Rename 'res' in line 514 to avoid with naming conflict with line 365

liugz18

2024-07-18 d80ac2fd2df4e7fb8a28acfa512bb11472b5cc99

 funasr/models/lora/layers.py

@@ -9,7 +9,8 @@
import math
from typing import Optional, List

class LoRALayer():

class LoRALayer:
    def __init__(
        self,
        r: int,
@@ -20,7 +21,7 @@
        self.r = r
        self.lora_alpha = lora_alpha
        # Optional dropout
        if lora_dropout > 0.:
        if lora_dropout > 0.0:
            self.lora_dropout = nn.Dropout(p=lora_dropout)
        else:
            self.lora_dropout = lambda x: x
@@ -41,8 +42,9 @@
        **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)
        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)))
@@ -54,7 +56,7 @@

    def reset_parameters(self):
        nn.Embedding.reset_parameters(self)
        if hasattr(self, 'lora_A'):
        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)
@@ -80,8 +82,13 @@
            result = nn.Embedding.forward(self, x)
            if self.r > 0:
                after_A = F.embedding(
                    x, self.lora_A.T, self.padding_idx, self.max_norm,
                    self.norm_type, self.scale_grad_by_freq, self.sparse
                    x,
                    self.lora_A.T,
                    self.padding_idx,
                    self.max_norm,
                    self.norm_type,
                    self.scale_grad_by_freq,
                    self.sparse,
                )
                result += (after_A @ self.lora_B.T) * self.scaling
            return result
@@ -97,14 +104,15 @@
        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)
        lora_dropout: float = 0.0,
        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)
        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
@@ -120,7 +128,7 @@

    def reset_parameters(self):
        nn.Linear.reset_parameters(self)
        if hasattr(self, 'lora_A'):
        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)
@@ -128,6 +136,7 @@
    def train(self, mode: bool = True):
        def T(w):
            return w.T if self.fan_in_fan_out else w

        nn.Linear.train(self, mode)
        if self.merge_weights and self.merged:
            # Make sure that the weights are not merged
@@ -138,6 +147,7 @@
    def eval(self):
        def T(w):
            return w.T if self.fan_in_fan_out else w

        nn.Linear.eval(self)
        if self.merge_weights and not self.merged:
            # Merge the weights and mark it
@@ -148,6 +158,7 @@
    def forward(self, x: torch.Tensor):
        def T(w):
            return w.T 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)
            if self.r > 0:
@@ -165,33 +176,34 @@
        out_features: int,
        r: int = 0,
        lora_alpha: int = 1,
        lora_dropout: float = 0.,
        lora_dropout: float = 0.0,
        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'
        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_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)
            )  # 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 = 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()
@@ -200,7 +212,7 @@

    def reset_parameters(self):
        nn.Linear.reset_parameters(self)
        if hasattr(self, 'lora_A'):
        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)
@@ -216,6 +228,7 @@
    def train(self, mode: bool = True):
        def T(w):
            return w.T if self.fan_in_fan_out else w

        nn.Linear.train(self, mode)
        if self.merge_weights and self.merged:
            # Make sure that the weights are not merged
@@ -223,7 +236,7 @@
                delta_w = F.conv1d(
                    self.lora_A.data.unsqueeze(0),
                    self.lora_B.data.unsqueeze(-1),
                    groups=sum(self.enable_lora)
                    groups=sum(self.enable_lora),
                ).squeeze(0)
                self.weight.data -= self.zero_pad(T(delta_w * self.scaling))
            self.merged = False
@@ -231,6 +244,7 @@
    def eval(self):
        def T(w):
            return w.T if self.fan_in_fan_out else w

        nn.Linear.eval(self)
        if self.merge_weights and not self.merged:
            # Merge the weights and mark it
@@ -238,7 +252,7 @@
                delta_w = F.conv1d(
                    self.lora_A.data.unsqueeze(0),
                    self.lora_B.data.unsqueeze(-1),
                    groups=sum(self.enable_lora)
                    groups=sum(self.enable_lora),
                ).squeeze(0)
                self.weight.data += self.zero_pad(T(delta_w * self.scaling))
            self.merged = True
@@ -246,6 +260,7 @@
    def forward(self, x: torch.Tensor):
        def T(w):
            return w.T if self.fan_in_fan_out else w

        if self.merged:
            return F.linear(x, T(self.weight), bias=self.bias)
        else:
@@ -255,7 +270,7 @@
                after_B = F.conv1d(
                    after_A.transpose(-2, -1),
                    self.lora_B.unsqueeze(-1),
                    groups=sum(self.enable_lora)
                    groups=sum(self.enable_lora),
                ).transpose(-2, -1)
                result += self.zero_pad(after_B) * self.scaling
            return result
@@ -270,21 +285,22 @@
        kernel_size: int,
        r: int = 0,
        lora_alpha: int = 1,
        lora_dropout: float = 0.,
        lora_dropout: float = 0.0,
        merge_weights: bool = True,
        **kwargs
    ):
        nn.Conv2d.__init__(self, in_channels, out_channels, kernel_size, **kwargs)
        LoRALayer.__init__(self, r=r, lora_alpha=lora_alpha, lora_dropout=lora_dropout,
                           merge_weights=merge_weights)
        LoRALayer.__init__(
            self, r=r, lora_alpha=lora_alpha, lora_dropout=lora_dropout, merge_weights=merge_weights
        )
        assert type(kernel_size) is int
        # Actual trainable parameters
        if r > 0:
            self.lora_A = nn.Parameter(
                self.weight.new_zeros((r*kernel_size, in_channels*kernel_size))
                self.weight.new_zeros((r * kernel_size, in_channels * kernel_size))
            )
            self.lora_B = nn.Parameter(
                self.weight.new_zeros((out_channels*kernel_size, r*kernel_size))
                self.weight.new_zeros((out_channels * kernel_size, r * kernel_size))
            )
            self.scaling = self.lora_alpha / self.r
            # Freezing the pre-trained weight matrix
@@ -293,7 +309,7 @@

    def reset_parameters(self):
        nn.Conv2d.reset_parameters(self)
        if hasattr(self, 'lora_A'):
        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)
@@ -317,7 +333,10 @@
            return F.conv2d(
                x,
                self.weight + (self.lora_B @ self.lora_A).view(self.weight.shape) * self.scaling,
                self.bias, self.stride, self.padding, self.dilation, self.groups
                self.bias,
                self.stride,
                self.padding,
                self.dilation,
                self.groups,
            )
        return nn.Conv2d.forward(self, x)