1、 环境搭建
conda create -n blip python=3.8 -y
conda activate blip
pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html
pip install torch==1.9.1+cu102 torchvision==0.10.1+cu102 torchaudio===0.9.1 -f https://download.pytorch.org/whl/torch_stable.html -i http://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com
# CUDA 10.2
conda install pytorch==1.9.1 torchvision==0.10.1 torchaudio==0.9.1 cudatoolkit=10.2 -c pytorch
pip3 install torch==1.8.2+cu102 torchvision==0.9.2+cu102 torchaudio===0.8.2 -f https://download.pytorch.org/whl/lts/1.8/torch_lts.html -i http://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com
# CUDA 117
pip install torch==1.13.0+cu117 torchvision==0.14.0+cu117 torchaudio==0.13.0 --extra-index-url https://download.pytorch.org/whl/cu117 -i http://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com
import torch
if torch.cuda.is_available():
print("GPU 可用")
else:
print("GPU 不可用")
cd /media/lhy/ImageRetrieval/BLIP-main
pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install ruamel.yaml==0.15.87 -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install opencv-python -i https://pypi.tuna.tsinghua.edu.cn/simple
2、 测试
1、下载预训练模型model_base_retrieval_coco.pth
2、下载语料库bert-base-uncased
3、下载2014coco数据val2014.zip
4、Image-Text Retrieval
Download COCO and Flickr30k datasets from the original websites, and set ‘image_root’ in configs/retrieval_{dataset}.yaml accordingly.
To evaluate the finetuned BLIP model on COCO, run:
python -m torch.distributed.run --nproc_per_node=8 train_retrieval.py \
--config ./configs/retrieval_coco.yaml \
--output_dir output/retrieval_coco \
--evaluate
我们的环境测试：
python -m torch.distributed.run --nproc_per_node=2 train_retrieval.py \
--config ./configs/retrieval_coco.yaml \
--output_dir output/retrieval_coco \
--evaluate
要使用 8 个 A100 GPU 微调预训练检查点，请首先在 configs/retrieval_coco.yaml 中将“预训练”设置为“ https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_base.pth ”。然后运行：
python -m torch.distributed.run --nproc_per_node=8 train_retrieval.py \
--config ./configs/retrieval_coco.yaml \
--output_dir 输出/retrieval_coco
3、算法代码解析
blip_retrieval模型
model = blip_retrieval(pretrained=config['pretrained'], image_size=config['image_size'], vit=config['vit'],
vit_grad_ckpt=config['vit_grad_ckpt'], vit_ckpt_layer=config['vit_ckpt_layer'],
queue_size=config['queue_size'], negative_all_rank=config['negative_all_rank'])
（1）图像编码器：提取图片特征
self.visual_encoder, vision_width = create_vit(vit,image_size, vit_grad_ckpt, vit_ckpt_layer)
进入create_vit函数
def create_vit(vit, image_size, use_grad_checkpointing=False, ckpt_layer=0, drop_path_rate=0):
assert vit in ['base', 'large'], "vit parameter must be base or large"
if vit=='base':
vision_width = 768
visual_encoder = VisionTransformer(img_size=image_size, patch_size=16, embed_dim=vision_width, depth=12,
num_heads=12, use_grad_checkpointing=use_grad_checkpointing, ckpt_layer=ckpt_layer,
drop_path_rate=0 or drop_path_rate
)
elif vit=='large':
vision_width = 1024
visual_encoder = VisionTransformer(img_size=image_size, patch_size=16, embed_dim=vision_width, depth=24,
num_heads=16, use_grad_checkpointing=use_grad_checkpointing, ckpt_layer=ckpt_layer,
drop_path_rate=0.1 or drop_path_rate
)
return visual_encoder, vision_width
进入VisionTransformer函数
class VisionTransformer(nn.Module):
""" Vision Transformer
A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale` -
https://arxiv.org/abs/2010.11929
"""
def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,
num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None, representation_size=None,
drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=None,
use_grad_checkpointing=False, ckpt_layer=0):
"""
Args:
img_size (int, tuple): input image size
patch_size (int, tuple): patch size
in_chans (int): number of input channels
num_classes (int): number of classes for classification head
embed_dim (int): embedding dimension
depth (int): depth of transformer
num_heads (int): number of attention heads
mlp_ratio (int): ratio of mlp hidden dim to embedding dim
qkv_bias (bool): enable bias for qkv if True
qk_scale (float): override default qk scale of head_dim ** -0.5 if set
representation_size (Optional[int]): enable and set representation layer (pre-logits) to this value if set
drop_rate (float): dropout rate
attn_drop_rate (float): attention dropout rate
drop_path_rate (float): stochastic depth rate
norm_layer: (nn.Module): normalization layer
"""
super().__init__()
self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models
norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
self.patch_embed = PatchEmbed(
img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
num_patches = self.patch_embed.num_patches
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
self.pos_drop = nn.Dropout(p=drop_rate)
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule
self.blocks = nn.ModuleList([
Block(
dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
use_grad_checkpointing=(use_grad_checkpointing and i>=depth-ckpt_layer)
)
for i in range(depth)])
self.norm = norm_layer(embed_dim)
trunc_normal_(self.pos_embed, std=.02)
trunc_normal_(self.cls_token, std=.02)
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
@torch.jit.ignore
def no_weight_decay(self):
return {'pos_embed', 'cls_token'}
def forward(self, x, register_blk=-1):
B = x.shape[0]
x = self.patch_embed(x)
cls_tokens = self.cls_token.expand(B, -1, -1) # stole cls_tokens impl from Phil Wang, thanks
x = torch.cat((cls_tokens, x), dim=1)
x = x + self.pos_embed[:,:x.size(1),:]
x = self.pos_drop(x)
for i,blk in enumerate(self.blocks):
x = blk(x, register_blk==i)
x = self.norm(x)
return x
@torch.jit.ignore()
def load_pretrained(self, checkpoint_path, prefix=''):
_load_weights(self, checkpoint_path, prefix)
ViT架构，如下图所示。ViT（Vision Transformer）是一种基于Transformer的图像分类模型，它将图像看作是一个序列，通过自注意力机制来提取图像的特征表示。ViT的图像特征提取过程可以分为以下几个步骤：1.图像划分：将输入的图像划分为一组固定大小的图像块，每个图像块都被视为一个序列元素。2.嵌入层：对于每个图像块，使用一个嵌入层将其转换为一个向量表示。这个嵌入层通常是一个全连接层或卷积层。3.位置编码：为了保留图像中像素之间的位置信息，ViT使用一种位置编码方法来为每个图像块添加位置信息。这个位置编码通常是一个正弦函数或余弦函数。4.Transformer编码器：为了执行分类，使用额外的[CLS] token来表示全局的图像特征，将嵌入向量和位置编码向量输入到Transformer编码器中，通过多层Transformer编码器来提取图像的特征表示。在每个Transformer编码器中，使用自注意力机制来捕捉图像块之间的关联信息，并使用前馈神经网络来提取高级特征。5.全局池化：最后，将所有Transformer编码器的输出向量进行平均池化或最大池化，得到整个图像的特征表示。通过这种方式，ViT可以将图像转换为一个序列，并使用Transformer编码器来提取图像的特征表示。这种方法不仅可以用于图像分类，还可以用于其他视觉-语言任务，如图像标注、视觉问答等。
1、图像划分
self.patch_embed = PatchEmbed(
img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
class PatchEmbed(nn.Module):
""" 2D Image to Patch Embedding
"""
def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, norm_layer=None, flatten=True):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
self.img_size = img_size
self.patch_size = patch_size
self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
self.num_patches = self.grid_size[0] * self.grid_size[1]
self.flatten = flatten
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
def forward(self, x):
B, C, H, W = x.shape
assert H == self.img_size[0] and W == self.img_size[1], \
f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
x = self.proj(x)
if self.flatten:
x = x.flatten(2).transpose(1, 2) # BCHW -> BNC
x = self.norm(x)
return x
2、Transformer编码器
self.blocks = nn.ModuleList([
Block(
dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
use_grad_checkpointing=(use_grad_checkpointing and i>=depth-ckpt_layer)
)
for i in range(depth)])
class Block(nn.Module):
def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, use_grad_checkpointing=False):
super().__init__()
self.norm1 = norm_layer(dim)
self.attn = Attention(
dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
# NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
if use_grad_checkpointing:
self.attn = checkpoint_wrapper(self.attn)
self.mlp = checkpoint_wrapper(self.mlp)
def forward(self, x, register_hook=False):
x = x + self.drop_path(self.attn(self.norm1(x), register_hook=register_hook))
x = x + self.drop_path(self.mlp(self.norm2(x)))
return x
进行Attention函数
class Attention(nn.Module):
def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
# NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
self.scale = qk_scale or head_dim ** -0.5
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.attn_gradients = None
self.attention_map = None
def save_attn_gradients(self, attn_gradients):
self.attn_gradients = attn_gradients
def get_attn_gradients(self):
return self.attn_gradients
def save_attention_map(self, attention_map):
self.attention_map = attention_map
def get_attention_map(self):
return self.attention_map
def forward(self, x, register_hook=False):
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
if register_hook:
self.save_attention_map(attn)
attn.register_hook(self.save_attn_gradients)
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
进行入MLP函数
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
class Mlp(nn.Module):
""" MLP as used in Vision Transformer, MLP-Mixer and related networks
"""
def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop = nn.Dropout(drop)
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
（2）文本编码器：提取文本特征
self.tokenizer = init_tokenizer()
def init_tokenizer():
tokenizer = BertTokenizer.from_pretrained('checkpoints/bert-base-uncased')#下载文件
tokenizer.add_special_tokens({'bos_token':'[DEC]'})
tokenizer.add_special_tokens({'additional_special_tokens':['[ENC]']})
tokenizer.enc_token_id = tokenizer.additional_special_tokens_ids[0]
return tokenizer
self.text_encoder = BertModel(config=med_config, add_pooling_layer=False)
class BertModel(BertPreTrainedModel):
"""
The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
cross-attention is added between the self-attention layers, following the architecture described in `Attention is
all you need <https://arxiv.org/abs/1706.03762>`__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an
input to the forward pass.
"""
def __init__(self, config, add_pooling_layer=True):
super().__init__(config)
self.config = config
self.embeddings = BertEmbeddings(config)
self.encoder = BertEncoder(config)
self.pooler = BertPooler(config) if add_pooling_layer else None
self.init_weights()
def get_input_embeddings(self):
return self.embeddings.word_embeddings
def set_input_embeddings(self, value):
self.embeddings.word_embeddings = value
def _prune_heads(self, heads_to_prune):
"""
Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
class PreTrainedModel
"""
for layer, heads in heads_to_prune.items():
self.encoder.layer[layer].attention.prune_heads(heads)
def get_extended_attention_mask(self, attention_mask: Tensor, input_shape: Tuple[int], device: device, is_decoder: bool) -> Tensor:
"""
Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
Arguments:
attention_mask (:obj:`torch.Tensor`):
Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
input_shape (:obj:`Tuple[int]`):
The shape of the input to the model.
device: (:obj:`torch.device`):
The device of the input to the model.
Returns:
:obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.
"""
# 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.
if attention_mask.dim() == 3:
extended_attention_mask = attention_mask[:, None, :, :]
elif attention_mask.dim() == 2:
# Provided a padding mask of dimensions [batch_size, seq_length]
# - if the model is a decoder, apply a causal mask in addition to the padding mask
# - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
if is_decoder:
batch_size, seq_length = input_shape
seq_ids = torch.arange(seq_length, device=device)
causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
# in case past_key_values are used we need to add a prefix ones mask to the causal mask
# causal and attention masks must have same type with pytorch version < 1.3
causal_mask = causal_mask.to(attention_mask.dtype)
if causal_mask.shape[1] < attention_mask.shape[1]:
prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]
causal_mask = torch.cat(
[
torch.ones((batch_size, seq_length, prefix_seq_len), device=device, dtype=causal_mask.dtype),
causal_mask,
],
axis=-1,
)
extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
else:
extended_attention_mask = attention_mask[:, None, None, :]
else:
raise ValueError(
"Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
input_shape, attention_mask.shape
)
)
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
# masked positions, this operation will create a tensor which is 0.0 for
# positions we want to attend and -10000.0 for masked positions.
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
extended_attention_mask = extended_attention_mask.to(dtype=self.dtype) # fp16 compatibility
extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
return extended_attention_mask
def forward(
self,
input_ids=None,
attention_mask=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
encoder_embeds=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_values=None,
use_cache=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
is_decoder=False,
mode='multimodal',
):
r"""
encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
the model is configured as a decoder.
encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
(those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
use_cache (:obj:`bool`, `optional`):
If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
decoding (see :obj:`past_key_values`).
"""
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 is_decoder:
use_cache = use_cache if use_cache is not None else self.config.use_cache
else:
use_cache = False
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:
input_shape = input_ids.size()
batch_size, seq_length = input_shape
device = input_ids.device
elif inputs_embeds is not None:
input_shape = inputs_embeds.size()[:-1]
batch_size, seq_length = input_shape
device = inputs_embeds.device
elif encoder_embeds is not None:
input_shape = encoder_embeds.size()[:-1]
batch_size, seq_length = input_shape
device = encoder_embeds.device
else:
raise ValueError("You have to specify either input_ids or inputs_embeds or encoder_embeds")
# past_key_values_length
past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
if attention_mask is None:
attention_mask = torch.ones(((batch_size, seq_length + past_key_values_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,
device, is_decoder)
# If a 2D or 3D attention mask is provided for the cross-attention
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
if encoder_hidden_states is not None:
if type(encoder_hidden_states) == list:
encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[0].size()
else:
encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
if type(encoder_attention_mask) == list:
encoder_extended_attention_mask = [self.invert_attention_mask(mask) for mask in encoder_attention_mask]
elif encoder_attention_mask is None:
encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
else:
encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
else:
encoder_extended_attention_mask = None
# Prepare head mask if needed
# 1.0 in head_mask indicate we keep the head
# attention_probs has shape bsz x n_heads x N x N
# input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
# and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
if encoder_embeds is None:
embedding_output = self.embeddings(
input_ids=input_ids,
position_ids=position_ids,
inputs_embeds=inputs_embeds,
past_key_values_length=past_key_values_length,
)
else:
embedding_output = encoder_embeds
encoder_outputs = self.encoder(
embedding_output,
attention_mask=extended_attention_mask,
head_mask=head_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_extended_attention_mask,
past_key_values=past_key_values,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
mode=mode,
)
sequence_output = encoder_outputs[0]
pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
if not return_dict:
return (sequence_output, pooled_output) + encoder_outputs[1:]
return BaseModelOutputWithPoolingAndCrossAttentions(
last_hidden_state=sequence_output,
pooler_output=pooled_output,
past_key_values=encoder_outputs.past_key_values,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
cross_attentions=encoder_outputs.cross_attentions,
)