Save event_threshold to checkpoint dict during training, restore it
during resume and evaluation. evaluate.py now reads from checkpoint
instead of hardcoding train_cfg default, so evaluation matches the
threshold used during training.
Generated by Mistral Vibe.
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
- stateless: each batch starts from zero GRU hidden state
- curriculum: random seq_len from [64,96,128,160,192] per epoch
- val fixed at seq_len=128 for consistent cross-epoch comparison
- remove 256 from choices (OOM at B=4 on 24GB VRAM)
Generated by Mistral Vibe.
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
NormalizeVelocity transform is disabled, so model outputs are already in
original m/s space. Denormalizing inflates RMSE by ~3.5x.
- evaluate.py: compute RMSE in model output space directly
- benchmark/evaluate.py: same, plus fix misleading "normalized" comment
Generated by Mistral Vibe.
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
- config: seq_len=128, batch_size=4 for long-sequence TBPTT
- dataset: create_tbptt_loader with non-overlapping windows, strict temporal order
- model: forward() accepts/exposes hidden state h; add step() for single-frame stateful inference
- train: carry detached hidden state across batches, reset at epoch boundary
- benchmark: fix model call for new (v_body, h) return signature
Generated by Mistral Vibe.
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
- model outputs (B, S, 2) instead of (B, 2) — GRU output at every timestep
- train/val loss computed over all S timesteps with reduction=none
- benchmark/evaluate.py takes pred[:, -1, :] for final-step evaluation
- added per-step loss logging (8 evenly spaced steps) to TensorBoard
Generated by Mistral Vibe.
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
Visualize raw temporal brightness change (threshold=0, log domain)
as green(+)/red(-) gradient overlay proportional to |change|.
Supports video output and live display modes.
Enables EventProcessor threshold=0 for raw mode without clipping.
Generated by Mistral Vibe.
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
- Activate CNNEncoder forward (replace zero placeholder with actual inference)
- Enable near-zero weight init for head final layer (weight*=0.01, bias=0)
- Disable NormalizeVelocity transform to train on raw velocity scale
- (BatchNorm remains commented out)
Generated by deepseek-v4-flash.
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
Multi-scene evaluation previously concatenated all scenes into one
continuous trace, causing scene boundary jumps to appear as glitches
in plots. Now evaluates each scene separately and inserts NaN
separators between scenes when concatenating for plotting.
Generated by Mistral Vibe (deepseek-v4-flash).
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
- Add --resume CLI arg to resume training from a checkpoint
- Restore model, optimizer, scheduler state; continue from saved epoch+1
- Preserve global_step and best_val_loss across resume
- Save run_id in checkpoints for TensorBoard log continuity
- Use logs/run_<timestamp>/ subdirectories to isolate experiment logs
- Fix: replace train_loss in checkpoint dict with global_step to avoid
KeyError when loading; track global_step through train_one_epoch
- Fix: use global_step (not batch_idx) as TensorBoard x-axis for batch loss
- Fix: print average loss at end of each epoch
Generated by Mistral Vibe (ds-v4-flash).
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
- Replace body_attitude() with body_up_vector(): rotate world-up [0,0,1]
by corrected world→body quaternion to get body up vector (pitch/roll only,
no yaw). Matches DiffPhysDrone's env.R[:, 2] approach.
- Update ComputeTilt transform to use body_up_vector_np
- Update visualize_dataset.py to display Euler angles and body up vector
- Update model.py comments and disable CNN (zero output)
- Sync AGENTS.md with new architecture description
Generated by Mistral Vibe (ds-v4-flash).
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
- Add body_attitude() that applies R_odom_to_body calibration then
converts directly to rotation vector (preserves yaw, unlike old
decompose_tilt which stripped it)
- Add quat_to_euler() for visualization display
- Update ComputeTilt transform to use body_attitude_np (also fixes
a bug where the old code omitted the static calibration)
- Update visualize_dataset.py to show Euler angles from body quaternion
instead of yaw-stripped tilt rotation vector
This aligns with the DiffPhysDrone approach: let the model decide
whether to use yaw information, rather than removing it upfront.
Generated by Mistral Vibe.
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
hexone2086