add lora finetune code

haoneng.lhn

2023-07-18 16c41542451f399bdb716f1d7cad31cf52f6f8c3

add lora finetune code

 funasr/bin/build_trainer.py

@@ -144,6 +144,8 @@
        args.patience = None
    args.local_rank = local_rank
    args.distributed = distributed
    for key, value in kwargs.items():
        args.key = value
    ASRTask.finetune_args = args

    return ASRTask

 funasr/bin/train.py

@@ -28,6 +28,7 @@
from funasr.utils.types import str2bool
from funasr.utils.types import str_or_none
from funasr.utils.yaml_no_alias_safe_dump import yaml_no_alias_safe_dump
from funasr.modules.lora.utils import mark_only_lora_as_trainable


def get_parser():
@@ -478,6 +479,18 @@
        default=None,
        help="oss bucket.",
    )
    parser.add_argument(
        "--enable_lora",
        type=str2bool,
        default=False,
        help="Apply lora for finetuning.",
    )
    parser.add_argument(
        "--lora_bias",
        type=str,
        default="none",
        help="oss bucket.",
    )

    return parser

@@ -521,6 +534,8 @@
        dtype=getattr(torch, args.train_dtype),
        device="cuda" if args.ngpu > 0 else "cpu",
    )
    if args.enable_lora:
        mark_only_lora_as_trainable(model, args.lora_bias)
    for t in args.freeze_param:
        for k, p in model.named_parameters():
            if k.startswith(t + ".") or k == t:

 funasr/modules/lora/__init__.py

 funasr/modules/lora/layers.py

New file
@@ -0,0 +1,319 @@
#  ------------------------------------------------------------------------------------------
#  Copyright (c) Microsoft Corporation. All rights reserved.
#  Licensed under the MIT License (MIT). See LICENSE in the repo root for license information.
#  ------------------------------------------------------------------------------------------
import torch
import torch.nn as nn
import torch.nn.functional as F

import math
from typing import Optional, List

class LoRALayer():
    def __init__(
        self, 
        r: int, 
        lora_alpha: int, 
        lora_dropout: float,
        merge_weights: bool,
    ):
        self.r = r
        self.lora_alpha = lora_alpha
        # Optional dropout
        if lora_dropout > 0.:
            self.lora_dropout = nn.Dropout(p=lora_dropout)
        else:
            self.lora_dropout = lambda x: x
        # Mark the weight as unmerged
        self.merged = False
        self.merge_weights = merge_weights


class Embedding(nn.Embedding, LoRALayer):
    # LoRA implemented in a dense layer
    def __init__(
        self,
        num_embeddings: int,
        embedding_dim: int,
        r: int = 0,
        lora_alpha: int = 1,
        merge_weights: bool = True,
        **kwargs
    ):
        nn.Embedding.__init__(self, num_embeddings, embedding_dim, **kwargs)
        LoRALayer.__init__(self, r=r, lora_alpha=lora_alpha, lora_dropout=0,
                           merge_weights=merge_weights)
        # Actual trainable parameters
        if r > 0:
            self.lora_A = nn.Parameter(self.weight.new_zeros((r, num_embeddings)))
            self.lora_B = nn.Parameter(self.weight.new_zeros((embedding_dim, r)))
            self.scaling = self.lora_alpha / self.r
            # Freezing the pre-trained weight matrix
            self.weight.requires_grad = False
        self.reset_parameters()

    def reset_parameters(self):
        nn.Embedding.reset_parameters(self)
        if hasattr(self, 'lora_A'):
            # initialize A the same way as the default for nn.Linear and B to zero
            nn.init.zeros_(self.lora_A)
            nn.init.normal_(self.lora_B)

    def train(self, mode: bool = True):
        nn.Embedding.train(self, mode)
        if mode:
            if self.merge_weights and self.merged:
                # Make sure that the weights are not merged
                if self.r > 0:
                    self.weight.data -= (self.lora_B @ self.lora_A).transpose(0, 1) * self.scaling
                self.merged = False
        else:
            if self.merge_weights and not self.merged:
                # Merge the weights and mark it
                if self.r > 0:
                    self.weight.data += (self.lora_B @ self.lora_A).transpose(0, 1) * self.scaling
                self.merged = True
        
    def forward(self, x: torch.Tensor):
        if self.r > 0 and not self.merged:
            result = nn.Embedding.forward(self, x)
            after_A = F.embedding(
                x, self.lora_A.transpose(0, 1), self.padding_idx, self.max_norm,
                self.norm_type, self.scale_grad_by_freq, self.sparse
            )
            result += (after_A @ self.lora_B.transpose(0, 1)) * self.scaling
            return result
        else:
            return nn.Embedding.forward(self, x)
            

class Linear(nn.Linear, LoRALayer):
    # LoRA implemented in a dense layer
    def __init__(
        self, 
        in_features: int, 
        out_features: int, 
        r: int = 0, 
        lora_alpha: int = 1, 
        lora_dropout: float = 0.,
        fan_in_fan_out: bool = False, # Set this to True if the layer to replace stores weight like (fan_in, fan_out)
        merge_weights: bool = True,
        **kwargs
    ):
        nn.Linear.__init__(self, in_features, out_features, **kwargs)
        LoRALayer.__init__(self, r=r, lora_alpha=lora_alpha, lora_dropout=lora_dropout,
                           merge_weights=merge_weights)

        self.fan_in_fan_out = fan_in_fan_out
        # Actual trainable parameters
        if r > 0:
            self.lora_A = nn.Parameter(self.weight.new_zeros((r, in_features)))
            self.lora_B = nn.Parameter(self.weight.new_zeros((out_features, r)))
            self.scaling = self.lora_alpha / self.r
            # Freezing the pre-trained weight matrix
            self.weight.requires_grad = False
        self.reset_parameters()
        if fan_in_fan_out:
            self.weight.data = self.weight.data.transpose(0, 1)

    def reset_parameters(self):
        nn.Linear.reset_parameters(self)
        if hasattr(self, 'lora_A'):
            # initialize A the same way as the default for nn.Linear and B to zero
            nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5))
            nn.init.zeros_(self.lora_B)

    def train(self, mode: bool = True):
        def T(w):
            return w.transpose(0, 1) if self.fan_in_fan_out else w
        nn.Linear.train(self, mode)
        if mode:
            if self.merge_weights and self.merged:
                # Make sure that the weights are not merged
                if self.r > 0:
                    self.weight.data -= T(self.lora_B @ self.lora_A) * self.scaling
                self.merged = False
        else:
            if self.merge_weights and not self.merged:
                # Merge the weights and mark it
                if self.r > 0:
                    self.weight.data += T(self.lora_B @ self.lora_A) * self.scaling
                self.merged = True       

    def forward(self, x: torch.Tensor):
        def T(w):
            return w.transpose(0, 1) if self.fan_in_fan_out else w
        if self.r > 0 and not self.merged:
            result = F.linear(x, T(self.weight), bias=self.bias)            
            result += (self.lora_dropout(x) @ self.lora_A.transpose(0, 1) @ self.lora_B.transpose(0, 1)) * self.scaling
            return result
        else:
            return F.linear(x, T(self.weight), bias=self.bias)


class MergedLinear(nn.Linear, LoRALayer):
    # LoRA implemented in a dense layer
    def __init__(
        self, 
        in_features: int, 
        out_features: int, 
        r: int = 0, 
        lora_alpha: int = 1, 
        lora_dropout: float = 0.,
        enable_lora: List[bool] = [False],
        fan_in_fan_out: bool = False,
        merge_weights: bool = True,
        **kwargs
    ):
        nn.Linear.__init__(self, in_features, out_features, **kwargs)
        LoRALayer.__init__(self, r=r, lora_alpha=lora_alpha, lora_dropout=lora_dropout,
                           merge_weights=merge_weights)
        assert out_features % len(enable_lora) == 0, \
            'The length of enable_lora must divide out_features'
        self.enable_lora = enable_lora
        self.fan_in_fan_out = fan_in_fan_out
        # Actual trainable parameters
        if r > 0 and any(enable_lora):
            self.lora_A = nn.Parameter(
                self.weight.new_zeros((r * sum(enable_lora), in_features)))
            self.lora_B = nn.Parameter(
                self.weight.new_zeros((out_features // len(enable_lora) * sum(enable_lora), r))
            ) # weights for Conv1D with groups=sum(enable_lora)
            self.scaling = self.lora_alpha / self.r
            # Freezing the pre-trained weight matrix
            self.weight.requires_grad = False
            # Compute the indices
            self.lora_ind = self.weight.new_zeros(
                (out_features, ), dtype=torch.bool
            ).view(len(enable_lora), -1)
            self.lora_ind[enable_lora, :] = True
            self.lora_ind = self.lora_ind.view(-1)
        self.reset_parameters()
        if fan_in_fan_out:
            self.weight.data = self.weight.data.transpose(0, 1)

    def reset_parameters(self):
        nn.Linear.reset_parameters(self)
        if hasattr(self, 'lora_A'):
            # initialize A the same way as the default for nn.Linear and B to zero
            nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5))
            nn.init.zeros_(self.lora_B)

    def zero_pad(self, x):
        result = x.new_zeros((*x.shape[:-1], self.out_features))
        result = result.view(-1, self.out_features)
        result[:, self.lora_ind] = x.reshape(
            -1, self.out_features // len(self.enable_lora) * sum(self.enable_lora)
        )
        return result.view((*x.shape[:-1], self.out_features))

    def train(self, mode: bool = True):
        def T(w):
            return w.transpose(0, 1) if self.fan_in_fan_out else w
        nn.Linear.train(self, mode)
        if mode:
            if self.merge_weights and self.merged:
                # Make sure that the weights are not merged
                if self.r > 0 and any(self.enable_lora):
                    delta_w = F.conv1d(
                        self.lora_A.data.unsqueeze(0), 
                        self.lora_B.data.unsqueeze(-1), 
                        groups=sum(self.enable_lora)
                    ).squeeze(0)
                    self.weight.data -= self.zero_pad(T(delta_w * self.scaling))
                self.merged = False
        else:
            if self.merge_weights and not self.merged:
                # Merge the weights and mark it
                if self.r > 0 and any(self.enable_lora):
                    delta_w = F.conv1d(
                        self.lora_A.data.unsqueeze(0), 
                        self.lora_B.data.unsqueeze(-1), 
                        groups=sum(self.enable_lora)
                    ).squeeze(0)
                    self.weight.data += self.zero_pad(T(delta_w * self.scaling))
                self.merged = True        

    def forward(self, x: torch.Tensor):
        def T(w):
            return w.transpose(0, 1) if self.fan_in_fan_out else w
        if self.merged:
            return F.linear(x, T(self.weight), bias=self.bias)
        else:
            result = F.linear(x, T(self.weight), bias=self.bias)
            if self.r > 0:
                after_A = F.linear(self.lora_dropout(x), self.lora_A)
                after_B = F.conv1d(
                    after_A.transpose(-2, -1), 
                    self.lora_B.unsqueeze(-1), 
                    groups=sum(self.enable_lora)
                ).transpose(-2, -1)
                result += self.zero_pad(after_B) * self.scaling
            return result
        

class ConvLoRA(nn.Module, LoRALayer):
    def __init__(self, conv_module, in_channels, out_channels, kernel_size, r=0, lora_alpha=1, lora_dropout=0., merge_weights=True, **kwargs):
        super(ConvLoRA, self).__init__()
        self.conv = conv_module(in_channels, out_channels, kernel_size, **kwargs)
        LoRALayer.__init__(self, r=r, lora_alpha=lora_alpha, lora_dropout=lora_dropout, merge_weights=merge_weights)
        assert isinstance(kernel_size, int)
        # Actual trainable parameters
        if r > 0:
            self.lora_A = nn.Parameter(
                self.conv.weight.new_zeros((r * kernel_size, in_channels * kernel_size))
            )
            self.lora_B = nn.Parameter(
              self.conv.weight.new_zeros((out_channels//self.conv.groups*kernel_size, r*kernel_size))
            )
            self.scaling = self.lora_alpha / self.r
            # Freezing the pre-trained weight matrix
            self.conv.weight.requires_grad = False
        self.reset_parameters()
        self.merged = False

    def reset_parameters(self):
        self.conv.reset_parameters()
        if hasattr(self, 'lora_A'):
            # initialize A the same way as the default for nn.Linear and B to zero
            nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5))
            nn.init.zeros_(self.lora_B)

    def train(self, mode=True):
        super(ConvLoRA, self).train(mode)
        if mode:
            if self.merge_weights and self.merged:
                if self.r > 0:
                    # Make sure that the weights are not merged
                    self.conv.weight.data -= (self.lora_B @ self.lora_A).view(self.conv.weight.shape) * self.scaling
                self.merged = False
        else:
            if self.merge_weights and not self.merged:
                if self.r > 0:
                    # Merge the weights and mark it
                    self.conv.weight.data += (self.lora_B @ self.lora_A).view(self.conv.weight.shape) * self.scaling
                self.merged = True

    def forward(self, x):
        if self.r > 0 and not self.merged:
            return self.conv._conv_forward(
                x, 
                self.conv.weight + (self.lora_B @ self.lora_A).view(self.conv.weight.shape) * self.scaling,
                self.conv.bias
            )
        return self.conv(x)

class Conv2d(ConvLoRA):
    def __init__(self, *args, **kwargs):
        super(Conv2d, self).__init__(nn.Conv2d, *args, **kwargs)

class Conv1d(ConvLoRA):
    def __init__(self, *args, **kwargs):
        super(Conv1d, self).__init__(nn.Conv1d, *args, **kwargs)

# Can Extend to other ones like this

class Conv3d(ConvLoRA):
    def __init__(self, *args, **kwargs):
        super(Conv3d, self).__init__(nn.Conv3d, *args, **kwargs)


 funasr/modules/lora/utils.py

New file
@@ -0,0 +1,50 @@
#  ------------------------------------------------------------------------------------------
#  Copyright (c) Microsoft Corporation. All rights reserved.
#  Licensed under the MIT License (MIT). See LICENSE in the repo root for license information.
#  ------------------------------------------------------------------------------------------
import torch
import torch.nn as nn

from typing import Dict

from .layers import LoRALayer


def mark_only_lora_as_trainable(model: nn.Module, bias: str = 'none') -> None:
    for n, p in model.named_parameters():
        if 'lora_' not in n and 'cif' not in n:
            p.requires_grad = False
    if bias == 'none':
        return
    elif bias == 'all':
        for n, p in model.named_parameters():
            if 'bias' in n:
                p.requires_grad = True
    elif bias == 'lora_only':
        for m in model.modules():
            if isinstance(m, LoRALayer) and \
                hasattr(m, 'bias') and \
                m.bias is not None:
                    m.bias.requires_grad = True
    else:
        raise NotImplementedError


def lora_state_dict(model: nn.Module, bias: str = 'none') -> Dict[str, torch.Tensor]:
    my_state_dict = model.state_dict()
    if bias == 'none':
        return {k: my_state_dict[k] for k in my_state_dict if 'lora_' in k}
    elif bias == 'all':
        return {k: my_state_dict[k] for k in my_state_dict if 'lora_' in k or 'bias' in k}
    elif bias == 'lora_only':
        to_return = {}
        for k in my_state_dict:
            if 'lora_' in k:
                to_return[k] = my_state_dict[k]
                bias_name = k.split('lora_')[0]+'bias'
                if bias_name in my_state_dict:
                    to_return[bias_name] = my_state_dict[bias_name]
        return to_return
    else:
        raise NotImplementedError