blob: 2e3784182f18098e96d35c44008baabc79f64b89 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
|
# Experiment Notes
## 2026-03-23: Initial Implementation and Experiments
### Setup
- GPU: NVIDIA RTX A6000 x4 (using GPU 1)
- PyTorch 2.10.0+cu128
- All code written from scratch following CLAUDE.md specifications
### Phase A: Toy LQ Sanity Check
- Status: Running...
- Config: d=64, m=10, L=12, sigma=0.03, 5000 steps, batch=256
- Methods: DFA, State Bridge, Credit Bridge
### Changes Log
- Created full project structure: models/, methods/, experiments/, metrics/, configs/
- models/residual_mlp.py: ResidualMLP with pre-LayerNorm residual blocks
- models/value_net.py: ValueNet V_phi with sinusoidal time embedding
- models/state_bridge.py: StateBridgeNet G_psi
- experiments/toy_lq.py: Linear-quadratic sanity check
- experiments/cifar_resmlp.py: CIFAR-10 main experiment
- metrics/credit_metrics.py: All diagnostic metrics
|
