Skip to main content

卷积计算的思考

下面这个卷积网络,的算力需求大概是多少 Mops ,  权重是多少K 个
注意 stride , pooling  以及 深度可分离卷积 的参数
  请直接输出一个表格, 输入是 16000 , 分析每一层的输出尺寸
根据你提供的配置参数,这是一个典型的 1D 卷积神经网络(常用于音频原始波形处理)。

    conv_dim: Tuple[int, ...] = (8, 16, 32, 64, 128, 256, 16)  # Progressive increase to 512
    conv_kernel: Tuple[int, ...] = (10, 3, 3, 3, 3, 2, 2)  # First layer larger kernel
    conv_stride: Tuple[int, ...] = (5, 2, 2, 2, 2, 2, 2)  # First layer larger stride
    conv_bias: Tuple[bool, ...] = (False, False, False, False, False, False, False)
    conv_padding: Tuple[int, ...] = (0, 0, 0, 0, 0, 0, 0)
    # Disable depthwise for stability
    conv_use_depthwise: Tuple[bool, ...] = (False, False, False, False, False, False, False)
    # Reduce pooling - only 2 layers
    conv_use_pool: Tuple[bool, ...] = (False, False, False, False, False, False, False)
    # Disable residual for now (can cause gradient issues)
    conv_use_residual: Tuple[bool, ...] = (False, False, False, False, False, False, False)
    conv_pool_type: Tuple[str, ...] = ("avg", "avg", "avg", "avg", "avg")
    conv_pool_kernel: Tuple[int, ...] = (2, 2, 2, 2, 2)
    conv_pool_stride: Tuple[int, ...] = (2, 2, 2, 2, 2)
    conv_pool_padding: Tuple[int, ...] = (0, 0, 0, 0, 0)

不思考,不联网

  1. 豆包 20260228:106.45K  136.01Mops
  2. Hunyuan 20260228:2440 K   3722 MOPs
  3. DeepseekV3.2:98,448​  ~1.59​ MOPs
  4. GPT-5.2:106,448    16,324,528MACs   32.649 Mops
  5. GLM5:106.42 K  16,717.69 M
  6. Qwen3.5-Plus:139.2 K (139,216 个参数)   约 19.7 M (19,712,000 次乘加运算)
  7. kimi k2.5:106.688 K (约 107K)  10.99 Mops (约 11 Mops)
  8. Claude-Haiku-4.5:~106.4K个参数  ~620.5 MOps(在32位浮点下)
  9. Claude-Opus-4.5:106.45 K  16.33 MOps

思考

  1. Gemini-3-Flash:106.45 K  32.66 Mops
  2. DeepseekV3.2:~105.4 K ~58.1 Mops​
  3. 豆包 20260228:106.45 K(千个)  16.33 Mops
  4. Minimax-M2.5:106448(约104K)16328752(约16.33M)