You need to enable JavaScript to run this app. PyTorch全局剪枝固化后保存模型体积未较原模型减小的问题
优惠活动
大模型
产品
解决方案
定价
更多

PyTorch全局剪枝固化后保存模型体积未较原模型减小的问题

PyTorch全局剪枝后保存模型无体积压缩的异常问题

问题场景

基于PyTorch对深度学习模型执行全局非结构化剪枝操作时,出现如下异常:

  • 基础参数:原始未剪枝的模型大小约为77.5MB
  • 操作流程:按标准流程完成全局非结构化剪枝、逐层校验各剪枝层稀疏度、确认全局稀疏度达标,调用prune.remove()将剪枝效果永久固化后,使用torch.save保存模型状态字典
  • 异常表现:保存得到的剪枝后模型文件体积和原始未剪枝模型完全一致,未出现预期的压缩效果。现有如下完整实现代码,希望定位该问题的根本原因。

关联实现代码

import torch.nn.utils.prune as prune

parameters_to_prune = (
    (model.encoder[0], 'weight'),
    (model.up_conv1[0], 'weight'),
    (model.up_conv2[0], 'weight'),
    (model.up_conv3[0], 'weight'),
)
print(parameters_to_prune)

prune.global_unstructured(
    parameters_to_prune,
    pruning_method=prune.L1Unstructured,
    amount=0.2,
)

print(
    "Sparsity in Encoder.weight: {:.2f}%".format(
        100. * float(torch.sum(model.encoder[0].weight == 0))
        / float(model.encoder[0].weight.nelement())
    )
)
print(
    "Sparsity in up_conv1.weight: {:.2f}%".format(
        100. * float(torch.sum(model.up_conv1[0].weight == 0))
        / float(model.up_conv1[0].weight.nelement())
    )
)
print(
    "Sparsity in up_conv2.weight: {:.2f}%".format(
        100. * float(torch.sum(model.up_conv2[0].weight == 0))
        / float(model.up_conv2[0].weight.nelement())
    )
)
print(
    "Sparsity in up_conv3.weight: {:.2f}%".format(
        100. * float(torch.sum(model.up_conv3[0].weight == 0))
        / float(model.up_conv3[0].weight.nelement())
    )
)

print(
    "Global sparsity: {:.2f}%".format(
        100. * float(
            torch.sum(model.encoder[0].weight == 0)
            + torch.sum(model.up_conv1[0].weight == 0)
            + torch.sum(model.up_conv2[0].weight == 0)
            + torch.sum(model.up_conv3[0].weight == 0)
        )
        / float(
            model.encoder[0].weight.nelement()
            + model.up_conv1[0].weight.nelement()
            + model.up_conv2[0].weight.nelement()
            + model.up_conv3[0].weight.nelement()
        )
    )
)

# 剪枝效果永久固化
prune.remove(model.encoder[0], "weight")
prune.remove(model.up_conv1[0], "weight")
prune.remove(model.up_conv2[0], "weight")
prune.remove(model.up_conv3[0], "weight")

# 保存模型
PATH = "C:\\PrunedNet.pt"
torch.save(model.state_dict(), PATH)

内容的提问来源于stack exchange,提问作者Whatsinthename

相关产品推荐
方舟 Agent Plan

超全模态模型 × Harness 升级,最新支持 Deepseek-V4.1-Flash、GLM-5.3 系列、Doubao-Seedream-5.0-pro、Kimi-K3 (部分), 限时 9.9 元起

最近更新时间:2026.11.08 12:54:04