIPSegModel(config)
|
|
self.extract_layers = config.extract_layers
|
|
|
|
self.decoder = CLIPSegDecoder(config)
|
|
|
|
def get_conditional_embeddings(
|
|
self,
|
|
batch_size: int = None,
|
|
input_ids: Optional[paddle.Tensor] = None,
|
|
attention_mask: Optional[paddle.Tensor] = None,
|
|
position_ids: Optional[paddle.Tensor] = None,
|
|
conditional_pixel_values: Optional[paddle.Tensor] = None,
|
|
):
|
|
if input_ids is not None:
|
|
# compute conditional embeddings from texts
|
|
if len(input_ids) != batch_size:
|
|
raise ValueError("Make sure to pass as many prompt texts as there are query images")
|
|
with paddle.no_grad():
|
|
conditional_embeddings = self.clip.get_text_features(
|
|
input_ids, attention_mask=attention_mask, position_ids=position_ids
|
|
)
|
|
elif conditional_pixel_values is not None:
|
|
# compute conditional embeddings from images
|
|
if len(conditional_pixel_values) != batch_size:
|
|
raise ValueError("Make sure to pass as many prompt images as there are query images")
|
|
with paddle.no_grad():
|
|
conditional_embeddings = self.clip.get_image_features(conditional_pixel_values)
|
|
else:
|
|
raise ValueError(
|
|
"Invalid conditional, should be either provided as `input_ids` or `conditional_pixel_values`"
|
|
)
|
|
|
|
return conditional_embeddings
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: Optional[paddle.Tensor] = None,
|
|
pixel_values: Optional[paddle.Tensor] = None,
|
|
conditional_pixel_values: Optional[paddle.Tensor] = None,
|
|
conditional_embeddings: Optional[paddle.Tensor] = None,
|
|
attention_mask: Optional[paddle.Tensor] = None,
|
|
position_ids: Optional[paddle.Tensor] = None,
|
|
labels: Optional[paddle.Tensor] = None,
|
|
output_attentions: Optional[bool] = None,
|
|
output_hidden_states: Optional[bool] = None,
|
|
return_dict: Optional[bool] = None,
|
|
) -> Union[Tuple, CLIPSegOutput]:
|
|
r"""
|
|
labels (`paddle.Tensor` of shape `(batch_size,)`, *optional*):
|
|
Labels for computing the sequence 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).
|
|
Returns:
|
|
Examples:
|
|
```python
|
|
>>> from paddlenlp.transformers import AutoProcessor, CLIPSegForImageSegmentation
|
|
>>> from PIL import Image
|
|
>>> import requests
|
|
>>> processor = AutoProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
|
|
>>> model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")
|
|
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
|
|
>>> image = Image.open(requests.get(url, stream=True).raw)
|
|
>>> texts = ["a cat", "a remote", "a blanket"]
|
|
>>> inputs = processor(text=texts, images=[image] * len(texts), padding=True, return_tensors="pd")
|
|
>>> outputs = model(**inputs)
|
|
>>> logits = outputs.logits
|
|
>>> print(logits.shape)
|
|
[3, 352, 352]
|
|
```"""
|
|
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
|
|
|
# step 1: forward the query images through the frozen CLIP vision encoder
|
|
with paddle.no_grad():
|
|
vision_outputs = self.clip.vision_model(
|
|
pixel_values=pixel_values,
|
|
output_attentions=output_attentions,
|
|
output_hidden_states=True, # we need the intermediate hidden states
|
|
return_dict=return_dict,
|
|
)
|
|
pooled_output = self.clip.visual_projection(vision_outputs[1])
|
|
|
|
hidden_states = vision_outputs.hidden_states if return_dict else vision_outputs[2]
|
|
# we add +1 here as the hidden states also include the initial embeddings
|
|
activations = [hidden_states[i + 1] for i in self.extract_layers]
|
|
|
|
# update vision_outputs
|
|
if return_dict:
|
|
vision_outputs = BaseModelOutputWithPooling(
|
|
last_hidden_state=vision_outputs.last_hidden_state,
|
|
pooler_output=vision_outputs.pooler_output,
|
|
hidden_states=vision_outputs.hidden_states if output_hidden_states else None,
|
|
attentions=vision_outputs.attentions,
|
|
)
|
|
else:
|
|
vision_outputs = (
|
|
vision_outputs[:2] + vision_outputs[3:] if not output_hidden_states else vision_outputs
|
|
)
|
|
|
|
# step 2: compute conditional embeddings, either from text, images or an own provided embedding
|
|
if conditional_embeddings is None:
|
|
conditional_embeddings = self.get_conditional_embeddings(
|
|
batch_size=pixel_values.shape[0],
|
|
input_ids=input_ids,
|
|
attention_mask=attention_mask,
|
|
position_ids=position_ids,
|
|
conditional_pixel_values=conditional_pixel_values,
|
|
)
|
|
else:
|
|
if conditional_embeddings.shape[0] != pixel_values.shape[0]:
|
|
raise ValueError(
|
|
"Make sure to pass as many conditional embeddings as there are query images in the batch"
|
|
)
|
|
if conditional_embeddings.shape[1] != self.config.projection_dim:
|
|
raise ValueError(
|
|
"Make sure that the feature dimension of the conditional embeddings matches"
|
|
" `config.projection_dim`."
|
|
)
|
|
|
|
# step 3: forward both the pooled output and the activations through the lightweight decoder to predict masks
|
|
decoder_outputs = self.decoder(
|
|
activations,
|
|
conditional_embeddings,
|
|
output_attentions=output_attentions,
|
|
output_hidden_states=output_hidden_states,
|
|
return_dict=return_dict,
|
|
)
|
|
logits = decoder_outputs.logits if return_dict else decoder_outputs[0]
|
|
|
|
loss = None
|
|
if labels is not None:
|
|
loss_fn = nn.BCEWithLogitsLoss()
|
|
loss = loss_fn(logits, labels)
|
|
|
|
if not return_dict:
|
|
output = (logits, conditional_embeddings, pooled_output, vision_outputs, decoder_outputs)
|
|
return ((loss,) + output) if loss is not None else output
|
|
|
|
return CLIPSegImageSegmentationOutput(
|
|
loss=loss,
|
|
logits=logits,
|
|
conditional_embeddings=conditional_embeddings,
|
|
pooled_output=pooled_output,
|
|
vision_model_output=vision_outputs,
|
|
decoder_output=decoder_outputs,
|
|
)
|