Publish Next-Drive model
Browse files- .gitattributes +1 -0
- README.md +77 -0
- drive_core.py +519 -0
- evaluation.json +0 -0
- inference.py +54 -0
- model.json +1 -0
- nlrt.py +226 -0
- policy.safetensors +3 -0
- preview.png +0 -0
- replay.mp4 +3 -0
- selftest.json +1 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
replay.mp4 filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
|
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: pytorch
|
| 4 |
+
tags:
|
| 5 |
+
- autonomous-driving
|
| 6 |
+
- imitation-learning
|
| 7 |
+
- simulation
|
| 8 |
+
- dagger
|
| 9 |
+
- next-lab
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# Next Drive
|
| 13 |
+
|
| 14 |
+
Next Drive is a small learned driver for a **simulated** city intersection. Every 0.1 s it chooses one of six
|
| 15 |
+
longitudinal actions (brake hard, brake, ease off, hold speed, speed up, accelerate) while following a straight, left or right route through a traffic
|
| 16 |
+
light, a crosswalk, a lead car that brakes suddenly, oncoming and cross traffic, and parked cars that hide a pedestrian.
|
| 17 |
+
|
| 18 |
+
**Simulation and education only. It is not connected to, and has never been tested on, a real vehicle, and nothing here
|
| 19 |
+
supports a claim about real-road safety.**
|
| 20 |
+
|
| 21 |
+
**Space:** https://huggingface.co/spaces/thelamapi/Next-Drive — pick a scene, density and driver; the drive is generated in your browser and can be replayed with its critical events.
|
| 22 |
+
|
| 23 |
+
## Results on unseen test scenes (450 per driver; seeds ≥ 10000; density 3 never seen in training)
|
| 24 |
+
|
| 25 |
+
| Driver | Collisions | Red-light violations | Goal reached | Hard brakes / drive | Time to goal (s) | Decision ms (mean / p95) |
|
| 26 |
+
|---|---:|---:|---:|---:|---:|---:|
|
| 27 |
+
| Learned + safety check | 4.4% | 0.4% | 89.1% | 1.67 | 21.0 | 0.34 / 0.47 |
|
| 28 |
+
| Learned | 4.4% | 10.0% | 89.8% | 0.95 | 19.3 | 0.28 / 0.40 |
|
| 29 |
+
| Rule-based | 27.1% | 0.0% | 71.1% | 0.95 | 23.3 | 0.02 / 0.07 |
|
| 30 |
+
| Expert (teacher) | 2.7% | 1.3% | 93.6% | 1.01 | 19.2 | 1.61 / 2.38 |
|
| 31 |
+
|
| 32 |
+
Collision / goal-reached rate by scene type:
|
| 33 |
+
|
| 34 |
+
| Scene | Learned + safety | Learned | Rule-based | Expert |
|
| 35 |
+
|---|---:|---:|---:|---:|
|
| 36 |
+
| light | 4% / 96% | 2% / 98% | 1% / 99% | 1% / 99% |
|
| 37 |
+
| crosswalk | 7% / 93% | 1% / 99% | 9% / 86% | 0% / 100% |
|
| 38 |
+
| lead_brake | 8% / 92% | 1% / 99% | 3% / 97% | 2% / 98% |
|
| 39 |
+
| occlusion | 2% / 79% | 10% / 73% | 68% / 32% | 7% / 82% |
|
| 40 |
+
| mixed | 1% / 86% | 8% / 80% | 54% / 42% | 3% / 89% |
|
| 41 |
+
|
| 42 |
+
Collision rate by traffic density:
|
| 43 |
+
|
| 44 |
+
| Density | Learned + safety | Learned | Rule-based | Expert |
|
| 45 |
+
|---|---:|---:|---:|---:|
|
| 46 |
+
| 1 | 1% | 2% | 23% | 0% |
|
| 47 |
+
| 2 | 5% | 3% | 23% | 3% |
|
| 48 |
+
| 3 (unseen) | 7% | 8% | 35% | 5% |
|
| 49 |
+
|
| 50 |
+
Rates count drives with at least one event. Decision latency was measured in Python on NVIDIA L4, single process; the
|
| 51 |
+
Space reports its own browser timing. All 138 failures of the learned drivers are listed in `evaluation.json`
|
| 52 |
+
and can be replayed in the Space.
|
| 53 |
+
|
| 54 |
+

|
| 55 |
+
|
| 56 |
+
[Replay video](replay.mp4): one unselected test scene, then failure replays.
|
| 57 |
+
|
| 58 |
+
## How it works
|
| 59 |
+
* **Simulator:** 1-lane roads, NS/EW signals with yellow and all-red, IDM-like traffic, pedestrians, and line-of-sight occlusion by parked cars. The ego car perceives cars always and pedestrians only when visible within 60 m.
|
| 60 |
+
* **Expert:** evaluates 6×6 two-phase action sequences (1 s + 2 s) at 0.25 s resolution with constant-velocity predictions, the known light schedule, a 2 m + 1.2 s following gap, and a speed cap that allows stopping before occluded gaps.
|
| 61 |
+
* **Learned driver:** MLP 21→64→64→6 on hand-made features (speed, light state, stop-line distance, lead gap, time-to-collision under three accelerations, visible-pedestrian and occlusion distances, route). Behaviour cloning on 300 expert drives (59750 states), then 3 DAgger round(s); iteration 2 was selected on validation seeds.
|
| 62 |
+
* **Safety check:** keeps an action only if braking at −6 m/s² after 0.5 s still avoids every predicted collision and red-light entry; otherwise picks the strongest safe deceleration.
|
| 63 |
+
* **Rule-based baseline:** IDM car following plus stop-line, visible-pedestrian and left-turn-yield rules.
|
| 64 |
+
* The browser runs a line-by-line JavaScript port; Node parity: 8 full drives, 0 mismatches.
|
| 65 |
+
|
| 66 |
+
## Usage
|
| 67 |
+
```bash
|
| 68 |
+
python inference.py --seed 10007 --kind mixed --density 2 --driver shield
|
| 69 |
+
python inference.py --selftest
|
| 70 |
+
```
|
| 71 |
+
`policy.safetensors` holds the PyTorch weights; `model.json` the same weights for NumPy and the browser.
|
| 72 |
+
|
| 73 |
+
## Limitations
|
| 74 |
+
* Speed-only control on fixed routes; no lane changes, no steering errors, no sensor noise beyond occlusion.
|
| 75 |
+
* Other road users follow simple scripted behaviour and never react to the ego car.
|
| 76 |
+
* The expert is imperfect too (see the table); the learned driver copies its habits, including frequent hard braking at yellow lights.
|
| 77 |
+
* All numbers come from this simulator and say nothing about real driving.
|
drive_core.py
ADDED
|
@@ -0,0 +1,519 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Next Drive: top-down urban intersection simulator, expert planner, rule-based driver, learned driver and safety check.
|
| 2 |
+
|
| 3 |
+
Simulation and training demo only. No physical vehicle, no claim about real-road safety.
|
| 4 |
+
The ego car follows a fixed route (straight, left or right turn) and only chooses its longitudinal
|
| 5 |
+
action every 0.1 s from ACTIONS. Everything is deterministic given (seed, kind, density) and is ported
|
| 6 |
+
line by line to web/core.js so the Space can generate new drives.
|
| 7 |
+
"""
|
| 8 |
+
import math
|
| 9 |
+
|
| 10 |
+
import numpy as np
|
| 11 |
+
|
| 12 |
+
from nlrt import Mulberry32, forward
|
| 13 |
+
|
| 14 |
+
DT, VMAX, V_LIMIT, T_MAX = 0.1, 14.0, 13.9, 45.0
|
| 15 |
+
ACTIONS = [-6.0, -3.0, -1.5, 0.0, 1.0, 2.0]
|
| 16 |
+
ACTION_NAMES = ["brake hard", "brake", "ease off", "hold speed", "speed up", "accelerate"]
|
| 17 |
+
LANE, S_STOP, HALF = 1.75, 60.0, 2.2
|
| 18 |
+
ROUTES = ["straight", "left", "right"]
|
| 19 |
+
KINDS = ["light", "crosswalk", "lead_brake", "occlusion", "mixed"]
|
| 20 |
+
TURN = {"right": (-5.75, 4.0), "left": (-5.25, 7.0)}
|
| 21 |
+
N_FEAT = 21
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
# ------------------------------------------------------------------ geometry
|
| 25 |
+
def path_len(route):
|
| 26 |
+
if route == "straight":
|
| 27 |
+
return 130.0
|
| 28 |
+
y0, r = TURN[route]
|
| 29 |
+
return (y0 + 70.0) + r * math.pi / 2 + 55.0
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
def pose(route, s):
|
| 33 |
+
"""Position and heading on the route at arc length s (clamped to the route)."""
|
| 34 |
+
s = max(0.0, min(s, path_len(route)))
|
| 35 |
+
if route == "straight":
|
| 36 |
+
return LANE, -70.0 + s, math.pi / 2
|
| 37 |
+
y0, r = TURN[route]
|
| 38 |
+
L1 = y0 + 70.0
|
| 39 |
+
if s < L1:
|
| 40 |
+
return LANE, -70.0 + s, math.pi / 2
|
| 41 |
+
arc = r * math.pi / 2
|
| 42 |
+
if s < L1 + arc:
|
| 43 |
+
th = (s - L1) / r
|
| 44 |
+
if route == "right":
|
| 45 |
+
return LANE + r - r * math.cos(th), y0 + r * math.sin(th), math.pi / 2 - th
|
| 46 |
+
return LANE - r + r * math.cos(th), y0 + r * math.sin(th), math.pi / 2 + th
|
| 47 |
+
d = s - L1 - arc
|
| 48 |
+
if route == "right":
|
| 49 |
+
return LANE + r + d, y0 + r, 0.0
|
| 50 |
+
return LANE - r - d, y0 + r, math.pi
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
def circles(x, y, h):
|
| 54 |
+
c, s = math.cos(h), math.sin(h)
|
| 55 |
+
return [(x + d * c, y + d * s) for d in (-1.4, 0.0, 1.4)]
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
def hits_car(ego, other, margin=0.0):
|
| 59 |
+
for ax, ay in ego:
|
| 60 |
+
for bx, by in other:
|
| 61 |
+
dx, dy, rr = ax - bx, ay - by, 2.0 + margin
|
| 62 |
+
if dx * dx + dy * dy < rr * rr:
|
| 63 |
+
return True
|
| 64 |
+
return False
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
def hits_ped(ego, px, py, margin=0.0):
|
| 68 |
+
for ax, ay in ego:
|
| 69 |
+
dx, dy, rr = ax - px, ay - py, 1.35 + margin
|
| 70 |
+
if dx * dx + dy * dy < rr * rr:
|
| 71 |
+
return True
|
| 72 |
+
return False
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
# ------------------------------------------------------------------ traffic lights
|
| 76 |
+
def light_ns(sc, t):
|
| 77 |
+
G, Y, R = sc["G"], 3.0, sc["R"]
|
| 78 |
+
u = (t + sc["phase"]) % (G + Y + R)
|
| 79 |
+
if u < G:
|
| 80 |
+
return 0, G - u
|
| 81 |
+
if u < G + Y:
|
| 82 |
+
return 1, G + Y - u
|
| 83 |
+
return 2, G + Y + R - u
|
| 84 |
+
|
| 85 |
+
|
| 86 |
+
def light_ew(sc, t):
|
| 87 |
+
G, Y, R = sc["G"], 3.0, sc["R"]
|
| 88 |
+
u = (t + sc["phase"]) % (G + Y + R)
|
| 89 |
+
if u < G + Y:
|
| 90 |
+
return 2
|
| 91 |
+
r = u - G - Y
|
| 92 |
+
if r < 1.0:
|
| 93 |
+
return 2
|
| 94 |
+
if r < R - 4.0:
|
| 95 |
+
return 0
|
| 96 |
+
if r < R - 1.0:
|
| 97 |
+
return 1
|
| 98 |
+
return 2
|
| 99 |
+
|
| 100 |
+
|
| 101 |
+
# ------------------------------------------------------------------ scenarios
|
| 102 |
+
def make_scenario(seed, kind="mixed", density=1):
|
| 103 |
+
rng = Mulberry32((seed * 7919 + KINDS.index(kind) * 101 + density * 13) % 4294967296)
|
| 104 |
+
u = rng.random()
|
| 105 |
+
route = "straight" if u < 0.5 else ("left" if u < 0.75 else "right")
|
| 106 |
+
if kind in ("occlusion", "crosswalk") and rng.random() < 0.5:
|
| 107 |
+
route = "straight"
|
| 108 |
+
G, R = 7.0 + 6.0 * rng.random(), 8.0 + 5.0 * rng.random()
|
| 109 |
+
sc = {"seed": seed, "kind": kind, "density": density, "route": route, "G": G, "R": R,
|
| 110 |
+
"phase": rng.random() * (G + 3.0 + R), "cars": [], "peds": [], "parked": [], "lead": None}
|
| 111 |
+
for k in range(density + 1): # cross traffic, both directions
|
| 112 |
+
for d in (1, -1):
|
| 113 |
+
sc["cars"].append({"axis": "x", "dir": d, "lane": -LANE * d, "p0": -70.0 * d, "v0": 8.0 + 4.0 * rng.random(),
|
| 114 |
+
"t0": k * 7.0 + 5.0 * rng.random()})
|
| 115 |
+
if route == "left" or kind == "mixed": # oncoming traffic (southbound)
|
| 116 |
+
for k in range(density):
|
| 117 |
+
sc["cars"].append({"axis": "y", "dir": -1, "lane": -LANE, "p0": 70.0, "v0": 8.0 + 4.0 * rng.random(),
|
| 118 |
+
"t0": k * 6.0 + 4.0 * rng.random()})
|
| 119 |
+
if kind in ("crosswalk", "mixed"):
|
| 120 |
+
for k in range(density + 1):
|
| 121 |
+
d = 1 if rng.random() < 0.5 else -1
|
| 122 |
+
sc["peds"].append({"x0": -9.0 * d, "y0": -7.5 + 1.6 * (rng.random() - 0.5), "vx": d * (1.0 + 0.6 * rng.random()),
|
| 123 |
+
"vy": 0.0, "t0": 1.0 + 11.0 * rng.random(), "x_end": 9.0 * d})
|
| 124 |
+
if kind in ("lead_brake", "mixed"):
|
| 125 |
+
sc["lead"] = {"s0": 14.0 + 10.0 * rng.random(), "v0": 9.0 + 3.0 * rng.random(),
|
| 126 |
+
"tb": 2.5 + 5.0 * rng.random(), "db": 1.5 + 1.5 * rng.random()}
|
| 127 |
+
if kind in ("occlusion", "mixed"):
|
| 128 |
+
y = -46.0
|
| 129 |
+
gaps = []
|
| 130 |
+
while y < -16.0:
|
| 131 |
+
L = 4.6
|
| 132 |
+
sc["parked"].append([4.6, y + L / 2, 2.0, L])
|
| 133 |
+
y += L + 0.4 + (2.6 if rng.random() < 0.45 else 0.0)
|
| 134 |
+
gaps.append(y - 0.2)
|
| 135 |
+
n_emerge = density
|
| 136 |
+
for k in range(min(n_emerge, len(gaps))):
|
| 137 |
+
gy = gaps[rng.randint(len(gaps))] - 1.0
|
| 138 |
+
D = 12.0 + 10.0 * rng.random()
|
| 139 |
+
sc["peds"].append({"x0": 5.8, "y0": gy, "vx": -(1.2 + 0.4 * rng.random()), "vy": 0.0,
|
| 140 |
+
"t0": max(0.0, (gy + 70.0 - HALF - D) / 10.0), "x_end": -6.0})
|
| 141 |
+
return sc
|
| 142 |
+
|
| 143 |
+
|
| 144 |
+
# ------------------------------------------------------------------ world state
|
| 145 |
+
def init_world(sc):
|
| 146 |
+
w = {"t": 0.0, "s": 0.0, "v": 10.0, "a": 0.0, "cars": [], "peds": [], "lead": None}
|
| 147 |
+
for c in sc["cars"]:
|
| 148 |
+
w["cars"].append({"p": c["p0"], "v": c["v0"], "active": False, "done": False, **{k: c[k] for k in ("axis", "dir", "lane", "v0", "t0")}})
|
| 149 |
+
for p in sc["peds"]:
|
| 150 |
+
w["peds"].append({"x": p["x0"], "y": p["y0"], "vx": 0.0, "vy": 0.0, "done": False, **{k: p[k] for k in ("t0", "x_end")}, "svx": p["vx"]})
|
| 151 |
+
if sc["lead"]:
|
| 152 |
+
L = sc["lead"]
|
| 153 |
+
w["lead"] = {"s": L["s0"], "v": L["v0"], **L}
|
| 154 |
+
return w
|
| 155 |
+
|
| 156 |
+
|
| 157 |
+
def car_xyh(c):
|
| 158 |
+
if c["axis"] == "x":
|
| 159 |
+
return c["p"], c["lane"], 0.0 if c["dir"] > 0 else math.pi
|
| 160 |
+
return c["lane"], c["p"], math.pi / 2 if c["dir"] > 0 else -math.pi / 2
|
| 161 |
+
|
| 162 |
+
|
| 163 |
+
def step_agents(sc, w):
|
| 164 |
+
t = w["t"]
|
| 165 |
+
ew, (ns, _) = light_ew(sc, t), light_ns(sc, t)
|
| 166 |
+
lanes = {}
|
| 167 |
+
for c in w["cars"]:
|
| 168 |
+
if not c["active"] and not c["done"] and t >= c["t0"]:
|
| 169 |
+
c["active"] = True
|
| 170 |
+
if c["active"]:
|
| 171 |
+
lanes.setdefault((c["axis"], c["dir"]), []).append(c)
|
| 172 |
+
for key, cs in lanes.items():
|
| 173 |
+
cs.sort(key=lambda c: -c["p"] * c["dir"]) # front-most first
|
| 174 |
+
for i, c in enumerate(cs):
|
| 175 |
+
light = ew if c["axis"] == "x" else ns
|
| 176 |
+
front = c["p"] * c["dir"] + HALF
|
| 177 |
+
d_stop = -10.0 - front
|
| 178 |
+
a = min(2.0, 1.5 * (c["v0"] - c["v"]))
|
| 179 |
+
if light != 0 and d_stop > -1.0 and c["v"] * c["v"] / (2 * 3.0) < d_stop + 1.0:
|
| 180 |
+
a = min(a, -c["v"] * c["v"] / (2 * max(d_stop - 0.5, 0.3)))
|
| 181 |
+
if i > 0:
|
| 182 |
+
gap = (cs[i - 1]["p"] - c["p"]) * c["dir"] - 4.4
|
| 183 |
+
if gap < 6.0 + c["v"] * 1.2:
|
| 184 |
+
a = min(a, -3.0 if gap > 3.0 else -6.0)
|
| 185 |
+
c["v"] = max(0.0, c["v"] + max(-6.0, a) * DT)
|
| 186 |
+
c["p"] += c["dir"] * c["v"] * DT
|
| 187 |
+
if c["p"] * c["dir"] > 75.0:
|
| 188 |
+
c["active"], c["done"] = False, True
|
| 189 |
+
for p in w["peds"]:
|
| 190 |
+
if p["done"]:
|
| 191 |
+
continue
|
| 192 |
+
if t >= p["t0"]:
|
| 193 |
+
p["vx"] = p["svx"]
|
| 194 |
+
p["x"] += p["vx"] * DT
|
| 195 |
+
if (p["x"] - p["x_end"]) * p["svx"] > 0:
|
| 196 |
+
p["done"] = True
|
| 197 |
+
L = w["lead"]
|
| 198 |
+
if L is not None:
|
| 199 |
+
if L["tb"] <= t < L["tb"] + L["db"]:
|
| 200 |
+
a = -6.0
|
| 201 |
+
else:
|
| 202 |
+
a = min(2.0, 1.5 * (L["v0"] - L["v"]))
|
| 203 |
+
d_stop = S_STOP - (L["s"] + HALF)
|
| 204 |
+
if ns != 0 and d_stop > -1.0 and L["v"] * L["v"] / (2 * 3.0) < d_stop + 1.0:
|
| 205 |
+
a = min(a, -L["v"] * L["v"] / (2 * max(d_stop - 0.5, 0.3)))
|
| 206 |
+
L["v"] = max(0.0, L["v"] + max(-6.0, a) * DT)
|
| 207 |
+
L["s"] += L["v"] * DT
|
| 208 |
+
|
| 209 |
+
|
| 210 |
+
# ------------------------------------------------------------------ perception
|
| 211 |
+
def seg_hits_rect(x0, y0, x1, y1, rect):
|
| 212 |
+
cx, cy, wd, ht = rect
|
| 213 |
+
xmin, xmax, ymin, ymax = cx - wd / 2, cx + wd / 2, cy - ht / 2, cy + ht / 2
|
| 214 |
+
t0, t1 = 0.0, 1.0
|
| 215 |
+
for p, q in ((-(x1 - x0), x0 - xmin), (x1 - x0, xmax - x0), (-(y1 - y0), y0 - ymin), (y1 - y0, ymax - y0)):
|
| 216 |
+
if p == 0:
|
| 217 |
+
if q < 0:
|
| 218 |
+
return False
|
| 219 |
+
else:
|
| 220 |
+
r = q / p
|
| 221 |
+
if p < 0:
|
| 222 |
+
if r > t1:
|
| 223 |
+
return False
|
| 224 |
+
if r > t0:
|
| 225 |
+
t0 = r
|
| 226 |
+
else:
|
| 227 |
+
if r < t0:
|
| 228 |
+
return False
|
| 229 |
+
if r < t1:
|
| 230 |
+
t1 = r
|
| 231 |
+
return True
|
| 232 |
+
|
| 233 |
+
|
| 234 |
+
def observe(sc, w):
|
| 235 |
+
ex, ey, eh = pose(sc["route"], w["s"])
|
| 236 |
+
eye = (ex + 1.0 * math.cos(eh), ey + 1.0 * math.sin(eh))
|
| 237 |
+
agents = []
|
| 238 |
+
for c in w["cars"]:
|
| 239 |
+
if c["active"]:
|
| 240 |
+
x, y, h = car_xyh(c)
|
| 241 |
+
agents.append({"kind": "car", "x": x, "y": y, "h": h, "vx": c["v"] * math.cos(h), "vy": c["v"] * math.sin(h)})
|
| 242 |
+
L = w["lead"]
|
| 243 |
+
if L is not None and L["s"] < path_len(sc["route"]):
|
| 244 |
+
x, y, h = pose(sc["route"], L["s"])
|
| 245 |
+
agents.append({"kind": "lead", "x": x, "y": y, "h": h, "vx": L["v"] * math.cos(h), "vy": L["v"] * math.sin(h), "s": L["s"], "v": L["v"]})
|
| 246 |
+
visible = []
|
| 247 |
+
for p in w["peds"]:
|
| 248 |
+
if p["done"]:
|
| 249 |
+
continue
|
| 250 |
+
dx, dy = p["x"] - eye[0], p["y"] - eye[1]
|
| 251 |
+
vis = dx * dx + dy * dy < 3600.0 and not any(seg_hits_rect(eye[0], eye[1], p["x"], p["y"], r) for r in sc["parked"])
|
| 252 |
+
visible.append(vis)
|
| 253 |
+
if vis:
|
| 254 |
+
agents.append({"kind": "ped", "x": p["x"], "y": p["y"], "h": 0.0, "vx": p["vx"], "vy": p["vy"]})
|
| 255 |
+
phantoms = []
|
| 256 |
+
s_front = w["s"] + HALF
|
| 257 |
+
for r in sc["parked"]:
|
| 258 |
+
gy = r[1] + r[3] / 2 + 0.3
|
| 259 |
+
ahead = (gy + 70.0) - s_front
|
| 260 |
+
if 0.0 < ahead < 35.0 and seg_hits_rect(eye[0], eye[1], 4.2, gy, r):
|
| 261 |
+
phantoms.append({"x": 4.2, "y": gy, "vx": -1.4, "vy": 0.0, "ahead": ahead})
|
| 262 |
+
ls, rem = light_ns(sc, w["t"])
|
| 263 |
+
return {"t": w["t"], "s": w["s"], "v": w["v"], "a": w["a"], "route": sc["route"], "agents": agents,
|
| 264 |
+
"phantoms": phantoms, "light": ls, "remaining": rem, "sc": sc, "visible": visible}
|
| 265 |
+
|
| 266 |
+
|
| 267 |
+
# ------------------------------------------------------------------ prediction helpers
|
| 268 |
+
def rollout_hits(obs, plan, horizon=3.0, dt=0.25, margin_car=0.4, margin_ped=0.6):
|
| 269 |
+
"""plan(k, tau) -> accel. Returns (first collision time or None, red-light crossing time or None, s_end, soft penalty).
|
| 270 |
+
Soft penalty: speed above the limit, speed above what allows stopping before an occluded gap
|
| 271 |
+
(where a hidden pedestrian could step out), and following closer than 2 m + 1.2 s to the lead car."""
|
| 272 |
+
s, v = obs["s"], obs["v"]
|
| 273 |
+
route = obs["route"]
|
| 274 |
+
n = int(round(horizon / dt))
|
| 275 |
+
excess = 0.0
|
| 276 |
+
red_t = None
|
| 277 |
+
lead = [ag for ag in obs["agents"] if ag["kind"] == "lead"]
|
| 278 |
+
for k in range(1, n + 1):
|
| 279 |
+
tau = k * dt
|
| 280 |
+
a = plan(k, tau)
|
| 281 |
+
v = min(VMAX, max(0.0, v + a * dt))
|
| 282 |
+
s_prev = s
|
| 283 |
+
s = s + v * dt
|
| 284 |
+
if v > V_LIMIT:
|
| 285 |
+
excess += v - V_LIMIT
|
| 286 |
+
for ph in obs["phantoms"]:
|
| 287 |
+
d = ph["y"] + 70.0 - (s + HALF)
|
| 288 |
+
if d > -1.0:
|
| 289 |
+
excess += max(0.0, v - (math.sqrt(2 * 4.0 * max(d - 2.0, 0.0)) + 3.0))
|
| 290 |
+
for ag in lead:
|
| 291 |
+
gap = ag["s"] + ag["v"] * tau - s - 4.4
|
| 292 |
+
excess += 0.5 * max(0.0, 2.0 + 1.2 * v - gap)
|
| 293 |
+
if red_t is None and s > s_prev:
|
| 294 |
+
lt, _ = light_ns(obs["sc"], obs["t"] + tau)
|
| 295 |
+
crossing = s_prev + HALF < S_STOP - 0.5 <= s + HALF
|
| 296 |
+
creeping = S_STOP - 0.5 <= obs["s"] + HALF < S_STOP + 6.0 and obs["v"] < 3.0
|
| 297 |
+
if lt == 2 and (crossing or creeping):
|
| 298 |
+
red_t = tau
|
| 299 |
+
ex, ey, eh = pose(route, s)
|
| 300 |
+
ec = circles(ex, ey, eh)
|
| 301 |
+
for ag in obs["agents"]:
|
| 302 |
+
px, py = ag["x"] + ag["vx"] * tau, ag["y"] + ag["vy"] * tau
|
| 303 |
+
if (px - ex) * (px - ex) + (py - ey) * (py - ey) > 64.0:
|
| 304 |
+
continue
|
| 305 |
+
if ag["kind"] == "ped":
|
| 306 |
+
if hits_ped(ec, px, py, margin_ped):
|
| 307 |
+
return tau, red_t, s, excess
|
| 308 |
+
elif hits_car(ec, circles(px, py, ag["h"]), margin_car):
|
| 309 |
+
return tau, red_t, s, excess
|
| 310 |
+
return None, red_t, s, excess
|
| 311 |
+
|
| 312 |
+
|
| 313 |
+
# ------------------------------------------------------------------ drivers
|
| 314 |
+
def expert(obs):
|
| 315 |
+
"""Search over two-phase action sequences (1 s then 2 s), constant-velocity predictions, phantom pedestrians."""
|
| 316 |
+
best, best_cost = 0, float("inf")
|
| 317 |
+
for i, a1 in enumerate(ACTIONS):
|
| 318 |
+
for a2 in ACTIONS:
|
| 319 |
+
plan = lambda k, tau, a1=a1, a2=a2: a1 if tau <= 1.0 + 1e-9 else a2
|
| 320 |
+
hit, red, s_end, excess = rollout_hits(obs, plan)
|
| 321 |
+
cost = 0.0
|
| 322 |
+
if hit is not None:
|
| 323 |
+
cost += 1e5 - 1e4 * hit
|
| 324 |
+
if red is not None:
|
| 325 |
+
cost += 2e4
|
| 326 |
+
cost += -10.0 * (s_end - obs["s"])
|
| 327 |
+
cost += 40.0 if a1 == -6.0 else (8.0 if a1 == -3.0 else 0.0)
|
| 328 |
+
cost += 3.0 * abs(a1 - obs["a"])
|
| 329 |
+
cost += 50.0 * excess
|
| 330 |
+
if cost < best_cost:
|
| 331 |
+
best, best_cost = i, cost
|
| 332 |
+
return best
|
| 333 |
+
|
| 334 |
+
|
| 335 |
+
def rule_driver(obs):
|
| 336 |
+
"""IDM-style rules: lights, lead car, visible pedestrians near the route, yield before a left turn."""
|
| 337 |
+
v, s = obs["v"], obs["s"]
|
| 338 |
+
route = obs["route"]
|
| 339 |
+
s_front = s + HALF
|
| 340 |
+
obst = []
|
| 341 |
+
for ag in obs["agents"]:
|
| 342 |
+
if ag["kind"] == "lead":
|
| 343 |
+
obst.append((ag["s"] - s - 4.4, ag["v"]))
|
| 344 |
+
d_stop = S_STOP - s_front
|
| 345 |
+
if d_stop > 0 and (obs["light"] == 2 or (obs["light"] == 1 and v * v / (2 * 3.0) < d_stop)):
|
| 346 |
+
obst.append((d_stop - 0.5, 0.0))
|
| 347 |
+
for ag in obs["agents"]:
|
| 348 |
+
if ag["kind"] != "ped":
|
| 349 |
+
continue
|
| 350 |
+
for j in range(0, 51, 2):
|
| 351 |
+
px, py, _ = pose(route, s + j)
|
| 352 |
+
fx, fy = ag["x"] + ag["vx"] * 1.5, ag["y"] + ag["vy"] * 1.5
|
| 353 |
+
d1 = (ag["x"] - px) * (ag["x"] - px) + (ag["y"] - py) * (ag["y"] - py)
|
| 354 |
+
d2 = (fx - px) * (fx - px) + (fy - py) * (fy - py)
|
| 355 |
+
if min(d1, d2) < 6.76:
|
| 356 |
+
obst.append((j - HALF - 1.5, 0.0))
|
| 357 |
+
break
|
| 358 |
+
if route == "left":
|
| 359 |
+
y0, _ = TURN["left"]
|
| 360 |
+
d_turn = (y0 + 70.0) - s_front
|
| 361 |
+
if d_turn > -1.0:
|
| 362 |
+
for ag in obs["agents"]:
|
| 363 |
+
if ag["kind"] == "car" and ag["vy"] < -0.5 and -4.0 < ag["y"] < 45.0 and ag["x"] < 0:
|
| 364 |
+
obst.append((d_turn - 0.5, 0.0))
|
| 365 |
+
break
|
| 366 |
+
q = v / 12.0
|
| 367 |
+
q4 = (q * q) * (q * q)
|
| 368 |
+
a = 1.5 * (1.0 - q4)
|
| 369 |
+
for gap, vo in obst:
|
| 370 |
+
gap = max(gap, 0.1)
|
| 371 |
+
sstar = 2.0 + v * 1.2 + v * (v - vo) / (2 * math.sqrt(1.5 * 3.0))
|
| 372 |
+
a = min(a, 1.5 * (1.0 - q4 - (sstar / gap) * (sstar / gap)))
|
| 373 |
+
a = max(-6.0, min(2.0, a))
|
| 374 |
+
return min(range(len(ACTIONS)), key=lambda i: (abs(ACTIONS[i] - a), i))
|
| 375 |
+
|
| 376 |
+
|
| 377 |
+
def features(obs):
|
| 378 |
+
v, s = obs["v"], obs["s"]
|
| 379 |
+
d_stop = S_STOP - (s + HALF)
|
| 380 |
+
f = [v / 14.0, obs["a"] / 6.0, min(max(d_stop, 0.0), 60.0) / 60.0, 1.0 if d_stop < 0 else 0.0,
|
| 381 |
+
1.0 if obs["light"] == 0 else 0.0, 1.0 if obs["light"] == 1 else 0.0, 1.0 if obs["light"] == 2 else 0.0,
|
| 382 |
+
min(obs["remaining"], 15.0) / 15.0]
|
| 383 |
+
lead = [ag for ag in obs["agents"] if ag["kind"] == "lead"]
|
| 384 |
+
if lead:
|
| 385 |
+
f += [min(max(lead[0]["s"] - s - 4.4, 0.0), 50.0) / 50.0, max(-1.0, min(1.0, (lead[0]["v"] - v) / 14.0))]
|
| 386 |
+
else:
|
| 387 |
+
f += [1.0, 0.0]
|
| 388 |
+
for acc in (0.0, -3.0, 1.0):
|
| 389 |
+
hit, _, _, _ = rollout_hits(obs, lambda k, tau, acc=acc: acc, horizon=5.0)
|
| 390 |
+
f.append((hit if hit is not None else 5.0) / 5.0)
|
| 391 |
+
ex, ey, _ = pose(obs["route"], s)
|
| 392 |
+
peds = [math.sqrt((ag["x"] - ex) * (ag["x"] - ex) + (ag["y"] - ey) * (ag["y"] - ey)) for ag in obs["agents"] if ag["kind"] == "ped"]
|
| 393 |
+
f.append(min(min(peds), 40.0) / 40.0 if peds else 1.0)
|
| 394 |
+
f.append(min(min(p["ahead"] for p in obs["phantoms"]), 40.0) / 40.0 if obs["phantoms"] else 1.0)
|
| 395 |
+
busy = any(ag["kind"] == "car" and abs(ag["x"]) < 12.0 and abs(ag["y"]) < 12.0 for ag in obs["agents"])
|
| 396 |
+
f.append(1.0 if busy else 0.0)
|
| 397 |
+
f += [1.0 if obs["route"] == r else 0.0 for r in ROUTES]
|
| 398 |
+
f.append(min(v * v / 12.0, 60.0) / 60.0)
|
| 399 |
+
f.append(1.0 if obs["light"] != 0 and d_stop > 0 and v * v / 6.0 < d_stop + 1.0 else 0.0)
|
| 400 |
+
return f
|
| 401 |
+
|
| 402 |
+
|
| 403 |
+
def learned(obs, layers):
|
| 404 |
+
z = forward(layers, np.asarray([features(obs)]))[0]
|
| 405 |
+
return int(np.argmax(z))
|
| 406 |
+
|
| 407 |
+
|
| 408 |
+
def shield(obs, i):
|
| 409 |
+
"""Safety check: keep action i only if braking hard after 0.5 s still avoids predicted collisions / red-light entry."""
|
| 410 |
+
def unsafe(a):
|
| 411 |
+
hit, red, _, _ = rollout_hits(obs, lambda k, tau: a if tau <= 0.5 + 1e-9 else -6.0, margin_car=0.2, margin_ped=0.4)
|
| 412 |
+
if hit is not None:
|
| 413 |
+
return True
|
| 414 |
+
if red is not None:
|
| 415 |
+
_, red_b, _, _ = rollout_hits(obs, lambda k, tau: -6.0, margin_car=0.2, margin_ped=0.4)
|
| 416 |
+
return red_b is None
|
| 417 |
+
return False
|
| 418 |
+
if not unsafe(ACTIONS[i]):
|
| 419 |
+
return i, False
|
| 420 |
+
for j in range(i - 1, -1, -1):
|
| 421 |
+
if not unsafe(ACTIONS[j]):
|
| 422 |
+
return j, True
|
| 423 |
+
return 0, True
|
| 424 |
+
|
| 425 |
+
|
| 426 |
+
# ------------------------------------------------------------------ episode
|
| 427 |
+
def run_episode(sc, driver, layers=None, record=True, timer=None, on_obs=None):
|
| 428 |
+
w = init_world(sc)
|
| 429 |
+
route = sc["route"]
|
| 430 |
+
goal = path_len(route) - 5.0
|
| 431 |
+
out = {"collision": False, "collision_with": None, "red_violations": 0, "reached": False, "hard_brakes": 0,
|
| 432 |
+
"overrides": 0, "steps": 0, "time": 0.0, "decision_ms": [], "trace": [], "events": [], "actions": []}
|
| 433 |
+
prev_a = 0.0
|
| 434 |
+
for step in range(int(round(T_MAX / DT))):
|
| 435 |
+
obs = observe(sc, w)
|
| 436 |
+
if on_obs is not None:
|
| 437 |
+
on_obs(obs)
|
| 438 |
+
t0 = timer() if timer else 0.0
|
| 439 |
+
override = False
|
| 440 |
+
if driver == "expert":
|
| 441 |
+
i = expert(obs)
|
| 442 |
+
elif driver == "rule":
|
| 443 |
+
i = rule_driver(obs)
|
| 444 |
+
elif driver == "learned":
|
| 445 |
+
i = learned(obs, layers)
|
| 446 |
+
else:
|
| 447 |
+
i, override = shield(obs, learned(obs, layers))
|
| 448 |
+
if timer:
|
| 449 |
+
out["decision_ms"].append(1000.0 * (timer() - t0))
|
| 450 |
+
a = ACTIONS[i]
|
| 451 |
+
out["actions"].append(i)
|
| 452 |
+
if override:
|
| 453 |
+
out["overrides"] += 1
|
| 454 |
+
if a <= -6.0 and prev_a > -6.0:
|
| 455 |
+
out["hard_brakes"] += 1
|
| 456 |
+
out["events"].append({"t": round(w["t"], 2), "type": "hard_brake"})
|
| 457 |
+
prev_a = a
|
| 458 |
+
s_front_prev = w["s"] + HALF
|
| 459 |
+
w["v"] = min(VMAX, max(0.0, w["v"] + a * DT))
|
| 460 |
+
w["s"] += w["v"] * DT
|
| 461 |
+
w["a"] = a
|
| 462 |
+
step_agents(sc, w)
|
| 463 |
+
w["t"] = (step + 1) * DT
|
| 464 |
+
if s_front_prev < S_STOP <= w["s"] + HALF and light_ns(sc, w["t"])[0] == 2:
|
| 465 |
+
out["red_violations"] += 1
|
| 466 |
+
out["events"].append({"t": round(w["t"], 2), "type": "red_light"})
|
| 467 |
+
ex, ey, eh = pose(route, w["s"])
|
| 468 |
+
ec = circles(ex, ey, eh)
|
| 469 |
+
hit = None
|
| 470 |
+
for c in w["cars"]:
|
| 471 |
+
if c["active"] and hits_car(ec, circles(*car_xyh(c))):
|
| 472 |
+
hit = "car"
|
| 473 |
+
if w["lead"] is not None and w["lead"]["s"] < path_len(route) and hits_car(ec, circles(*pose(route, w["lead"]["s"]))):
|
| 474 |
+
hit = "lead car"
|
| 475 |
+
for p in w["peds"]:
|
| 476 |
+
if not p["done"] and hits_ped(ec, p["x"], p["y"]):
|
| 477 |
+
hit = "pedestrian"
|
| 478 |
+
if record:
|
| 479 |
+
out["trace"].append([round(w["t"], 2), round(ex, 3), round(ey, 3), round(eh, 4), round(w["v"], 3), i, int(override),
|
| 480 |
+
[[round(v_, 2) for v_ in car_xyh(c)] for c in w["cars"] if c["active"]] +
|
| 481 |
+
([[round(v_, 2) for v_ in pose(route, w["lead"]["s"])]] if w["lead"] is not None and w["lead"]["s"] < path_len(route) else []),
|
| 482 |
+
[[round(p["x"], 2), round(p["y"], 2)] for p in w["peds"] if not p["done"]],
|
| 483 |
+
obs["visible"], light_ns(sc, w["t"])[0], light_ew(sc, w["t"])])
|
| 484 |
+
out["steps"] = step + 1
|
| 485 |
+
out["time"] = w["t"]
|
| 486 |
+
if hit:
|
| 487 |
+
out["collision"], out["collision_with"] = True, hit
|
| 488 |
+
out["events"].append({"t": round(w["t"], 2), "type": "collision", "with": hit})
|
| 489 |
+
break
|
| 490 |
+
if w["s"] >= goal:
|
| 491 |
+
out["reached"] = True
|
| 492 |
+
break
|
| 493 |
+
return out
|
| 494 |
+
|
| 495 |
+
|
| 496 |
+
# ------------------------------------------------------------------ batch jobs (used with multiprocessing)
|
| 497 |
+
def summarize(r):
|
| 498 |
+
return {k: r[k] for k in ("collision", "collision_with", "red_violations", "reached", "hard_brakes", "overrides", "steps", "time")}
|
| 499 |
+
|
| 500 |
+
|
| 501 |
+
def collect_job(job):
|
| 502 |
+
"""job = (seed, kind, density, driver, layers). Runs the driver and labels every visited state with the expert."""
|
| 503 |
+
seed, kind, density, driver, layers = job
|
| 504 |
+
X, y = [], []
|
| 505 |
+
|
| 506 |
+
def label(obs):
|
| 507 |
+
X.append(features(obs))
|
| 508 |
+
y.append(expert(obs))
|
| 509 |
+
|
| 510 |
+
r = run_episode(make_scenario(seed, kind, density), driver, layers, record=False, on_obs=label)
|
| 511 |
+
return X, y, summarize(r)
|
| 512 |
+
|
| 513 |
+
|
| 514 |
+
def eval_job(job):
|
| 515 |
+
seed, kind, density, driver, layers = job
|
| 516 |
+
sc = make_scenario(seed, kind, density)
|
| 517 |
+
r = run_episode(sc, driver, layers, record=False)
|
| 518 |
+
return {"seed": seed, "kind": kind, "density": density, "driver": driver, "route": sc["route"], **summarize(r),
|
| 519 |
+
"events": r["events"][-4:]}
|
evaluation.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
inference.py
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Next Drive: run the learned driver (optionally with the safety check) on a simulated intersection scene.
|
| 2 |
+
|
| 3 |
+
Simulation and education only. Usage:
|
| 4 |
+
python inference.py --seed 10007 --kind mixed --density 2 --driver shield
|
| 5 |
+
python inference.py --selftest
|
| 6 |
+
"""
|
| 7 |
+
import argparse
|
| 8 |
+
import json
|
| 9 |
+
import os
|
| 10 |
+
import sys
|
| 11 |
+
|
| 12 |
+
HERE = os.path.dirname(os.path.abspath(__file__))
|
| 13 |
+
sys.path.insert(0, HERE)
|
| 14 |
+
|
| 15 |
+
import drive_core as dc # noqa: E402
|
| 16 |
+
from nlrt import load_model # noqa: E402
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def load(path=os.path.join(HERE, "model.json")):
|
| 20 |
+
return load_model(path)["nets"]["policy"]
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
def drive(seed, kind="mixed", density=2, driver="shield", layers=None):
|
| 24 |
+
sc = dc.make_scenario(seed, kind, density)
|
| 25 |
+
return sc, dc.run_episode(sc, driver, layers if layers is not None else load(), record=False)
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
def _selftest():
|
| 29 |
+
layers = load()
|
| 30 |
+
with open(os.path.join(HERE, "selftest.json"), encoding="utf-8") as f:
|
| 31 |
+
cases = json.load(f)["cases"]
|
| 32 |
+
bad = []
|
| 33 |
+
for c in cases:
|
| 34 |
+
_, r = drive(layers=layers, **c["input"])
|
| 35 |
+
got = {k: r[k] for k in c["expected"]}
|
| 36 |
+
if got != c["expected"]:
|
| 37 |
+
bad.append({"input": c["input"], "got": got, "expected": c["expected"]})
|
| 38 |
+
print(json.dumps({"passed": not bad, "cases": len(cases), "mismatches": bad[:3]}))
|
| 39 |
+
return not bad
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
if __name__ == "__main__":
|
| 43 |
+
ap = argparse.ArgumentParser()
|
| 44 |
+
ap.add_argument("--selftest", action="store_true")
|
| 45 |
+
ap.add_argument("--seed", type=int, default=10007)
|
| 46 |
+
ap.add_argument("--kind", default="mixed", choices=dc.KINDS)
|
| 47 |
+
ap.add_argument("--density", type=int, default=2)
|
| 48 |
+
ap.add_argument("--driver", default="shield", choices=["shield", "learned", "rule", "expert"])
|
| 49 |
+
a = ap.parse_args()
|
| 50 |
+
if a.selftest:
|
| 51 |
+
sys.exit(0 if _selftest() else 1)
|
| 52 |
+
sc, r = drive(a.seed, a.kind, a.density, a.driver)
|
| 53 |
+
print(json.dumps({"route": sc["route"], **{k: r[k] for k in ("collision", "collision_with", "red_violations", "reached", "hard_brakes", "overrides", "time")},
|
| 54 |
+
"events": r["events"]}, indent=1))
|
model.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"format":"nlrt-1","meta":{"name":"Next Drive","actions":[-6.0,-3.0,-1.5,0.0,1.0,2.0],"action_names":["brake hard","brake","ease off","hold speed","speed up","accelerate"],"n_features":21,"selected_dagger_iteration":2,"note":"simulation and education only"},"nets":{"policy":[{"type":"linear","w":"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","b":"4FS3Pjs0Db3i13++Y3z6jvqkAr4mKe6+fhS2PRMojr3IjvU9RLufPjmHwrwvx2m+OLRKvCgyub3pRvo9038jvjeiAT0bQ4m+DXUxvsfYgD5zQUk+Eui9vuBCuL7d0wU+FxWTPqFtRr4RaKQ++tfUvVzuBD66ULg9UiXFvTsVMD4qCPO9GD2KvuxGzr0s2S4+CDz8vcvCtj7c77a91AhDvtU+Er2hXXU8MdxePRXpO7yl7K69+DerviKv577aEAc96YCJPQAuRD6G9+G+74A6Pt4iur7OSbW9Skg7vvkZEz4YlE4+CoGWvPvJbb353Za9X7O5PlHZQT1frbq+19EYvQ==","w_shape":[64,21],"b_shape":[64]},{"type":"relu"},{"type":"linear","w":"OVV6v+GOzz6YvM2/0F0FgFvlDb9w2C2+0jQpvzcawr6VWKq+sPNYv5I/v76kSmY+oyiTvgfPhr/soNk+2RTuPkwawrz1t52+3e6dP2MY7T7V94c/JzHrPlEV4j7hutg9mnr2vnsiMT8OCSy/IAVBv3WBl7/AY2s+rlwwPzA5LTyp+cy/Gi6uPpvC+D5zFT2/J0CWvgIi475WQ1S/IxzoPYfJkD6FDDG/itPdPYvHcr6VYfG9PBtQP02yFD6PFmU+x9vGvSJPiDztf5A+G2QdPJjfRD8wC4O9B9isPkdiEr5jeem/OeYbP541Zr8klLk+WlcPvwsu8z3B4iy/cOymvmkYRb94ZTI9OoEmwNZXBQCvJIi+mB7EvTIkLD+XNBs/roYQPanfIj31lmC9dVibPvGp+LwhKZA+PbCWPkU99D1LqjY/F4IIwNeD8b5gKlI/sE6QPk24Ej3TZas+bZSYPqF64b75wvE+QBGlvw6to740RSW/32emPHcsDz4chw6+JcoHwDnGHz10Xr+99xw2vrson763SZ++8okevww2373tqaa9d4jevvKQLD7pLOA+PleIPdNxvz7HSjE+tFn5P+nTxrw1m1u9WMPXPgtbkz917QO+GFUGPdZE8j3pSzc9R0NCv0fliT4rxCi+Jr1ZPvQ3I79Ak2O+OrPBvj+ZI72Ob3M+zsNBu4eMcj+3XAUA0ZkUPWesPb8fIlY+yajwPp+BRD4AckW9gIGRvutryT72G7E9mlVTP4IadL2nqzU846ArP2CjRb3oDYW+aZJIvXQ5Fj49RwM8xMi0PhenNb0/0LI+ENIFvjIhlb2OsfA+wcaAPYMy2bwLU4S+L6kavjuA1L7pxq6+VYACvaoUiz65rJE+40JfPk2Nz70alx6+VgNiPh1jkb+PKI89uVkaP75YyT4o+QU+97gfwMHuhz58EZ0+sxZmPl3ymr7JXAu/Yh9Ovqjcgj3KxKK9D27uvddThz1aAOY9RwtcPqWxk74zFUo+624EP4Z4E79BtSg+d1YiPgMeaz2jnDa+2VgFABfelz4q3oy/jxHUvBOxmD6XzCa92gomP3/hvb6+Mdc+fJ9xPY+rQj48d5g73tjXvWvZOr7hkha/ntSlvQlVrz09W8I+lD+LvES6SD2JROw9MPkHP0olLz7uEZ082ICaPsqP1r1z5Bk+e6yYvmOcgz4soky/o/2xPiszSb6+fCI/RFgEvm6UF75Qjak9pDcwvE+DAz+qeKe/WbcPPjmptL5dd5G/5w1pPZJ0er/RBR0+8PhuvjBzoL710Ok9evsBPoPCCD4vB+U9/B+cPJdfmT3ES7M+8vv1Po4Fvb7/0c0+SbkevWhHFj59WSy/5qyFPg3ZQ72rQKU+S9GhPwhgBYC1z2W9cjoBPxDoHr8/6Gq9bfBWPg7rwr79Frk+MfEQvfV1QD1d6KW9faa7PZGchj3I6c68ElVSP366kr7UElK/mTubvofE7z7EGpc+rPEePgW2DL7axLS+IuFjvfoz8j6/8Ps+Xj6+vq6IdL6JfxS+P3W3P8Xtxb4oU66+VI2rPlyzC77tCg6+cuyKPjB9eTsDCqc+7TYwPtY87L1IAfC9Al7Zvqt0Qj48nQK/u3OBvkQ65T0EGdC9jduVPkX3Zb0U/Eg+PP4dvYDpIb5hLMa+G1Z+PNcbBT2L0Rk/Wj6GvmIn4D55i0k9yT7NPX26bj4N5smYtSjdmw5eBQCUvAWAsGjfnMVYBQAFfoudy10FAFf5lgYBWgWAbdnVm69LT6J7uB+H/18FAMilfqycVwWACF4FgOrKtgNamzwHvmMFAKpbBQAR/J+ep18FALXCVKwpYAUAKloFgLBcBQC/WwUA9yIEkgLKcwUqXAUAAfGuBvJeBQDvVwUAp1gFADpfBYCkWQUAZMoCB2+l9ZSZYwUAm9YcmKdXBQAb+RKfcFkFANxiBQDOXwUAKV0FAKRkBYBXXgUArGEFALJhBIkAXQUAMdm5BTWNBgBlTLCezTPaoBx2kJLSGpWMYFsFAPM70AGyWgUAsWIFgEhOCInJIJGPQM8BvyvPIz4kW2O/oV4FgPQgK77bg548fl13PoA4Oj4Lclc+unAKv58NxL1m5rI9dULZPXbo5r52lX897W3QPn9a/T5UdPS+BCaPvuqlgT9DaDs/iTyrPPtBDb2hJ9m9tCFGPatGu71YYHG/lDAFPlydir6tI+w+NiK6Ppmhx71nYGy/HoT8PX6ytT4cjGi+zd5BPbhVMT+V5x+/q9r7PcGSuT5iwMK+PuWAPSpR4j1q5OM+pg3HvrCWwT525j4/eV+ave7FMD1aJX0+KBPKP4hUpD7k/4M9DTClvc68U77mKYa+6UeRPqbcVT7zMj4+47g7vwFQsz4/tV6/LsouvoOn7b6nY58+4+/vvideBQAgVYk+PR64PvCA5r550Iq/NJkSPxmqur7LPTO+6olhPasbjb29Uhq/uP64PvJVhz//L4W8tDqKv4CnID8zqys/Si7KPuXwWD9k5yq+XdjLPhSM5T6i542/KPUQv9UZzD6f9Eq+B05xPmUKhb1RkWg9578JwImspz7XXF8+8fQEv7+bRb94yz4/iBbtva0nir6X6ho/96+nPzlV6b2ddS+9eK/bvgCfHL7n77Y+kLuNvpQLjb7oNdC9esT1Pl895r/ZV4A+QrR9PXi6pT1e+4c+AsFcPehL4T6tZ5i9egP2vDiJZb9ldT+9bNbHvxGDHL31ASE+EruGvKuG7j7+YQWAT/qJPYEncD/atTM/b1RyP5D3mL5EnB4/c0WpPKjidb5i1VK8tTcoPxciN77avIS/8Lw8v7YLHL39x9k+GbsKv4PYk70uots+zq7RvHma1z1s1u+8zghUPxWs5z7zKI0+ORQJPy4kN7611yu/epqKPlJdHz+zRXW+Kou6vnT/0b7jzHW/1LUWvxeILz6innY+zYIHvw5URD+DRgu+w7thv/49Oj/cawU/uQHTPpLHYb8ke+I+i1imvjvq1b5/28e/NV9svTa2obyz1vY9NM3HPr4b3T6x0Pe+S+vRPlDU2z291rI+5nevvizj2D58L+077gQIv7QN4T0PKHPAs2EFAA9WJD4xh6a+2cTOPrgJtj4+qYS+WC+dvqe9gb4OK8i9CwGEPMSBtz39eaw+qFkDPxSKCD/ddYe/qJ2mPB3XVD4XPa89zEZmPV4zqj6KLGw+8HoDvwILMb4l83K/mqsfve/fN72qbJw+obt2vohZ/T0IQBnAKXwfvY+36ruv3RU/yNUoPYHjx77CwFQ+euaNPUZNBj7p+v6+sOQBvldXuL2n1U0+0AvAvjaIJD5V35s/5czYPQrlTb7Fh2c+LafpP9lhAz3meLc9Cc8hPnPM/LyvIns+ekPRPa96ArzUfGs+lbffvsOOybw7WHI+O0XAPvp3EL/APnc+vpfVv1BiBYDF/GG+5JDrvXIW6z6WTLg+Yhi+PMpJBL+zmwO9TUtjPiavaT4etPK+PaFfPg6ChT1HXfy9yHQmvyTWcb5eRjo//iE1PvaWQT7QyFA+BLaoPd9iBj18O5w+kdwlv943ib72Pb2+QH8TvFDK0z4d7Zy9QnxWv//Wlj6KXf49+M5Lv4dslr74vag+/4PuvtX/kz4iGdA+awVIvx++bL4w/mE+94/zvtMCGz9zw5m93gCjP4FeDz68ALY+J3fAPjVJybob02M+gtBMvPg/9j24TbG8EAvmvlQCnT28Ges+DzEMPtoj9b4Lwic+ga6ZvxyXbT4ldIO+qFmhu70sdT5UWwUA8HdJPFvF2z5Euui+bxcaPtWgGL6UizG/QdaAPl5PJr2hadY8Zm4wP5MSLj6t8EA+U+MHvrlr270F7Si/OrliP8DDnL6qhQE/WkOrPh+VAj4STBY+h4/jPrp58T34owM/oidmvmNA1b7Sf788NRS4vRnYlD4X5Pu+jjI3vB9UuT74E+O+Xp/tPSTbUT4OCCk+6waePmNYoD2bXiS9wJKjvqPnW7+m8RO8XLWLPgKKBD8lyf89x9lbPUAC6z5vucE+p6K7Pq/UU71rGea9gWfCvnWC4z1v3nI9EfntPcgeeb76K4q96v7QPi02Br+awA0+OnrRvbTpLT6sqgbA3N7AEsiypr1njay+Va5cPker/L4ud+S895Pgv3Itab6loPE9kJYwPeVzwb7mprA+PZHgPm+tZr7LZz++Z6Yguz1Zxj5op3Y+0rmqv+bQhb2RV2o9Hk+nOzeWQz/4RkG+T6UJv2Gaqr5LgyY+di8ePln9Bz9fNSDAR++kPlr7Er+cNhK/xXXSvhtKqL7psMQ9XtJIPo0wMzz0tgO+6r1GPVQBAz+5VpK+6t9BP6zkO7ysH3c+Aq+bPuP8Gj8gLLE9/lsFgJFhRT4twa+98CvTPmBwHL6PPIy/8ndyPSXEGb4lq1S/LtIdv517N73U4bk+4/onvj6QHLT+5v2434cGNo1aBQB9x62xem5fMAnOj4agxQaA1Zn6txzB5qM7UY6x80U7ucSaBri3aQWAFPcDu66Wt7MpVo+2LcRwqFh1gLOyWgWAdRSXCxydyaQ5dK2mTLwqu+yC06j3l9W3v6OCtUIeHrBV0G62kx8ysxwGpzdvHze33R8MqTK5obYdYZGql0A6snZEzbZ0JkGtZcvCrmeWzLT6iT43OJsKsIAV3LjaRga4SBqpsx5JiIhqSUe272QFgO2FIDbwB8G5Mp3+NQgCkJLvXzC0vpIGAEj3mbcniwW4Eb23swwxMLZ2WBygwpiMtgzWurZLojizwONGrlMqKbfmFZ89zoJMPiEKiz9+XQUAMnidvab28T0CwRu+A8wmPwqV4b1dV7s8jFYiPplutT5Gr0C+NK70PF58IT0+fAm/e501PysRXj4ddM8+R92evoXBBL4wASY+vhC6vBujNL7LJEG+wO2Av2Rfqb5jNZA+q52dPv6qKj6hQY49R4OXvVDn/D4nbso7vOzMveh/qj4nf/w+G8LWvcIPnT39f8A+IYtFPzXyHb9p3/u8DOYEPJffEz83wng++AYVv3rUN754DQQ+pE/jPhzTJj88j+K+5y2MPslM2ztsfq69oyXIviAdfr6g9rc9TYcaP9f0KD53EiI/BzUSPu/ajz54B+4+o0lCtHe85bWQddiwi78GgDrPV7NkVp+tIfC8r/nFLAIW7xC4fBW0rEGCm7alylC3X4++txHCgqSSPV26E8I6tBy71rftwc8033g1sAuVH6sf55SxjecPrnBAWa/GPwq52sLuL7GiprQ9g7CrPg2ptJF3FLTw7gS17G07tZr0FLWPUk2v5bTmtM90H7PddXuvl7/VtdLukrg4fL+u92NjLkW+ljQX6zmsAeuruTBZy7dT7gSxN2QFgAQT8LT4cxMpHcr7tr+P9rXjCQmzYxHMlLvTg7H1lQWAH5QluA4ehbSdI2u18Gi4tiNIMQDa1aiykHRatBvgerfwKTM01q/DtP7vlrmRm0U4nT7xsnOZBYCAsJG5WmdvNTg92biQ4xk5H1Z5unEk+jkPUvq48iBMONMssbknonA33vsmuv/LPDlmcmm6AopFuaHCHjcDVAs58Lwuuk6z/7gc3TE4sznmudDlBrrOdQI3CjWOuaNpTLfdxWm5jCEmuWr3Z7e2epS5Fk3HOWCDWzgrcLc275MwuUq7lTholo66EtaROLZ9s7ilAhI5/MDGuc3GnbnObCa63aRROoZ0jLl69AC2iYqxt/E4yLldus24GpH7N/S6gLg8wzI4WIUFgA0upTi+LR632qiBuU0gBDgrm6y50OKAt+g+q7kMhRU6GvVSNbhX4bmvI6S8rOEJvTEk77wMkQWAokFCvQXnVzsdNne8gEEIuo3LFb2WDIy7HY8lva0raL2Y6uO8R/caO+Oxnb03s/S7fHeNuWDwsbwJx8i7fD2euzxpb7yQg968+AWBvHQGgr34Hp68y4tIu8YA3rvp4Qq9OCXTvDb5C733VxG9YxSlvCLcYbugGwy8sMhmvMvKebwtruK8ruFkvUmy9LxC6668y8Q5vev1A7xOfxq9gWU4vYvTGryEC3a5Ne0wvK+LjLnWqb68qSERvb1SA70HxbM358ujvH1eBQBnmk29H5vzvMFUqrxJXCi9D7ujuiGperx35w+850uyvLvsxbyIkji9BrHmum35tLt3LwW7WGQFgFuhg7r9qYq7dguIuzZdLbtKurm8DPvJuGZ3NbxNB7y81Yr0vKlILrq8DsO8lujpuVvRvru9OUK6Wv0susgCwrsTD2K7buM1uz3K7rt8TLq8I9BJuw3xjLu63Fu6DViDusv6sbuYNIC7MG0LvFTxmbysbR+73FE6vAqWWLx4a+G5sFKAvOeH2rvc8AS8gZoIOhJGdbzaDRe66100vdD1TbzeOZi75LkFu9d/arzdNba5DqQzuyUj8LsJU+O7zZBKsF7XoLsdVAyDTUvYvAiNqrzDxv67WBdLvFf/SrtFM9i7RTMyvHBpFrxyvKi7M6zXuwfqCL5aLZi9NzKLP1ZZBYCIhec9uVgdPwVnmL5zPdC+PUXwvKBofr758fS65UyuvZ79cz7xVd49hJ4bvgNWPL7Cc/i9oeXfPhYzgz4L9am+z7+QPbi6tj3TMmA+hCiCOofE/j3eSyq+ZPDCvaD89T5hBDi+tHRBvv7pKj7VquI9rT3DP0HWsL5Nf5W+/d/yPoISrbwwYqQ8F/NFPkCCjz7ksEU+1KliPaozDr3eMpi99+ABvlmKBz8zGZ++xGjzvknIBj4Ezbk9As8wPlgKVb8ELYQ9JdW1vRNuJ764L4u+YjC6vn+uPD7XTSy9JmedvaF41j4y9fM+XOwEPKObED5ZYgUAhnUTlA1cBQAcqQYAADIWletiBYCMOFWgsPcSgSB/rZEwXQUApsqAkr6nMqDfXQUAAGQFgPO2Q6ywXAWA2WEFgDMBAYPVYwUA0me8mrZsd41KbMuZdVwFAIfoEKk+sx2APFsFAOZXBQBG6NyTTj9pA9E56QTRXQUAwu0UBfVXBQDzYQUAZgKokuhhBYAxxwEAF6Iao5NEBACrYQUAg/EyjthiBYB8UKqd3h/3m7ZfBQDcWwUAa2EFAF1hBQAuXQUAjl8FAI1gBQAvXQWAClgFAF1lBYBVPSqL+SKrBUuCcAeq5tWWSWIFALJYBQCJVwUAOl0FAChkBYCHWAUAjQJNPnpwFj4cHNa+7lcFAD745T5oxFO8opWCv3HXEMAd6x+9Il+NPIxGmT2d/J4+jluqu8Vssb8Eo00+bI58P86uG78D86G/S0j0vUa9kT/vw5Y+pMc7PhIsUj/ghwE/vtarPiN8QD8HlW2+b9H0Pk2lpb6jX6k+HH97vs+di72VwwjAz5TFPlVICD/4lPC+fcUHv09KTD7d/BQ9I55hvnfNWT4bB029fMCGPhLRBj7+5bC/1iZ9P+VZ/L6XMwC/XZDlO53fS77qZ/w8p6myv3FOaj6PnkM8lADCvWCwGz4gxtk+DiUNPxGE/r5GB78+vT8QvxgvOb59/gXAs54ivm2B4b66N2S+qZQHv0RfBYCVRRC9Ik4GPlRRVD/cpco/bekTvwnxCj6bCyC/LNRAvuTnfr4rcOe+RX3svckGFT9EDo0+wf/MPtjuYb+2TIQ/NZZCPmvIHD5IAO29DkyrvsAJAz7liHm+yluuvo2PHT/ks7m+FQLtvN3w9D1UCQW+0us8v7F9sL3W5XK+GNjOPmSY7L4RGso+T9J9vr3RW7wrj9E9/7Nqv7ZhWL5QG3e+/yOsvVn9k71KHQI/7ZuaP+21A7+JHZK9ydGxvkOI0L/3PB6/+K6nvOJcz74pody+I9MXv/L03j489Mc+Bz6mvrRXj7+UpxK+eCoDv4+PK75yij8+FyAMv2wWZr5RWwWAKzqEvlPFnD2vP9Y//ZQaP7XUmj1/faU+OPWhvp7iUz5H3Sy/7cHKvNQ6HL8C4yM9RTkbPVwY6T3H84o+s+C2Py5Gg74mibE+hgACPytMjb6KzNo+rnSivB1w/z6CRSI/fdklPmdfQD4BS3A/WeIyv3I/v73rkRm/pq1APoxnY74mZaW+RY5wP20wJr4dzBg+8GuiPdWDDr7ZzRW/efjgPTwkKj2VfuA+bd5JvxHHfT/YYX6/ImTEPQjY575S7es+k0mEvv+GAb4rR2e/MsETvrbZ7D3vVzi9z5BTP7nOOr/cU9s+GYEFPgtZh79VjJq+reuouFLlA7ptKZe4ebkFAMW3mrkm37m1KbqatsBStKlLpuq7vGy7t397gLqzfHq7jfnUuwxvorgQ04G8WvlTul+UjbteU+i6O5foNhL31LSfEhC3jokcuCJggLrNcz28+OdJufkzKLuVNp+4ZBvstwoWU7m2VXu559+iuYoKkLpo6UK2xY74uSoEjrhFHdy5n4O8ueI1CbwhHjmytAiTuuxGabd+RwmvWWsivK433Lm29r+vdcuZrdDOL7iJJjWxK09yuyCDmLrE23610Zt6mUoUsrd3sgUA8Gvdu5CcurqSuhG5yKBpusgIRDEG152646FRukwICbteIU20gbndur9fA72fJvg9SbEmPr5hBYAGtcm8NNb1PbaEYr+2AyO/L1ScPpQIGT1RBbw7cxtPPhywOj1m4Mg9LYPLvCp0qr/50Qu+eCdkv7VtPr/+R3s/UH0gP4pBpL2SmB0/ETSJPlFgUz6swK86VZKfPkji/z4WwmG+7fxTPghT+72GeD4+3HxGv+Qfqz60qlE/EuRDv4qc472TkHs/w9LgvsfJt75dFhc+A2bcPmmdgj2LDOq8iCV1v+Wvgj6qH/M+SnUnvrzKHj26KTK+nVClPqjwuj4OUsi9nqwkPWBEMz73rxm+I778PgIXGD4nxTK/43KUvsHkob7JssO+F/RpvoxqGTzc6OKSFHI0ri8LKyDgewUAHxl6ESpkIK0Qt2SODl4FgE4BCLLyXwUAQYEPsYWU8KeTxYazPmEFgJS2LbaGqbmdLLRcqzYSPJxQYAUACCcTABGoKZJc5eQDoMW8lXzDx7U0ykGfvwohri6SYKVa8ZuMa9ZbqniilRby2wWc1+T3r7u2OZwxsoyuUgczB4TSjxjfqBOnJu3VnYhYBYBitf+gKayVpIjSnJTuh2e1OWv1pS42FJhmXwWAGdCIsjVXBYD/Xjqux/hlrzGsuZuWZQWAukcGJDVcBYCCxIGzDVXprz5pTKexpW+mLGIFgM92L6NLVNSoqJ0lqQiQxZPZWsuqVChKnLDB9qmas7iaSKIGgCd58atdXQUArc2Gr7WUH6rxHzClaFgFAPSzi60qyeqpwwYCpwULBBZ0XRK1nGMFgIcwPSbWz54svd8jlQ4eM6zgziGlDa6drI4ZgahjvkazY+PGF57YLaCEWgUAuWtRqFe9Z5eVDgGmvgkXnF80lJ93BYSV2DJRoszCVqpHSwOAsgVInIfU9ywjmVunD0qTlSJ+iqB/VwUAE/cQsJReOJsVRsCS8PgUnRqdfprgCAIbdLamoBe3R6jh5/Ka7loFgLAC26C5YQWAdtmpsJlbLqZjLACjyBUGq5+KEaLJT62fV8OIl/Ssi5vBX4Shwbiim6Zp2D7obbA+53lHPs8tro/Rw34+XNRMvx+k2b60WN4+ph4wOlonNj+ErRY+6Z6fvTmQmL5b6lK/cNVDPrnzO79skq2+F7LXvEMwwz6Y/p++hPQLvjMjzb04Hfy+cFUsvEzswD0EBbm+gsGTvWTw2D572So/6myRPm4eEj1IkOw9u/+qP+g5i7x2IiS/U6GTP2wGyL5aHSK+CyhtPnBq4j7SY/s94v63PRJLq74zrZO+Gj3sv+WhtL5U3qC+o6x/vKWslb7IUQE+6D8HvWd18T4H5o8+ji4uPeXzkj29DNU+6F3FPiLwbz16xWS9/FAXPzfNvz7FmVy95XJUPUEKIT9mXwUA7hvPgfpXBYAmtgUAxs78kiNnBQAC1S+hhHPCB4dAsItEYAWAMrS7ndddBQCtYwWAs10FANyax6usZAWAZ2AFABlZBQB5WwWAHclkkjlbBQCIusOVHvvYBi7TV6lSYgWALGQFAFpaBYB6KdaX11sFAJFYBQAoWwUAsVsFgLlfBQCKYgUAmFwFALxZBQDjYQYAb6iGBYEqaYh0XgUAP90EgOtjBYBraMwbdUKWBUddBQAQXgUANVQFAAhdBYBUVwUAoF8FgM1hBQCcYQWAXVsFANtXBQB/mEyXJVzvlL0jOwR0I1iB56pFA0hjBQBfZAWAJGQFgBEUtQGMWAUAPQ0HP7ProT3ARwe9+mAFgIdn/z5F7gi/tfILPwdvSz/sSkE9WRimPiQHdT46xq09GUBiPjC2MroC2Ig9StD4vhlusD0JRWI+cIYpPzc3kb8eE0e++MKfPm/GTb4fpnM+uR/FPW1KFz7vxp4+PeDavkVBOz+KE/E9TV6ivnTZGT4BBOU8aei4vSUdAr8hXZg/W2uIvmxdO796Tj8/cekEP5bGfL4fGRS/0bVmvuWyx77C4Z6+p1M0v1MUIr9xvwW+gcbrvmVvRb4n9QG/+sWPvqUg6z2bBlO9cvYbvtX5Lz9Pvqg+yBqHvHJyBL4jlFY+a6t1P/VWHL7333w+TSvIPjirMr/XjBq+g7IbwBhcBQCSb5W+B1oyvlZ69T6lrQM/PmY2vk8zt77+qVA+n2yIPnu2Bj6QCjG+PZ2rPv3SHj7Wzqk+aVGavx6uBcDct4c/gvoVP3Tzmz4TvwI/oCSvPfF7yL7ZmFQ/Y8WIv2qrDT7d2lW/DrlAv/7tCr43Quk9L4LUv2GQ/L3w/cU+THvcvYcXm73CoFs8enghvwje8T2gDbu7zyElv5O2iD6pjog9HmdVPi0XJr63PDI+UfXNP8g/VD6IFo49I92jvLwJgD9/ev6+utlpvY5SYj60siI9FfSHvx2M8D4jcek+QnUcPkcHmr8oDuW8+5uFv+/mnL6Sfzi8AtpSvXERrb/nXwUAIvuivRIWA0AB3nk/P1C+PsuRQL5DENg+FhMSP5Sm5r5QoB8+201Pvj0Kzr2YgTe/GGoDv6Lmij+krYK+ssJ5PQfyYD58CFw+XGCfPlrj+D1xLdm+PRF8PuIOQj51Vok+5D+pPuwjor4hQ8y/J5kIP9DYNj+QLp6+GzEDv8hYtL7vWaO8+6JSv0HeFDxFiK49B4OwvACGrj89fMQ93B2Fvw5NO73rTwq/AqC4P0zVrL61XOu+GCV7PhCYiT6RjVi+aYZxPj6IqLs5J6o8NTHfPuHwgj7+pZu+zmEhP+lwWj105OY9xUltv3T0yL1kEvI9GrsyP+7Yk72L+ig/9FwFAJJJg70S8nc9EwL6vt01wL4ffks9eKqWPoGgiT7snbU9UQyUPfZYnz5UUR+9KdzOvguYfr6277I+7U46vxSfWT0F/oC9T7JyvhzwOz0ZlTO9vlNXvh61kTsryAw7J9KCPvknKD7jTY094Wm0vv6n7T3RCAJAqTgQPlyiiL7BDWy9MI7FPmHyMT+tjcy8Q6H6PWOfID4tF7i+wz+QPtZuATwVJSa/Ow7Uvjj3v7zZJqs/Cq9ePXniGj/aGME9T5k+v0k/mb5B77U8IjK/PQNkjr5a+Sg+sRevPE7XIL+BgPq+bXIKvt2sv73s98o98qISPuoLS4GYrxKYK603AFegBYC0XQUAvV0FgMxfBYAmXgWAd/O7nN5hBYBNePMHMP8bm66zvgQEZQUAkPP3qvBiBQC9TmOBCl8FgFJeBYBUYQWA+4tJDNhkBYC6YgWAkX9CpxLIsY7bYgUAqGEFAPRXBQCkrQCAMzmPgCKUpQZ9XQWAkl4FAFpXBYDnXgUAwloFAIVoKBhqVyMc32IFgNgPF4K2XAUAIMMGANhVGJ+XKokb6mMFANxYBYA7YwUAQWMFAKhcBYB5MR2byV0FAMxoBYBeYwUAgwIGANxALptGnoWBc18FAAD2AgDlYwWAw9KYgk1VSgbFYAWA9lsFAHXpRZJmZAWA8KlGo7ZYBYBHfwUAH7OmguxfiKdfXQWACVwFgEAuZau/YwWAIVS9q/J+opCgwyuxW14FAGx6OrOgYgWAbryco/xhBYAWZAUAbGMFgBxkBQD8ZAUAr77VCdY6tbGZVwWA1wwyrkkeIgreXwWA6PUcoED3BYALWwWAPUISrSbhv441TEehqGMFAFRiBQAPALwGVZLGCetaBYBfWAWAQd8mo2RiBYBRfbSz7V0FAMBcBYAiaAUAZXzzrZVhBQA4mSuj8bfAgUBaBQCFZAWA3QXmoiu5BQCVtG2weLeBrS3jrQbQh6qUdWIFgDGkBZINH9ackxEjmrpZBQCTP4uiEl8Cv1JkEDydeAfAoWMFAM9jR75gUfq98AjxPiHWhT+ksW6+uDthPqXRW774tvs+K9YQPfavBL9dSJg+dX/rPlsu2T6ufMm/jkdcvrS0bj8jezg/M0eDPtc7Kz9JaYc+QIJsvlfVpz4izUi/sVJ9PrYIVb+0E689W68EPpjyTr1rMKG/sdCrPdfXPj0yhGy+xYvMvj2crz3vL/++EVK2Ptpcj76QpL6+RVPDPfEhAz7HFI+9fxaLPlFtmD77di0/2skwPOinl74qAm4+vM+pPp/+Jz6DhTi9sRN0Pg+hE75P4ma/AuUFP9QdvjyJCN09TjVSv/0QhL0WiFu/UHKDvlTPhL+zyhs+8gsTv8YtVqYuGps+LS+Ov+IQN77POXs/PY40vwHGM79xbFo+Qx2mPqVAJr5r/eO+NcI4PqBkvjw9V5W+sM6Hv3TK+T7lQT6/hGKdPha46z5NVr4+HIokvuWP076/QYu+LmWFvjYiJD7mhg+/osKOPoc3DD8SSAi+rFSowMQ4FD9jzow+yj98vh6/Tj6+jqu+K/4GvIyaaj4XFpY+JeDevvxSwz25Q6I931Cjvjx9Kj9CcFM+TobXv++Ixj66YpQ9YKttP+z7k77Fy4w+tlpVPYd17T4DSiu/n1O7vzeGEz8fCBA/+Uy+vWT1ir9uEdE+NG2TP32YzrxiEmO/Px2CPqO5s77iC8+GJQx/Pm1kjT4F1au+fzcVP4wrHzsAAo2/AS3jPNi2aj69fs++/oCRv7lE0z4Qi1I+CmejvUA5Ab9VdwA/pcEnvkbphz7UeQE/qhWwPh1q4z7uHsY+UfFYvt0qSb8s9528GbmYvnVRdj55GtU+bdZAvhRMMMCajwE+X4RBP+g05b7rsTy/0ypZvZsbqb6XYM69uejlPn+KTT/OhA++QD5gPgR3jb4Bso+9uE7NPg1G378Z5fa+DTwVv1dbGT7mFzG/OI45PzMtCz0jZHE9Ep7XvMBYxL4xWgk/yJ9zvPVMGj83Iwm/Ne3kPtVsxb91hRe9w9EGgFu8Q6I4YgWAqOQFANAMZ66RYAWAVZUVrvAxu6HlFUqwWF0FgGnKRKYoy/mw/CjPpzFnBQCtU2u2UAqbkswcDauTIUanBGkFAOIUNafkhBGm9JkvqewXzYYKheSy6+ORmFfPapmmbYSBFN4Wr9fjhZrNT+Sep97yoW8zj6C1YgWAxFRYpSXlm5+drp2VEUekBMycVLNqkQaevmnDqCIuK6cZYAWA5oejsb7Qm6WBWAUAfu7dgPo1mZi9YAWAFL2GqPqtkadhZAWA1GUFAMzv8wxjdwWA6OmOsCJxzaaDXpWYJ/CIqn5ya4VrOsiHBVsFgEv2rIidHGWhP8vapg1kIb/obUc+KfzFv61cBQDEvEs+v57yvqP+hb5J1Gg+vtqZPQ7DuL4lRRy/KfiePspaVT51bbC/Zd2/PVJTBz8U8N49Ky09vhKcVD7aHA8+d7YYPga1WD6Br4k+uvHUPQMpPz5rxBk/1xbVvtDs0b5DVeS+SZ/8vRnxFj2NQq69A6KkvxEn0T4runY8tLX8vpCSSb6kMlO+tI8Yv9C6mD5FNjM9s6rfPayLur1aJ7s+X3mHvtLzrD6poLW+OgOEvhh40D5a3fY++XEKPTFYTsD6mtI9FViOvbABQT5l1Dk+jjJhvaEs/j3EvAs/xFYJvv4WXT4MWx0+wn7LvigWNT5hR8K+dRFXPYZcGcATYAUAxL7NPAacQL4NLoA++hpMPxqN4T5hwD2/VLS4vow6mj5j0JY+S+QIvdPoEb5LYgI/LYRdvetv975NdCK/W2+gPuZOoT604vo992yePr05dz7ugfQ+m+2cPgNe/L5fUzq+6HepvhIHgb6RXum+E4B1u3T5DL4CM6U91Ai2PiQMwb56++S+PZOZPvtf0b5PV/E9XuLkPsdmLz6AYwK9o6b9Pvb0iD6cHps+u8WUvmr22T4qL10+/6FfPiyYLDyT90w/lI/4vAdJZDyWPdK9LR4wv2hNrL6gIj6+Kv4fvjW62b21hE6/GZ4MvjO9iz3+qNq774lFv7f3wr0YZSTAj2MFgEcga73iuKI+OqexPuygYz+mSii/DbX9PCzmWr+0rfw+M30av/hS277kdEA9a7ERP4Xntb7aK6m+bVuqPxeu0b04kfI9FIdRPkVdHD//Anm9NJMjP345UL+jTfu+5ZfAPl8ErL9pmdo+EzFHP+1Va7/0kVLACKXCvoqVAj+tILC+cVTzvMWiQ76HDy6/cSwdvoT/NT9qDTM/FBWAvxYMGD/Oais76lOGPzSsnD5BkADAFRMmvrIow7+pDZc/kOIsv1bn0T4b37W9lv4iPdSGfL/VcI2/UyHQPrDH3D7gpQs/zUNbv+yWUz7MeCy/xPinvnnVjz3qYr48mx3Ivm9fBYCW/Dm9urjLPfjHADpX9uS9ZNEnPgYZ7L73xyG+x42bPhnKIj7zAMK+TWJ+PtlV2L6uVvO9oZHDvuAI5D4UuoQ+YpVXPkC4ZT6Nrr0+NJOKu6C5Ajzt5IQ+GfsHv9HhWj639cO9lfpVvr3m3j4NtkI+t4g/vzmSeT6pRFI+6RAYv5tCKL9Ppkg+fY+fPC1+GT54SKG+88SqPmlfAz7SF748o/qHv6/KgT5rkWs/61cGv1vEGD44uQC9qkQlv1KSHME8Q9i+fF30vCV0Oj54OZK8/ZKovR5laj52UP89m5NZPNe7Vb9RyvU9b2Cpv86IoL4h9dG3aUAHuWNtKjoFWAUAcThTuuzCk7jvJOC2IkAQrrIatbokM5q2BB1cuoc8LboDzZS5IPhQNkkC9bqq9Q6z4kdZuKOdRjh5D9u47hyNthibRbfZnBm5933VN6Z0oroQdRM08vm4uIB4FrOvkjc2Ag0KtyHVELjlWhu5C3TSuV7RZLcegxm5scWCNKlwjDmtW7C53GqztyShH7qVpM80CFQhugm0kbdpbaK6Ne99ukwbE7h2qrqzL1vkubLesrBN/8m4XNMYuSjw27jGpk+164IKuf2HBQAHvry5XUzguY3roLeV1rS4eW+7tPxrHDqGMKa3OxFEuakayLjyit+3SxHJPjD06D3uqQY+r2IFgMQJ4T6qsle/JPbVPmiMhD7vgxu+CasjPid0mr4h1rs+WDEHvp7IQz1wzmw+3kiCvj+uFz5VZg2/8n3KO7H1yT562sM9fhwLPjfKmT4OccC9EFuvvG/ktj3enPE92NmsPr4jf76XbPQ96K9ZPf4M4D0x8Xu/n0avPrGcNT4AZhW+mRw1P6UMNT/Z87U9yAOIPpujeD4Lii6/Hj2XPFQ/Ej8b+cO+U7oavj0Mer9or8E+sADEvRFcmj414Ae9PyAivCvprD2TNJw9TmhjvUasd729d+s+LjUnP1XDhL2Sjee9kGAmv03o2D4a30O/dUW3PjWApb2I/9i9hOg8PldkBQDWRxU/LFo2P45tET9LN+W+O7awvRBjFz4b5Tm+RVlgPsFcpL0U9u++wazVPZ5iBL/Sq5a+QQlNv1ImAz49tac/NEYmPswNpz7AEPY+9fZIvSyyMjxGxhU/IgnDPjX0gT+clui+znwZPs6mPr02iqu9TLuEvz4G0L7Qd60+ObwlvzE7sr1wTCw/HlG6PaeChj5/m3A+SW1iP46osD2NVpC+6XQQwM78zT4P3YU+si2QvTTZAr95cZc9iVUePln/Er25w2o+VGnWPXGo4bvdb0O9ANz9PiLj6j7KcHq+3yyePlBNgL+CruO+k61ev/8OpT1LJWM+fdbtPeGkxz+iXAUAzcGlPD2qlr4ayJu+VMEEPhfNjbxDuLM+1k2WPqXciT033IW+8XEhP/s1hz59roQ+tONyP/XjuDpT2ks/4/m5vs2VS759LRI+JSukPuhwNT25u8U9wAEQvgaZzLvEVQ++IS8IP1xSAr40agq+7vi3vphkRb+XHia/RqvGvufQdD0wiAE8XfGFvQ6HKD+aM5o+KW7fPgmOOL+KbhG+cbAAP5cz7b5TSX8+gCBiwJwJPz4Khra8KCE6Ppa7Ar3uQQXA+KtEPl3XdTs1JlM+q1ilPNgJv7yrHNy9A/bBvYjeYL6r4Io+Kc0pPtIX0bxGYzw+PiJbqpS/j6tIo8aT5pcFgKM3OLQNlJMGg2ngtF+9/6pAPHqwU3gNo8V9k7G5UcWzH/Y9r583Ao7J0zq4TV0FgFtIN6D37Eynfr9WnhbFSq0asE6t9WyIsahcRakp/p+1YNVmp0muAp6BeUmESfuNsadr3qZYSYiudndFo1DKIapiB0SXxAs5qH79YK0hoR0Kj05SqP2ENbR7WZepi5W5pryAVa9+PvoFXUgLtW9+fK6CYwWA5lxWnHQlG6POSg8INVEXmbkjUJ0uZm2fsF0FAGecDqwgdAWAP4NZs6xMNbHroxarHJMosX8Dcai5mbSgE6HKnXCIpKGVTjuIkXJarWUrTj5cZRU+A2dTPhtfBYBltJM+XCWpPT5nQT7kZ1899MakPj4UqD1DjUq+2ApePlJ4jj441PK9dRWrPdCJPD88Nwg+qqWoPd36nj2qkgY+kVPOPpygPj27Y3Q+Foa9PSRchz7rboG9aUL3vGuJOj/vP44+FoeVPfJ8pD1zWIq+dirhPm/JaL3tX6U+WivCvchtTD44O44+jv0vv3CHqb4lzP+99HcuPuuYqTytryU+kQ+2viHoIz8J+Y0+a/ibPhFxMT7n5+w9Zf01vnJEjT7JSRs+qJTQPbATgT5swie7j58/P4GQWL3jOyg+UhUMP9pYBjxBKdS9R/sav/FtMj6eVS4/TUiwvpNBWD8gYgWAagGePVie+z0pIBk/YAo8PvrJlj5Uzjc+Fn9UPmK+Bb/FsIE99NgqPxRNBb7mOdQ+OzPtPu/jEz/ZIaq9ne8PPxiHgz5+y9S9XVbaPXNWhT2QlwK/yhKGPj5V8L8Gdo8+9h4zPl9ca78sOZG+peZQvtiInj8Gel6+9HpAvsf9+r5qPJU+SHjNPsmmij62ItC9yWfkvhIoYT8t14m+eS0KvqKM1z86q6C8VzBDvkBYNj+rN1u+2kInv8I6NL9nNdA8/D+CvvFMmbviIyi/8mmUPW5AGT/Y8rC+rn2ZPtoahj0c5Ze7LHm1vYJVhD/IYSe/wGQMP2B+hTwhsve/Jp4RgVL3ST6W2ww+slOUvu2A6j4b2KM+MctEPsr0uT4isUk+tPPGPHJVIL6ZWSq+OSrHvwB3vr7LfYq/RFe+vh0zGT7/yDM+/eK+Pi9QmL6BSV69D3U+vuvB+b5u2wE+sRfgPmzLdD6gvlk+u+0zv2E2hD4nJj0/v17evqj3Uz5sQJI/MT+TPCLf3T5Gd3Q+zgusPq7Osj1FmC8/wli6Pf7iA7+HClq/f02zvW5tDz8A42S/yPzovvGEiL45Kms+gISpv7FeKz6+tqg89HYCPuQmiT4wxUE/LzASvjjm3b59wZ4+TafXvYqpV78/t82+VxBjPjUJ1D4pIzw+jeaGvjNgBQDEymE+CBuJvgBgBT9gpvc+WKWJPS8fEL+Teg29BUTBPq/+Bz6Qh2S+k9qiPhZ7Gz0G5Fg9CV87v7np5bwM6zg8PKeNvnLVAL6kN9Y+47OfPqaeYL1kWkI/T71vPQZrHr5w0YS+7ZVVvqXxU76oXx4+azTmvn32KT5u4gQ9tVawvv8xoL00FEM+lbOlPMColT089TW9A2TfvmniCD2BIj4+DVMLPs0aG76o0AU/5SS/vjE107xqxPw+4OYHvuJkBMDgrv89juhPPZttQj7BzB0+2yRiPyW53D7d9NG+JfOtPtBour9oa8I+Tmt4vmHEvT62UCQ9LydMPnZmOz5XXgWAnXylvt2eBj+WiUO+vox1Ptlrkz7OEo2+2jhdPc+BOT4qdY8+G4rDPbkGxT3+K3K/wX+Zv/hMwL2iwbW+H6gQP+D3Dz7JZ+48o50UP3Ffbz74lIs+aFo3Pzx7/z6hKiC+glNfPmj5qDy9ld29DbD4Pi/P6729U5w+PeWsPsPmKr+235e+8p27vZe1Ob9c6z2+gjs3vrTJBD/nWra9/UWwvkJtjL82udA+Pn6UP2WA4L1maOQ9Imbnvhxwu74Claa/WwzSvCkphD12suE+3O8NPI7ouz5BG169r21ivQOksb23nsG+ZdtJPlb0nb886Fs+y9kjvkLbxj7nl/C/SBavAK0qhT6quJK+38gOP7cY+b6DeIW+Hc88vzDpvj2CR24+au/LPXmHi76NfPM+4wocvj2csjtfzdE8t6GvPuO4hb8BGnU9iigRPyDlyL0XbIs+l6WIvsewtT5265k+R/j/vjXJnL7VjSc/TCzFPn13yj4cyUfAI13WPtdaNz9uzlm/1JaHvhlrkr8kRxS/+uw6PRa+Mz5Jm+y+yHStPmWJWr7gKhM/TNSSPpxG8z4FqAXA9D8yveh2xjwcL+Q+eCNBwG7lqz6m7qu8r3DrPmCTxLsQYze/gCYoP4bKd78V9rM8Qb8Tvzs2W76GFxE/h+fIvpXZOryJPDO8tEcMvICBBYAbH0C7MR+funFlxrpwdV03cKcHvA135zk5NDy8QdRSvN2PzLswe5C4RprqvF1wejfMeBO7408muyQJW7vuwh26G5nCuTk3j7uoiZG7/P2WvDm5l7u+6FA6kt6CuqU6l7mx6x+8h/ZSvL84B7wkw3W715sWuUiHPLvC15a6402DuT2Ve7zv92A7x+Zluo1QJbsXFp+70EVbuYkLIrzeaBq8y4UEuQrEpreZWQS8o8N0uesYVjuAf/u7wV8CvNfg4rR/DTu8u7oFgCb3zLziZVC8UXusu2BC97uKODM5DawbvDAfq7vX9Aq7dMXcuCjHOLzmgoI+eqswPsRMn73jWgWAJkX7PpVlkr028wU/BA6av4eCrz6AhH++AFk3vsy2Yrsu4Gg+4ZzRvkCY7j2bPl0/m+Apv4H0mr/B+MO+x9ZKPxZTFz+uU7S+CviUPRynaT49Dy0+PaWSP+0qkj5eyIg+WaeJvVOrir3SP+A+d9FkPgwqsr+Copw9n9o+P01/n76W3/u+AAssPwPta76Stpi++GqNvqHbDz/eC3Y+vSsuPs6gKb9BaEU+nZsKP+mhCT6iaZA99E6lPjRFvr2cGy+/U5RhviN5prw2WRw80pVpPg+40D4Tax2829Rbv5CkhT2HCHC//7pJvo97yr6l5Qu+pd44Pply5T0vcw4/Bl4FAJUp8b2T5zK/OvnJve312z7dZ1K+B83LvMx2rbyQDb0+IfBeveqgcz6qBSM9QS/MvlDfnD5gzpU9EmKDPoZbCL4MX7S8nj0PvdaB3z4PMjA8KhhZPR5dEb+dZ1K+Sz1GPvgPMD625y69RpwjvqZpHD6zRog+8mVoPrY7Z76ZLoo+dGv9Ps48Tj20BSI+zGf0PX2iOD8skXe/xo88PXDlRj5UbYi8BlBHPg10Ur+xP5s95QmCPtP0/D3PDMk+4tyRv2d+3j0XRXo9uCQyveaw1b7NjkE9euSLPgB23j5+k4u+jBl7PqCjgD51rlE+ORqoPvRTqT45CB8+CBV7vg1bBYA30ek+n1wIv9Ber70Da+o/IEhivvgoVz49LKk+f81vPZTlUDyfYCO+UpEWPtbfVb/52mS+ZRSlPYL57D5/lgrAJq51Pd3xGT/+mgG/MU1TPhCyQL47+dK95R/nPfmaTj53u6E+6WcePa0WGr9rXaC8CE+BP0p11r2ANGi+x8x3PgryLL4iyUG/a3y0P47UMz86Kkc+jzUwv+oPFz6+QLG9i9Jbv5JggL4q1Aa/VDfrv0oegb3Sy4S+Q3EpvifXxb5CvSe++LN5vdqblr4TLQk/MJuoPoz2LD71fhY+1O/0vaB76D0IkBC+59hBP3IP97zyDbs9eFKIPoAQfr7fXQWAWwu2vQ2HNj/Bjsk+/b4sPbFRRT55T1O/9xeSPm/7Iz4hNFs9K2hNvcisDb4CpZS/2TLnPlOFuj50SvQ++boRvz3XwL0nFz88fOMGPlJr7z2M0DS+WqiRvhG2gz6ulN8+udcOP5kvB75ToRi/PwL3PdE8Mz8+UPe9SH9fv3vRhj9REMs+VffvvlVPUj7RARQ/qUR/PiNB9z7bUla9PcS/vpQdr76mPSK/6etTPrNIrLuGtI082LGRPfqYjz5dZ4U80VaJvltcHr1IGBU+8Y+qPmzeEj8jucW+a2ZTP4YYIjx4cS0/gep3v3rXFb5zsTI/qqA4i77BCpMN2XAFVQ0GAC7Y3pPvWgUAMG7gjPFjBQDEWJqK21cFgAsxoIm3UkCZId3sAZphBYAImLmmxmUFAN9YBYC3XiKG0+27jUbbJYGVWgUAQauCAq5gBQCVjRqo1GIFgNBXBQCeWwUAlJs5gZLkfosaEAUAmWIFAPJoA4B4XAUAm9QEAAdjBQAOZAUAWV8FgL8ZIJKQhUCBElsFAHejuqDGYAWAfyornlRdBYDeYgWAR1gFACAOMIEhYwUAmF8FgMkqkgG9cAIA52AFgJ2/BABWpgUA+OA0lu9z55+SVCuM5F8FAFtdBQDnXgUA22EFAKxdBYA9WwUA6jsrgcNLJ774Wfg9Z2tUv7heBYDmZpC7qBOhvasLjz8Defy+LxXpPvHtV77skuW+JbjuPgclJr6XM+w+ESVbPprvcz6+xIc+kkQVvtnYiT79tyc+Mit+PkA9QL02ATW+GA41PizLqj4KeUG+HBECv+otRj3Wn/a+cD7APswd1j5hIdq+TiMMwFm5Iz8+QNM+ZWjyPsfeij7jcs+8M3aBvykMY77LbLu9tzvGvZWYMj3aFK+8Z/SbPniYLb9ccCI/n9IDQC9vYT65/QQ9TwT2PeOAkL/QYJk8wBDfvfkbMjxnepm+35mHveUw1D75jwA/Wh60viUBur6NBb8+svcrv6oGW77WAeS+1fRmPg0gSMA5XQUA52JuPqnZy75vbg8/ZtEFPswmRD5rUFq+C+9+PEIWqz2eZgI+BFMvvo98Zz6ih0M+nnsQu8A6TL/DfIG+Nd+DPp53qD69rno+Mi2zPrquvz4lbN++ccK+PomM1b+TGHi+ltaEvnzlejyeFtE+Zaf6PHzvN78dSXU+1mGfPls/Vbwt1xC/zRoHPT7F2zrUgwE+5l30vTFNUz5CGuM9XqW/PgMHyT7kmHk+FH3kvvs6uz52FzC+8EG2vTcunT6bS70+kTWovm+bjj3rTou+MuO4vhY+sb78aiQ+lGKovZ5sdj1Xnyi/QJvhPeadrL7Xo6+9y3F/v7JmMz4e8RHAn2IFAPXjLL6lXmC+GxlBP7CXhz6G6i8+Rv4ivrC92r7KjU+8o5oiPvsZgb76GAs+rpTsPonwrj1dlv++8Ewavfaqqz5srw8+wLIhPiu0mz5CizE+Ps6DPbSJDr60wbW+bxXxvv4Ohr/MoDG+GP6Tu505Kz5q8BrA8iLwPUYlXD7TiKq74C7hPtF0972Yy3m+U86SPr9dwT70uMi+yS40vnCbcD5e8Pm82bAzvoL1kTunkmk/49BwvTKo9r1cx7i9ZEzTwDD8Qb5U4kG8IBGUPoaSh7kWk8++INv8PSc0DT9ylqq+HRyVvnkunjzlCqi+dnpxvg==","b":"3OtDPotrEz4dvuy8oYUFvjpOab6LqZCqNpoCPhaCgj5WQW29kechPltknj0Lvha9iTPXPVKWSrrCJUO8PTr4uaW7irpwOZS9wdcNvX1XHb7+pUCrzCCgPpYSVb4XD1E9wIt5vN0/gD6RvxS16b6ZsLanhj5r89akhhJkPoGfG71YYwO+LBBxPaA/DqmiQJGxN1L6PXfGN76FiP09W143tahIgD5pdcI97Qdmvn5A/D0oKte6Zi5hvRszcr5iXpA8CPydt6bPkb1GHUg96LqdPARY1LxJ6/Y7PPKePlVNx7yT9Gw+t61ZPUGKID6miUW+AqsFprqPsT7ybTs+IGBlPg==","w_shape":[64,64],"b_shape":[64]},{"type":"relu"},{"type":"linear","w":"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","b":"+VLGvE1s+74qGzS+UMSfPTfYsDtc7pY+","w_shape":[6,64],"b_shape":[6]}]},"tensors":{}}
|
nlrt.py
ADDED
|
@@ -0,0 +1,226 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""nlrt: tiny portable runtime shared by Next Lab models.
|
| 2 |
+
|
| 3 |
+
The same layer list runs in NumPy (this file) and in the browser (nlrt.js).
|
| 4 |
+
Weights are stored as base64 little-endian float32 inside model.json.
|
| 5 |
+
Only NumPy is required to load and run a model.
|
| 6 |
+
"""
|
| 7 |
+
import base64
|
| 8 |
+
import json
|
| 9 |
+
import math
|
| 10 |
+
|
| 11 |
+
import numpy as np
|
| 12 |
+
|
| 13 |
+
__all__ = ["b64e", "b64d", "forward", "load_model", "Mulberry32", "export_sequential"]
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
def b64e(a):
|
| 17 |
+
return base64.b64encode(np.ascontiguousarray(a, dtype="<f4").tobytes()).decode("ascii")
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def b64d(s, shape=None):
|
| 21 |
+
a = np.frombuffer(base64.b64decode(s), dtype="<f4").astype(np.float64)
|
| 22 |
+
return a.reshape(shape) if shape is not None else a
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
def load_model(path):
|
| 26 |
+
with open(path, "r", encoding="utf-8") as f:
|
| 27 |
+
spec = json.load(f)
|
| 28 |
+
for net in spec.get("nets", {}).values():
|
| 29 |
+
_decode(net)
|
| 30 |
+
for k, v in list(spec.get("tensors", {}).items()):
|
| 31 |
+
spec["tensors"][k] = b64d(v["data"], v["shape"])
|
| 32 |
+
return spec
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
def _decode(layers):
|
| 36 |
+
for L in layers:
|
| 37 |
+
for key in ("w", "b"):
|
| 38 |
+
if key in L and isinstance(L[key], str):
|
| 39 |
+
L[key] = b64d(L[key], L.get(key + "_shape"))
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
class Mulberry32:
|
| 43 |
+
"""Deterministic PRNG with an identical JavaScript twin in nlrt.js."""
|
| 44 |
+
|
| 45 |
+
def __init__(self, seed):
|
| 46 |
+
self.a = seed & 0xFFFFFFFF
|
| 47 |
+
|
| 48 |
+
def random(self):
|
| 49 |
+
self.a = (self.a + 0x6D2B79F5) & 0xFFFFFFFF
|
| 50 |
+
t = self.a
|
| 51 |
+
t = ((t ^ (t >> 15)) * (t | 1)) & 0xFFFFFFFF
|
| 52 |
+
t ^= (t + (((t ^ (t >> 7)) * (t | 61)) & 0xFFFFFFFF)) & 0xFFFFFFFF
|
| 53 |
+
return ((t ^ (t >> 14)) & 0xFFFFFFFF) / 4294967296.0
|
| 54 |
+
|
| 55 |
+
def uniform(self, lo, hi):
|
| 56 |
+
return lo + (hi - lo) * self.random()
|
| 57 |
+
|
| 58 |
+
def randint(self, n):
|
| 59 |
+
return int(math.floor(self.random() * n))
|
| 60 |
+
|
| 61 |
+
def normal(self):
|
| 62 |
+
u = max(self.random(), 1e-12)
|
| 63 |
+
v = self.random()
|
| 64 |
+
return math.sqrt(-2.0 * math.log(u)) * math.cos(2.0 * math.pi * v)
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
def _pair(v):
|
| 68 |
+
return list(v) if isinstance(v, (list, tuple)) else [v, v]
|
| 69 |
+
|
| 70 |
+
|
| 71 |
+
def _conv2d(x, w, b, stride, pad):
|
| 72 |
+
n, c, h, wd = x.shape
|
| 73 |
+
o, _, kh, kw = w.shape
|
| 74 |
+
sh, sw = stride
|
| 75 |
+
ph, pw = pad
|
| 76 |
+
xp = np.pad(x, ((0, 0), (0, 0), (ph, ph), (pw, pw)))
|
| 77 |
+
oh = (h + 2 * ph - kh) // sh + 1
|
| 78 |
+
ow = (wd + 2 * pw - kw) // sw + 1
|
| 79 |
+
cols = np.empty((n, c, kh, kw, oh, ow), dtype=x.dtype)
|
| 80 |
+
for i in range(kh):
|
| 81 |
+
for j in range(kw):
|
| 82 |
+
cols[:, :, i, j] = xp[:, :, i:i + sh * oh:sh, j:j + sw * ow:sw]
|
| 83 |
+
out = np.tensordot(cols, w, axes=([1, 2, 3], [1, 2, 3])) # n, oh, ow, o
|
| 84 |
+
out = out.transpose(0, 3, 1, 2)
|
| 85 |
+
if b is not None:
|
| 86 |
+
out = out + b[None, :, None, None]
|
| 87 |
+
return out
|
| 88 |
+
|
| 89 |
+
|
| 90 |
+
def _maxpool(x, k):
|
| 91 |
+
kh, kw = k
|
| 92 |
+
n, c, h, w = x.shape
|
| 93 |
+
oh, ow = h // kh, w // kw
|
| 94 |
+
x = x[:, :, :oh * kh, :ow * kw].reshape(n, c, oh, kh, ow, kw)
|
| 95 |
+
return x.max(axis=(3, 5))
|
| 96 |
+
|
| 97 |
+
|
| 98 |
+
def forward(layers, x):
|
| 99 |
+
"""Run a layer list on a batch. x: (N, D) or (N, C, H, W) or (N, C, L)."""
|
| 100 |
+
x = np.asarray(x, dtype=np.float64)
|
| 101 |
+
for L in layers:
|
| 102 |
+
t = L["type"]
|
| 103 |
+
if t == "linear":
|
| 104 |
+
x = x @ L["w"].T + L["b"]
|
| 105 |
+
elif t == "conv2d":
|
| 106 |
+
squeeze = x.ndim == 3
|
| 107 |
+
if squeeze:
|
| 108 |
+
x = x[:, :, None, :]
|
| 109 |
+
x = _conv2d(x, L["w"], L["b"], _pair(L["stride"]), _pair(L["pad"]))
|
| 110 |
+
if squeeze:
|
| 111 |
+
x = x[:, :, 0, :]
|
| 112 |
+
elif t == "relu":
|
| 113 |
+
x = np.maximum(x, 0)
|
| 114 |
+
elif t == "tanh":
|
| 115 |
+
x = np.tanh(x)
|
| 116 |
+
elif t == "sigmoid":
|
| 117 |
+
x = 1 / (1 + np.exp(-x))
|
| 118 |
+
elif t == "maxpool":
|
| 119 |
+
squeeze = x.ndim == 3
|
| 120 |
+
if squeeze:
|
| 121 |
+
x = x[:, :, None, :]
|
| 122 |
+
x = _maxpool(x, _pair(L["k"]))
|
| 123 |
+
if squeeze:
|
| 124 |
+
x = x[:, :, 0, :]
|
| 125 |
+
elif t == "gap":
|
| 126 |
+
x = x.reshape(x.shape[0], x.shape[1], -1).mean(axis=2)
|
| 127 |
+
elif t == "flatten":
|
| 128 |
+
x = x.reshape(x.shape[0], -1)
|
| 129 |
+
elif t == "upsample":
|
| 130 |
+
x = x.repeat(L["s"], axis=-1)
|
| 131 |
+
if x.ndim == 4:
|
| 132 |
+
x = x.repeat(L["s"], axis=-2)
|
| 133 |
+
elif t == "l2norm":
|
| 134 |
+
x = x / np.maximum(np.linalg.norm(x, axis=1, keepdims=True), 1e-12)
|
| 135 |
+
elif t == "reshape":
|
| 136 |
+
x = x.reshape([x.shape[0]] + list(L["shape"]))
|
| 137 |
+
else:
|
| 138 |
+
raise ValueError("unknown layer " + t)
|
| 139 |
+
return x
|
| 140 |
+
|
| 141 |
+
|
| 142 |
+
def export_sequential(seq):
|
| 143 |
+
"""Convert a torch nn.Sequential into a portable layer list (BatchNorm folded)."""
|
| 144 |
+
import torch.nn as nn
|
| 145 |
+
|
| 146 |
+
mods = list(seq.modules())[1:] if not isinstance(seq, (list, tuple)) else list(seq)
|
| 147 |
+
mods = [m for m in mods if not isinstance(m, nn.Sequential)]
|
| 148 |
+
layers = []
|
| 149 |
+
|
| 150 |
+
def tensor(t):
|
| 151 |
+
return t.detach().cpu().double().numpy()
|
| 152 |
+
|
| 153 |
+
for m in mods:
|
| 154 |
+
if isinstance(m, nn.Linear):
|
| 155 |
+
b = tensor(m.bias) if m.bias is not None else np.zeros(m.out_features)
|
| 156 |
+
layers.append({"type": "linear", "w": tensor(m.weight), "b": b})
|
| 157 |
+
elif isinstance(m, (nn.Conv2d, nn.Conv1d)):
|
| 158 |
+
w = tensor(m.weight)
|
| 159 |
+
b = tensor(m.bias) if m.bias is not None else np.zeros(w.shape[0])
|
| 160 |
+
if isinstance(m, nn.Conv1d):
|
| 161 |
+
w = w[:, :, None, :]
|
| 162 |
+
stride, pad = [1, m.stride[0]], [0, m.padding[0]]
|
| 163 |
+
else:
|
| 164 |
+
stride, pad = list(m.stride), list(m.padding)
|
| 165 |
+
layers.append({"type": "conv2d", "w": w, "b": b, "stride": stride, "pad": pad})
|
| 166 |
+
elif isinstance(m, (nn.BatchNorm1d, nn.BatchNorm2d)):
|
| 167 |
+
prev = layers[-1]
|
| 168 |
+
assert prev["type"] in ("conv2d", "linear"), "BatchNorm must follow conv/linear"
|
| 169 |
+
scale = tensor(m.weight) / np.sqrt(tensor(m.running_var) + m.eps)
|
| 170 |
+
shift = tensor(m.bias) - tensor(m.running_mean) * scale
|
| 171 |
+
if prev["type"] == "conv2d":
|
| 172 |
+
prev["w"] = prev["w"] * scale[:, None, None, None]
|
| 173 |
+
else:
|
| 174 |
+
prev["w"] = prev["w"] * scale[:, None]
|
| 175 |
+
prev["b"] = prev["b"] * scale + shift
|
| 176 |
+
elif isinstance(m, nn.ReLU):
|
| 177 |
+
layers.append({"type": "relu"})
|
| 178 |
+
elif isinstance(m, nn.Tanh):
|
| 179 |
+
layers.append({"type": "tanh"})
|
| 180 |
+
elif isinstance(m, nn.Sigmoid):
|
| 181 |
+
layers.append({"type": "sigmoid"})
|
| 182 |
+
elif isinstance(m, nn.MaxPool2d):
|
| 183 |
+
layers.append({"type": "maxpool", "k": _pair(m.kernel_size)})
|
| 184 |
+
elif isinstance(m, nn.MaxPool1d):
|
| 185 |
+
k = m.kernel_size if isinstance(m, int) else m.kernel_size
|
| 186 |
+
layers.append({"type": "maxpool", "k": [1, k if isinstance(k, int) else k[0]]})
|
| 187 |
+
elif isinstance(m, (nn.AdaptiveAvgPool2d, nn.AdaptiveAvgPool1d)):
|
| 188 |
+
layers.append({"type": "gap"})
|
| 189 |
+
elif isinstance(m, nn.Flatten):
|
| 190 |
+
layers.append({"type": "flatten"})
|
| 191 |
+
elif isinstance(m, nn.Upsample):
|
| 192 |
+
layers.append({"type": "upsample", "s": int(m.scale_factor)})
|
| 193 |
+
elif isinstance(m, nn.Dropout):
|
| 194 |
+
continue
|
| 195 |
+
elif m.__class__.__name__ == "L2Norm":
|
| 196 |
+
layers.append({"type": "l2norm"})
|
| 197 |
+
elif m.__class__.__name__ == "Reshape":
|
| 198 |
+
layers.append({"type": "reshape", "shape": list(m.shape)})
|
| 199 |
+
else:
|
| 200 |
+
raise ValueError("cannot export " + m.__class__.__name__)
|
| 201 |
+
return layers
|
| 202 |
+
|
| 203 |
+
|
| 204 |
+
def encode_layers(layers):
|
| 205 |
+
out = []
|
| 206 |
+
for L in layers:
|
| 207 |
+
L2 = dict(L)
|
| 208 |
+
for key in ("w", "b"):
|
| 209 |
+
if key in L2:
|
| 210 |
+
arr = np.asarray(L2[key])
|
| 211 |
+
L2[key + "_shape"] = list(arr.shape)
|
| 212 |
+
L2[key] = b64e(arr)
|
| 213 |
+
out.append(L2)
|
| 214 |
+
return out
|
| 215 |
+
|
| 216 |
+
|
| 217 |
+
def save_model(path, nets=None, tensors=None, meta=None):
|
| 218 |
+
spec = {"format": "nlrt-1", "meta": meta or {}, "nets": {}, "tensors": {}}
|
| 219 |
+
for name, layers in (nets or {}).items():
|
| 220 |
+
spec["nets"][name] = encode_layers(layers)
|
| 221 |
+
for name, arr in (tensors or {}).items():
|
| 222 |
+
arr = np.asarray(arr)
|
| 223 |
+
spec["tensors"][name] = {"shape": list(arr.shape), "data": b64e(arr)}
|
| 224 |
+
with open(path, "w", encoding="utf-8") as f:
|
| 225 |
+
json.dump(spec, f, separators=(",", ":"))
|
| 226 |
+
return spec
|
policy.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f9622f738b576a4511e84f93510a3b5042aa7719f82a8e49178a333f6239a240
|
| 3 |
+
size 24248
|
preview.png
ADDED
|
replay.mp4
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2a1ac18bbb78c61d07c6013e9ce42285c0aefc9181d5fb3c7211468e9d30fa09
|
| 3 |
+
size 153163
|
selftest.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"cases": [{"input": {"seed": 10007, "kind": "mixed", "density": 2, "driver": "shield"}, "expected": {"collision": false, "red_violations": 0, "reached": true, "hard_brakes": 3, "overrides": 39, "steps": 344}}, {"input": {"seed": 10007, "kind": "mixed", "density": 2, "driver": "learned"}, "expected": {"collision": false, "red_violations": 0, "reached": true, "hard_brakes": 0, "overrides": 0, "steps": 189}}, {"input": {"seed": 12345, "kind": "occlusion", "density": 3, "driver": "shield"}, "expected": {"collision": false, "red_violations": 0, "reached": true, "hard_brakes": 0, "overrides": 0, "steps": 312}}, {"input": {"seed": 10111, "kind": "lead_brake", "density": 1, "driver": "learned"}, "expected": {"collision": false, "red_violations": 0, "reached": true, "hard_brakes": 2, "overrides": 0, "steps": 149}}, {"input": {"seed": 10222, "kind": "crosswalk", "density": 2, "driver": "rule"}, "expected": {"collision": false, "red_violations": 0, "reached": true, "hard_brakes": 4, "overrides": 0, "steps": 398}}, {"input": {"seed": 10333, "kind": "light", "density": 3, "driver": "expert"}, "expected": {"collision": false, "red_violations": 0, "reached": true, "hard_brakes": 7, "overrides": 0, "steps": 244}}, {"input": {"seed": 10444, "kind": "occlusion", "density": 2, "driver": "rule"}, "expected": {"collision": true, "red_violations": 0, "reached": false, "hard_brakes": 1, "overrides": 0, "steps": 41}}, {"input": {"seed": 10555, "kind": "mixed", "density": 3, "driver": "shield"}, "expected": {"collision": false, "red_violations": 0, "reached": true, "hard_brakes": 0, "overrides": 0, "steps": 234}}]}
|