论文阅读与源码解析：ConvLoRA

论文阅读与源码解析：Conv-LoRA: Convolution Meets LoRA: Parameter Efficient Finetuning for Segment Anything Model (ICLR 2024)

论文地址：https://arxiv.org/abs/2401.17868
GitHub项目地址：https://github.com/autogluon/autogluon/tree/master/examples/automm/Conv-LoRA
源码：https://github.com/autogluon/autogluon/issues/4165

Motivation

预训练SAM微调下游任务以提高性能
现有的工作未能分析或解决 SAM 固有的某些限制。

1. SAM的图像编码器是一个普通的ViT，缺乏对密集预测有用的特定于视觉的归纳偏差。
2. SAM的预训练本质上是一个二进制掩码预测任务，低级掩码预测预训练阻碍了 SAM 捕获对多类语义分割等任务至关重要的高级图像语义信息的能力。

目的：解决上述限制，仍然保留 SAM 在预训练期间获得的有价值的分割知识，我们在冻结大部分 SAM 的预训练权重的同时微调一小组（额外）模型参数，微调模块能增强SAM编码器与图像相关的局部先验，促进高级语义信息的获取。

Method

遵循LoRA设计策略，在预训练权重旁边添加一个旁路

1. 将输入降维
2. 将输入上采样，经过不同的专家，专家本质就是一个3x3卷积操作，然后下采样作为这个专家的输出
   
3. 由门控网络选择前k个的专家按比例相加作为输出
4. 将输出升回原来的维度

ConvLoRA模块作用：

1. 将细小的可训练线性投影层引入 SAM 编码器的每个变换器层，从而帮助恢复其提取高级语义信息的能力。
2. Conv-LoRA 在其瓶颈结构中集成了轻量级卷积层。卷积可以通过局部空间操作引入与图像相关的局部先验(即像素与其相邻像素的相关性更强)
3. 结合了多个并行卷积专家，每个专家都专门针对不同的特征尺度。将局部先验注入到图像特征的适当尺度中。

源码解读

# Modified based on https://github.com/microsoft/LoRA/blob/main/loralib/layers.py

#  ------------------------------------------------------------------------------------------
#  Copyright (c) Microsoft Corporation. All rights reserved.
#  Licensed under the MIT License (MIT). See LICENSE in the repo root for license information.
#  ------------------------------------------------------------------------------------------

import math
from typing import List, Optional

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.distributions.normal import Normal


def identity(x):
    return x


class LoRALayer:
    """
    Abstract LoRA Layer.

    Parameters
    ----------
    r
        rank r of the low-rank decomposition.
    lora_alpha
        Scaling factor. Can be simply set to same value as r as initialization is scaled already.
    merge_weights
        Merging weights during inference to reduce latency.

    References
    ----------
    1. Edward J. Hu*, Yelong Shen*, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen,
    "LoRA: Low-Rank Adaptation of Large Language Models", 2021
    https://arxiv.org/abs/2106.09685
    2. Code: https://github.com/microsoft/LoRA
    """

    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 > 0.0:
            self.lora_dropout = nn.Dropout(p=lora_dropout)
        else:
            self.lora_dropout = identity
        # Mark the weight as unmerged
        self.merged = False
        self.merge_weights = merge_weights


class ConvLoRALinear(nn.Linear, LoRALayer):
    """
    Conv-LoRA incorporated in Linear Layer. Weights of linear layer are set to be frozen per default.

    Parameters
    ----------
    in_features
        input dimension, set to the original linear layer input dimension LoRA is replacing.
    out_features
        output dimension, set to the original linear layer output dimension LoRA is replacing.
    r
        rank r of the low-rank decomposition.
    lora_alpha
        Scaling factor. Can be simply set to same value as r as initialization is scaled already.
    lora_dropout
        Dropout probability.
    fan_in_fan_out
        Set this to True if the layer to replace stores weight like (fan_in, fan_out).
    merge_weights
        Merging weights during inference to reduce latency.
    conv_lora_expert_num
        The number of experts in MoE-Conv.

    References
    ----------
    1. Zihan Zhong, Zhiqiang Tang, Tong He, Haoyang Fang, Chun Yuan,
    "Convolution Meets LoRA: Parameter Efficient Finetuning for Segment Anything Model", 2024
    https://arxiv.org/abs/2401.17868
    """

    def __init__(
        self,
        in_features: int,
        out_features: int,
        r: int = 0,
        lora_alpha: int = 1,
        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 = False,
        conv_lora_expert_num: Optional[int] = None,
        **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

            # MoE-Conv
            topk = 1
            self.lora_moe_gating = MoEGate(M=conv_lora_expert_num, d=self.r, K=topk)
            self.lora_moe_experts = nn.ModuleList([])
            self.upsample_ratios = list(range(1, conv_lora_expert_num + 1))
            for upsample_ratio in self.upsample_ratios:
                # 卷积专家本质就是3x3卷积加上一个非线性函数
                expert = nn.Conv2d(in_channels=r, out_channels=r, kernel_size=3, stride=1, padding=1, bias=True)
                expert.bias.data.zero_()
                self.lora_moe_experts.append(nn.Sequential(expert, nn.GELU()))
            self.num_experts = conv_lora_expert_num
            self.multiply_by_gates = False

        self.reset_parameters()
        if fan_in_fan_out:
            self.weight.data = self.weight.data.T

    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 T(self, w):
        return w.T if self.fan_in_fan_out else w

    def forward(self, x: torch.Tensor):
        # 输入与原来的固定权重做Linear映射
        result = F.linear(x, self.T(self.weight), bias=self.bias)
        if self.r > 0:
            # 对输入tensor经过Linear把通道维度降低
            lora_res = self.lora_dropout(x) @ self.lora_A.T
            # 转化成4维（b,c,h,w）
            dim = lora_res.dim()
            if dim == 3:
                B, L, C = lora_res.size()
                H = W = int(math.sqrt(L))
                lora_res = lora_res.reshape(B, H, W, C)
            else:
                H, W = lora_res.size()[1:3]

            # Calculate the gating values.
            lora_res = lora_res.permute(0, 3, 1, 2).contiguous()
            # 获得门控网络，并返回一个loss,避免门控网络训练时总是选择较好的那几个
            gates, moe_loss = self.lora_moe_gating(lora_res)

            # Distribute data samples to experts.
            dispatcher = SparseDispatcher(self.num_experts, gates)
            expert_inputs = dispatcher.dispatch(lora_res)
            expert_outputs = []
            # 经过不同的专家
            for i in range(self.num_experts):
                if len(expert_inputs[i]) == 0:
                    continue
                upsample_ratio = self.upsample_ratios[i]
                cur_res = expert_inputs[i]
                # 插值方式上采样
                if upsample_ratio != 1:
                    cur_res = F.interpolate(cur_res, scale_factor=upsample_ratio, mode="bicubic")
                # 对上采样的tensor输入到卷积专家中处理
                cur_res = self.lora_moe_experts[i](cur_res)
                # 下采样回原来的尺度大小
                if upsample_ratio != 1:
                    cur_res = F.interpolate(cur_res, size=(int(H), int(W)), mode="bicubic")
                expert_outputs.append(cur_res)

            # Combine data samples after processing by each expert.
            temp_lora_res = dispatcher.combine(expert_outputs, multiply_by_gates=self.multiply_by_gates)
            # 残差
            lora_res = lora_res + temp_lora_res

            lora_res = lora_res.permute(0, 2, 3, 1).contiguous()
            if dim == 3:
                lora_res = lora_res.reshape(B, L, C)
            # 线性层升回原来的通道维度，再与原来的结果相加
            result += (lora_res @ self.lora_B.T) * self.scaling

        return result, moe_loss


class MoEGate(nn.Module):
    def __init__(self, d, M=4, K=1, noisy_gating=True):
        """Constructor
        Args:
            d: input channel dimensionality.
            M: the number of experts.
            K: the number of chosen experts for each forward pass.
        """
        super(MoEGate, self).__init__()
        self.M = M
        self.k = K
        self.gap = nn.AdaptiveAvgPool2d((1, 1))  # global average pooling

        self.noisy_gating = noisy_gating

        self.w_gate = nn.Parameter(torch.zeros(d, M), requires_grad=True)
        self.w_noise = nn.Parameter(torch.zeros(d, M), requires_grad=True)

        self.softplus = nn.Softplus()
        self.softmax = nn.Softmax(1)
        self.register_buffer("mean", torch.tensor([0.0]))
        self.register_buffer("std", torch.tensor([1.0]))
        assert self.k <= self.M

    def forward(self, feats, loss_coef=1e-2, noise_epsilon=1e-2):
        batch_size = feats.shape[0]

        feats_S = self.gap(feats).view(batch_size, -1)

        clean_logits = feats_S @ self.w_gate
        if self.noisy_gating and self.training:
            raw_noise_stddev = feats_S @ self.w_noise
            noise_stddev = self.softplus(raw_noise_stddev) + noise_epsilon
            noisy_logits = clean_logits + (torch.randn_like(clean_logits) * noise_stddev)
            logits = noisy_logits
        else:
            logits = clean_logits

        top_logits, top_indices = logits.topk(min(self.k + 1, self.M), dim=1)
        top_k_logits = top_logits[:, : self.k]
        top_k_indices = top_indices[:, : self.k]
        top_k_gates = self.softmax(top_k_logits)
        zeros = torch.zeros_like(logits, requires_grad=True).float()
        gates = zeros.scatter(1, top_k_indices, top_k_gates).to(logits.dtype)

        if self.noisy_gating and self.k < self.M and self.training:
            load = (self._prob_in_top_k(clean_logits, noisy_logits, noise_stddev, top_logits)).sum(0)
        else:
            load = self._gates_to_load(gates)

        importance = gates.sum(0)
        loss = self.cv_squared(importance) + self.cv_squared(load)
        loss *= loss_coef

        return gates, loss

    def _gates_to_load(self, gates):
        """Compute the true load per expert, given the gates.
        The load is the number of examples for which the corresponding gate is >0.
        Args:
        gates: a `Tensor` of shape [batch_size, n]
        Returns:
        a float32 `Tensor` of shape [n]
        """
        return (gates > 0).sum(0)

    def cv_squared(self, x):
        """The squared coefficient of variation of a sample.
        Useful as a loss to encourage a positive distribution to be more uniform.
        Epsilons added for numerical stability.
        Returns 0 for an empty Tensor.
        Args:
        x: a `Tensor`.
        Returns:
        a `Scalar`.
        """
        eps = 1e-10

        if x.shape[0] == 1:
            return torch.tensor([0], device=x.device, dtype=x.dtype)
        return x.float().var() / (x.float().mean() ** 2 + eps)

    def _prob_in_top_k(self, clean_values, noisy_values, noise_stddev, noisy_top_values):
        """Helper function to NoisyTopKGating.
        Computes the probability that value is in top k, given different random noise.
        This gives us a way of backpropagating from a loss that balances the number
        of times each expert is in the top k experts per example.
        In the case of no noise, pass in None for noise_stddev, and the result will
        not be differentiable.
        Args:
        clean_values: a `Tensor` of shape [batch, n].
        noisy_values: a `Tensor` of shape [batch, n].  Equal to clean values plus
          normally distributed noise with standard deviation noise_stddev.
        noise_stddev: a `Tensor` of shape [batch, n], or None
        noisy_top_values: a `Tensor` of shape [batch, m].
           "values" Output of tf.top_k(noisy_top_values, m).  m >= k+1
        Returns:
        a `Tensor` of shape [batch, n].
        """
        batch = clean_values.size(0)
        m = noisy_top_values.size(1)
        top_values_flat = noisy_top_values.flatten()

        threshold_positions_if_in = torch.arange(batch, device=clean_values.device) * m + self.k
        threshold_if_in = torch.unsqueeze(torch.gather(top_values_flat, 0, threshold_positions_if_in), 1)
        is_in = torch.gt(noisy_values, threshold_if_in)
        threshold_positions_if_out = threshold_positions_if_in - 1
        threshold_if_out = torch.unsqueeze(torch.gather(top_values_flat, 0, threshold_positions_if_out), 1)
        # is each value currently in the top k.
        normal = Normal(self.mean, self.std)
        prob_if_in = normal.cdf((clean_values - threshold_if_in) / noise_stddev)
        prob_if_out = normal.cdf((clean_values - threshold_if_out) / noise_stddev)
        prob = torch.where(is_in, prob_if_in, prob_if_out)
        return prob


class SparseDispatcher(object):
    """Helper for implementing a mixture of experts.
    The purpose of this class is to create input minibatches for the
    experts and to combine the results of the experts to form a unified
    output tensor.
    There are two functions:
    dispatch - take an input Tensor and create input Tensors for each expert.
    combine - take output Tensors from each expert and form a combined output
      Tensor.  Outputs from different experts for the same batch element are
      summed together, weighted by the provided "gates".
    The class is initialized with a "gates" Tensor, which specifies which
    batch elements go to which experts, and the weights to use when combining
    the outputs.  Batch element b is sent to expert e iff gates[b, e] != 0.
    The inputs and outputs are all two-dimensional [batch, depth].
    Caller is responsible for collapsing additional dimensions prior to
    calling this class and reshaping the output to the original shape.
    See common_layers.reshape_like().
    Example use:
    gates: a float32 `Tensor` with shape `[batch_size, num_experts]`
    inputs: a float32 `Tensor` with shape `[batch_size, input_size]`
    experts: a list of length `num_experts` containing sub-networks.
    dispatcher = SparseDispatcher(num_experts, gates)
    expert_inputs = dispatcher.dispatch(inputs)
    expert_outputs = [experts[i](expert_inputs[i]) for i in range(num_experts)]
    outputs = dispatcher.combine(expert_outputs)
    The preceding code sets the output for a particular example b to:
    output[b] = Sum_i(gates[b, i] * experts[i](inputs[b]))
    This class takes advantage of sparsity in the gate matrix by including in the
    `Tensor`s for expert i only the batch elements for which `gates[b, i] > 0`.

    References
    ----------
    1. Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, Jeff Dean,
    "Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer", 2017
    https://arxiv.org/abs/1701.06538
    2. Code: https://github.com/davidmrau/mixture-of-experts/blob/master/moe.py
    """

    def __init__(self, num_experts, gates):
        """Create a SparseDispatcher."""
        self._gates = gates
        self._num_experts = num_experts
        # sort experts
        sorted_experts, index_sorted_experts = torch.nonzero(gates).sort(0)
        # drop indices
        _, self._expert_index = sorted_experts.split(1, dim=1)
        # get according batch index for each expert
        self._batch_index = torch.nonzero(gates)[index_sorted_experts[:, 1], 0]
        # calculate num samples that each expert gets
        self._part_sizes = (gates > 0).sum(0).tolist()
        # expand gates to match with self._batch_index
        gates_exp = gates[self._batch_index.flatten()]
        self._nonzero_gates = torch.gather(gates_exp, 1, self._expert_index)

    def dispatch(self, inp):
        """Create one input Tensor for each expert.
        The `Tensor` for a expert `i` contains the slices of `inp` corresponding
        to the batch elements `b` where `gates[b, i] > 0`.
        Args:
          inp: a `Tensor` of shape "[batch_size, <extra_input_dims>]`
        Returns:
          a list of `num_experts` `Tensor`s with shapes
            `[expert_batch_size_i, <extra_input_dims>]`.
        """

        # assigns samples to experts whose gate is nonzero

        # expand according to batch index so we can just split by _part_sizes
        inp_exp = inp[self._batch_index].squeeze(1)  # [bs * num_of chosen experts, dim]
        return torch.split(inp_exp, self._part_sizes, dim=0)

    def combine(self, expert_out, multiply_by_gates=True):
        """Sum together the expert output, weighted by the gates.
        The slice corresponding to a particular batch element `b` is computed
        as the sum over all experts `i` of the expert output, weighted by the
        corresponding gate values.  If `multiply_by_gates` is set to False, the
        gate values are ignored.
        Args:
          expert_out: a list of `num_experts` `Tensor`s, each with shape
            `[expert_batch_size_i, <extra_output_dims>]`.
          multiply_by_gates: a boolean
        Returns:
          a `Tensor` with shape `[batch_size, <extra_output_dims>]`.
        """
        # apply exp to expert outputs, so we are not longer in log space
        stitched = torch.cat(expert_out, 0).exp()

        if multiply_by_gates:
            # stitched = stitched.mul(self._nonzero_gates)
            stitched = stitched.mul(self._nonzero_gates.unsqueeze(-1).unsqueeze(-1))
        # zeros = torch.zeros(self._gates.size(0), expert_out[-1].size(1), requires_grad=True, device=stitched.device)
        zeros = torch.zeros(
            self._gates.size(0),
            expert_out[-1].size()[1],
            expert_out[-1].size()[2],
            expert_out[-1].size()[3],
            requires_grad=True,
            device=stitched.device,
        )
        # combine samples that have been processed by the same k experts
        combined = zeros.index_add(0, self._batch_index, stitched.float())
        # add eps to all zero values in order to avoid nans when going back to log space
        combined[combined == 0] = np.finfo(float).eps
        # back to log space
        return combined.log()

    def expert_to_gates(self):
        """Gate values corresponding to the examples in the per-expert `Tensor`s.
        Returns:
          a list of `num_experts` one-dimensional `Tensor`s with type `tf.float32`
              and shapes `[expert_batch_size_i]`
        """
        # split nonzero gates for each expert
        return torch.split(self._nonzero_gates, self._part_sizes, dim=0)
      # back to log space
        return combined.log()

    def expert_to_gates(self):
        """Gate values corresponding to the examples in the per-expert `Tensor`s.
        Returns:
          a list of `num_experts` one-dimensional `Tensor`s with type `tf.float32`
              and shapes `[expert_batch_size_i]`
        """
        # split nonzero gates for each expert
        return torch.split(self._nonzero_gates, self._part_sizes, dim=0)