Upload your SLP file to get intelligent configuration recommendations for sleap-nn training.
1
Upload Data
2
Model Type
3
Parameters
4
Export
Step 1: Upload SLP File
Drop your .slp file here or click to browse
Supports SLEAP labels files (.slp)
Loading SLP file...
Video not embedded: Upload the video file to preview frames and visualize keypoints.
Drop video file here or click to browse
Supports .mp4, .avi, .mov, .mkv
Load Existing Config (Optional)
Loaded
Loaded
Loaded
Load a previous config to use as baseline for your new training.
Data Summary
▼
Frame 0 / 0
Step 2: Select Model Type
▼
Top-Down Pipeline: This requires training two models sequentially:
Centroid Model - Detects animal centers in full images
Centered Instance Model - Detects keypoints in cropped regions
You'll configure both models and get two YAML configs.
Step 3: Configure Parameters
Data Configuration
▼
0.1251.01.0
▼Additional Data Parameters
5%10%30%
1024x768
Original Image
(max_height/max_width)
→
1024x768
Input to Model
(after scale)
→
1024x768
Model Output
(confmaps)
▼Augmentation Preview
Geometric
Rotation±15°
Scale0.9 - 1.1
Translate0%
Intensity
Brightness0%
Contrast0%
Original
Augmented
Anchor Part
?
The anchor part is used to center the crop around each instance. Choose a body part that is consistently visible and near the center of the animal. The percentage shows how often each keypoint is labeled (not NaN) across all instances.
Click a keypoint on the skeleton to select it as anchor. The percentage shows visibility (% of labeled/non-NaN instances):
Model Configuration
▼
Receptive Field Tip: The receptive field (RF) determines how much context the model sees.
Adjust Input Scale and Max Stride so the RF covers approximately the size of one animal.
Lower scale → larger effective RF (good for large animals)
Higher max_stride → larger RF but more parameters
RF: 63px | Effective RF: 63px
Receptive Field
Avg Animal Size
📖 Guide: Choosing Stride & Sigma▶
Output Stride (precision vs speed tradeoff)
Stride
Resolution
Memory
Best For
1
Full
High
High precision (small keypoints)
2
Half
Medium
Balanced (default for most)
4
Quarter
Low
Large animals, fast training
Sigma (Gaussian spread for confidence maps)
Rule of thumb: sigma ≈ keypoint_size / (2 × output_stride)
Scenario
Recommended σ
Small keypoints, high precision
1.5 - 2.5
Standard pose estimation
3.0 - 5.0
Large animals, easier training
5.0 - 7.5
💡 Tips:
If using output_stride > 1, reduce sigma proportionally
When adjusting sigma at stride=2, use half the value you'd use at stride=1
Start with defaults and adjust if keypoints are jittery (increase σ) or imprecise (decrease σ)
Part Affinity Fields (PAFs)
Identity Classification
Gaussian Spread Guide
Effective radius: 10px (2σ covers 95%)
Min distance between animal centroids: -px (keep σ < -px to avoid overlap between animals)Each keypoint has its own confidence map channel - no overlap concern between keypoints.
Min distance between connected keypoints: -px (from skeleton edges)
Model Architecture
~1.3M
Total Parameters
4
Encoder Blocks
4
Decoder Blocks
Model Input Size
?
Images are padded to be divisible by max_stride. UNet uses repeated 2× downsampling/upsampling, so dimensions must be divisible by 2^num_blocks to avoid size mismatches.
Learning Rate Scheduler
?
Adjusts the learning rate during training to improve convergence. A well-tuned scheduler can significantly improve final model accuracy.
Checkpoint Options
Additional Settings
💾 Image Cache Memory-
Raw image data:
-
With overhead (1.2×):
-
Per-worker replication:
-
⚠️ With memory caching + num_workers > 0, each worker gets a copy of the cache.
Speed Tip: To improve training speed, set Num Workers to 2 or more and change Data Pipeline to "Cache to Memory" or "Cache to Disk".
Disk Cache Settings
Hard Keypoint Mining
Logging
Estimated GPU Memory?
Rough estimate based on batch size, image dimensions, model architecture. Actual usage varies by GPU and framework.
Single Instance
~2.0 GB
Model Params~1.1M
Weights~4 MB
Batch Images~16 MB
Feature Maps~512 MB
Conf Maps~64 MB
Gradients~1.0 GB
Should fit on most GPUs (8GB+)
Configure the Centered Instance model that detects keypoints in cropped regions around each detected centroid.
The anchor point is inherited from the Centroid model configuration.
Data Configuration
▼
0.51.01.0
Crop region preview
Crop Info:
Crop size: -
Model input: -
Model output: -
Note: The orange box shows the crop region. Ensure it fully contains your animal in all poses. The anchor point (where crops are centered) is set in the Centroid tab.
▼Augmentation Settings
Geometric
Rotation±15°
Scale0.9 - 1.1
Translate0%
Intensity
Brightness0%
Contrast0%
Original Crop
Augmented
Model Configuration
▼
Head Configuration
Gaussian spread: 5px (2σ covers 95%)
Each keypoint has its own channel - no overlap concern
View Model Architecture
~1.1M
Parameters
4
Encoder Blocks
4
Decoder Blocks
Trainer Configuration
▼
Learning Rate Scheduler
?
Adjusts the learning rate during training to improve convergence.
Checkpoint Options
Additional Settings
Image Cache Memory-
Raw image data:
-
With overhead (1.2×):
-
Per-worker replication:
-
With memory caching + num_workers > 0, each worker gets a copy of the cache.
Speed Tip: To improve training speed, set Num Workers to 2+ and use "Cache to Memory" or "Cache to Disk".
Disk Cache Settings
Hard Keypoint Mining
Logging
Evaluation
Estimated GPU Memory?
Rough estimate based on batch size, crop dimensions, model architecture.
~1.0 GB
Model Params~1.1M
Weights~4 MB
Batch Crops~8 MB
Feature Maps~256 MB
Conf Maps~32 MB
Gradients~0.5 GB
Should fit on most GPUs (8GB+)
Generated Configuration
This preview updates as you change parameters. Switch back to adjust settings.
Step 4: Export Configuration
CLI Commands
?
Copy these commands to train your model using the sleap-nn command line interface. Make sure to update the paths to match your system.
Basic training command:
With labels path override:
Train Centered Instance model (after centroid):
Video Path Replacement Examples
If your video files have moved, use these options:
# Replace by order (if videos moved to new location):