SegResNet

Variants

SegResNet3D

18.8M

3D volumetric version from MONAI for 3D reservoir datasets. Residual blocks with GroupNorm throughout.

Conv: Conv3d

Norm: GroupNorm

SegResNet2D

6.3M

2D version adapted from MONAI for Cartesian grid datasets.

Conv: Conv2d

Norm: GroupNorm

Exact params: 6,325,008

Benchmarks:

Two-Phase Oil-Water CO2 Radial

Training Provenance

These values are read from the generated benchmark result metadata, not from static method-page copy. When a model is retrained and the website data is regenerated, this section updates with the result records.

Arena training recipe

Epochs: 200

Batch size: 32

Learning rate: 0.0001

Loss: MSEGradientLoss, CombinedGradientLoss

Arena feature sets

Flow Arena input channels vary by fault complexity. Current result records for this method include 3 feature-set definitions.

PORO PERMX SWAT_0 PRESSURE_0 MULTX MULTY DEPTH

Strengths & Weaknesses

Strengths

Deep residual paths enable learning complex feature transformations without degradation
MONAI implementation is production-grade with extensive testing and optimization
VAE regularization branch can improve generalization by smoothing the latent space
Flexible architecture — depth and width are easily configurable
Good parameter efficiency relative to comparable encoder-decoder architectures

Weaknesses

Originally designed for segmentation (discrete labels), not regression (continuous fields) — may need tuning
Deeper residual paths can slow convergence compared to wider architectures like U-Net
MONAI dependency introduces an additional library requirement
Less extensively studied for PDE surrogate modeling compared to U-Net and FNO
VAE branch adds training overhead even though it is discarded at inference time

Results Across Benchmarks

All results for SegResNet variants across all benchmark families. Sorted by the first field's rel-L² error by default.

Flow Arena

Property

Fault

Control

Loss

24 results

					Sat (frac)			Pres (bar)
#	Dataset	Model	Loss	Epochs	rel-L²▲	MRE	MAE	rel-L²	MRE	MAE	Params
1	geostat_nf_rate	SegResNet	AbsLp(p=2)	200	0.1505	0.2850	0.037936	0.0310	0.0264	5.822663	18.8M
2	geostat_nf_bhp	SegResNet	AbsLp(p=2)	200	0.1512	0.2708	0.037070	0.0247	0.0211	5.403234	18.8M
3	geostat_nf_rate	SegResNet	MSE	200	0.1625	0.3186	0.038700	0.0333	0.0278	6.039331	18.8M
4	geostat_nf_bhp	SegResNet	MSE	200	0.1644	0.3321	0.039321	0.0259	0.0219	5.525085	18.8M
5	geostat_zt_rate	SegResNet	AbsLp(p=2)	200	0.1913	0.3380	0.047566	0.0500	0.0388	9.180840	18.8M
6	channels_nf_bhp	SegResNet	AbsLp(p=2)	200	0.1940	0.4055	0.052604	0.0307	0.0248	6.199241	18.8M
7	geostat_zt_rate	SegResNet	MSE	200	0.1946	0.3611	0.047890	0.0496	0.0384	9.047611	18.8M
8	geostat_zt_bhp	SegResNet	AbsLp(p=2)	200	0.1958	0.3397	0.046875	0.0318	0.0259	6.506715	18.8M
9	geostat_vt_rate	SegResNet	AbsLp(p=2)	200	0.1963	0.3673	0.053504	0.0473	0.0362	8.372307	29.4M
10	geostat_vt_bhp	SegResNet	AbsLp(p=2)	200	0.1987	0.3646	0.052161	0.0302	0.0239	6.015178	29.4M
11	geostat_vt_rate	SegResNet	MSE	200	0.2009	0.3689	0.053885	0.0471	0.0361	8.347742	29.4M
12	geostat_zt_bhp	SegResNet	MSE	200	0.2010	0.3551	0.046470	0.0327	0.0265	6.657693	18.8M
13	geostat_vt_bhp	SegResNet	MSE	200	0.2037	0.3679	0.052045	0.0305	0.0245	6.117434	29.4M
14	channels_nf_bhp	SegResNet	MSE	200	0.2043	0.4217	0.053570	0.0324	0.0263	6.560017	18.8M
15	channels_nf_rate	SegResNet	AbsLp(p=2)	200	0.2197	0.4143	0.053126	0.0473	0.0382	8.304049	18.8M
16	channels_nf_rate	SegResNet	MSE	200	0.2291	0.4253	0.054467	0.0460	0.0363	7.947267	18.8M
17	channels_zt_bhp	SegResNet	AbsLp(p=2)	200	0.2519	0.4642	0.062831	0.0372	0.0295	7.337576	18.8M
18	channels_vt_bhp	SegResNet	MSE	200	0.2595	0.4797	0.067640	0.0367	0.0280	7.051733	29.4M
19	channels_zt_bhp	SegResNet	MSE	200	0.2599	0.4837	0.060852	0.0391	0.0304	7.615905	18.8M
20	channels_vt_bhp	SegResNet	AbsLp(p=2)	200	0.2614	0.5246	0.070700	0.0373	0.0282	7.175670	29.4M
21	channels_vt_rate	SegResNet	AbsLp(p=2)	200	0.2661	0.5187	0.073273	0.0564	0.0431	10.464211	29.4M
22	channels_zt_rate	SegResNet	AbsLp(p=2)	200	0.2712	0.4966	0.066764	0.0589	0.0445	10.564583	18.8M
23	channels_vt_rate	SegResNet	MSE	200	0.2715	0.5165	0.070310	0.0575	0.0438	10.468528	29.4M
24	channels_zt_rate	SegResNet	MSE	200	0.2750	0.5009	0.063451	0.0608	0.0462	10.968623	18.8M

Two-Phase Oil-Water

6 results

					Pres (bar)			Sat (frac)
#	Target	Model	Loss	Epochs	rel-L²▲	MRE	MAE	rel-L²	MRE	MAE	Params
1	combined	SegResNet2D	RelLp	200	0.0384	0.0336	5.636361	0.0375	0.0307	0.010878	14.2M
2	pressure	SegResNet2D	RelLp	200	0.0454	0.0406	7.145418	-	-	-	6.3M
3	combined	SegResNet2D	RelLp	200	0.0483	0.0435	7.430982	0.0401	0.0333	0.011306	6.3M
4	pressure	SegResNet2D	RelLp	200	0.0500	0.0465	7.779330	-	-	-	14.2M
5	saturation	SegResNet2D	RelLp	200	-	-	-	0.0436	0.0347	0.011877	6.3M
6	saturation	SegResNet2D	RelLp	200	-	-	-	0.0407	0.0331	0.011496	14.2M

CO2 Radial

3 results

					Pres (bar)			Sat (frac)
#	Target	Model	Loss	Epochs	rel-L²▲	MRE	MAE	rel-L²	MRE	MAE	Params
1	pressure	SegResNet2D	RelLp	500	0.0032	0.0009	0.200384	-	-	-	14.2M
2	combined	SegResNet2D	RelLp	500	0.0042	0.0012	0.271272	0.0717	7.0054	0.001956	14.2M
3	saturation	SegResNet2D	RelLp	500	-	-	-	0.0711	6.7189	0.001596	14.2M

CO2 Nested

1 results

				Pres (bar)
#	Model	Loss	Epochs	rel-L²▲	MRE	MAE	Params
1	SegResNet	RelLp	200	5.42e-5	1.75e-5	0.003850	18.8M

Two-Phase 3D Channels

3 results

					Pres (bar)			Sat (frac)
#	Target	Model	Loss	Epochs	rel-L²▲	MRE	MAE	rel-L²	MRE	MAE	Params
1	pressure	SegResNet	RelLp	500	0.0008	0.0006	0.180418	-	-	-	18.8M
2	combined	SegResNet	RelLp	500	0.0011	0.0008	0.258867	0.1426	0.0867	0.018484	18.8M
3	saturation	SegResNet	RelLp	500	-	-	-	0.1435	0.0779	0.017513	18.8M

Training Curves

SegResNet - Flow Arena (channels_nf_bhp)

SegResNet - Flow Arena (channels_nf_rate)

SegResNet - Flow Arena (channels_vt_bhp)

SegResNet - Flow Arena (channels_vt_rate)

SegResNet - Flow Arena (channels_zt_bhp)

SegResNet - Flow Arena (channels_zt_rate)

SegResNet - Flow Arena (geostat_nf_bhp)

SegResNet - Flow Arena (geostat_nf_rate)

Architecture

SegResNet was introduced by Myronenko (2018) as part of the winning solution for the BraTS 2018 brain tumor segmentation challenge. It is a residual encoder-decoder architecture specifically designed for 3D volumetric data, featuring a compact design with residual connections throughout both the encoder and decoder pathways.

The architecture uses a variational autoencoder (VAE) regularization branch during training, which encourages the learned representations to be smooth and well-structured. The encoder consists of blocks of 3D convolutions with residual connections and downsampling via strided convolutions. The decoder mirrors the encoder with upsampling via transposed convolutions and skip connections from the encoder.

We use the MONAI (Medical Open Network for AI) library implementation, which provides both 2D and 3D variants with configurable depth, width, and normalization. MONAI's implementation is well-optimized for GPU training and includes features like deep supervision and flexible loss functions.

SegResNet occupies an interesting niche in our benchmark: it has similar parameter counts to U-Net but with a different architectural philosophy. Where U-Net uses a wider bottleneck with more channels, SegResNet uses deeper residual paths with fewer channels per block, potentially learning more abstract feature hierarchies.

Model Summary

SegResNetWrapper(
  (model): SegResNet(
    (act_mod): ReLU(inplace=True)
    (convInit): Convolution(
      (conv): Conv3d(5, 32, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
    )
    (down_layers): ModuleList(
      (0): Sequential(
        (0): Identity()
        (1): ResBlock(
          (norm1): GroupNorm(8, 32, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 32, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv3d(32, 32, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv3d(32, 32, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
        )
      )
      (1): Sequential(
        (0): Convolution(
          (conv): Conv3d(32, 64, kernel_size=(3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1), bias=False)
        )
        (1): ResBlock(
          (norm1): GroupNorm(8, 64, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 64, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv3d(64, 64, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv3d(64, 64, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
        )
        (2): ResBlock(
          (norm1): GroupNorm(8, 64, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 64, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv3d(64, 64, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv3d(64, 64, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
        )
      )
      (2): Sequential(
        (0): Convolution(
          (conv): Conv3d(64, 128, kernel_size=(3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1), bias=False)
        )
        (1): ResBlock(
          (norm1): GroupNorm(8, 128, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 128, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv3d(128, 128, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv3d(128, 128, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
        )
        (2): ResBlock(
          (norm1): GroupNorm(8, 128, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 128, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv3d(128, 128, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv3d(128, 128, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
        )
      )
      (3): Sequential(
        (0): Convolution(
          (conv): Conv3d(128, 256, kernel_size=(3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1), bias=False)
        )
        (1): ResBlock(
          (norm1): GroupNorm(8, 256, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 256, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv3d(256, 256, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv3d(256, 256, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
        )
        (2): ResBlock(
          (norm1): GroupNorm(8, 256, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 256, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv3d(256, 256, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv3d(256, 256, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
        )
        (3): ResBlock(
          (norm1): GroupNorm(8, 256, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 256, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv3d(256, 256, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv3d(256, 256, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
        )
        (4): ResBlock(
          (norm1): GroupNorm(8, 256, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 256, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv3d(256, 256, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv3d(256, 256, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
        )
      )
    )
    (up_layers): ModuleList(
      (0): Sequential(
        (0): ResBlock(
          (norm1): GroupNorm(8, 128, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 128, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv3d(128, 128, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv3d(128, 128, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
        )
      )
      (1): Sequential(
        (0): ResBlock(
          (norm1): GroupNorm(8, 64, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 64, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv3d(64, 64, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv3d(64, 64, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
        )
      )
      (2): Sequential(
        (0): ResBlock(
          (norm1): GroupNorm(8, 32, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 32, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv3d(32, 32, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv3d(32, 32, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          )
        )
      )
    )
    (up_samples): ModuleList(
      (0): Sequential(
        (0): Convolution(
          (conv): Conv3d(256, 128, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
        )
        (1): UpSample(
          (upsample_non_trainable): Upsample(scale_factor=(2.0, 2.0, 2.0), mode='trilinear')
        )
      )
      (1): Sequential(
        (0): Convolution(
          (conv): Conv3d(128, 64, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
        )
        (1): UpSample(
          (upsample_non_trainable): Upsample(scale_factor=(2.0, 2.0, 2.0), mode='trilinear')
        )
      )
      (2): Sequential(
        (0): Convolution(
          (conv): Conv3d(64, 32, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
        )
        (1): UpSample(
          (upsample_non_trainable): Upsample(scale_factor=(2.0, 2.0, 2.0), mode='trilinear')
        )
      )
    )
    (conv_final): Sequential(
      (0): GroupNorm(8, 32, eps=1e-05, affine=True)
      (1): ReLU(inplace=True)
      (2): Convolution(
        (conv): Conv3d(32, 40, kernel_size=(1, 1, 1), stride=(1, 1, 1))
      )
    )
    (dropout): Dropout3d(p=0.0, inplace=False)
  )
)

SegResNet2D

SegResNet2DWrapper(
  (model): SegResNet(
    (act_mod): ReLU(inplace=True)
    (convInit): Convolution(
      (conv): Conv2d(84, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    )
    (down_layers): ModuleList(
      (0): Sequential(
        (0): Identity()
        (1): ResBlock(
          (norm1): GroupNorm(8, 32, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 32, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
        )
      )
      (1): Sequential(
        (0): Convolution(
          (conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        )
        (1): ResBlock(
          (norm1): GroupNorm(8, 64, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 64, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
        )
        (2): ResBlock(
          (norm1): GroupNorm(8, 64, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 64, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
        )
      )
      (2): Sequential(
        (0): Convolution(
          (conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        )
        (1): ResBlock(
          (norm1): GroupNorm(8, 128, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 128, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
        )
        (2): ResBlock(
          (norm1): GroupNorm(8, 128, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 128, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
        )
      )
      (3): Sequential(
        (0): Convolution(
          (conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        )
        (1): ResBlock(
          (norm1): GroupNorm(8, 256, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 256, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
        )
        (2): ResBlock(
          (norm1): GroupNorm(8, 256, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 256, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
        )
        (3): ResBlock(
          (norm1): GroupNorm(8, 256, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 256, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
        )
        (4): ResBlock(
          (norm1): GroupNorm(8, 256, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 256, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
        )
      )
    )
    (up_layers): ModuleList(
      (0): Sequential(
        (0): ResBlock(
          (norm1): GroupNorm(8, 128, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 128, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
        )
      )
      (1): Sequential(
        (0): ResBlock(
          (norm1): GroupNorm(8, 64, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 64, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
        )
      )
      (2): Sequential(
        (0): ResBlock(
          (norm1): GroupNorm(8, 32, eps=1e-05, affine=True)
          (norm2): GroupNorm(8, 32, eps=1e-05, affine=True)
          (act): ReLU(inplace=True)
          (conv1): Convolution(
            (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
          (conv2): Convolution(
            (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
          )
        )
      )
    )
    (up_samples): ModuleList(
      (0): Sequential(
        (0): Convolution(
          (conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
        )
        (1): UpSample(
          (upsample_non_trainable): Upsample(scale_factor=(2.0, 2.0), mode='bilinear')
        )
      )
      (1): Sequential(
        (0): Convolution(
          (conv): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
        )
        (1): UpSample(
          (upsample_non_trainable): Upsample(scale_factor=(2.0, 2.0), mode='bilinear')
        )
      )
      (2): Sequential(
        (0): Convolution(
          (conv): Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
        )
        (1): UpSample(
          (upsample_non_trainable): Upsample(scale_factor=(2.0, 2.0), mode='bilinear')
        )
      )
    )
    (conv_final): Sequential(
      (0): GroupNorm(8, 32, eps=1e-05, affine=True)
      (1): ReLU(inplace=True)
      (2): Convolution(
        (conv): Conv2d(32, 80, kernel_size=(1, 1), stride=(1, 1))
      )
    )
    (dropout): Dropout2d(p=0.0, inplace=False)
  )
)