Upload model files

app.py

@@ -1,7 +1,45 @@
 import gradio as gr
+import torch
+import numpy as np
+from modules.models import *
+from util import get_prompt_template
+from PIL import Image
+
 
 def greet(name):
     return "Hello " + name + "!!"
 
-iface = gr.Interface(fn=greet, inputs="text", outputs="text")
-iface.launch()
+
+def main():
+    device = "cuda:0" if torch.cuda.is_available() else "cpu"
+
+    # Get model
+    model_conf_file = f'./config/model/ACL_ViT16.yaml'
+    model = ACL(model_conf_file, device)
+    model.train(False)
+    model.load('./pretrain/Param_best.pth')
+
+    # Get placeholder text
+    prompt_template, text_pos_at_prompt, prompt_length = get_prompt_template()
+
+    # Input pre processing
+
+    # Inference
+    placeholder_tokens = model.get_placeholder_token(prompt_template.replace('{}', ''))
+    # audio_driven_embedding = model.encode_audio(audios.to(model.device), placeholder_tokens, text_pos_at_prompt,
+    #                                             prompt_length)
+
+    # Localization result
+    # out_dict = model(images.to(model.device), audio_driven_embedding, 352)
+    # seg = out_dict['heatmap'][j:j + 1]
+    # seg_image = ((1 - seg.squeeze().detach().cpu().numpy()) * 255).astype(np.uint8)
+    # seg_image = Image.fromarray(seg_image)
+    heatmap_image = cv2.applyColorMap(np.array(seg_image), cv2.COLORMAP_JET)
+    # overlaid_image = cv2.addWeighted(np.array(original_image), 0.5, heatmap_image, 0.5, 0)
+
+
+if __name__ == "__main__":
+    iface = gr.Interface(fn=greet, inputs="text", outputs="text")
+    iface.launch()
+
+    main()

config/model/ACL_ViT16.yaml

@@ -0,0 +1,89 @@
+model:
+  clip: ViT16
+  vision_backbone: null
+  audio_backbone: BEATs
+  audio_proj: FGA512
+
+pretrain:
+  vision_backbone: null
+  audio_backbone: ./pretrain/BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt
+  audio_proj: null
+
+fga_conf:
+  FGA:
+    input_size: 768
+    output_size: 768
+
+  FGA512:
+    input_size: 768
+    output_size: 512
+
+clip_conf:
+  RN50:
+    name: RN50
+    vision:
+      image_resolution: 224
+      vision_layers: [3, 4, 6, 3]
+      vision_width: 64
+      heads: 8
+      vision_patch_size: null
+    text:
+      transformer_layers: 12
+      transformer_width: 512
+      transformer_heads: 8
+      vocab_size: 49408
+      context_length: 77
+    embedding_dim: 1024
+
+  ViT16:
+    name: ViT-B/16
+    vision:
+      image_resolution: 224
+      vision_layers: 12
+      vision_width: 768
+      heads: 12
+      vision_patch_size: 16
+    text:
+      transformer_layers: 12
+      transformer_width: 512
+      transformer_heads: 8
+      vocab_size: 49408
+      context_length: 77
+    embedding_dim: 512
+
+  ViT14:
+    name: ViT-L/14
+    vision:
+      image_resolution: 224
+      vision_layers: 24
+      vision_width: 1024
+      heads: 16
+      vision_patch_size: 14
+    text:
+      transformer_layers: 12
+      transformer_width: 768
+      transformer_heads: 12
+      vocab_size: 49408
+      context_length: 77
+    embedding_dim: 768
+
+vision_backbone_conf:
+  maskclip_plus_rn50_512:
+    name: maskclip_plus_rn50_512
+    image_resolution: 512
+    vision_layers: [ 3, 4, 6, 3 ]
+    vision_width: 2048
+    aspp:
+      dilations: [ 6, 12, 18, 24 ]
+      in_channels: 2048
+      channels: 512
+
+  maskclip_plus_rn101_512:
+    name: maskclip_plus_rn101_512
+    image_resolution: 512
+    vision_layers: [ 3, 4, 23, 3 ]
+    vision_width: 2048
+    aspp:
+      dilations: [ 6, 12, 18, 24 ]
+      in_channels: 2048
+      channels: 1024

config/train/Exp_ACL_v1.yaml

@@ -0,0 +1,45 @@
+model: ACL
+
+common:
+  train_data: vggss
+  epoch: 20
+  batch_size: 8
+  input_resolution: 352
+  num_workers: 4
+  seed: 0
+  loss:
+    - acl_i
+    - acl_f
+    - area_reg
+  loss_w:
+    - 1
+    - 1
+    - 1
+
+optimizer: Adam
+scheduler: null
+amp: True
+
+optim_conf:
+  Adam:
+    module_path: torch.optim
+    module_name: Adam
+    lr: 0.0001
+    weight_decay: 0.0001
+
+  AdamW:
+    module_path: torch.optim
+    module_name: AdamW
+    lr: 0.001
+
+  SGDR:
+    module_path: torch.optim
+    module_name: SGD
+    lr: 0.5
+    weight_decay: 0.00001
+
+sched_conf:
+  Cosine:
+    module_path: torch.optim.lr_scheduler
+    module_name: CosineAnnealingLR
+    eta_ratio: 0.0

modules/AudioToken/AudioToken.py

@@ -0,0 +1,100 @@
+import torch
+from diffusers.loaders import AttnProcsLayers
+
+from modules.BEATs.BEATs import BEATs, BEATsConfig
+from modules.AudioToken.embedder import FGAEmbedder
+from diffusers import AutoencoderKL, UNet2DConditionModel
+from diffusers.models.attention_processor import LoRAAttnProcessor
+
+
+class AudioTokenWrapper(torch.nn.Module):
+    """Simple wrapper module for Stable Diffusion that holds all the models together"""
+
+    def __init__(
+        self,
+        args,
+        accelerator,
+    ):
+
+        super().__init__()
+        # Load scheduler and models
+        from modules.clip_text_model.modeling_clip import CLIPTextModel
+        self.text_encoder = CLIPTextModel.from_pretrained(
+            args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
+        )
+        self.unet = UNet2DConditionModel.from_pretrained(
+            args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision
+        )
+        self.vae = AutoencoderKL.from_pretrained(
+            args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision
+        )
+
+        checkpoint = torch.load(
+            'models/BEATs/BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt')
+        cfg = BEATsConfig(checkpoint['cfg'])
+        self.aud_encoder = BEATs(cfg)
+        self.aud_encoder.load_state_dict(checkpoint['model'])
+        self.aud_encoder.predictor = None
+        input_size = 768 * 3
+
+        if args.pretrained_model_name_or_path == "CompVis/stable-diffusion-v1-4":
+            self.embedder = FGAEmbedder(input_size=input_size, output_size=768)
+
+        else:
+            self.embedder = FGAEmbedder(input_size=input_size, output_size=1024)
+
+        self.vae.eval()
+        self.unet.eval()
+        self.text_encoder.eval()
+        self.aud_encoder.eval()
+
+        if 'lora' in args and args.lora:
+            # Set correct lora layers
+            lora_attn_procs = {}
+            for name in self.unet.attn_processors.keys():
+                cross_attention_dim = None if name.endswith(
+                    "attn1.processor") else self.unet.config.cross_attention_dim
+                if name.startswith("mid_block"):
+                    hidden_size = self.unet.config.block_out_channels[-1]
+                elif name.startswith("up_blocks"):
+                    block_id = int(name[len("up_blocks.")])
+                    hidden_size = list(reversed(self.unet.config.block_out_channels))[block_id]
+                elif name.startswith("down_blocks"):
+                    block_id = int(name[len("down_blocks.")])
+                    hidden_size = self.unet.config.block_out_channels[block_id]
+
+                lora_attn_procs[name] = LoRAAttnProcessor(hidden_size=hidden_size,
+                                                          cross_attention_dim=cross_attention_dim)
+
+            self.unet.set_attn_processor(lora_attn_procs)
+            self.lora_layers = AttnProcsLayers(self.unet.attn_processors)
+
+        if args.data_set == 'train':
+
+            # Freeze vae, unet, text_enc and aud_encoder
+            self.vae.requires_grad_(False)
+            self.unet.requires_grad_(False)
+            self.text_encoder.requires_grad_(False)
+            self.aud_encoder.requires_grad_(False)
+            self.embedder.requires_grad_(True)
+            self.embedder.train()
+
+            if 'lora' in args and args.lora:
+                self.unet.train()
+
+        if args.data_set == 'test':
+
+            from transformers import CLIPTextModel
+            self.text_encoder = CLIPTextModel.from_pretrained(
+                args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
+            )
+
+            self.embedder.eval()
+            embedder_learned_embeds = args.learned_embeds
+            self.embedder.load_state_dict(torch.load(embedder_learned_embeds, map_location=accelerator.device))
+
+            if 'lora' in args and args.lora:
+                self.lora_layers.eval()
+                lora_layers_learned_embeds = args.lora_learned_embeds
+                self.lora_layers.load_state_dict(torch.load(lora_layers_learned_embeds, map_location=accelerator.device))
+                self.unet.load_attn_procs(lora_layers_learned_embeds)

modules/AudioToken/embedder.py

@@ -0,0 +1,17 @@
+import torch.nn as nn
+from modules.FGA.atten import Atten
+
+class FGAEmbedder(nn.Module):
+    def __init__(self, input_size=768*3, output_size=768):
+        super(FGAEmbedder, self).__init__()
+        self.fc1 = nn.Linear(input_size, input_size)
+        self.fc2 = nn.Linear(input_size, output_size)
+        self.gelu = nn.GELU()
+        self.fga = Atten(util_e=[output_size], pairwise_flag=False)
+
+    def forward(self, audio_embs):
+        audio_embs = self.fc1(audio_embs)
+        audio_embs = self.gelu(audio_embs)
+        audio_embs = self.fc2(audio_embs)
+        attend = self.fga([audio_embs])[0]
+        return attend

modules/BEATs/BEATs.py

@@ -0,0 +1,180 @@
+# --------------------------------------------------------
+# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)
+# Github source: https://github.com/microsoft/unilm/tree/master/beats
+# Copyright (c) 2022 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Based on fairseq code bases
+# https://github.com/pytorch/fairseq
+# --------------------------------------------------------
+
+
+import torch
+import torch.nn as nn
+from torch.nn import LayerNorm
+import torchaudio.compliance.kaldi as ta_kaldi
+from torch.cuda.amp import autocast
+
+from modules.BEATs.backbone import (
+    TransformerEncoder,
+)
+
+import logging
+from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+
+class BEATsConfig:
+    def __init__(self, cfg=None):
+        self.input_patch_size: int = -1  # path size of patch embedding
+        self.embed_dim: int = 512  # patch embedding dimension
+        self.conv_bias: bool = False  # include bias in conv encoder
+
+        self.encoder_layers: int = 12  # num encoder layers in the transformer
+        self.encoder_embed_dim: int = 768  # encoder embedding dimension
+        self.encoder_ffn_embed_dim: int = 3072  # encoder embedding dimension for FFN
+        self.encoder_attention_heads: int = 12  # num encoder attention heads
+        self.activation_fn: str = "gelu"  # activation function to use
+
+        self.layer_wise_gradient_decay_ratio: float = 1.0  # ratio for layer-wise gradient decay
+        self.layer_norm_first: bool = False  # apply layernorm first in the transformer
+        self.deep_norm: bool = False  # apply deep_norm first in the transformer
+
+        # dropouts
+        self.dropout: float = 0.1  # dropout probability for the transformer
+        self.attention_dropout: float = 0.1  # dropout probability for attention weights
+        self.activation_dropout: float = 0.0  # dropout probability after activation in FFN
+        self.encoder_layerdrop: float = 0.0  # probability of dropping a tarnsformer layer
+        self.dropout_input: float = 0.0  # dropout to apply to the input (after feat extr)
+
+        # positional embeddings
+        self.conv_pos: int = 128  # number of filters for convolutional positional embeddings
+        self.conv_pos_groups: int = 16  # number of groups for convolutional positional embedding
+
+        # relative position embedding
+        self.relative_position_embedding: bool = False  # apply relative position embedding
+        self.num_buckets: int = 320  # number of buckets for relative position embedding
+        self.max_distance: int = 1280  # maximum distance for relative position embedding
+        self.gru_rel_pos: bool = False  # apply gated relative position embedding
+
+        # label predictor
+        self.finetuned_model: bool = False  # whether the model is a fine-tuned model.
+        self.predictor_dropout: float = 0.1  # dropout probability for the predictor
+        self.predictor_class: int = 527  # target class number for the predictor
+
+        if cfg is not None:
+            self.update(cfg)
+
+    def update(self, cfg: dict):
+        self.__dict__.update(cfg)
+
+
+class BEATs(nn.Module):
+    def __init__(
+            self,
+            cfg: BEATsConfig,
+    ) -> None:
+        super().__init__()
+        logger.info(f"BEATs Config: {cfg.__dict__}")
+
+        self.cfg = cfg
+
+        self.embed = cfg.embed_dim
+        self.post_extract_proj = (
+            nn.Linear(self.embed, cfg.encoder_embed_dim)
+            if self.embed != cfg.encoder_embed_dim
+            else None
+        )
+
+        self.input_patch_size = cfg.input_patch_size
+        self.patch_embedding = nn.Conv2d(1, self.embed, kernel_size=self.input_patch_size, stride=self.input_patch_size,
+                                         bias=cfg.conv_bias)
+
+        self.dropout_input = nn.Dropout(cfg.dropout_input)
+
+        assert not cfg.deep_norm or not cfg.layer_norm_first
+        self.encoder = TransformerEncoder(cfg)
+        self.layer_norm = LayerNorm(self.embed)
+
+        if cfg.finetuned_model:
+            self.predictor_dropout = nn.Dropout(cfg.predictor_dropout)
+            self.predictor = nn.Linear(cfg.encoder_embed_dim, cfg.predictor_class)
+        else:
+            self.predictor = None
+
+    def forward_padding_mask(
+            self,
+            features: torch.Tensor,
+            padding_mask: torch.Tensor,
+    ) -> torch.Tensor:
+        extra = padding_mask.size(1) % features.size(1)
+        if extra > 0:
+            padding_mask = padding_mask[:, :-extra]
+        padding_mask = padding_mask.view(
+            padding_mask.size(0), features.size(1), -1
+        )
+        padding_mask = padding_mask.all(-1)
+        return padding_mask
+
+    @autocast(enabled=False)
+    def preprocess(
+            self,
+            source: torch.Tensor,
+            fbank_mean: float = 15.41663,
+            fbank_std: float = 6.55582,
+    ) -> torch.Tensor:
+        fbanks = []
+        for waveform in source:
+            waveform = waveform.unsqueeze(0) * 2 ** 15
+            fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)
+            fbanks.append(fbank)
+        fbank = torch.stack(fbanks, dim=0)
+        fbank = (fbank - fbank_mean) / (2 * fbank_std)
+        return fbank
+
+    def extract_features(
+            self,
+            source: torch.Tensor,
+            padding_mask: Optional[torch.Tensor] = None,
+            fbank_mean: float = 15.41663,
+            fbank_std: float = 6.55582,
+    ):
+        fbank = self.preprocess(source, fbank_mean=fbank_mean, fbank_std=fbank_std)
+        if padding_mask is not None:
+            padding_mask = self.forward_padding_mask(fbank, padding_mask)
+        # ToDo Aug here
+        fbank = fbank.unsqueeze(1)
+        features = self.patch_embedding(fbank)
+        features = features.reshape(features.shape[0], features.shape[1], -1)
+        features = features.transpose(1, 2)
+        features = self.layer_norm(features)
+
+        if padding_mask is not None:
+            padding_mask = self.forward_padding_mask(features, padding_mask)
+
+        if self.post_extract_proj is not None:
+            features = self.post_extract_proj(features)
+
+        x = self.dropout_input(features)
+
+        x, layers_sum, layers = self.encoder(
+            x,
+            padding_mask=padding_mask,
+        )
+
+        if self.predictor is not None:
+            x = self.predictor_dropout(x)
+            logits = self.predictor(x)
+
+            if padding_mask is not None and padding_mask.any():
+                logits[padding_mask] = 0
+                logits = logits.sum(dim=1)
+                logits = logits / (~padding_mask).sum(dim=1).unsqueeze(-1).expand_as(logits)
+            else:
+                logits = logits.mean(dim=1)
+
+            lprobs = torch.sigmoid(logits)
+
+            return lprobs, padding_mask
+        else:
+            return x, layers_sum, layers

modules/BEATs/Tokenizers.py

@@ -0,0 +1,172 @@
+# --------------------------------------------------------
+# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)
+# Github source: https://github.com/microsoft/unilm/tree/master/beats
+# Copyright (c) 2022 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Based on fairseq code bases
+# https://github.com/pytorch/fairseq
+# --------------------------------------------------------
+
+
+import torch
+import torch.nn as nn
+from torch.nn import LayerNorm
+import torchaudio.compliance.kaldi as ta_kaldi
+
+from modules.BEATs.backbone import (
+    TransformerEncoder,
+)
+from modules.BEATs.quantizer import (
+    NormEMAVectorQuantizer,
+)
+
+import logging
+from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+
+class TokenizersConfig:
+    def __init__(self, cfg=None):
+        self.input_patch_size: int = -1  # path size of patch embedding
+        self.embed_dim: int = 512  # patch embedding dimension
+        self.conv_bias: bool = False  # include bias in conv encoder
+
+        self.encoder_layers: int = 12  # num encoder layers in the transformer
+        self.encoder_embed_dim: int = 768  # encoder embedding dimension
+        self.encoder_ffn_embed_dim: int = 3072  # encoder embedding dimension for FFN
+        self.encoder_attention_heads: int = 12  # num encoder attention heads
+        self.activation_fn: str = "gelu"  # activation function to use
+
+        self.layer_norm_first: bool = False  # apply layernorm first in the transformer
+        self.deep_norm: bool = False  # apply deep_norm first in the transformer
+
+        # dropouts
+        self.dropout: float = 0.1  # dropout probability for the transformer
+        self.attention_dropout: float = 0.1  # dropout probability for attention weights
+        self.activation_dropout: float = 0.0  # dropout probability after activation in FFN
+        self.encoder_layerdrop: float = 0.0  # probability of dropping a tarnsformer layer
+        self.dropout_input: float = 0.0  # dropout to apply to the input (after feat extr)
+
+        # positional embeddings
+        self.conv_pos: int = 128  # number of filters for convolutional positional embeddings
+        self.conv_pos_groups: int = 16  # number of groups for convolutional positional embedding
+
+        # relative position embedding
+        self.relative_position_embedding: bool = False  # apply relative position embedding
+        self.num_buckets: int = 320  # number of buckets for relative position embedding
+        self.max_distance: int = 1280  # maximum distance for relative position embedding
+        self.gru_rel_pos: bool = False  # apply gated relative position embedding
+
+        # quantizer
+        self.quant_n: int = 1024 # codebook number in quantizer
+        self.quant_dim: int = 256    # codebook dimension in quantizer
+
+        if cfg is not None:
+            self.update(cfg)
+
+    def update(self, cfg: dict):
+        self.__dict__.update(cfg)
+
+
+class Tokenizers(nn.Module):
+    def __init__(
+            self,
+            cfg: TokenizersConfig,
+    ) -> None:
+        super().__init__()
+        logger.info(f"Tokenizers Config: {cfg.__dict__}")
+
+        self.cfg = cfg
+
+        self.embed = cfg.embed_dim
+        self.post_extract_proj = (
+            nn.Linear(self.embed, cfg.encoder_embed_dim)
+            if self.embed != cfg.encoder_embed_dim
+            else None
+        )
+
+        self.input_patch_size = cfg.input_patch_size
+        self.patch_embedding = nn.Conv2d(1, self.embed, kernel_size=self.input_patch_size, stride=self.input_patch_size,
+                                         bias=cfg.conv_bias)
+
+        self.dropout_input = nn.Dropout(cfg.dropout_input)
+
+        assert not cfg.deep_norm or not cfg.layer_norm_first
+        self.encoder = TransformerEncoder(cfg)
+        self.layer_norm = LayerNorm(self.embed)
+
+        self.quantize = NormEMAVectorQuantizer(
+            n_embed=cfg.quant_n, embedding_dim=cfg.quant_dim, beta=1.0, kmeans_init=True, decay=0.99,
+        )
+        self.quant_n = cfg.quant_n
+        self.quantize_layer = nn.Sequential(
+            nn.Linear(cfg.encoder_embed_dim, cfg.encoder_embed_dim),
+            nn.Tanh(),
+            nn.Linear(cfg.encoder_embed_dim, cfg.quant_dim)  # for quantize
+        )
+
+    def forward_padding_mask(
+            self,
+            features: torch.Tensor,
+            padding_mask: torch.Tensor,
+    ) -> torch.Tensor:
+        extra = padding_mask.size(1) % features.size(1)
+        if extra > 0:
+            padding_mask = padding_mask[:, :-extra]
+        padding_mask = padding_mask.view(
+            padding_mask.size(0), features.size(1), -1
+        )
+        padding_mask = padding_mask.all(-1)
+        return padding_mask
+
+    def preprocess(
+            self,
+            source: torch.Tensor,
+            fbank_mean: float = 15.41663,
+            fbank_std: float = 6.55582,
+    ) -> torch.Tensor:
+        fbanks = []
+        for waveform in source:
+            waveform = waveform.unsqueeze(0) * 2 ** 15
+            fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)
+            fbanks.append(fbank)
+        fbank = torch.stack(fbanks, dim=0)
+        fbank = (fbank - fbank_mean) / (2 * fbank_std)
+        return fbank
+
+    def extract_labels(
+            self,
+            source: torch.Tensor,
+            padding_mask: Optional[torch.Tensor] = None,
+            fbank_mean: float = 15.41663,
+            fbank_std: float = 6.55582,
+    ):
+        fbank = self.preprocess(source, fbank_mean=fbank_mean, fbank_std=fbank_std)
+
+        if padding_mask is not None:
+            padding_mask = self.forward_padding_mask(fbank, padding_mask)
+
+        fbank = fbank.unsqueeze(1)
+        features = self.patch_embedding(fbank)
+        features = features.reshape(features.shape[0], features.shape[1], -1)
+        features = features.transpose(1, 2)
+        features = self.layer_norm(features)
+
+        if padding_mask is not None:
+            padding_mask = self.forward_padding_mask(features, padding_mask)
+
+        if self.post_extract_proj is not None:
+            features = self.post_extract_proj(features)
+
+        x = self.dropout_input(features)
+
+        x, layer_results = self.encoder(
+            x,
+            padding_mask=padding_mask,
+        )
+
+        quantize_input = self.quantize_layer(x)
+        quantize_feature, embed_loss, embed_ind = self.quantize(quantize_input)
+
+        return embed_ind

modules/BEATs/backbone.py

@@ -0,0 +1,788 @@
+# --------------------------------------------------------
+# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)
+# Github source: https://github.com/microsoft/unilm/tree/master/beats
+# Copyright (c) 2022 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Based on fairseq code bases
+# https://github.com/pytorch/fairseq
+# --------------------------------------------------------
+
+import math
+import numpy as np
+from typing import Dict, Optional, Tuple
+import torch
+from torch import Tensor, nn
+import torch.nn.functional as F
+from torch.nn import LayerNorm, Parameter
+from modules.BEATs.modules import (
+    GradMultiply,
+    SamePad,
+    get_activation_fn,
+    GLU_Linear,
+    quant_noise,
+)
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+
+        self.dropout = args.dropout
+        self.embedding_dim = args.encoder_embed_dim
+
+        self.pos_conv = nn.Conv1d(
+            self.embedding_dim,
+            self.embedding_dim,
+            kernel_size=args.conv_pos,
+            padding=args.conv_pos // 2,
+            groups=args.conv_pos_groups,
+        )
+        dropout = 0
+        std = math.sqrt((4 * (1.0 - dropout)) / (args.conv_pos * self.embedding_dim))
+        nn.init.normal_(self.pos_conv.weight, mean=0, std=std)
+        nn.init.constant_(self.pos_conv.bias, 0)
+
+        self.pos_conv = nn.utils.weight_norm(self.pos_conv, name="weight", dim=2)
+        self.pos_conv = nn.Sequential(self.pos_conv, SamePad(args.conv_pos), nn.GELU())
+
+        if hasattr(args, "relative_position_embedding"):
+            self.relative_position_embedding = args.relative_position_embedding
+            self.num_buckets = args.num_buckets
+            self.max_distance = args.max_distance
+        else:
+            self.relative_position_embedding = False
+            self.num_buckets = 0
+            self.max_distance = 0
+
+        self.layers = nn.ModuleList(
+            [
+                TransformerSentenceEncoderLayer(
+                    embedding_dim=self.embedding_dim,
+                    ffn_embedding_dim=args.encoder_ffn_embed_dim,
+                    num_attention_heads=args.encoder_attention_heads,
+                    dropout=self.dropout,
+                    attention_dropout=args.attention_dropout,
+                    activation_dropout=args.activation_dropout,
+                    activation_fn=args.activation_fn,
+                    layer_norm_first=args.layer_norm_first,
+                    deep_norm=args.deep_norm,
+                    has_relative_attention_bias=self.relative_position_embedding,
+                    num_buckets=self.num_buckets,
+                    max_distance=self.max_distance,
+                    gru_rel_pos=args.gru_rel_pos,
+                    encoder_layers=args.encoder_layers,
+                )
+                for i in range(args.encoder_layers)
+            ]
+        )
+        if self.relative_position_embedding:
+            for i in range(1, args.encoder_layers):
+                del self.layers[i].self_attn.relative_attention_bias
+                self.layers[i].self_attn.relative_attention_bias = self.layers[0].self_attn.relative_attention_bias
+
+        self.layer_norm_first = args.layer_norm_first
+        self.layer_norm = LayerNorm(self.embedding_dim)
+        self.layerdrop = args.encoder_layerdrop
+
+        self.apply(init_bert_params)
+
+        if args.deep_norm:
+            deep_norm_beta = math.pow(8 * args.encoder_layers, -1 / 4)
+            for i in range(args.encoder_layers):
+                nn.init.xavier_normal_(self.layers[i].self_attn.k_proj.weight, gain=1)
+                nn.init.xavier_normal_(self.layers[i].self_attn.v_proj.weight, gain=deep_norm_beta)
+                nn.init.xavier_normal_(self.layers[i].self_attn.q_proj.weight, gain=1)
+                nn.init.xavier_normal_(self.layers[i].self_attn.out_proj.weight, gain=deep_norm_beta)
+                nn.init.xavier_normal_(self.layers[i].fc1.weight, gain=deep_norm_beta)
+                nn.init.xavier_normal_(self.layers[i].fc2.weight, gain=deep_norm_beta)
+
+        self.layer_wise_gradient_decay_ratio = getattr(args, "layer_wise_gradient_decay_ratio", 1)
+
+    def forward(self, x, padding_mask=None, layer=None):
+        x, layers_sum, layers = self.extract_features(x, padding_mask, layer)
+
+        if self.layer_norm_first and layer is None:
+            x = self.layer_norm(x)
+
+        return x, layers_sum, layers
+
+    def extract_features(self, x, padding_mask=None, tgt_layer=None):
+
+        if padding_mask is not None:
+            x[padding_mask] = 0
+
+        x_conv = self.pos_conv(x.transpose(1, 2))
+        x_conv = x_conv.transpose(1, 2)
+        x += x_conv
+
+        if not self.layer_norm_first:
+            x = self.layer_norm(x)
+
+        x = F.dropout(x, p=self.dropout, training=self.training)
+
+        # B x T x C -> T x B x C
+        x = x.transpose(0, 1)
+        layers = []
+
+        layer_results = []
+        z = None
+        if tgt_layer is not None:
+            layer_results.append((x, z))
+        r = None
+        pos_bias = None
+        for i, layer in enumerate(self.layers):
+            if self.layer_wise_gradient_decay_ratio != 1.0:
+                x = GradMultiply.apply(x, self.layer_wise_gradient_decay_ratio)
+            dropout_probability = np.random.random()
+            if not self.training or (dropout_probability > self.layerdrop):
+                x, z, pos_bias = layer(x, self_attn_padding_mask=padding_mask, need_weights=False, pos_bias=pos_bias)
+            if tgt_layer is not None:
+                layer_results.append((x, z))
+            if i == tgt_layer:
+                r = x
+                break
+            if i in [3, 7, 11]:
+                layers.append(x.transpose(0, 1))
+
+        if r is not None:
+            x = r
+
+        # T x B x C -> B x T x C
+        x = x.transpose(0, 1)
+        layers_cat = torch.cat(layers, dim=2)
+        # layers = layers[0] + layers[1] + layers[2]
+
+        return x, layers_cat, layers
+
+
+class TransformerSentenceEncoderLayer(nn.Module):
+    def __init__(
+            self,
+            embedding_dim: float = 768,
+            ffn_embedding_dim: float = 3072,
+            num_attention_heads: float = 8,
+            dropout: float = 0.1,
+            attention_dropout: float = 0.1,
+            activation_dropout: float = 0.1,
+            activation_fn: str = "relu",
+            layer_norm_first: bool = False,
+            deep_norm: bool = False,
+            has_relative_attention_bias: bool = False,
+            num_buckets: int = 0,
+            max_distance: int = 0,
+            rescale_init: bool = False,
+            gru_rel_pos: bool = False,
+            encoder_layers: int = 0,
+    ) -> None:
+
+        super().__init__()
+        self.embedding_dim = embedding_dim
+        self.dropout = dropout
+        self.activation_dropout = activation_dropout
+
+        self.activation_name = activation_fn
+        self.activation_fn = get_activation_fn(activation_fn)
+        self.self_attn = MultiheadAttention(
+            self.embedding_dim,
+            num_attention_heads,
+            dropout=attention_dropout,
+            self_attention=True,
+            has_relative_attention_bias=has_relative_attention_bias,
+            num_buckets=num_buckets,
+            max_distance=max_distance,
+            rescale_init=rescale_init,
+            gru_rel_pos=gru_rel_pos,
+        )
+
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(self.activation_dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.layer_norm_first = layer_norm_first
+
+        self.self_attn_layer_norm = LayerNorm(self.embedding_dim)
+
+        if self.activation_name == "glu":
+            self.fc1 = GLU_Linear(self.embedding_dim, ffn_embedding_dim, "swish")
+        else:
+            self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim)
+        self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim)
+
+        self.final_layer_norm = LayerNorm(self.embedding_dim)
+
+        self.deep_norm = deep_norm
+        if self.deep_norm:
+            self.deep_norm_alpha = math.pow(2 * encoder_layers, 1 / 4)
+        else:
+            self.deep_norm_alpha = 1
+
+    def forward(
+            self,
+            x: torch.Tensor,
+            self_attn_mask: torch.Tensor = None,
+            self_attn_padding_mask: torch.Tensor = None,
+            need_weights: bool = False,
+            pos_bias=None
+    ):
+        residual = x
+
+        if self.layer_norm_first:
+            x = self.self_attn_layer_norm(x)
+            x, attn, pos_bias = self.self_attn(
+                query=x,
+                key=x,
+                value=x,
+                key_padding_mask=self_attn_padding_mask,
+                need_weights=False,
+                attn_mask=self_attn_mask,
+                position_bias=pos_bias
+            )
+            x = self.dropout1(x)
+            x = residual + x
+
+            residual = x
+            x = self.final_layer_norm(x)
+            if self.activation_name == "glu":
+                x = self.fc1(x)
+            else:
+                x = self.activation_fn(self.fc1(x))
+            x = self.dropout2(x)
+            x = self.fc2(x)
+            x = self.dropout3(x)
+            x = residual + x
+        else:
+            x, attn, pos_bias = self.self_attn(
+                query=x,
+                key=x,
+                value=x,
+                key_padding_mask=self_attn_padding_mask,
+                need_weights=need_weights,
+                attn_mask=self_attn_mask,
+                position_bias=pos_bias
+            )
+
+            x = self.dropout1(x)
+            x = residual * self.deep_norm_alpha + x
+
+            x = self.self_attn_layer_norm(x)
+
+            residual = x
+            if self.activation_name == "glu":
+                x = self.fc1(x)
+            else:
+                x = self.activation_fn(self.fc1(x))
+            x = self.dropout2(x)
+            x = self.fc2(x)
+            x = self.dropout3(x)
+            x = residual * self.deep_norm_alpha + x
+            x = self.final_layer_norm(x)
+
+        return x, attn, pos_bias
+
+
+class MultiheadAttention(nn.Module):
+    """Multi-headed attention.
+
+    See "Attention Is All You Need" for more details.
+    """
+
+    def __init__(
+            self,
+            embed_dim,
+            num_heads,
+            kdim=None,
+            vdim=None,
+            dropout=0.0,
+            bias=True,
+            add_bias_kv=False,
+            add_zero_attn=False,
+            self_attention=False,
+            encoder_decoder_attention=False,
+            q_noise=0.0,
+            qn_block_size=8,
+            has_relative_attention_bias=False,
+            num_buckets=32,
+            max_distance=128,
+            gru_rel_pos=False,
+            rescale_init=False,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.kdim = kdim if kdim is not None else embed_dim
+        self.vdim = vdim if vdim is not None else embed_dim
+        self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim
+
+        self.num_heads = num_heads
+        self.dropout_module = nn.Dropout(dropout)
+
+        self.has_relative_attention_bias = has_relative_attention_bias
+        self.num_buckets = num_buckets
+        self.max_distance = max_distance
+        if self.has_relative_attention_bias:
+            self.relative_attention_bias = nn.Embedding(num_buckets, num_heads)
+
+        self.head_dim = embed_dim // num_heads
+        self.q_head_dim = self.head_dim
+        self.k_head_dim = self.head_dim
+        assert (
+                self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+        self.scaling = self.head_dim ** -0.5
+
+        self.self_attention = self_attention
+        self.encoder_decoder_attention = encoder_decoder_attention
+
+        assert not self.self_attention or self.qkv_same_dim, (
+            "Self-attention requires query, key and " "value to be of the same size"
+        )
+
+        k_bias = True
+        if rescale_init:
+            k_bias = False
+
+        k_embed_dim = embed_dim
+        q_embed_dim = embed_dim
+
+        self.k_proj = quant_noise(
+            nn.Linear(self.kdim, k_embed_dim, bias=k_bias), q_noise, qn_block_size
+        )
+        self.v_proj = quant_noise(
+            nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size
+        )
+        self.q_proj = quant_noise(
+            nn.Linear(embed_dim, q_embed_dim, bias=bias), q_noise, qn_block_size
+        )
+
+        self.out_proj = quant_noise(
+            nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size
+        )
+
+        if add_bias_kv:
+            self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim))
+            self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim))
+        else:
+            self.bias_k = self.bias_v = None
+
+        self.add_zero_attn = add_zero_attn
+
+        self.gru_rel_pos = gru_rel_pos
+        if self.gru_rel_pos:
+            self.grep_linear = nn.Linear(self.q_head_dim, 8)
+            self.grep_a = nn.Parameter(torch.ones(1, num_heads, 1, 1))
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        if self.qkv_same_dim:
+            # Empirically observed the convergence to be much better with
+            # the scaled initialization
+            nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2))
+            nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2))
+            nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2))
+        else:
+            nn.init.xavier_uniform_(self.k_proj.weight)
+            nn.init.xavier_uniform_(self.v_proj.weight)
+            nn.init.xavier_uniform_(self.q_proj.weight)
+
+        nn.init.xavier_uniform_(self.out_proj.weight)
+        if self.out_proj.bias is not None:
+            nn.init.constant_(self.out_proj.bias, 0.0)
+        if self.bias_k is not None:
+            nn.init.xavier_normal_(self.bias_k)
+        if self.bias_v is not None:
+            nn.init.xavier_normal_(self.bias_v)
+        if self.has_relative_attention_bias:
+            nn.init.xavier_normal_(self.relative_attention_bias.weight)
+
+    def _relative_positions_bucket(self, relative_positions, bidirectional=True):
+        num_buckets = self.num_buckets
+        max_distance = self.max_distance
+        relative_buckets = 0
+
+        if bidirectional:
+            num_buckets = num_buckets // 2
+            relative_buckets += (relative_positions > 0).to(torch.long) * num_buckets
+            relative_positions = torch.abs(relative_positions)
+        else:
+            relative_positions = -torch.min(relative_positions, torch.zeros_like(relative_positions))
+
+        max_exact = num_buckets // 2
+        is_small = relative_positions < max_exact
+
+        relative_postion_if_large = max_exact + (
+                torch.log(relative_positions.float() / max_exact)
+                / math.log(max_distance / max_exact)
+                * (num_buckets - max_exact)
+        ).to(torch.long)
+        relative_postion_if_large = torch.min(
+            relative_postion_if_large, torch.full_like(relative_postion_if_large, num_buckets - 1)
+        )
+
+        relative_buckets += torch.where(is_small, relative_positions, relative_postion_if_large)
+        return relative_buckets
+
+    def compute_bias(self, query_length, key_length):
+        context_position = torch.arange(query_length, dtype=torch.long)[:, None]
+        memory_position = torch.arange(key_length, dtype=torch.long)[None, :]
+        relative_position = memory_position - context_position
+        relative_position_bucket = self._relative_positions_bucket(
+            relative_position,
+            bidirectional=True
+        )
+        relative_position_bucket = relative_position_bucket.to(self.relative_attention_bias.weight.device)
+        values = self.relative_attention_bias(relative_position_bucket)
+        values = values.permute([2, 0, 1])
+        return values
+
+    def forward(
+            self,
+            query,
+            key: Optional[Tensor],
+            value: Optional[Tensor],
+            key_padding_mask: Optional[Tensor] = None,
+            incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,
+            need_weights: bool = True,
+            static_kv: bool = False,
+            attn_mask: Optional[Tensor] = None,
+            before_softmax: bool = False,
+            need_head_weights: bool = False,
+            position_bias: Optional[Tensor] = None
+    ) -> Tuple[Tensor, Optional[Tensor], Optional[Tensor]]:
+        """Input shape: Time x Batch x Channel
+
+        Args:
+            key_padding_mask (ByteTensor, optional): mask to exclude
+                keys that are pads, of shape `(batch, src_len)`, where
+                padding elements are indicated by 1s.
+            need_weights (bool, optional): return the attention weights,
+                averaged over heads (default: False).
+            attn_mask (ByteTensor, optional): typically used to
+                implement causal attention, where the mask prevents the
+                attention from looking forward in time (default: None).
+            before_softmax (bool, optional): return the raw attention
+                weights and values before the attention softmax.
+            need_head_weights (bool, optional): return the attention
+                weights for each head. Implies *need_weights*. Default:
+                return the average attention weights over all heads.
+        """
+        if need_head_weights:
+            need_weights = True
+
+        is_tpu = query.device.type == "xla"
+
+        tgt_len, bsz, embed_dim = query.size()
+        src_len = tgt_len
+        assert embed_dim == self.embed_dim
+        assert list(query.size()) == [tgt_len, bsz, embed_dim]
+        if key is not None:
+            src_len, key_bsz, _ = key.size()
+            if not torch.jit.is_scripting():
+                assert key_bsz == bsz
+                assert value is not None
+                assert src_len, bsz == value.shape[:2]
+
+        if self.has_relative_attention_bias and position_bias is None:
+            position_bias = self.compute_bias(tgt_len, src_len)
+            position_bias = position_bias.unsqueeze(0).repeat(bsz, 1, 1, 1).view(bsz * self.num_heads, tgt_len, src_len)
+
+        if incremental_state is not None:
+            saved_state = self._get_input_buffer(incremental_state)
+            if saved_state is not None and "prev_key" in saved_state:
+                # previous time steps are cached - no need to recompute
+                # key and value if they are static
+                if static_kv:
+                    assert self.encoder_decoder_attention and not self.self_attention
+                    key = value = None
+        else:
+            saved_state = None
+
+        if self.self_attention:
+            q = self.q_proj(query)
+            k = self.k_proj(query)
+            v = self.v_proj(query)
+        elif self.encoder_decoder_attention:
+            # encoder-decoder attention
+            q = self.q_proj(query)
+            if key is None:
+                assert value is None
+                k = v = None
+            else:
+                k = self.k_proj(key)
+                v = self.v_proj(key)
+
+        else:
+            assert key is not None and value is not None
+            q = self.q_proj(query)
+            k = self.k_proj(key)
+            v = self.v_proj(value)
+        q *= self.scaling
+        alpha = 32
+        q *= 1 / alpha
+
+        if self.bias_k is not None:
+            assert self.bias_v is not None
+            k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)])
+            v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)])
+            if attn_mask is not None:
+                attn_mask = torch.cat(
+                    [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1
+                )
+            if key_padding_mask is not None:
+                key_padding_mask = torch.cat(
+                    [
+                        key_padding_mask,
+                        key_padding_mask.new_zeros(key_padding_mask.size(0), 1),
+                    ],
+                    dim=1,
+                )
+
+        q = (
+            q.contiguous()
+                .view(tgt_len, bsz * self.num_heads, self.q_head_dim)
+                .transpose(0, 1)
+        )
+        if k is not None:
+            k = (
+                k.contiguous()
+                    .view(-1, bsz * self.num_heads, self.k_head_dim)
+                    .transpose(0, 1)
+            )
+        if v is not None:
+            v = (
+                v.contiguous()
+                    .view(-1, bsz * self.num_heads, self.head_dim)
+                    .transpose(0, 1)
+            )
+
+        if saved_state is not None:
+            # saved states are stored with shape (bsz, num_heads, seq_len, head_dim)
+            if "prev_key" in saved_state:
+                _prev_key = saved_state["prev_key"]
+                assert _prev_key is not None
+                prev_key = _prev_key.view(bsz * self.num_heads, -1, self.head_dim)
+                if static_kv:
+                    k = prev_key
+                else:
+                    assert k is not None
+                    k = torch.cat([prev_key, k], dim=1)
+                src_len = k.size(1)
+            if "prev_value" in saved_state:
+                _prev_value = saved_state["prev_value"]
+                assert _prev_value is not None
+                prev_value = _prev_value.view(bsz * self.num_heads, -1, self.head_dim)
+                if static_kv:
+                    v = prev_value
+                else:
+                    assert v is not None
+                    v = torch.cat([prev_value, v], dim=1)
+            prev_key_padding_mask: Optional[Tensor] = None
+            if "prev_key_padding_mask" in saved_state:
+                prev_key_padding_mask = saved_state["prev_key_padding_mask"]
+            assert k is not None and v is not None
+            key_padding_mask = MultiheadAttention._append_prev_key_padding_mask(
+                key_padding_mask=key_padding_mask,
+                prev_key_padding_mask=prev_key_padding_mask,
+                batch_size=bsz,
+                src_len=k.size(1),
+                static_kv=static_kv,
+            )
+
+            saved_state["prev_key"] = k.view(bsz, self.num_heads, -1, self.head_dim)
+            saved_state["prev_value"] = v.view(bsz, self.num_heads, -1, self.head_dim)
+            saved_state["prev_key_padding_mask"] = key_padding_mask
+            # In this branch incremental_state is never None
+            assert incremental_state is not None
+            incremental_state = self._set_input_buffer(incremental_state, saved_state)
+        assert k is not None
+        assert k.size(1) == src_len
+
+        # This is part of a workaround to get around fork/join parallelism
+        # not supporting Optional types.
+        if key_padding_mask is not None and key_padding_mask.dim() == 0:
+            key_padding_mask = None
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz
+            assert key_padding_mask.size(1) == src_len
+
+        if self.add_zero_attn:
+            assert v is not None
+            src_len += 1
+            k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])], dim=1)
+            v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])], dim=1)
+            if attn_mask is not None:
+                attn_mask = torch.cat(
+                    [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1
+                )
+            if key_padding_mask is not None:
+                key_padding_mask = torch.cat(
+                    [
+                        key_padding_mask,
+                        torch.zeros(key_padding_mask.size(0), 1).type_as(
+                            key_padding_mask
+                        ),
+                    ],
+                    dim=1,
+                )
+
+        attn_weights = torch.bmm(q, k.transpose(1, 2))
+        attn_weights = (attn_weights - attn_weights.max(dim=-1, keepdim=True)[0]) * alpha
+        attn_weights = self.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz)
+
+        assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len]
+
+        if attn_mask is not None:
+            attn_mask = attn_mask.unsqueeze(0)
+            attn_weights += attn_mask
+
+        if key_padding_mask is not None:
+            # don't attend to padding symbols
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            if not is_tpu:
+                attn_weights = attn_weights.masked_fill(
+                    key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),
+                    float("-inf"),
+                )
+            else:
+                attn_weights = attn_weights.transpose(0, 2)
+                attn_weights = attn_weights.masked_fill(key_padding_mask, float("-inf"))
+                attn_weights = attn_weights.transpose(0, 2)
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        if before_softmax:
+            return attn_weights, v, position_bias
+
+        if position_bias is not None:
+            attn_mask_rel_pos = position_bias
+            if self.gru_rel_pos == 1:
+                query_layer = q.view(bsz, self.num_heads, tgt_len, self.q_head_dim) * alpha / self.scaling
+                _B, _H, _L, __ = query_layer.size()
+                gate_a, gate_b = torch.sigmoid(self.grep_linear(query_layer).view(
+                    _B, _H, _L, 2, 4).sum(-1, keepdim=False)).chunk(2, dim=-1)
+                gate_a_1 = gate_a * (gate_b * self.grep_a - 1.0) + 2.0
+                attn_mask_rel_pos = gate_a_1.view(bsz * self.num_heads, tgt_len, 1) * position_bias
+
+            attn_mask_rel_pos = attn_mask_rel_pos.view(attn_weights.size())
+
+            attn_weights = attn_weights + attn_mask_rel_pos
+
+        attn_weights_float = F.softmax(
+            attn_weights, dim=-1
+        )
+        attn_weights = attn_weights_float.type_as(attn_weights)
+        attn_probs = self.dropout_module(attn_weights)
+
+        assert v is not None
+        attn = torch.bmm(attn_probs, v)
+        assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim]
+        attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        attn = self.out_proj(attn)
+        attn_weights: Optional[Tensor] = None
+        if need_weights:
+            attn_weights = attn_weights_float.view(
+                bsz, self.num_heads, tgt_len, src_len
+            ).transpose(1, 0)
+            if not need_head_weights:
+                # average attention weights over heads
+                attn_weights = attn_weights.mean(dim=0)
+
+        return attn, attn_weights, position_bias
+
+    @staticmethod
+    def _append_prev_key_padding_mask(
+            key_padding_mask: Optional[Tensor],
+            prev_key_padding_mask: Optional[Tensor],
+            batch_size: int,
+            src_len: int,
+            static_kv: bool,
+    ) -> Optional[Tensor]:
+        # saved key padding masks have shape (bsz, seq_len)
+        if prev_key_padding_mask is not None and static_kv:
+            new_key_padding_mask = prev_key_padding_mask
+        elif prev_key_padding_mask is not None and key_padding_mask is not None:
+            new_key_padding_mask = torch.cat(
+                [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1
+            )
+        # During incremental decoding, as the padding token enters and
+707