Spaces:
Sleeping
Sleeping
Commit History
-
fix: adjust figure size for scatter plot in generate_scatter_plot function
Signed-off-by: Simone Rossi <simone.rossi@eurecom.fr>a8d5869 -
refactor: format code for readability in generate_scatter_plot function
Signed-off-by: Simone Rossi <simone.rossi@eurecom.fr>3d7b01d
-
feat: add depth slider and cache figures in BNN prior demo
- Add depth slider (3-10) to select number of layers - Move plot code into cached functions (plot_prior, plot_history, plot_posterior) - Fix SLURM qos to backfill and use MIG GPU config Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>4078c11 -
feat: add SLURM array job for BNN prior depth sweep
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>b28cdfa -
feat: add BNN prior optimization demo page
Add interactive demo page visualizing the optimization of BNN prior standard deviations to match a Gaussian process prior. Includes prior comparison and posterior predictive tabs. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>1bdf5da
-
fix: update save directory default and optimize inference for CPU device
Signed-off-by: Simone Rossi <simone.rossi@eurecom.fr>1d20dc2 -
refactor: inline inference into main(), add CLI args and save all artifacts to parameterized dir
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>1dfbfff -
fix: update bnn-priors git source from SSH to HTTPS in requirements and lock files
Signed-off-by: Simone Rossi <simone.rossi@eurecom.fr>fbba934 -
fix: update SLURM array range and reduce job time; add steps parameter to script
Signed-off-by: Simone Rossi <simone.rossi@eurecom.fr>afe4e2b -
feat: work-in-progress on prior demo
Signed-off-by: Simone Rossi <simone.rossi@eurecom.fr>5e43ad9
-
fix: adjust training set size range and default value in UI
Signed-off-by: Simone Rossi <simone.rossi@eurecom.fr>79f1186 -
fix: use to_numpy() to fix Pylance type error on comparison
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>4a9e73e -
feat: use log scale x-axis for grokking plot with numbered ticks
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>13f7fca
-
feat: add SLURM array job for grokking sweep
- Sweeps train_size from 1000 to 9000 (step 1000) - Array job with 9 tasks on h200 partition4b096f4 -
fix: config adam correctly
Signed-off-by: Simone Rossi <simone.rossi@eurecom.fr>5cb1d0d -
fix: include pyarrow
Signed-off-by: Simone Rossi <simone.rossi@eurecom.fr>de88999 -
revert: restore PEP 723 script for grokking training
- Revert pyproject.toml to CPU PyTorch - Grokking script uses PEP 723 with CUDA 12.1 index4ee84e3 -
feat: switch to CUDA 12.1 for PyTorch
- Remove PEP 723 script from grokking training - Update pyproject.toml to use cu121 index instead of cpu6532e49 -
feat: add CUDA 12.1 support for grokking script
- Use PyTorch cu121 index for GPU acceleration - Let litgpt pull torch dependencyc31d89a -
cleanup: remove compute workspace and adjust layout
- Remove compute/ pyproject.toml and src/ (now using PEP 723 scripts) - Widen column layout in aleatoric_epistemic_uncertainty.py3f68c70 -
refactor: convert grokking script to PEP 723 format
- Add inline dependency metadata for uv run - Allows running directly with: uv run compute/grokking_modular_arithmetic.py7e17c63 -
feat: add grokking demo page with chat interface
- Grokking curve plot showing train/test accuracy - Chat interface to test the model on modular addition - Probability distribution bar chart for predictions - New 'Deep Learning' section in navigationfd0b0b7 -
feat: add grokking training script for modular arithmetic
- Small transformer (2 layers, 128 emb, 4 heads) using LitGPT - Task: (a, b) -> (a + b) mod 97 - Training: AdamW, lr=0.01, 10000 steps - Saves metrics as Parquet and model as .pt5902f32
-
feat: add DEMO_DATA_DIR env var for flexible data directory
- Use DEMO_DATA_DIR env var to configure data path (default: ./data) - Set DEMO_DATA_DIR=/data in Dockerfile for deployed environment - Enables local development with ./data and production with /data mount Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>c8530a3 -
feat: add Bayesian neural network uncertainty decomposition with Parquet caching
- Add aleatoric vs epistemic uncertainty visualization using HMC inference - Implement disk-based caching with Parquet for persistent results storage - Add data/ directory to gitignore for synced Hugging Face data - Update AGENTS.md with data directory and sync command Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>94d0706
-
feat: improve covariance matrix visualization with color-coded crosshairs and scatter points
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>1e11244 -
feat: reduce contour levels in GP joint distribution plot
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>f6021a2 -
feat: add Gaussian Process joint distribution visualization
- Add GP prior joint demo with three-panel layout (GridSpec 3:1:1) - Register demo in navigation under "Gaussian Processes" section - Add scipy dependency for multivariate normal PDF - Minor CSS styling updates for typography Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>7e6b171
-
feat: add deep net prior pathologies visualization
- New demo showing pathologies in deep neural network priors with NTK parameterization - Tanh activation, width=32, N(0,1) weight and bias prior - Interactive depth slider (2-30 layers) - 512 samples cached with @st.cache_resource - 95% and 99% confidence interval bands - 1D input range [-12, 12] - PyTorch CPU dependency added Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>6eac6a7 -
feat: add compute workspace for heavy model training
- Create compute/ workspace with pyproject.toml for artifact generation - Configure uv workspace with demos as root and compute as member - Add numpy and scipy dependencies to compute workspace - Update AGENTS.md with compute workspace documentation This sets up a dedicated workspace for training heavy models and generating artifacts (.npy, .npz) that demos can load as packages.41a3e04 -
feat: add activation function selector for function space manifold plot
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>48a3214 -
refactor: rename "ASI Demos" to "Demos" and finalize CLAUDE.md
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>40cb5c7
-
fix: make sidebar expand button transparent but clickable
0612f92 -
feat: upgrade Python version to 3.13 in Dockerfile and streamline application code copying
Signed-off-by: Simone Rossi <simone.rossi@eurecom.fr>e090600 -
fix: use python -m streamlit to avoid shebang issue in multi-stage build
a641bb7 -
feat: update Dockerfile port to 7860 and include config.py
46a0251 -
feat: add function space manifold visualization for Bayesian deep learning
Visualize how the function space prior of a neural network lives on a low-dimensional manifold. Monte Carlo samples of (sin(θ·x1+b), sin(θ·x2+b)) lie on a 1D sinusoidal curve in [-1,1]². New files: - config.py: shared matplotlib/tueplots styling - pages/bayesian-deep-learning/00_function_space_manifold.py - Updated app.py with navigation6cf0dad