funasr pages

游雁

2023-11-16 4ace5a95b052d338947fc88809a440ccd55cf6b4

 funasr/models/decoder/sanm_decoder.py

@@ -1,11 +1,12 @@
from typing import List
from typing import Tuple

import logging
import torch
import torch.nn as nn
import numpy as np

from funasr.modules.streaming_utils import utils as myutils
from funasr.models.decoder.transformer_decoder import BaseTransformerDecoder
from typeguard import check_argument_types

from funasr.modules.attention import MultiHeadedAttentionSANMDecoder, MultiHeadedAttentionCrossAtt
from funasr.modules.embedding import PositionalEncoding
@@ -92,6 +93,46 @@
        if self.self_attn:
            if self.normalize_before:
                tgt = self.norm2(tgt)
            x, _ = self.self_attn(tgt, tgt_mask)
            x = residual + self.dropout(x)

        if self.src_attn is not None:
            residual = x
            if self.normalize_before:
                x = self.norm3(x)

            x = residual + self.dropout(self.src_attn(x, memory, memory_mask))

        return x, tgt_mask, memory, memory_mask, cache

    def forward_one_step(self, tgt, tgt_mask, memory, memory_mask=None, cache=None):
        """Compute decoded features.

        Args:
            tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size).
            tgt_mask (torch.Tensor): Mask for input tensor (#batch, maxlen_out).
            memory (torch.Tensor): Encoded memory, float32 (#batch, maxlen_in, size).
            memory_mask (torch.Tensor): Encoded memory mask (#batch, maxlen_in).
            cache (List[torch.Tensor]): List of cached tensors.
                Each tensor shape should be (#batch, maxlen_out - 1, size).

        Returns:
            torch.Tensor: Output tensor(#batch, maxlen_out, size).
            torch.Tensor: Mask for output tensor (#batch, maxlen_out).
            torch.Tensor: Encoded memory (#batch, maxlen_in, size).
            torch.Tensor: Encoded memory mask (#batch, maxlen_in).

        """
        # tgt = self.dropout(tgt)
        residual = tgt
        if self.normalize_before:
            tgt = self.norm1(tgt)
        tgt = self.feed_forward(tgt)

        x = tgt
        if self.self_attn:
            if self.normalize_before:
                tgt = self.norm2(tgt)
            if self.training:
                cache = None
            x, cache = self.self_attn(tgt, tgt_mask, cache=cache)
@@ -107,10 +148,50 @@

        return x, tgt_mask, memory, memory_mask, cache

    def forward_chunk(self, tgt, memory, fsmn_cache=None, opt_cache=None, chunk_size=None, look_back=0):
        """Compute decoded features.

        Args:
            tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size).
            tgt_mask (torch.Tensor): Mask for input tensor (#batch, maxlen_out).
            memory (torch.Tensor): Encoded memory, float32 (#batch, maxlen_in, size).
            memory_mask (torch.Tensor): Encoded memory mask (#batch, maxlen_in).
            cache (List[torch.Tensor]): List of cached tensors.
                Each tensor shape should be (#batch, maxlen_out - 1, size).

        Returns:
            torch.Tensor: Output tensor(#batch, maxlen_out, size).
            torch.Tensor: Mask for output tensor (#batch, maxlen_out).
            torch.Tensor: Encoded memory (#batch, maxlen_in, size).
            torch.Tensor: Encoded memory mask (#batch, maxlen_in).

        """
        residual = tgt
        if self.normalize_before:
            tgt = self.norm1(tgt)
        tgt = self.feed_forward(tgt)

        x = tgt
        if self.self_attn:
            if self.normalize_before:
                tgt = self.norm2(tgt)
            x, fsmn_cache = self.self_attn(tgt, None, fsmn_cache)
            x = residual + self.dropout(x)

        if self.src_attn is not None:
            residual = x
            if self.normalize_before:
                x = self.norm3(x)

            x, opt_cache = self.src_attn.forward_chunk(x, memory, opt_cache, chunk_size, look_back)
            x = residual + x

        return x, memory, fsmn_cache, opt_cache


class FsmnDecoderSCAMAOpt(BaseTransformerDecoder):
    """
    author: Speech Lab, Alibaba Group, China
    Author: Speech Lab of DAMO Academy, Alibaba Group
    SCAMA: Streaming chunk-aware multihead attention for online end-to-end speech recognition
    https://arxiv.org/abs/2006.01713

@@ -136,8 +217,10 @@
            sanm_shfit: int = None,
            concat_embeds: bool = False,
            attention_dim: int = None,
            tf2torch_tensor_name_prefix_torch: str = "decoder",
            tf2torch_tensor_name_prefix_tf: str = "seq2seq/decoder",
            embed_tensor_name_prefix_tf: str = None,
    ):
        assert check_argument_types()
        super().__init__(
            vocab_size=vocab_size,
            encoder_output_size=encoder_output_size,
@@ -241,6 +324,9 @@
        else:
            self.embed_concat_ffn = None
        self.concat_embeds = concat_embeds
        self.tf2torch_tensor_name_prefix_torch = tf2torch_tensor_name_prefix_torch
        self.tf2torch_tensor_name_prefix_tf = tf2torch_tensor_name_prefix_tf
        self.embed_tensor_name_prefix_tf = embed_tensor_name_prefix_tf

    def forward(
            self,
@@ -352,7 +438,7 @@
        for i in range(self.att_layer_num):
            decoder = self.decoders[i]
            c = cache[i]
            x, tgt_mask, memory, memory_mask, c_ret = decoder(
            x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_one_step(
                x, tgt_mask, memory, memory_mask, cache=c
            )
            new_cache.append(c_ret)
@@ -362,13 +448,13 @@
                j = i + self.att_layer_num
                decoder = self.decoders2[i]
                c = cache[j]
                x, tgt_mask, memory, memory_mask, c_ret = decoder(
                x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_one_step(
                    x, tgt_mask, memory, memory_mask, cache=c
                )
                new_cache.append(c_ret)

        for decoder in self.decoders3:
            x, tgt_mask, memory, memory_mask, _ = decoder(
            x, tgt_mask, memory, memory_mask, _ = decoder.forward_one_step(
                x, tgt_mask, memory, None, cache=None
            )

@@ -382,9 +468,390 @@

        return y, new_cache

    def gen_tf2torch_map_dict(self):
    
        tensor_name_prefix_torch = self.tf2torch_tensor_name_prefix_torch
        tensor_name_prefix_tf = self.tf2torch_tensor_name_prefix_tf
        embed_tensor_name_prefix_tf = self.embed_tensor_name_prefix_tf if self.embed_tensor_name_prefix_tf is not None else tensor_name_prefix_tf
        map_dict_local = {
        
            ## decoder
            # ffn
            "{}.decoders.layeridx.norm1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/LayerNorm/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders.layeridx.norm1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/LayerNorm/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders.layeridx.feed_forward.w_1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/conv1d/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (1024,256),(1,256,1024)
            "{}.decoders.layeridx.feed_forward.w_1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/conv1d/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders.layeridx.feed_forward.norm.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/LayerNorm_1/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders.layeridx.feed_forward.norm.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/LayerNorm_1/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders.layeridx.feed_forward.w_2.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/conv1d_1/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (256,1024),(1,1024,256)
        
            # fsmn
            "{}.decoders.layeridx.norm2.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_memory_block/LayerNorm/gamma".format(
                    tensor_name_prefix_tf),
                    "squeeze": None,
                    "transpose": None,
                },  # (256,),(256,)
            "{}.decoders.layeridx.norm2.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_memory_block/LayerNorm/beta".format(
                    tensor_name_prefix_tf),
                    "squeeze": None,
                    "transpose": None,
                },  # (256,),(256,)
            "{}.decoders.layeridx.self_attn.fsmn_block.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_memory_block/depth_conv_w".format(
                    tensor_name_prefix_tf),
                    "squeeze": 0,
                    "transpose": (1, 2, 0),
                },  # (256,1,31),(1,31,256,1)
            # src att
            "{}.decoders.layeridx.norm3.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/LayerNorm/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders.layeridx.norm3.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/LayerNorm/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders.layeridx.src_attn.linear_q.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/conv1d/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (256,256),(1,256,256)
            "{}.decoders.layeridx.src_attn.linear_q.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/conv1d/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders.layeridx.src_attn.linear_k_v.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/conv1d_1/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (1024,256),(1,256,1024)
            "{}.decoders.layeridx.src_attn.linear_k_v.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/conv1d_1/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders.layeridx.src_attn.linear_out.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/conv1d_2/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (256,256),(1,256,256)
            "{}.decoders.layeridx.src_attn.linear_out.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/conv1d_2/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            # dnn
            "{}.decoders3.layeridx.norm1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/LayerNorm/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders3.layeridx.norm1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/LayerNorm/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders3.layeridx.feed_forward.w_1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/conv1d/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (1024,256),(1,256,1024)
            "{}.decoders3.layeridx.feed_forward.w_1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/conv1d/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders3.layeridx.feed_forward.norm.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/LayerNorm_1/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders3.layeridx.feed_forward.norm.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/LayerNorm_1/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders3.layeridx.feed_forward.w_2.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/conv1d_1/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (256,1024),(1,1024,256)
        
            # embed_concat_ffn
            "{}.embed_concat_ffn.layeridx.norm1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/LayerNorm/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.embed_concat_ffn.layeridx.norm1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/LayerNorm/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.embed_concat_ffn.layeridx.feed_forward.w_1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/conv1d/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (1024,256),(1,256,1024)
            "{}.embed_concat_ffn.layeridx.feed_forward.w_1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/conv1d/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.embed_concat_ffn.layeridx.feed_forward.norm.weight".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/LayerNorm_1/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.embed_concat_ffn.layeridx.feed_forward.norm.bias".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/LayerNorm_1/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.embed_concat_ffn.layeridx.feed_forward.w_2.weight".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/conv1d_1/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (256,1024),(1,1024,256)
        
            # out norm
            "{}.after_norm.weight".format(tensor_name_prefix_torch):
                {"name": "{}/LayerNorm/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.after_norm.bias".format(tensor_name_prefix_torch):
                {"name": "{}/LayerNorm/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
        
            # in embed
            "{}.embed.0.weight".format(tensor_name_prefix_torch):
                {"name": "{}/w_embs".format(embed_tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (4235,256),(4235,256)
        
            # out layer
            "{}.output_layer.weight".format(tensor_name_prefix_torch):
                {"name": ["{}/dense/kernel".format(tensor_name_prefix_tf),
                          "{}/w_embs".format(embed_tensor_name_prefix_tf)],
                 "squeeze": [None, None],
                 "transpose": [(1, 0), None],
                 },  # (4235,256),(256,4235)
            "{}.output_layer.bias".format(tensor_name_prefix_torch):
                {"name": ["{}/dense/bias".format(tensor_name_prefix_tf),
                          "seq2seq/2bias" if tensor_name_prefix_tf == "seq2seq/decoder/inputter_1" else "seq2seq/bias"],
                 "squeeze": [None, None],
                 "transpose": [None, None],
                 },  # (4235,),(4235,)
        
        }
        return map_dict_local

    def convert_tf2torch(self,
                         var_dict_tf,
                         var_dict_torch,
                         ):
    
        map_dict = self.gen_tf2torch_map_dict()
        var_dict_torch_update = dict()
        decoder_layeridx_sets = set()
        for name in sorted(var_dict_torch.keys(), reverse=False):
            names = name.split('.')
            if names[0] == self.tf2torch_tensor_name_prefix_torch:
                if names[1] == "decoders":
                    layeridx = int(names[2])
                    name_q = name.replace(".{}.".format(layeridx), ".layeridx.")
                    layeridx_bias = 0
                    layeridx += layeridx_bias
                    decoder_layeridx_sets.add(layeridx)
                    if name_q in map_dict.keys():
                        name_v = map_dict[name_q]["name"]
                        name_tf = name_v.replace("layeridx", "{}".format(layeridx))
                        data_tf = var_dict_tf[name_tf]
                        if map_dict[name_q]["squeeze"] is not None:
                            data_tf = np.squeeze(data_tf, axis=map_dict[name_q]["squeeze"])
                        if map_dict[name_q]["transpose"] is not None:
                            data_tf = np.transpose(data_tf, map_dict[name_q]["transpose"])
                        data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                        assert var_dict_torch[name].size() == data_tf.size(), "{}, {}, {} != {}".format(name, name_tf,
                                                                                                        var_dict_torch[
                                                                                                            name].size(),
                                                                                                        data_tf.size())
                        var_dict_torch_update[name] = data_tf
                        logging.info(
                            "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(), name_v,
                                                                                          var_dict_tf[name_tf].shape))
            
                elif names[1] == "decoders2":
                    layeridx = int(names[2])
                    name_q = name.replace(".{}.".format(layeridx), ".layeridx.")
                    name_q = name_q.replace("decoders2", "decoders")
                    layeridx_bias = len(decoder_layeridx_sets)
                
                    layeridx += layeridx_bias
                    if "decoders." in name:
                        decoder_layeridx_sets.add(layeridx)
                    if name_q in map_dict.keys():
                        name_v = map_dict[name_q]["name"]
                        name_tf = name_v.replace("layeridx", "{}".format(layeridx))
                        data_tf = var_dict_tf[name_tf]
                        if map_dict[name_q]["squeeze"] is not None:
                            data_tf = np.squeeze(data_tf, axis=map_dict[name_q]["squeeze"])
                        if map_dict[name_q]["transpose"] is not None:
                            data_tf = np.transpose(data_tf, map_dict[name_q]["transpose"])
                        data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                        assert var_dict_torch[name].size() == data_tf.size(), "{}, {}, {} != {}".format(name, name_tf,
                                                                                                        var_dict_torch[
                                                                                                            name].size(),
                                                                                                        data_tf.size())
                        var_dict_torch_update[name] = data_tf
                        logging.info(
                            "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(), name_v,
                                                                                          var_dict_tf[name_tf].shape))
            
                elif names[1] == "decoders3":
                    layeridx = int(names[2])
                    name_q = name.replace(".{}.".format(layeridx), ".layeridx.")
                
                    layeridx_bias = 0
                    layeridx += layeridx_bias
                    if "decoders." in name:
                        decoder_layeridx_sets.add(layeridx)
                    if name_q in map_dict.keys():
                        name_v = map_dict[name_q]["name"]
                        name_tf = name_v.replace("layeridx", "{}".format(layeridx))
                        data_tf = var_dict_tf[name_tf]
                        if map_dict[name_q]["squeeze"] is not None:
                            data_tf = np.squeeze(data_tf, axis=map_dict[name_q]["squeeze"])
                        if map_dict[name_q]["transpose"] is not None:
                            data_tf = np.transpose(data_tf, map_dict[name_q]["transpose"])
                        data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                        assert var_dict_torch[name].size() == data_tf.size(), "{}, {}, {} != {}".format(name, name_tf,
                                                                                                        var_dict_torch[
                                                                                                            name].size(),
                                                                                                        data_tf.size())
                        var_dict_torch_update[name] = data_tf
                        logging.info(
                            "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(), name_v,
                                                                                          var_dict_tf[name_tf].shape))
            
                elif names[1] == "embed" or names[1] == "output_layer":
                    name_tf = map_dict[name]["name"]
                    if isinstance(name_tf, list):
                        idx_list = 0
                        if name_tf[idx_list] in var_dict_tf.keys():
                            pass
                        else:
                            idx_list = 1
                        data_tf = var_dict_tf[name_tf[idx_list]]
                        if map_dict[name]["squeeze"][idx_list] is not None:
                            data_tf = np.squeeze(data_tf, axis=map_dict[name]["squeeze"][idx_list])
                        if map_dict[name]["transpose"][idx_list] is not None:
                            data_tf = np.transpose(data_tf, map_dict[name]["transpose"][idx_list])
                        data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                        assert var_dict_torch[name].size() == data_tf.size(), "{}, {}, {} != {}".format(name, name_tf,
                                                                                                        var_dict_torch[
                                                                                                            name].size(),
                                                                                                        data_tf.size())
                        var_dict_torch_update[name] = data_tf
                        logging.info("torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(),
                                                                                                   name_tf[idx_list],
                                                                                                   var_dict_tf[name_tf[
                                                                                                       idx_list]].shape))
                
                    else:
                        data_tf = var_dict_tf[name_tf]
                        if map_dict[name]["squeeze"] is not None:
                            data_tf = np.squeeze(data_tf, axis=map_dict[name]["squeeze"])
                        if map_dict[name]["transpose"] is not None:
                            data_tf = np.transpose(data_tf, map_dict[name]["transpose"])
                        data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                        assert var_dict_torch[name].size() == data_tf.size(), "{}, {}, {} != {}".format(name, name_tf,
                                                                                                        var_dict_torch[
                                                                                                            name].size(),
                                                                                                        data_tf.size())
                        var_dict_torch_update[name] = data_tf
                        logging.info(
                            "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(), name_tf,
                                                                                          var_dict_tf[name_tf].shape))
            
                elif names[1] == "after_norm":
                    name_tf = map_dict[name]["name"]
                    data_tf = var_dict_tf[name_tf]
                    data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                    var_dict_torch_update[name] = data_tf
                    logging.info(
                        "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(), name_tf,
                                                                                      var_dict_tf[name_tf].shape))
            
                elif names[1] == "embed_concat_ffn":
                    layeridx = int(names[2])
                    name_q = name.replace(".{}.".format(layeridx), ".layeridx.")
                
                    layeridx_bias = 0
                    layeridx += layeridx_bias
                    if "decoders." in name:
                        decoder_layeridx_sets.add(layeridx)
                    if name_q in map_dict.keys():
                        name_v = map_dict[name_q]["name"]
                        name_tf = name_v.replace("layeridx", "{}".format(layeridx))
                        data_tf = var_dict_tf[name_tf]
                        if map_dict[name_q]["squeeze"] is not None:
                            data_tf = np.squeeze(data_tf, axis=map_dict[name_q]["squeeze"])
                        if map_dict[name_q]["transpose"] is not None:
                            data_tf = np.transpose(data_tf, map_dict[name_q]["transpose"])
                        data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                        assert var_dict_torch[name].size() == data_tf.size(), "{}, {}, {} != {}".format(name, name_tf,
                                                                                                        var_dict_torch[
                                                                                                            name].size(),
                                                                                                        data_tf.size())
                        var_dict_torch_update[name] = data_tf
                        logging.info(
                            "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(), name_v,
                                                                                          var_dict_tf[name_tf].shape))
    
        return var_dict_torch_update


class ParaformerSANMDecoder(BaseTransformerDecoder):
    """
    author: Speech Lab, Alibaba Group, China
    Author: Speech Lab of DAMO Academy, Alibaba Group
    Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition
    https://arxiv.org/abs/2006.01713
    """
@@ -407,8 +874,14 @@
        att_layer_num: int = 6,
        kernel_size: int = 21,
        sanm_shfit: int = 0,
        lora_list: List[str] = None,
        lora_rank: int = 8,
        lora_alpha: int = 16,
        lora_dropout: float = 0.1,
        chunk_multiply_factor: tuple = (1,),
        tf2torch_tensor_name_prefix_torch: str = "decoder",
        tf2torch_tensor_name_prefix_tf: str = "seq2seq/decoder",
    ):
        assert check_argument_types()
        super().__init__(
            vocab_size=vocab_size,
            encoder_output_size=encoder_output_size,
@@ -458,7 +931,7 @@
                    attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=sanm_shfit
                ),
                MultiHeadedAttentionCrossAtt(
                    attention_heads, attention_dim, src_attention_dropout_rate
                    attention_heads, attention_dim, src_attention_dropout_rate, lora_list, lora_rank, lora_alpha, lora_dropout
                ),
                PositionwiseFeedForwardDecoderSANM(attention_dim, linear_units, dropout_rate),
                dropout_rate,
@@ -496,6 +969,9 @@
                concat_after,
            ),
        )
        self.tf2torch_tensor_name_prefix_torch = tf2torch_tensor_name_prefix_torch
        self.tf2torch_tensor_name_prefix_tf = tf2torch_tensor_name_prefix_tf
        self.chunk_multiply_factor = chunk_multiply_factor

    def forward(
        self,
@@ -503,6 +979,7 @@
        hlens: torch.Tensor,
        ys_in_pad: torch.Tensor,
        ys_in_lens: torch.Tensor,
        chunk_mask: torch.Tensor = None,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Forward decoder.

@@ -523,9 +1000,13 @@
        """
        tgt = ys_in_pad
        tgt_mask = myutils.sequence_mask(ys_in_lens, device=tgt.device)[:, :, None]

        
        memory = hs_pad
        memory_mask = myutils.sequence_mask(hlens, device=memory.device)[:, None, :]
        if chunk_mask is not None:
            memory_mask = memory_mask * chunk_mask
            if tgt_mask.size(1) != memory_mask.size(1):
                memory_mask = torch.cat((memory_mask, memory_mask[:, -2:-1, :]), dim=1)

        x = tgt
        x, tgt_mask, memory, memory_mask, _ = self.decoders(
@@ -553,6 +1034,73 @@
            ys.unsqueeze(0), ys_mask, x.unsqueeze(0), cache=state
        )
        return logp.squeeze(0), state

    def forward_chunk(
        self,
        memory: torch.Tensor,
        tgt: torch.Tensor,
        cache: dict = None,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Forward decoder.

        Args:
            hs_pad: encoded memory, float32  (batch, maxlen_in, feat)
            hlens: (batch)
            ys_in_pad:
                input token ids, int64 (batch, maxlen_out)
                if input_layer == "embed"
                input tensor (batch, maxlen_out, #mels) in the other cases
            ys_in_lens: (batch)
        Returns:
            (tuple): tuple containing:

            x: decoded token score before softmax (batch, maxlen_out, token)
                if use_output_layer is True,
            olens: (batch, )
        """
        x = tgt
        if cache["decode_fsmn"] is None:
            cache_layer_num = len(self.decoders)
            if self.decoders2 is not None:
                cache_layer_num += len(self.decoders2)
            fsmn_cache = [None] * cache_layer_num
        else:
            fsmn_cache = cache["decode_fsmn"]

        if cache["opt"] is None:
            cache_layer_num = len(self.decoders)
            opt_cache = [None] * cache_layer_num
        else:
            opt_cache = cache["opt"]

        for i in range(self.att_layer_num):
            decoder = self.decoders[i]
            x, memory, fsmn_cache[i], opt_cache[i] = decoder.forward_chunk(
                x, memory, fsmn_cache=fsmn_cache[i], opt_cache=opt_cache[i],
                chunk_size=cache["chunk_size"], look_back=cache["decoder_chunk_look_back"]
            )

        if self.num_blocks - self.att_layer_num > 1:
            for i in range(self.num_blocks - self.att_layer_num):
                j = i + self.att_layer_num
                decoder = self.decoders2[i]
                x, memory, fsmn_cache[j], _  = decoder.forward_chunk(
                    x, memory, fsmn_cache=fsmn_cache[j]
                )

        for decoder in self.decoders3:
            x, memory, _, _ = decoder.forward_chunk(
                x, memory
            )
        if self.normalize_before:
            x = self.after_norm(x)
        if self.output_layer is not None:
            x = self.output_layer(x)

        cache["decode_fsmn"] = fsmn_cache
        if cache["decoder_chunk_look_back"] > 0 or cache["decoder_chunk_look_back"] == -1:
            cache["opt"] = opt_cache
        return x

    def forward_one_step(
        self,
@@ -585,7 +1133,7 @@
        for i in range(self.att_layer_num):
            decoder = self.decoders[i]
            c = cache[i]
            x, tgt_mask, memory, memory_mask, c_ret = decoder(
            x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_one_step(
                x, tgt_mask, memory, None, cache=c
            )
            new_cache.append(c_ret)
@@ -595,14 +1143,14 @@
                j = i + self.att_layer_num
                decoder = self.decoders2[i]
                c = cache[j]
                x, tgt_mask, memory, memory_mask, c_ret = decoder(
                x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_one_step(
                    x, tgt_mask, memory, None, cache=c
                )
                new_cache.append(c_ret)

        for decoder in self.decoders3:

            x, tgt_mask, memory, memory_mask, _ = decoder(
            x, tgt_mask, memory, memory_mask, _ = decoder.forward_one_step(
                x, tgt_mask, memory, None, cache=None
            )

@@ -613,4 +1161,381 @@
        if self.output_layer is not None:
            y = torch.log_softmax(self.output_layer(y), dim=-1)

        return y, new_cache
        return y, new_cache

    def gen_tf2torch_map_dict(self):
    
        tensor_name_prefix_torch = self.tf2torch_tensor_name_prefix_torch
        tensor_name_prefix_tf = self.tf2torch_tensor_name_prefix_tf
        map_dict_local = {
        
            ## decoder
            # ffn
            "{}.decoders.layeridx.norm1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/LayerNorm/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders.layeridx.norm1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/LayerNorm/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders.layeridx.feed_forward.w_1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/conv1d/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (1024,256),(1,256,1024)
            "{}.decoders.layeridx.feed_forward.w_1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/conv1d/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders.layeridx.feed_forward.norm.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/LayerNorm_1/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders.layeridx.feed_forward.norm.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/LayerNorm_1/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders.layeridx.feed_forward.w_2.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_ffn/conv1d_1/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (256,1024),(1,1024,256)
        
            # fsmn
            "{}.decoders.layeridx.norm2.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_memory_block/LayerNorm/gamma".format(
                    tensor_name_prefix_tf),
                    "squeeze": None,
                    "transpose": None,
                },  # (256,),(256,)
            "{}.decoders.layeridx.norm2.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_memory_block/LayerNorm/beta".format(
                    tensor_name_prefix_tf),
                    "squeeze": None,
                    "transpose": None,
                },  # (256,),(256,)
            "{}.decoders.layeridx.self_attn.fsmn_block.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/decoder_memory_block/depth_conv_w".format(
                    tensor_name_prefix_tf),
                    "squeeze": 0,
                    "transpose": (1, 2, 0),
                },  # (256,1,31),(1,31,256,1)
            # src att
            "{}.decoders.layeridx.norm3.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/LayerNorm/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders.layeridx.norm3.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/LayerNorm/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders.layeridx.src_attn.linear_q.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/conv1d/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (256,256),(1,256,256)
            "{}.decoders.layeridx.src_attn.linear_q.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/conv1d/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders.layeridx.src_attn.linear_k_v.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/conv1d_1/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (1024,256),(1,256,1024)
            "{}.decoders.layeridx.src_attn.linear_k_v.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/conv1d_1/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders.layeridx.src_attn.linear_out.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/conv1d_2/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (256,256),(1,256,256)
            "{}.decoders.layeridx.src_attn.linear_out.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_fsmn_layer_layeridx/multi_head/conv1d_2/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            # dnn
            "{}.decoders3.layeridx.norm1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/LayerNorm/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders3.layeridx.norm1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/LayerNorm/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.decoders3.layeridx.feed_forward.w_1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/conv1d/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (1024,256),(1,256,1024)
            "{}.decoders3.layeridx.feed_forward.w_1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/conv1d/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders3.layeridx.feed_forward.norm.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/LayerNorm_1/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders3.layeridx.feed_forward.norm.bias".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/LayerNorm_1/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.decoders3.layeridx.feed_forward.w_2.weight".format(tensor_name_prefix_torch):
                {"name": "{}/decoder_dnn_layer_layeridx/conv1d_1/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (256,1024),(1,1024,256)
        
            # embed_concat_ffn
            "{}.embed_concat_ffn.layeridx.norm1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/LayerNorm/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.embed_concat_ffn.layeridx.norm1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/LayerNorm/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.embed_concat_ffn.layeridx.feed_forward.w_1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/conv1d/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (1024,256),(1,256,1024)
            "{}.embed_concat_ffn.layeridx.feed_forward.w_1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/conv1d/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.embed_concat_ffn.layeridx.feed_forward.norm.weight".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/LayerNorm_1/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.embed_concat_ffn.layeridx.feed_forward.norm.bias".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/LayerNorm_1/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (1024,),(1024,)
            "{}.embed_concat_ffn.layeridx.feed_forward.w_2.weight".format(tensor_name_prefix_torch):
                {"name": "{}/cif_concat/conv1d_1/kernel".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 0),
                 },  # (256,1024),(1,1024,256)
        
            # out norm
            "{}.after_norm.weight".format(tensor_name_prefix_torch):
                {"name": "{}/LayerNorm/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
            "{}.after_norm.bias".format(tensor_name_prefix_torch):
                {"name": "{}/LayerNorm/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (256,),(256,)
        
            # in embed
            "{}.embed.0.weight".format(tensor_name_prefix_torch):
                {"name": "{}/w_embs".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },  # (4235,256),(4235,256)
        
            # out layer
            "{}.output_layer.weight".format(tensor_name_prefix_torch):
                {"name": ["{}/dense/kernel".format(tensor_name_prefix_tf), "{}/w_embs".format(tensor_name_prefix_tf)],
                 "squeeze": [None, None],
                 "transpose": [(1, 0), None],
                 },  # (4235,256),(256,4235)
            "{}.output_layer.bias".format(tensor_name_prefix_torch):
                {"name": ["{}/dense/bias".format(tensor_name_prefix_tf),
                          "seq2seq/2bias" if tensor_name_prefix_tf == "seq2seq/decoder/inputter_1" else "seq2seq/bias"],
                 "squeeze": [None, None],
                 "transpose": [None, None],
                 },  # (4235,),(4235,)
        
        }
        return map_dict_local

    def convert_tf2torch(self,
                         var_dict_tf,
                         var_dict_torch,
                         ):
        map_dict = self.gen_tf2torch_map_dict()
        var_dict_torch_update = dict()
        decoder_layeridx_sets = set()
        for name in sorted(var_dict_torch.keys(), reverse=False):
            names = name.split('.')
            if names[0] == self.tf2torch_tensor_name_prefix_torch:
                if names[1] == "decoders":
                    layeridx = int(names[2])
                    name_q = name.replace(".{}.".format(layeridx), ".layeridx.")
                    layeridx_bias = 0
                    layeridx += layeridx_bias
                    decoder_layeridx_sets.add(layeridx)
                    if name_q in map_dict.keys():
                        name_v = map_dict[name_q]["name"]
                        name_tf = name_v.replace("layeridx", "{}".format(layeridx))
                        data_tf = var_dict_tf[name_tf]
                        if map_dict[name_q]["squeeze"] is not None:
                            data_tf = np.squeeze(data_tf, axis=map_dict[name_q]["squeeze"])
                        if map_dict[name_q]["transpose"] is not None:
                            data_tf = np.transpose(data_tf, map_dict[name_q]["transpose"])
                        data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                        assert var_dict_torch[name].size() == data_tf.size(), "{}, {}, {} != {}".format(name, name_tf,
                                                                                                        var_dict_torch[
                                                                                                            name].size(),
                                                                                                        data_tf.size())
                        var_dict_torch_update[name] = data_tf
                        logging.info(
                            "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(), name_v,
                                                                                          var_dict_tf[name_tf].shape))
            
                elif names[1] == "decoders2":
                    layeridx = int(names[2])
                    name_q = name.replace(".{}.".format(layeridx), ".layeridx.")
                    name_q = name_q.replace("decoders2", "decoders")
                    layeridx_bias = len(decoder_layeridx_sets)
                
                    layeridx += layeridx_bias
                    if "decoders." in name:
                        decoder_layeridx_sets.add(layeridx)
                    if name_q in map_dict.keys():
                        name_v = map_dict[name_q]["name"]
                        name_tf = name_v.replace("layeridx", "{}".format(layeridx))
                        data_tf = var_dict_tf[name_tf]
                        if map_dict[name_q]["squeeze"] is not None:
                            data_tf = np.squeeze(data_tf, axis=map_dict[name_q]["squeeze"])
                        if map_dict[name_q]["transpose"] is not None:
                            data_tf = np.transpose(data_tf, map_dict[name_q]["transpose"])
                        data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                        assert var_dict_torch[name].size() == data_tf.size(), "{}, {}, {} != {}".format(name, name_tf,
                                                                                                        var_dict_torch[
                                                                                                            name].size(),
                                                                                                        data_tf.size())
                        var_dict_torch_update[name] = data_tf
                        logging.info(
                            "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(), name_v,
                                                                                          var_dict_tf[name_tf].shape))
            
                elif names[1] == "decoders3":
                    layeridx = int(names[2])
                    name_q = name.replace(".{}.".format(layeridx), ".layeridx.")
                
                    layeridx_bias = 0
                    layeridx += layeridx_bias
                    if "decoders." in name:
                        decoder_layeridx_sets.add(layeridx)
                    if name_q in map_dict.keys():
                        name_v = map_dict[name_q]["name"]
                        name_tf = name_v.replace("layeridx", "{}".format(layeridx))
                        data_tf = var_dict_tf[name_tf]
                        if map_dict[name_q]["squeeze"] is not None:
                            data_tf = np.squeeze(data_tf, axis=map_dict[name_q]["squeeze"])
                        if map_dict[name_q]["transpose"] is not None:
                            data_tf = np.transpose(data_tf, map_dict[name_q]["transpose"])
                        data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                        assert var_dict_torch[name].size() == data_tf.size(), "{}, {}, {} != {}".format(name, name_tf,
                                                                                                        var_dict_torch[
                                                                                                            name].size(),
                                                                                                        data_tf.size())
                        var_dict_torch_update[name] = data_tf
                        logging.info(
                            "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(), name_v,
                                                                                          var_dict_tf[name_tf].shape))
            
                elif names[1] == "embed" or names[1] == "output_layer":
                    name_tf = map_dict[name]["name"]
                    if isinstance(name_tf, list):
                        idx_list = 0
                        if name_tf[idx_list] in var_dict_tf.keys():
                            pass
                        else:
                            idx_list = 1
                        data_tf = var_dict_tf[name_tf[idx_list]]
                        if map_dict[name]["squeeze"][idx_list] is not None:
                            data_tf = np.squeeze(data_tf, axis=map_dict[name]["squeeze"][idx_list])
                        if map_dict[name]["transpose"][idx_list] is not None:
                            data_tf = np.transpose(data_tf, map_dict[name]["transpose"][idx_list])
                        data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                        assert var_dict_torch[name].size() == data_tf.size(), "{}, {}, {} != {}".format(name, name_tf,
                                                                                                        var_dict_torch[
                                                                                                            name].size(),
                                                                                                        data_tf.size())
                        var_dict_torch_update[name] = data_tf
                        logging.info("torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(),
                                                                                                   name_tf[idx_list],
                                                                                                   var_dict_tf[name_tf[
                                                                                                       idx_list]].shape))
                
                    else:
                        data_tf = var_dict_tf[name_tf]
                        if map_dict[name]["squeeze"] is not None:
                            data_tf = np.squeeze(data_tf, axis=map_dict[name]["squeeze"])
                        if map_dict[name]["transpose"] is not None:
                            data_tf = np.transpose(data_tf, map_dict[name]["transpose"])
                        data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                        assert var_dict_torch[name].size() == data_tf.size(), "{}, {}, {} != {}".format(name, name_tf,
                                                                                                        var_dict_torch[
                                                                                                            name].size(),
                                                                                                        data_tf.size())
                        var_dict_torch_update[name] = data_tf
                        logging.info(
                            "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(), name_tf,
                                                                                          var_dict_tf[name_tf].shape))
            
                elif names[1] == "after_norm":
                    name_tf = map_dict[name]["name"]
                    data_tf = var_dict_tf[name_tf]
                    data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                    var_dict_torch_update[name] = data_tf
                    logging.info(
                        "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(), name_tf,
                                                                                      var_dict_tf[name_tf].shape))
            
                elif names[1] == "embed_concat_ffn":
                    layeridx = int(names[2])
                    name_q = name.replace(".{}.".format(layeridx), ".layeridx.")
                
                    layeridx_bias = 0
                    layeridx += layeridx_bias
                    if "decoders." in name:
                        decoder_layeridx_sets.add(layeridx)
                    if name_q in map_dict.keys():
                        name_v = map_dict[name_q]["name"]
                        name_tf = name_v.replace("layeridx", "{}".format(layeridx))
                        data_tf = var_dict_tf[name_tf]
                        if map_dict[name_q]["squeeze"] is not None:
                            data_tf = np.squeeze(data_tf, axis=map_dict[name_q]["squeeze"])
                        if map_dict[name_q]["transpose"] is not None:
                            data_tf = np.transpose(data_tf, map_dict[name_q]["transpose"])
                        data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                        assert var_dict_torch[name].size() == data_tf.size(), "{}, {}, {} != {}".format(name, name_tf,
                                                                                                        var_dict_torch[
                                                                                                            name].size(),
                                                                                                        data_tf.size())
                        var_dict_torch_update[name] = data_tf
                        logging.info(
                            "torch tensor: {}, {}, loading from tf tensor: {}, {}".format(name, data_tf.size(), name_v,
                                                                                          var_dict_tf[name_tf].shape))
    
        return var_dict_torch_update