CSATv2 is a lightweight high-resolution vision backbone designed to maximize throughput at 512×512 resolution.
By applying frequency-domain compression at the input stage, the model suppresses redundant spatial information and achieves extremely fast inference.
Highlights
🚀 2,800 images/s at 512×512 resolution (A6000 1×GPU)
⚡ Frequency-domain compression for lightweight and efficient modeling
🎯 80.02% ImageNet-1K Top-1 Accuracy
🪶 Only 11M parameters
🧩 Suitable for image classification or as a high-throughput detection backbone
This model is an improved version of the architecture used in the paper
Special thanks to Juno for contributing ideas and feedback that greatly helped in lightweighting and optimizing the model.
Background and Motivation
In computational pathology, a single whole-slide image (WSI) is typically
partitioned into thousands to tens of thousands of high-resolution image
patches (e.g., 512×512 pixels) for analysis.
This setting places strong constraints on both throughput and latency:
even small inefficiencies in patch-level inference can lead to
prohibitively long end-to-end processing times at the slide level.
CSATv2 was originally designed to address this constraint by enabling
high-throughput, high-resolution inference while preserving classification
accuracy. In practical deployments, this design reduced slide-level
processing time from tens of minutes to approximately one minute,
making near–real-time pathological analysis feasible at scale.
Model description
image
Training Details
The model was trained on ImageNet-1K using a high-resolution training
pipeline adapted from common ImageNet training practices.
Dataset: ImageNet-1K
Input resolution: 512×512
Model: CSATv2
Optimizer: AdamW
Learning rate: 2e-3
Learning rate schedule: Cosine
Epochs: 300
Warmup epochs: 5
Weight decay: 2e-2
Batch size: 128 (per GPU)
Mixed precision training: Enabled (AMP)
Data Augmentation
Random resized crop (scale: 0.08–1.0, ratio: 3/4–4/3)
Horizontal flip (p = 0.5)
RandAugment (rand-m7-mstd0.5-inc1)
Mixup (α = 0.8)
CutMix (α = 1.0)
Bicubic interpolation
Regularization
Label smoothing: Disabled (handled implicitly via Mixup / CutMix)
Dropout / DropPath: Disabled
Random erase: Disabled
Optimization Details
Exponential Moving Average (EMA): Enabled
EMA decay: 0.99996
Gradient clipping: Disabled
Channels-last memory format: Optional
The training pipeline was adapted from publicly available ImageNet
training repositories (Solving ImageNet),
with task-specific modifications for high-resolution and
high-throughput training.
