wenmin-wu/cv-nested-unet-dense-skip-connections
skills/cv/nested-unet-dense-skip-connections/SKILL.md
UNet++ dense cross-depth skip connections that propagate deeper decoder features into all shallower decoder levels
$ install --global
skillsauthnpx skillsauth add wenmin-wu/ds-skills cv-nested-unet-dense-skip-connectionsInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
Security Scan Results
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
3
Scanners Passed
9
Scanners in report
Clean
TrivyContainer and dependency vulnerability scanner
95%
Clean
SemgrepStatic code analysis for vulnerabilities
95%
Clean
mcp-scan (Snyk)Model Context Protocol security validation
95%
Skipped
Snyk (dep)Open source security scanning
50%
Skipped
Socket.devSupply chain security analysis
50%
Skipped
VirusTotalMulti-engine malware detection
50%
Skipped
CrowdStrikeAdvanced threat intelligence
50%
Skipped
OSV-ScannerOpen Source Vulnerability database check
50%
Skipped
OWASP Dep-Check
50%
Last scanned: Apr 18, 2026, 2:55 AM29.2s1 file scanned
SKILL.md
- name:
- cv-nested-unet-dense-skip-connections
- description:
- UNet++ dense cross-depth skip connections that propagate deeper decoder features into all shallower decoder levels
Overview
Vanilla U-Net concatenates only one encoder feature into each decoder stage. UNet++ instead upsamples features from each decoder stage multiple times, concatenating them into all shallower decoder levels. This dense pattern gives every decoder output access to multi-scale features at the original resolution, improving segmentation of small and fine-grained structures.
Quick Start
from keras.layers import Conv2DTranspose, concatenate
# Deepest decoder (level 4)
deconv4 = Conv2DTranspose(filters*16, (3,3), strides=(2,2), padding='same')(convm)
# Upsample deconv4 to shallower levels for cross-depth skips
deconv4_up1 = Conv2DTranspose(filters*16, (3,3), strides=(2,2), padding='same')(deconv4)
deconv4_up2 = Conv2DTranspose(filters*16, (3,3), strides=(2,2), padding='same')(deconv4_up1)
deconv4_up3 = Conv2DTranspose(filters*16, (3,3), strides=(2,2), padding='same')(deconv4_up2)
deconv3 = Conv2DTranspose(filters*8, (3,3), strides=(2,2), padding='same')(uconv4)
deconv3_up1 = Conv2DTranspose(filters*8, (3,3), strides=(2,2), padding='same')(deconv3)
deconv3_up2 = Conv2DTranspose(filters*8, (3,3), strides=(2,2), padding='same')(deconv3_up1)
# Shallower decoders concatenate all upstream upsampled features
uconv3 = concatenate([deconv3, deconv4_up1, conv3])
uconv2 = concatenate([deconv2, deconv3_up1, deconv4_up2, conv2])
uconv1 = concatenate([deconv1, deconv2_up1, deconv3_up2, deconv4_up3, conv1])
Workflow
- Build encoder and bottleneck as usual
- For each decoder level, generate the primary upsampled feature
- Additionally upsample that feature K more times to match shallower resolutions
- At each shallower decoder level, concatenate: its own upsampled feature + all upsampled-extras from deeper levels + encoder skip
- Apply conv blocks on the concatenated tensor as in standard U-Net
Key Decisions
- vs. plain U-Net: More parameters and memory, but richer multi-scale context at every resolution.
- Feature count: Deeper features get upsampled multiple times — keep channel counts modest to avoid OOM.
- Deep supervision: Optionally attach segmentation heads at each decoder level for auxiliary losses.
References
- Nested Unet with EfficientNet Encoder
Related Skills
wenmin-wu/timeseries-scaled-pinball-loss
data-ai
VerifiedTrustedCommunity
Scaled Pinball Loss (SPL) metric for evaluating quantile forecasts, normalized by mean absolute successive differences of training data
31SKILL.mdUpdated Apr 23, 2026
wenmin-wu/timeseries-retroactive-outlier-rescaling
data-ai
VerifiedTrustedCommunity
Walk backward through a time series and multiplicatively rescale segments when jumps exceed a fraction of the running mean to correct data collection anomalies
31SKILL.mdUpdated Apr 23, 2026
wenmin-wu/timeseries-ratio-target-for-smape
testing
VerifiedTrustedCommunity
Transform forecasting target to next/current ratio minus one so that optimizing MAE or squared error implicitly minimizes SMAPE
31SKILL.mdUpdated Apr 23, 2026
wenmin-wu/timeseries-quantile-ratio-scaling
tools
VerifiedTrustedCommunity
Convert point forecasts to prediction intervals by scaling with logit-transformed quantile ratios passed through a Normal CDF
31SKILL.mdUpdated Apr 23, 2026