查看 pytorch 网络结构¶
约 41 个字 417 行代码 预计阅读时间 5 分钟
法①：print()¶
实例化好模型即可 print()
Python
# 实例化模型
model = Model(configs)
print("模型结构:")
print(model)
# 计算模型参数数量
total_params = sum(p.numel() for p in model.parameters())
print(f"模型参数数量: {total_params}")
输出：
输出：
Python
模型结构:
Model(
(revin_layer): RevIN()
(Maxpools): ModuleList(
(0-3): 4 x AvgPool1d(kernel_size=(3,), stride=(2,), padding=(1,))
)
(down_blocks): ModuleList(
(0): block_model(
(Linear_channel): ModuleList(
(0-6): 7 x Linear(in_features=96, out_features=720, bias=True)
)
(ln): LayerNorm((720,), eps=1e-05, elementwise_affine=True)
(relu): ReLU(inplace=True)
)
(1): block_model(
(Linear_channel): ModuleList(
(0-6): 7 x Linear(in_features=48, out_features=360, bias=True)
)
(ln): LayerNorm((360,), eps=1e-05, elementwise_affine=True)
(relu): ReLU(inplace=True)
)
(2): block_model(
(Linear_channel): ModuleList(
(0-6): 7 x Linear(in_features=24, out_features=180, bias=True)
)
(ln): LayerNorm((180,), eps=1e-05, elementwise_affine=True)
(relu): ReLU(inplace=True)
)
(3): block_model(
(Linear_channel): ModuleList(
(0-6): 7 x Linear(in_features=12, out_features=90, bias=True)
)
(ln): LayerNorm((90,), eps=1e-05, elementwise_affine=True)
(relu): ReLU(inplace=True)
)
)
(dct_models): ModuleList(
(0): dct_channel_block(
(fc): Sequential(
(0): Linear(in_features=96, out_features=192, bias=False)
(1): Dropout(p=0.5, inplace=False)
(2): ReLU(inplace=True)
(3): Linear(in_features=192, out_features=96, bias=False)
(4): Dropout(p=0.5, inplace=False)
(5): Sigmoid()
)
(dct_norm): LayerNorm((96,), eps=1e-06, elementwise_affine=True)
)
(1): dct_channel_block(
(fc): Sequential(
(0): Linear(in_features=48, out_features=96, bias=False)
(1): Dropout(p=0.5, inplace=False)
(2): ReLU(inplace=True)
(3): Linear(in_features=96, out_features=48, bias=False)
(4): Dropout(p=0.5, inplace=False)
(5): Sigmoid()
)
(dct_norm): LayerNorm((48,), eps=1e-06, elementwise_affine=True)
)
(2): dct_channel_block(
(fc): Sequential(
(0): Linear(in_features=24, out_features=48, bias=False)
(1): Dropout(p=0.5, inplace=False)
(2): ReLU(inplace=True)
(3): Linear(in_features=48, out_features=24, bias=False)
(4): Dropout(p=0.5, inplace=False)
(5): Sigmoid()
)
(dct_norm): LayerNorm((24,), eps=1e-06, elementwise_affine=True)
)
(3): dct_channel_block(
(fc): Sequential(
(0): Linear(in_features=12, out_features=24, bias=False)
(1): Dropout(p=0.5, inplace=False)
(2): ReLU(inplace=True)
(3): Linear(in_features=24, out_features=12, bias=False)
(4): Dropout(p=0.5, inplace=False)
(5): Sigmoid()
)
(dct_norm): LayerNorm((12,), eps=1e-06, elementwise_affine=True)
)
)
(icbs): ModuleList(
(0): ICB(
(conv1): Conv1d(270, 270, kernel_size=(1,), stride=(1,))
(conv3): Conv1d(270, 270, kernel_size=(3,), stride=(1,), padding=(1,))
(conv5): Conv1d(270, 270, kernel_size=(5,), stride=(1,), padding=(2,))
(conv7): Conv1d(270, 270, kernel_size=(7,), stride=(1,), padding=(3,))
(conv_out): Conv1d(1080, 180, kernel_size=(1,), stride=(1,))
(drop): Dropout(p=0.5, inplace=False)
(act): Tanh()
(sigmoid): Sigmoid()
(relu): ReLU()
)
(1): ICB(
(conv1): Conv1d(540, 540, kernel_size=(1,), stride=(1,))
(conv3): Conv1d(540, 540, kernel_size=(3,), stride=(1,), padding=(1,))
(conv5): Conv1d(540, 540, kernel_size=(5,), stride=(1,), padding=(2,))
(conv7): Conv1d(540, 540, kernel_size=(7,), stride=(1,), padding=(3,))
(conv_out): Conv1d(2160, 360, kernel_size=(1,), stride=(1,))
(drop): Dropout(p=0.5, inplace=False)
(act): Tanh()
(sigmoid): Sigmoid()
(relu): ReLU()
)
(2): ICB(
(conv1): Conv1d(1080, 1080, kernel_size=(1,), stride=(1,))
(conv3): Conv1d(1080, 1080, kernel_size=(3,), stride=(1,), padding=(1,))
(conv5): Conv1d(1080, 1080, kernel_size=(5,), stride=(1,), padding=(2,))
(conv7): Conv1d(1080, 1080, kernel_size=(7,), stride=(1,), padding=(3,))
(conv_out): Conv1d(4320, 720, kernel_size=(1,), stride=(1,))
(drop): Dropout(p=0.5, inplace=False)
(act): Tanh()
(sigmoid): Sigmoid()
(relu): ReLU()
)
)
(channel_mixer): ChebyKANLayer(
(fc1): ChebyKANLinear()
)
)
模型参数数量: 29289904
输入形状: torch.Size([16, 96, 7])
输出形状: torch.Size([16, 720, 7])
预期输出形状: [batch_size=16, pred_len=720, enc_in=7]
法②：torchinfo summary()¶
注意在同一个设备上
Python
pip install torchinfo
from torchinfo import summary
model = Model(configs)
x = torch.randn(batch_size, seq_len, enc_in)
summary(model, input_data=x)
输出：
Python
==========================================================================================
Layer (type:depth-idx) Output Shape Param #
==========================================================================================
Model [16, 720, 7] --
├─RevIN: 1-1 [16, 96, 7] 14
├─ChebyKANLayer: 1-2 [16, 96, 7] --
│ └─ChebyKANLinear: 2-1 [1536, 7] 196
├─ModuleList: 1-12 -- (recursive)
│ └─block_model: 2-2 [16, 7, 720] 1,440
│ │ └─ModuleList: 3-1 -- 488,880
├─ModuleList: 1-13 -- (recursive)
│ └─dct_channel_block: 2-3 [16, 7, 96] 1
│ │ └─LayerNorm: 3-2 [16, 7, 96] 192
│ │ └─Sequential: 3-3 [16, 7, 96] 36,864
│ │ └─LayerNorm: 3-4 [16, 7, 96] (recursive)
├─ModuleList: 1-14 -- --
│ └─AvgPool1d: 2-4 [16, 7, 48] --
├─ModuleList: 1-12 -- (recursive)
│ └─block_model: 2-5 [16, 7, 360] 720
│ │ └─ModuleList: 3-5 -- 123,480
├─ModuleList: 1-13 -- (recursive)
│ └─dct_channel_block: 2-6 [16, 7, 48] 1
│ │ └─LayerNorm: 3-6 [16, 7, 48] 96
│ │ └─Sequential: 3-7 [16, 7, 48] 9,216
│ │ └─LayerNorm: 3-8 [16, 7, 48] (recursive)
├─ModuleList: 1-14 -- --
│ └─AvgPool1d: 2-7 [16, 7, 24] --
├─ModuleList: 1-12 -- (recursive)
│ └─block_model: 2-8 [16, 7, 180] 360
│ │ └─ModuleList: 3-9 -- 31,500
├─ModuleList: 1-13 -- (recursive)
│ └─dct_channel_block: 2-9 [16, 7, 24] 1
│ │ └─LayerNorm: 3-10 [16, 7, 24] 48
│ │ └─Sequential: 3-11 [16, 7, 24] 2,304
│ │ └─LayerNorm: 3-12 [16, 7, 24] (recursive)
├─ModuleList: 1-14 -- --
│ └─AvgPool1d: 2-10 [16, 7, 12] --
├─ModuleList: 1-12 -- (recursive)
│ └─block_model: 2-11 [16, 7, 90] 180
│ │ └─ModuleList: 3-13 -- 8,190
├─ModuleList: 1-13 -- (recursive)
│ └─dct_channel_block: 2-12 [16, 7, 12] 1
│ │ └─LayerNorm: 3-14 [16, 7, 12] 24
│ │ └─Sequential: 3-15 [16, 7, 12] 576
│ │ └─LayerNorm: 3-16 [16, 7, 12] (recursive)
├─ModuleList: 1-14 -- --
│ └─AvgPool1d: 2-13 [16, 7, 6] --
├─ChebyKANLayer: 1-15 [16, 90, 7] (recursive)
│ └─ChebyKANLinear: 2-14 [1440, 7] (recursive)
├─ModuleList: 1-16 -- --
│ └─ICB: 2-15 [16, 7, 180] --
│ │ └─Conv1d: 3-17 [16, 270, 7] 73,170
│ │ └─Conv1d: 3-18 [16, 270, 7] 218,970
│ │ └─Conv1d: 3-19 [16, 270, 7] 364,770
│ │ └─Conv1d: 3-20 [16, 270, 7] 510,570
│ │ └─Tanh: 3-21 [16, 270, 7] --
│ │ └─Sigmoid: 3-22 [16, 270, 7] --
│ │ └─ReLU: 3-23 [16, 270, 7] --
│ │ └─Tanh: 3-24 [16, 270, 7] --
│ │ └─Sigmoid: 3-25 [16, 270, 7] --
│ │ └─ReLU: 3-26 [16, 270, 7] --
│ │ └─Tanh: 3-27 [16, 270, 7] --
│ │ └─Sigmoid: 3-28 [16, 270, 7] --
│ │ └─ReLU: 3-29 [16, 270, 7] --
│ │ └─Tanh: 3-30 [16, 270, 7] --
│ │ └─Sigmoid: 3-31 [16, 270, 7] --
│ │ └─ReLU: 3-32 [16, 270, 7] --
│ │ └─Dropout: 3-33 [16, 1080, 7] --
│ │ └─Conv1d: 3-34 [16, 180, 7] 194,580
│ └─ICB: 2-16 [16, 7, 360] --
│ │ └─Conv1d: 3-35 [16, 540, 7] 292,140
│ │ └─Conv1d: 3-36 [16, 540, 7] 875,340
│ │ └─Conv1d: 3-37 [16, 540, 7] 1,458,540
│ │ └─Conv1d: 3-38 [16, 540, 7] 2,041,740
│ │ └─Tanh: 3-39 [16, 540, 7] --
│ │ └─Sigmoid: 3-40 [16, 540, 7] --
│ │ └─ReLU: 3-41 [16, 540, 7] --
│ │ └─Tanh: 3-42 [16, 540, 7] --
│ │ └─Sigmoid: 3-43 [16, 540, 7] --
│ │ └─ReLU: 3-44 [16, 540, 7] --
│ │ └─Tanh: 3-45 [16, 540, 7] --
│ │ └─Sigmoid: 3-46 [16, 540, 7] --
│ │ └─ReLU: 3-47 [16, 540, 7] --
│ │ └─Tanh: 3-48 [16, 540, 7] --
│ │ └─Sigmoid: 3-49 [16, 540, 7] --
│ │ └─ReLU: 3-50 [16, 540, 7] --
│ │ └─Dropout: 3-51 [16, 2160, 7] --
│ │ └─Conv1d: 3-52 [16, 360, 7] 777,960
│ └─ICB: 2-17 [16, 7, 720] --
│ │ └─Conv1d: 3-53 [16, 1080, 7] 1,167,480
│ │ └─Conv1d: 3-54 [16, 1080, 7] 3,500,280
│ │ └─Conv1d: 3-55 [16, 1080, 7] 5,833,080
│ │ └─Conv1d: 3-56 [16, 1080, 7] 8,165,880
│ │ └─Tanh: 3-57 [16, 1080, 7] --
│ │ └─Sigmoid: 3-58 [16, 1080, 7] --
│ │ └─ReLU: 3-59 [16, 1080, 7] --
│ │ └─Tanh: 3-60 [16, 1080, 7] --
│ │ └─Sigmoid: 3-61 [16, 1080, 7] --
│ │ └─ReLU: 3-62 [16, 1080, 7] --
│ │ └─Tanh: 3-63 [16, 1080, 7] --
│ │ └─Sigmoid: 3-64 [16, 1080, 7] --
│ │ └─ReLU: 3-65 [16, 1080, 7] --
│ │ └─Tanh: 3-66 [16, 1080, 7] --
│ │ └─Sigmoid: 3-67 [16, 1080, 7] --
│ │ └─ReLU: 3-68 [16, 1080, 7] --
│ │ └─Dropout: 3-69 [16, 4320, 7] --
│ │ └─Conv1d: 3-70 [16, 720, 7] 3,111,120
├─RevIN: 1-17 [16, 720, 7] (recursive)
==========================================================================================
Total params: 29,289,904
Trainable params: 29,289,904
Non-trainable params: 0
Total mult-adds (G): 3.21
==========================================================================================
Input size (MB): 0.04
Forward/backward pass size (MB): 10.82
Params size (MB): 117.15
Estimated Total Size (MB): 128.01
==========================================================================================
法③：torchsummary summary()¶
Python
from torchsummary import summary
# 实例化模型
model = Model(configs)
# 生成随机输入数据
batch_size = 16
seq_len = configs.seq_len
enc_in = configs.enc_in
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
x = torch.randn(batch_size, seq_len, enc_in).to(device)
model = model.to(device)
# summary(model, input_data=x)
summary(model, input_size=(configs.seq_len, configs.enc_in), batch_size=batch_size, device=device.type)
输出：
Python
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
RevIN-1 [16, 96, 7] 0
ChebyKANLinear-2 [16, 7] 0
ChebyKANLayer-3 [16, 96, 7] 0
Linear-4 [16, 720] 69,840
Linear-5 [16, 720] 69,840
Linear-6 [16, 720] 69,840
Linear-7 [16, 720] 69,840
Linear-8 [16, 720] 69,840
Linear-9 [16, 720] 69,840
Linear-10 [16, 720] 69,840
block_model-11 [16, 7, 720] 0
LayerNorm-12 [16, 7, 96] 192
Linear-13 [16, 7, 192] 18,432
Dropout-14 [16, 7, 192] 0
ReLU-15 [16, 7, 192] 0
Linear-16 [16, 7, 96] 18,432
Dropout-17 [16, 7, 96] 0
Sigmoid-18 [16, 7, 96] 0
LayerNorm-19 [16, 7, 96] 192
dct_channel_block-20 [16, 7, 96] 0
AvgPool1d-21 [16, 7, 48] 0
Linear-22 [16, 360] 17,640
Linear-23 [16, 360] 17,640
Linear-24 [16, 360] 17,640
Linear-25 [16, 360] 17,640
Linear-26 [16, 360] 17,640
Linear-27 [16, 360] 17,640
Linear-28 [16, 360] 17,640
block_model-29 [16, 7, 360] 0
LayerNorm-30 [16, 7, 48] 96
Linear-31 [16, 7, 96] 4,608
Dropout-32 [16, 7, 96] 0
ReLU-33 [16, 7, 96] 0
Linear-34 [16, 7, 48] 4,608
Dropout-35 [16, 7, 48] 0
Sigmoid-36 [16, 7, 48] 0
LayerNorm-37 [16, 7, 48] 96
dct_channel_block-38 [16, 7, 48] 0
AvgPool1d-39 [16, 7, 24] 0
Linear-40 [16, 180] 4,500
Linear-41 [16, 180] 4,500
Linear-42 [16, 180] 4,500
Linear-43 [16, 180] 4,500
Linear-44 [16, 180] 4,500
Linear-45 [16, 180] 4,500
Linear-46 [16, 180] 4,500
block_model-47 [16, 7, 180] 0
LayerNorm-48 [16, 7, 24] 48
Linear-49 [16, 7, 48] 1,152
Dropout-50 [16, 7, 48] 0
ReLU-51 [16, 7, 48] 0
Linear-52 [16, 7, 24] 1,152
Dropout-53 [16, 7, 24] 0
Sigmoid-54 [16, 7, 24] 0
LayerNorm-55 [16, 7, 24] 48
dct_channel_block-56 [16, 7, 24] 0
AvgPool1d-57 [16, 7, 12] 0
Linear-58 [16, 90] 1,170
Linear-59 [16, 90] 1,170
Linear-60 [16, 90] 1,170
Linear-61 [16, 90] 1,170
Linear-62 [16, 90] 1,170
Linear-63 [16, 90] 1,170
Linear-64 [16, 90] 1,170
block_model-65 [16, 7, 90] 0
LayerNorm-66 [16, 7, 12] 24
Linear-67 [16, 7, 24] 288
Dropout-68 [16, 7, 24] 0
ReLU-69 [16, 7, 24] 0
Linear-70 [16, 7, 12] 288
Dropout-71 [16, 7, 12] 0
Sigmoid-72 [16, 7, 12] 0
LayerNorm-73 [16, 7, 12] 24
dct_channel_block-74 [16, 7, 12] 0
AvgPool1d-75 [16, 7, 6] 0
ChebyKANLinear-76 [16, 7] 0
ChebyKANLayer-77 [16, 90, 7] 0
Conv1d-78 [16, 270, 7] 73,170
Conv1d-79 [16, 270, 7] 218,970
Conv1d-80 [16, 270, 7] 364,770
Conv1d-81 [16, 270, 7] 510,570
Tanh-82 [16, 270, 7] 0
Sigmoid-83 [16, 270, 7] 0
ReLU-84 [16, 270, 7] 0
Tanh-85 [16, 270, 7] 0
Sigmoid-86 [16, 270, 7] 0
ReLU-87 [16, 270, 7] 0
Tanh-88 [16, 270, 7] 0
Sigmoid-89 [16, 270, 7] 0
ReLU-90 [16, 270, 7] 0
Tanh-91 [16, 270, 7] 0
Sigmoid-92 [16, 270, 7] 0
ReLU-93 [16, 270, 7] 0
Dropout-94 [16, 1080, 7] 0
Conv1d-95 [16, 180, 7] 194,580
ICB-96 [16, 7, 180] 0
Conv1d-97 [16, 540, 7] 292,140
Conv1d-98 [16, 540, 7] 875,340
Conv1d-99 [16, 540, 7] 1,458,540
Conv1d-100 [16, 540, 7] 2,041,740
Tanh-101 [16, 540, 7] 0
Sigmoid-102 [16, 540, 7] 0
ReLU-103 [16, 540, 7] 0
Tanh-104 [16, 540, 7] 0
Sigmoid-105 [16, 540, 7] 0
ReLU-106 [16, 540, 7] 0
Tanh-107 [16, 540, 7] 0
Sigmoid-108 [16, 540, 7] 0
ReLU-109 [16, 540, 7] 0
Tanh-110 [16, 540, 7] 0
Sigmoid-111 [16, 540, 7] 0
ReLU-112 [16, 540, 7] 0
Dropout-113 [16, 2160, 7] 0
Conv1d-114 [16, 360, 7] 777,960
ICB-115 [16, 7, 360] 0
Conv1d-116 [16, 1080, 7] 1,167,480
Conv1d-117 [16, 1080, 7] 3,500,280
Conv1d-118 [16, 1080, 7] 5,833,080
Conv1d-119 [16, 1080, 7] 8,165,880
Tanh-120 [16, 1080, 7] 0
Sigmoid-121 [16, 1080, 7] 0
ReLU-122 [16, 1080, 7] 0
Tanh-123 [16, 1080, 7] 0
Sigmoid-124 [16, 1080, 7] 0
ReLU-125 [16, 1080, 7] 0
Tanh-126 [16, 1080, 7] 0
Sigmoid-127 [16, 1080, 7] 0
ReLU-128 [16, 1080, 7] 0
Tanh-129 [16, 1080, 7] 0
Sigmoid-130 [16, 1080, 7] 0
ReLU-131 [16, 1080, 7] 0
Dropout-132 [16, 4320, 7] 0
Conv1d-133 [16, 720, 7] 3,111,120
ICB-134 [16, 7, 720] 0
RevIN-135 [16, 720, 7] 0
================================================================
Total params: 29,287,350
Trainable params: 29,287,350
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.04
Forward/backward pass size (MB): 39.54
Params size (MB): 111.72
Estimated Total Size (MB): 151.30
----------------------------------------------------------------
输入形状: torch.Size([16, 96, 7])
2025-04-18 14:03:332025-04-18 15:34:07