Source code for ofasys.module.transformer_layer

# Copyright 2022 The OFA-Sys Team. All rights reserved.
# This source code is licensed under the Apache 2.0 license
# found in the LICENSE file in the root directory.

import copy
from typing import Dict, List, Optional

import torch
import torch.nn as nn
from torch import Tensor

from ofasys import ModalityType
from ofasys.module import utils
from ofasys.module.sparse_dispatcher import SparseDispatcher
from . import Dropout, DropPath, LayerNorm, MultiheadAttention, TransformerConfig


[docs]class TransformerEncoderLayer(nn.Module): """Encoder layer block. In the original paper each operation (multi-head attention or FFN) is postprocessed with: `dropout -> add residual -> layernorm`. In the tensor2tensor code they suggest that learning is more robust when preprocessing each layer with layernorm and postprocessing with: `dropout -> add residual`. We default to the approach in the paper, but the tensor2tensor approach can be enabled by setting *cfg.encoder.normalize_before* to ``True``. Args: args (argparse.Namespace): parsed command-line arguments """ def __init__(self, args, drop_path_rate=0.0): cfg = TransformerConfig.from_namespace(args) super().__init__() self.cfg = cfg self.embed_dim = cfg.encoder.embed_dim self.self_attn = self.build_self_attention(self.embed_dim, cfg) self.self_attn_layer_norm = LayerNorm(self.embed_dim) self.dropout_module = Dropout(cfg.dropout, module_name=self.__class__.__name__) self.activation_fn = utils.get_activation_fn(activation=cfg.activation_fn) activation_dropout_p = cfg.activation_dropout if activation_dropout_p == 0: # for backwards compatibility with models that use cfg.relu_dropout activation_dropout_p = cfg.relu_dropout or 0 self.activation_dropout_module = Dropout(float(activation_dropout_p), module_name=self.__class__.__name__) self.normalize_before = cfg.encoder.normalize_before self.fc1 = nn.Linear(self.embed_dim, cfg.encoder.ffn_embed_dim) self.fc2 = nn.Linear(cfg.encoder.ffn_embed_dim, self.embed_dim) if cfg.modal_ffn: self.experts_num = int(len(ModalityType)) self.experts_fc1 = nn.ModuleList([copy.deepcopy(self.fc1) for i in range(self.experts_num)]) self.experts_fc2 = nn.ModuleList([copy.deepcopy(self.fc2) for i in range(self.experts_num)]) self.gates_way = torch.eye(self.experts_num) self.attn_ln = LayerNorm(self.embed_dim) if cfg.scale_attn else None self.nh = self.self_attn.num_heads self.head_dim = self.self_attn.head_dim self.ffn_layernorm = LayerNorm(cfg.encoder_ffn_embed_dim) if cfg.scale_fc else None self.w_resid = ( nn.Parameter( torch.ones( self.embed_dim, ), requires_grad=True, ) if cfg.scale_resids else None ) self.final_layer_norm = LayerNorm(self.embed_dim) self.drop_path = DropPath(drop_path_rate, batch_axis=1) def build_self_attention(self, embed_dim, cfg): return MultiheadAttention( embed_dim, cfg.encoder.attention_heads, dropout=cfg.attention_dropout, self_attention=True, scale_factor=cfg.attn_scale_factor, scale_heads=cfg.scale_heads, use_fused=cfg.use_fused, ) def residual_connection(self, x, residual): return residual + self.drop_path(x)
[docs] def upgrade_state_dict_named(self, state_dict, name): """ Rename layer norm states from `...layer_norms.0.weight` to `...self_attn_layer_norm.weight` and `...layer_norms.1.weight` to `...final_layer_norm.weight` """ layer_norm_map = {"0": "self_attn_layer_norm", "1": "final_layer_norm"} for old, new in layer_norm_map.items(): for m in ("weight", "bias"): k = "{}.layer_norms.{}.{}".format(name, old, m) if k in state_dict: state_dict["{}.{}.{}".format(name, new, m)] = state_dict[k] del state_dict[k] if "{}.{}.{}".format(name, new, m) not in state_dict and "{}.{}".format(new, m) in self.state_dict(): state_dict["{}.{}.{}".format(name, new, m)] = self.state_dict()["{}.{}".format(new, m)] prefix = name + "." if name != "" else "" for param_name, param_tensor in self.state_dict().items(): if (prefix + param_name) not in state_dict: for i in range(self.experts_num): if param_name.find("experts_fc1"+"."+str(i)) != -1: state_dict[prefix + param_name] = self.state_dict()[param_name.replace("experts_fc1"+"."+str(i), "fc1")] if param_name.find("experts_fc2"+"."+str(i)) != -1: state_dict[prefix + param_name] = self.state_dict()[param_name.replace("experts_fc2"+"."+str(i), "fc2")] if param_name.find("experts_fc1") == -1 and param_name.find("experts_fc2") == -1: state_dict[prefix + param_name] = self.state_dict()[param_name]
def modal_for_ffn(self, modal_mask, x, experts_fc): modal_mask_fc = modal_mask bs, seq_len = modal_mask_fc.shape modal_mask_fc = modal_mask_fc.view(bs * seq_len) gates = torch.index_select(self.gates_way.to(modal_mask_fc.device) , 0, modal_mask_fc) gates = gates.view(bs * seq_len, self.experts_num) dispatcher = SparseDispatcher(self.experts_num, gates) seq_len, bs, dim = x.shape x = x.view(seq_len * bs, dim) expert_inputs = dispatcher.dispatch(x) expert_outputs = [experts_fc[i](expert_inputs[i]) for i in range(self.experts_num)] x = dispatcher.combine(expert_outputs) x = x.view(seq_len, bs, -1) x = x.half() return x
[docs] def forward( self, x, encoder_padding_mask: Optional[Tensor], attn_mask: Optional[Tensor] = None, self_attn_bias: Optional[Tensor] = None, need_attn: bool = False, modal_mask = None, ): """ Args: x (Tensor): input to the layer of shape ``(seq_len, batch, embed_dim)`` encoder_padding_mask (ByteTensor): binary ByteTensor of shape ``(batch, seq_len)`` where padding elements are indicated by ``1``. attn_mask (ByteTensor): binary tensor of shape ``(tgt_len, src_len)``, where `tgt_len` is the length of output and `src_len` is the length of input, though here both are equal to `seq_len`. `attn_mask[tgt_i, src_j] = 1` means that when calculating the embedding for `tgt_i`, we exclude (mask out) `src_j`. This is useful for strided self-attention. self_attn_bias (Tensor): need_attn (bool): Returns: encoded output of shape ``(seq_len, batch, embed_dim)`` """ # anything in original attn_mask = 1, becomes -1e8 # anything in original attn_mask = 0, becomes 0 # Note that we cannot use -inf here, because at some edge cases, # the attention weight (before softmax) for some padded element in query # will become -inf, which results in NaN in model parameters if attn_mask is not None: attn_mask = attn_mask.masked_fill(attn_mask.to(torch.bool), -1e8 if x.dtype == torch.float32 else -1e4) residual = x if self.normalize_before: x = self.self_attn_layer_norm(x) x, self_attn_weights = self.self_attn( query=x, key=x, value=x, key_padding_mask=encoder_padding_mask, need_weights=need_attn, attn_mask=attn_mask, attn_bias=self_attn_bias, ) if self.attn_ln is not None: x = self.attn_ln(x) x = self.dropout_module(x) x = self.residual_connection(x, residual) if not self.normalize_before: x = self.self_attn_layer_norm(x) residual = x if self.normalize_before: x = self.final_layer_norm(x) # ffn1 modal if self.cfg.modal_ffn: x = self.modal_for_ffn(modal_mask, x, self.experts_fc1) x = self.activation_fn(x) else: x = self.activation_fn(self.fc1(x)) x = self.activation_dropout_module(x) if self.ffn_layernorm is not None: x = self.ffn_layernorm(x) # ffn2 modal if self.cfg.modal_ffn: x = self.modal_for_ffn(modal_mask, x, self.experts_fc2) else: x = self.fc2(x) x = self.dropout_module(x) if self.w_resid is not None: residual = torch.mul(self.w_resid, residual) x = self.residual_connection(x, residual) if not self.normalize_before: x = self.final_layer_norm(x) return x, self_attn_weights
[docs]class TransformerDecoderLayer(nn.Module): """Decoder layer block. In the original paper each operation (multi-head attention, encoder attention or FFN) is postprocessed with: `dropout -> add residual -> layernorm`. In the tensor2tensor code they suggest that learning is more robust when preprocessing each layer with layernorm and postprocessing with: `dropout -> add residual`. We default to the approach in the paper, but the tensor2tensor approach can be enabled by setting *cfg.decoder_normalize_before* to ``True``. Args: args (argparse.Namespace): parsed command-line arguments no_encoder_attn (bool, optional): whether to attend to encoder outputs (default: False). add_bias_kv (bool, optional): (default: False). add_zero_attn (bool, optional): (default: False). drop_path_rate (float, optional): (default: 0.0). """ def __init__( self, args, no_encoder_attn=False, add_bias_kv=False, add_zero_attn=False, drop_path_rate=0.0, ): cfg = TransformerConfig.from_namespace(args) self.cfg = cfg super().__init__() self.embed_dim = cfg.decoder.embed_dim self.dropout_module = Dropout(cfg.dropout, module_name=self.__class__.__name__) self.cross_self_attention = cfg.cross_self_attention self.self_attn = self.build_self_attention( self.embed_dim, cfg, add_bias_kv=add_bias_kv, add_zero_attn=add_zero_attn, ) self.self_attn_ln = LayerNorm(self.embed_dim) if cfg.scale_attn else None self.cross_attn_ln = LayerNorm(self.embed_dim) if cfg.scale_attn else None self.nh = self.self_attn.num_heads self.head_dim = self.self_attn.head_dim self.activation_fn = utils.get_activation_fn(activation=cfg.activation_fn) activation_dropout_p = cfg.activation_dropout if activation_dropout_p == 0: # for backwards compatibility with models that use cfg.relu_dropout activation_dropout_p = cfg.relu_dropout or 0 self.activation_dropout_module = Dropout(float(activation_dropout_p), module_name=self.__class__.__name__) self.normalize_before = cfg.decoder.normalize_before # use layerNorm rather than FusedLayerNorm for exporting. # char_inputs can be used to determint this. # TODO remove this once we update apex with the fix export = getattr(cfg, "char_inputs", False) self.self_attn_layer_norm = LayerNorm(self.embed_dim, export=export) if no_encoder_attn: self.encoder_attn = None self.encoder_attn_layer_norm = None else: self.encoder_attn = self.build_encoder_attention(self.embed_dim, cfg) self.encoder_attn_layer_norm = LayerNorm(self.embed_dim, export=export) self.ffn_layernorm = LayerNorm(cfg.decoder_ffn_embed_dim) if cfg.scale_fc else None self.w_resid = ( nn.Parameter( torch.ones( self.embed_dim, ), requires_grad=True, ) if cfg.scale_resids else None ) self.fc1 = nn.Linear(self.embed_dim, cfg.decoder.ffn_embed_dim) self.fc2 = nn.Linear(cfg.decoder.ffn_embed_dim, self.embed_dim) self.final_layer_norm = LayerNorm(self.embed_dim, export=export) self.need_attn = True self.drop_path = DropPath(drop_path_rate, batch_axis=1) if cfg.modal_ffn: self.experts_num = int(len(ModalityType)) self.experts_fc1 = nn.ModuleList([copy.deepcopy(self.fc1) for i in range(self.experts_num)]) self.experts_fc2 = nn.ModuleList([copy.deepcopy(self.fc2) for i in range(self.experts_num)]) self.gates_way = torch.eye(self.experts_num) def build_self_attention(self, embed_dim, cfg, add_bias_kv=False, add_zero_attn=False): return MultiheadAttention( embed_dim, cfg.decoder.attention_heads, dropout=cfg.attention_dropout, add_bias_kv=add_bias_kv, add_zero_attn=add_zero_attn, self_attention=not cfg.cross_self_attention, scale_factor=cfg.attn_scale_factor, scale_heads=cfg.scale_heads, use_fused=cfg.use_fused, ) def build_encoder_attention(self, embed_dim, cfg): return MultiheadAttention( embed_dim, cfg.decoder.attention_heads, kdim=cfg.encoder.embed_dim, vdim=cfg.encoder.embed_dim, dropout=cfg.attention_dropout, encoder_decoder_attention=True, scale_factor=cfg.attn_scale_factor, scale_heads=cfg.scale_heads, use_fused=cfg.use_fused, ) def residual_connection(self, x, residual): return residual + self.drop_path(x) def modal_for_ffn(self, modal_mask, x, experts_fc): modal_mask_fc = modal_mask bs, seq_len = modal_mask_fc.shape modal_mask_fc = modal_mask_fc.view(bs * seq_len) gates = torch.index_select(self.gates_way.to(modal_mask_fc.device) , 0, modal_mask_fc) gates = gates.view(bs * seq_len, self.experts_num) dispatcher = SparseDispatcher(self.experts_num, gates) seq_len, bs, dim = x.shape x = x.view(seq_len * bs, dim) expert_inputs = dispatcher.dispatch(x) expert_outputs = [experts_fc[i](expert_inputs[i]) for i in range(self.experts_num)] x = dispatcher.combine(expert_outputs) x = x.view(seq_len, bs, -1) x = x.half() return x
[docs] def forward( self, x, encoder_out: Optional[torch.Tensor] = None, encoder_padding_mask: Optional[torch.Tensor] = None, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, prev_self_attn_state: Optional[List[torch.Tensor]] = None, prev_attn_state: Optional[List[torch.Tensor]] = None, self_attn_mask: Optional[torch.Tensor] = None, self_attn_padding_mask: Optional[torch.Tensor] = None, need_attn: bool = False, need_head_weights: bool = False, self_attn_bias: Optional[Tensor] = None, cross_attn_bias: Optional[Tensor] = None, modal_mask = None, ): """ Args: x (Tensor): input to the layer of shape ``(seq_len, batch, embed_dim)`` encoder_out: encoder_padding_mask (ByteTensor, optional): binary ByteTensor of shape ``(batch, src_len)`` where padding elements are indicated by ``1``. incremental_state: self_attn_mask: self_attn_padding_mask: need_attn (bool, optional): return attention weights need_head_weights (bool, optional): return attention weights for each head (default: return average over heads). self_attn_bias (Tensor, optional): attention bias for self attention. cross_attn_bias (Tensor, optional): attenion bias for cross attention. Returns: encoded output of shape ``(seq_len, batch, embed_dim)`` """ if need_head_weights: need_attn = True residual = x if self.normalize_before: x = self.self_attn_layer_norm(x) if prev_self_attn_state is not None: prev_key, prev_value = prev_self_attn_state[:2] saved_state: Dict[str, Optional[Tensor]] = { "prev_key": prev_key, "prev_value": prev_value, } if len(prev_self_attn_state) >= 3: saved_state["prev_key_padding_mask"] = prev_self_attn_state[2] assert incremental_state is not None self.self_attn._set_input_buffer(incremental_state, saved_state) _self_attn_input_buffer = self.self_attn._get_input_buffer(incremental_state) if self.cross_self_attention and not ( incremental_state is not None and _self_attn_input_buffer is not None and "prev_key" in _self_attn_input_buffer ): if self_attn_mask is not None: assert encoder_out is not None self_attn_mask = torch.cat((x.new_zeros(x.size(0), encoder_out.size(0)), self_attn_mask), dim=1) if self_attn_padding_mask is not None: if encoder_padding_mask is None: assert encoder_out is not None encoder_padding_mask = self_attn_padding_mask.new_zeros(encoder_out.size(1), encoder_out.size(0)) self_attn_padding_mask = torch.cat((encoder_padding_mask, self_attn_padding_mask), dim=1) assert encoder_out is not None y = torch.cat((encoder_out, x), dim=0) else: y = x x, self_attn_weights = self.self_attn( query=x, key=y, value=y, key_padding_mask=self_attn_padding_mask, incremental_state=incremental_state, need_weights=need_attn, attn_mask=self_attn_mask, attn_bias=self_attn_bias, ) if self.self_attn_ln is not None: x = self.self_attn_ln(x) x = self.dropout_module(x) x = self.residual_connection(x, residual) if not self.normalize_before: x = self.self_attn_layer_norm(x) cross_attn_weights = None if self.encoder_attn is not None and encoder_out is not None: residual = x if self.normalize_before: x = self.encoder_attn_layer_norm(x) if prev_attn_state is not None: prev_key, prev_value = prev_attn_state[:2] saved_state: Dict[str, Optional[Tensor]] = { "prev_key": prev_key, "prev_value": prev_value, } if len(prev_attn_state) >= 3: saved_state["prev_key_padding_mask"] = prev_attn_state[2] assert incremental_state is not None self.encoder_attn._set_input_buffer(incremental_state, saved_state) x, cross_attn_weights = self.encoder_attn( query=x, key=encoder_out, value=encoder_out, key_padding_mask=encoder_padding_mask, incremental_state=incremental_state, static_kv=True, need_weights=need_attn or (not self.training and self.need_attn), need_head_weights=need_head_weights, attn_bias=cross_attn_bias, ) if self.cross_attn_ln is not None: x = self.cross_attn_ln(x) x = self.dropout_module(x) x = self.residual_connection(x, residual) if not self.normalize_before: x = self.encoder_attn_layer_norm(x) residual = x if self.normalize_before: x = self.final_layer_norm(x) # ffn1 modal if self.cfg.modal_ffn: x = self.modal_for_ffn(modal_mask[:x.shape[1],:x.shape[0]], x, self.experts_fc1) x = self.activation_fn(x) else: x = self.activation_fn(self.fc1(x)) x = self.activation_dropout_module(x) if self.ffn_layernorm is not None: x = self.ffn_layernorm(x) # ffn2 modal if self.cfg.modal_ffn: x = self.modal_for_ffn(modal_mask[:x.shape[1],:x.shape[0]], x, self.experts_fc2) else: x = self.fc2(x) x = self.dropout_module(x) if self.w_resid is not None: residual = torch.mul(self.w_resid, residual) x = self.residual_connection(x, residual) if not self.normalize_before: x = self.final_layer_norm(x) return x, cross_attn_weights, self_attn_weights
def make_generation_fast_(self, need_attn: bool = False, **kwargs): self.need_attn = need_attn
[docs] def upgrade_state_dict_named(self, state_dict, name): """ Rename layer norm states from `...layer_norms.0.weight` to `...self_attn_layer_norm.weight` and `...layer_norms.1.weight` to `...final_layer_norm.weight` """ # update layer norms layer_norm_map = { "0": "self_attn_layer_norm", "1": "encoder_attn_layer_norm", "2": "final_layer_norm", } for old, new in layer_norm_map.items(): for m in ("weight", "bias"): k = "{}.layer_norms.{}.{}".format(name, old, m) if k in state_dict: state_dict["{}.{}.{}".format(name, new, m)] = state_dict[k] del state_dict[k] if "{}.{}.{}".format(name, new, m) not in state_dict and "{}.{}".format(new, m) in self.state_dict(): state_dict["{}.{}.{}".format(name, new, m)] = self.state_dict()["{}.{}".format(new, m)] prefix = name + "." if name != "" else "" for param_name, param_tensor in self.state_dict().items(): if (prefix + param_name) not in state_dict: for i in range(self.experts_num): if param_name.find("experts_fc1"+"."+str(i)) != -1: state_dict[prefix + param_name] = self.state_dict()[param_name.replace("experts_fc1"+"."+str(i), "fc1")] if param_name.find("experts_fc2"+"."+str(i)) != -1: state_dict[prefix + param_name] = self.state_dict()[param_name.replace("experts_fc2"+"."+str(i), "fc2")] if param_name.find("experts_fc1") == -1 and param_name.find("experts_fc2") == -1: state_dict[prefix + param_name] = self.state_dict()[param_name]