Download code/adaspe_train/tests/smoke_model.py from cognaclee/Frank-Insert-Wipe: direct link, hf CLI and curl.
- Browser
- Download file 2.69 kB
-
https://huggingface.co/cognaclee/Frank-Insert-Wipe/resolve/main/code/adaspe_train/tests/smoke_model.py
- Command line
-
hf download hf://cognaclee/Frank-Insert-Wipe/code/adaspe_train/tests/smoke_model.py
-
curl -L -o smoke_model.py https://huggingface.co/cognaclee/Frank-Insert-Wipe/resolve/main/code/adaspe_train/tests/smoke_model.py
2.69 kB
| import os, sys; os.environ.setdefault("JAX_PLATFORMS", "cpu") | |
| sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) # adaspe_train/ | |
| import jax, jax.numpy as jnp, flax.nnx as nnx | |
| import tasks | |
| from adaspe.model import AdaSpePi05Config | |
| H = tasks.ACTION_HORIZON | |
| for generator, stage in [(g, s) for g in ["flow_head", "pi05_expert"] for s in ["action", "timing"]]: | |
| print("== generator", generator) | |
| cfg = AdaSpePi05Config(paligemma_variant="dummy", action_expert_variant="dummy", generator=generator, action_horizon=H, num_control_points=tasks.NUM_CONTROL_POINTS, | |
| min_query_interval=tasks.MIN_QUERY_INTERVAL, max_query_interval=tasks.MAX_QUERY_INTERVAL, | |
| stage=stage, flow_hidden_dim=48, flow_num_layers=1, timing_label_mode=tasks.LABEL_MODE, | |
| knot_label_mode=tasks.LABEL_MODE, knot_motion_norm=tasks.KNOT_MOTION_NORM, | |
| timing_motion_scale=(0.02,) * 6, e_crit=10.0, | |
| **(dict(num_ode_steps=10, timing_num_ode_steps=10) if generator == "pi05_expert" else {})) | |
| model = cfg.create(jax.random.key(0)) | |
| obs, act = cfg.fake_obs(2), cfg.fake_act(2) | |
| grip = jnp.array(([1.] * 3 + [-1.] * 4 + [1.] * 3 + [0.] * 2 + [1.] * H)[:H]) # open/close/open + "no command" zeros | |
| act = act.at[..., 6].set(grip) | |
| loss = model.compute_loss(jax.random.key(1), obs, act, train=True) | |
| print(stage, "H", H, "K", cfg.num_control_points, "loss", loss.shape, loss) | |
| assert bool(jnp.all(jnp.isfinite(loss))) | |
| filt = nnx.All(nnx.Param, nnx.Not(cfg.get_freeze_filter())) | |
| graphdef, params, rest = nnx.split(model, filt, ...) | |
| def f(p): | |
| m = nnx.merge(graphdef, p, rest); return m.compute_loss(jax.random.key(1), obs, act, train=True).mean() | |
| g = jax.grad(f)(params) | |
| flat = {"/".join(map(str, k)): float(jnp.abs(v.value).sum()) for k, v in nnx.to_flat_state(g)} if hasattr(nnx, "to_flat_state") else {"/".join(map(str,k)): float(jnp.abs(v.value).sum()) for k, v in g.flat_state().items()} | |
| nz = sorted({k.split("/")[0] for k, v in flat.items() if v > 0}); print(stage, "params with grad:", len(flat), "top-level nonzero:", nz) | |
| if stage == "action": | |
| sp = model.sample_spline(jax.random.key(2), obs) | |
| print("sample", model.sample_actions(jax.random.key(2), obs).shape, {k: v.shape for k, v in sp.items()}, | |
| "query_interval", sp["query_interval"][:, 0], "execute_steps", sp["execute_steps"]) | |
| assert sp["control_points"].shape[1] == cfg.num_control_points and sp["dense_actions"].shape[1] == H | |
| assert bool(jnp.all((sp["execute_steps"] >= 1) & (sp["execute_steps"] <= H))) | |