"""
Full-episode GT-anchored replay using a finetuned Ctrl-World on bimanual/multiview.
Reads from the SAME prepared training dataset dir (dataset_example/bimanual_multiview)
whose annotations already contain: 3-view resized videos, `states` (14-D qpos,
frame-aligned), `video_length`, `texts`. No separate inference prep needed.
For each episode:
1. interact_num computed from episode length to cover the full clip.
2. Replay GT 14-D qpos as action condition (no policy).
3. Autoregressive generation with Ctrl-World.
4. Save predicted video + PSNR/SSIM/LPIPS per episode.
Mirrors scripts/infer_single_arm_multiview_ctrlworld.py but for bimanual:
- 3 views: cam_high, cam_left_wrist, cam_right_wrist
- 14-D state, bimanual stat.json
- down_sample consistent with training (native, since arrays are frame-aligned)
"""
import sys
import os
import json
import time
import importlib
from pathlib import Path
from argparse import ArgumentParser
import numpy as np
import torch
import einops
import mediapy
import piq
ctrl_world_dir = os.path.join(
os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
"models", "Ctrl-World"
)
sys.path.insert(0, ctrl_world_dir)
from models.pipeline_ctrl_world import CtrlWorldDiffusionPipeline
from models.ctrl_world import CrtlWorld
from decord import VideoReader, cpu
from accelerate import Accelerator
class CtrlWorldAgent:
def __init__(self, args):
args.val_model_path = args.ckpt_path
self.args = args
self.accelerator = Accelerator()
self.device = self.accelerator.device
self.dtype = args.dtype
self.model = CrtlWorld(args)
state_dict = torch.load(args.val_model_path, map_location="cpu")
model_sd = self.model.state_dict()
filtered = {k: v for k, v in state_dict.items()
if k in model_sd and model_sd[k].shape == v.shape}
dropped = [k for k in state_dict if k not in filtered]
missing, unexpected = self.model.load_state_dict(filtered, strict=False)
if dropped or missing:
print(f"[load] dropped {len(dropped)} shape-mismatch tensors "
f"(e.g. {dropped[:2]}); missing {len(missing)} (reinit). "
f"A finetuned 14-D bimanual checkpoint should load with 0 dropped.")
self.model.to(self.device).to(self.dtype)
self.model.eval()
print("Ctrl-World (bimanual) model loaded")
with open(args.data_stat_path, "r") as f:
data_stat = json.load(f)
self.state_p01 = np.array(data_stat["state_01"])[None, :]
self.state_p99 = np.array(data_stat["state_99"])[None, :]
# View layout. Default = 3-view vertical strip (3 rows x 1 col).
# Grid variants set num_views/grid_rows/grid_cols on the config.
self.num_views = int(getattr(args, "num_views", 3))
self.grid_rows = int(getattr(args, "grid_rows", self.num_views))
self.grid_cols = int(getattr(args, "grid_cols", 1))
def stack_views_to_canvas(self, per_view_latents):
"""Place a list of per-view latents (each (..., h, w)) ROW-MAJOR into one
(..., grid_rows*h, grid_cols*w) canvas. For the 3-view default this is a
pure vertical concat; for the 2x2 grid it is a proper grid layout."""
lat_h, lat_w = per_view_latents[0].shape[-2:]
canvas = torch.zeros(
(*per_view_latents[0].shape[:-2], self.grid_rows * lat_h, self.grid_cols * lat_w),
dtype=per_view_latents[0].dtype, device=per_view_latents[0].device,
)
for v_i, v in enumerate(per_view_latents):
r = v_i // self.grid_cols
c = v_i % self.grid_cols
canvas[..., r * lat_h:(r + 1) * lat_h, c * lat_w:(c + 1) * lat_w] = v
return canvas
def normalize_bound(self, data, data_min, data_max, clip_min=-1, clip_max=1, eps=1e-8):
ndata = 2 * (data - data_min) / (data_max - data_min + eps) - 1
return np.clip(ndata, clip_min, clip_max)
def get_traj_info(self, episode_id, start_idx=0, steps=8):
val_dataset_dir = self.args.val_dataset_dir
annotation_path = f"{val_dataset_dir}/annotation/val/{episode_id}.json"
if not os.path.exists(annotation_path):
annotation_path = f"{val_dataset_dir}/annotation/train/{episode_id}.json"
with open(annotation_path) as f:
anno = json.load(f)
length = anno["video_length"]
frames_ids = np.arange(start_idx, start_idx + steps)
max_ids = np.ones_like(frames_ids) * (length - 1)
frames_ids = np.min([frames_ids, max_ids], axis=0).astype(int)
instruction = anno["texts"][0]
qpos_action = np.array(anno["states"]) # 14-D, frame-aligned
qpos_action = qpos_action[frames_ids]
video_latent = []
video_dict = []
for vid_info in anno["videos"]:
video_path = f"{val_dataset_dir}/{vid_info['video_path']}"
vr = VideoReader(video_path, ctx=cpu(0), num_threads=2)
actual_video_len = len(vr)
if length > actual_video_len:
length = actual_video_len
frames_ids = np.clip(frames_ids, 0, length - 1)
try:
true_video = vr.get_batch(range(length)).asnumpy()
except Exception:
true_video = vr.get_batch(range(length)).numpy()
true_video = true_video[frames_ids]
video_dict.append(true_video)
device = self.device
true_video_t = torch.from_numpy(true_video).to(self.dtype).to(device)
x = true_video_t.permute(0, 3, 1, 2) / 255.0 * 2 - 1
vae = self.model.pipeline.vae
with torch.no_grad():
latents = []
for i in range(0, len(x), 32):
batch = x[i:i + 32]
latent = vae.encode(batch).latent_dist.sample().mul_(vae.config.scaling_factor)
latents.append(latent)
x = torch.cat(latents, dim=0)
video_latent.append(x)
return qpos_action, video_dict, video_latent, instruction
def forward_wm(self, action_cond, video_latent_true, video_latent_cond, his_cond=None, text=None):
args = self.args
image_cond = video_latent_cond
action_cond = self.normalize_bound(action_cond, self.state_p01, self.state_p99)
action_cond = torch.tensor(action_cond).unsqueeze(0).to(self.device).to(self.dtype)
with torch.no_grad():
if text is not None:
text_token = self.model.action_encoder(
action_cond, text, self.model.tokenizer, self.model.text_encoder
)
else:
text_token = self.model.action_encoder(action_cond)
pipeline = self.model.pipeline
_, latents = CtrlWorldDiffusionPipeline.__call__(
pipeline,
image=image_cond,
text=text_token,
width=int(args.width * self.grid_cols),
height=int(args.height * self.grid_rows),
num_frames=args.num_frames,
history=his_cond,
num_inference_steps=args.num_inference_steps,
decode_chunk_size=args.decode_chunk_size,
max_guidance_scale=args.guidance_scale,
fps=args.fps,
motion_bucket_id=args.motion_bucket_id,
mask=None,
output_type="latent",
return_dict=False,
frame_level_cond=True,
)
latents = einops.rearrange(latents, "b f c (m h) (n w) -> (b m n) f c h w", m=self.grid_rows, n=self.grid_cols)
true_video = torch.stack(video_latent_true, dim=0)
decoded_video = []
bsz, frame_num = true_video.shape[:2]
true_video_flat = true_video.flatten(0, 1)
for i in range(0, true_video_flat.shape[0], args.decode_chunk_size):
chunk = true_video_flat[i:i + args.decode_chunk_size] / pipeline.vae.config.scaling_factor
decoded_video.append(pipeline.vae.decode(chunk, num_frames=chunk.shape[0]).sample)
true_video_dec = torch.cat(decoded_video, dim=0)
true_video_dec = true_video_dec.reshape(bsz, frame_num, *true_video_dec.shape[1:])
true_video_dec = ((true_video_dec / 2.0 + 0.5).clamp(0, 1) * 255)
true_video_dec = true_video_dec.detach().to(torch.float32).cpu().numpy().transpose(0, 1, 3, 4, 2).astype(np.uint8)
decoded_video = []
bsz, frame_num = latents.shape[:2]
x = latents.flatten(0, 1)
for i in range(0, x.shape[0], args.decode_chunk_size):
chunk = x[i:i + args.decode_chunk_size] / pipeline.vae.config.scaling_factor
decoded_video.append(pipeline.vae.decode(chunk, num_frames=chunk.shape[0]).sample)
videos = torch.cat(decoded_video, dim=0)
videos = videos.reshape(bsz, frame_num, *videos.shape[1:])
videos = ((videos / 2.0 + 0.5).clamp(0, 1) * 255)
videos = videos.detach().to(torch.float32).cpu().numpy().transpose(0, 1, 3, 4, 2).astype(np.uint8)
return true_video_dec, videos, latents
def compute_metrics(pred_frames, gt_frames):
x = torch.clamp(torch.from_numpy(pred_frames.copy()) / 255.0, 0, 1).permute(0, 3, 1, 2)
y = torch.clamp(torch.from_numpy(gt_frames.copy()) / 255.0, 0, 1).permute(0, 3, 1, 2)
psnr_val = piq.psnr(x, y).mean().item()
ssim_val = piq.ssim(x, y).mean().item()
lpips_val = piq.LPIPS()(x, y).mean().item()
return {"psnr": psnr_val, "ssim": ssim_val, "lpips": lpips_val}
def find_annotation(val_dataset_dir, episode_id):
for split in ["val", "train"]:
p = f"{val_dataset_dir}/annotation/{split}/{episode_id}.json"
if os.path.exists(p):
return p, split
return None, None
def split_category(episode_id):
"""Split a `{category}_{task}__{idx}` annotation stem into (category, name).
Bimanual/humanoid annotation stems embed the makovian/non_makovian category
(e.g. `non_makovian_fold_blue_towel__000070`). Per the sampling-dataset-layout
rule, `{category}` must be its own path segment, not baked into the episode
dir name. Returns (category, stripped_episode_name). If no known category
prefix is found, category is None and the name is returned unchanged.
"""
if episode_id.startswith("non_makovian_"):
return "non_makovian", episode_id[len("non_makovian_"):]
if episode_id.startswith("makovian_"):
return "makovian", episode_id[len("makovian_"):]
return None, episode_id
def episode_output_dir(base_dir, episode_id):
"""Path //episode_ (falls back to flat if no category)."""
category, name = split_category(episode_id)
base = Path(base_dir)
if category is not None:
base = base / category
return base / f"episode_{name}"
def run_episode(agent, episode_id, save_dir, input_save_dir=None):
args = agent.args
pred_step = args.num_frames # 5
num_history = args.num_history # 6
anno_path, split = find_annotation(args.val_dataset_dir, episode_id)
if anno_path is None:
print(f" no annotation for {episode_id}")
return None
with open(anno_path) as f:
anno = json.load(f)
episode_length = anno["video_length"]
interact_num = (episode_length - 1) // (pred_step - 1)
if interact_num < 1:
print(f" Episode {episode_id} too short ({episode_length} frames), skipping")
return None
total_steps_needed = (interact_num - 1) * (pred_step - 1) + pred_step
eef_gt, video_dict, video_latents, instruction = agent.get_traj_info(
episode_id, start_idx=0, steps=min(total_steps_needed, episode_length)
)
his_cond = []
his_eef = []
first_latent = agent.stack_views_to_canvas([v[0] for v in video_latents]).unsqueeze(0)
for _ in range(num_history * 4):
his_cond.append(first_latent)
his_eef.append(eef_gt[0:1])
video_to_save_pred = []
video_to_save_gt = []
history_idx = [0, 0, -8, -6, -4, -2]
for i in range(interact_num):
start_id = int(i * (pred_step - 1))
end_id = start_id + pred_step
if end_id > len(eef_gt):
break
video_latent_true = [v[start_id:end_id] for v in video_latents]
qpos_pose = eef_gt[start_id:end_id]
his_pose = np.concatenate([his_eef[idx] for idx in history_idx], axis=0)
action_cond = np.concatenate([his_pose, qpos_pose], axis=0)
his_cond_input = torch.cat([his_cond[idx] for idx in history_idx], dim=0).unsqueeze(0)
current_latent = his_cond[-1]
true_videos, pred_videos, predicted_latents = agent.forward_wm(
action_cond, video_latent_true, current_latent,
his_cond=his_cond_input,
text=instruction if args.text_cond else None,
)
his_eef.append(qpos_pose[pred_step - 1:pred_step])
his_cond.append(
agent.stack_views_to_canvas([v[pred_step - 1] for v in predicted_latents]).unsqueeze(0)
)
if i == interact_num - 1:
video_to_save_pred.append(pred_videos)
video_to_save_gt.append(true_videos)
else:
video_to_save_pred.append(pred_videos[:, :pred_step - 1])
video_to_save_gt.append(true_videos[:, :pred_step - 1])
if (i + 1) % 20 == 0:
print(f" Step {i+1}/{interact_num}")
if not video_to_save_pred:
return None
concat_pred = np.concatenate(video_to_save_pred, axis=1)
concat_gt = np.concatenate(video_to_save_gt, axis=1)
num_views = concat_pred.shape[0]
min_len = min(concat_pred.shape[1], concat_gt.shape[1])
concat_pred = concat_pred[:, :min_len]
concat_gt = concat_gt[:, :min_len]
pred_strip = np.concatenate([concat_pred[v] for v in range(num_views)], axis=2)
gt_strip = np.concatenate([concat_gt[v] for v in range(num_views)], axis=2)
full_pred = np.concatenate([gt_strip, pred_strip], axis=1)
out_fps = max(1, int(round(30 / args.infer_down_sample)))
ep_save_dir = episode_output_dir(save_dir, episode_id)
ep_save_dir.mkdir(parents=True, exist_ok=True)
mediapy.write_video(str(ep_save_dir / "full_pred.mp4"), full_pred, fps=out_fps)
mediapy.write_video(str(ep_save_dir / "pred_all_views.mp4"), pred_strip, fps=out_fps)
mediapy.write_video(str(ep_save_dir / "gt_all_views.mp4"), gt_strip, fps=out_fps)
if input_save_dir is not None:
ep_input_dir = episode_output_dir(input_save_dir, episode_id)
ep_input_dir.mkdir(parents=True, exist_ok=True)
mediapy.write_video(str(ep_input_dir / "full_gt.mp4"), gt_strip, fps=out_fps)
for vid_idx, vid_info in enumerate(anno["videos"]):
src_video = Path(args.val_dataset_dir) / vid_info["video_path"]
dst_video = ep_input_dir / f"view_{vid_idx}.mp4"
if src_video.exists() and not dst_video.exists():
import shutil
shutil.copy2(str(src_video), str(dst_video))
input_meta = {
"episode_id": episode_id,
"instruction": instruction,
"num_frames": min_len,
"episode_length_original": episode_length,
"fps": out_fps,
"interact_num": interact_num,
"pred_step": pred_step,
"mode": "replay",
"states": anno["states"][:min_len],
}
with open(ep_input_dir / "metadata.json", "w") as f:
json.dump(input_meta, f, indent=2)
# Prefer the real camera keys recorded at prep time (humanoid names vary per
# task); fall back to the bimanual defaults for the 3/4-view layouts.
view_names = anno.get("view_keys")
if not view_names or len(view_names) != num_views:
if getattr(agent, "num_views", 3) == 4:
# 2x2 grid: row-major [TL, TR, BL, BR]
view_names = ["cam_high", "cam_low", "cam_left_wrist", "cam_right_wrist"]
else:
view_names = ["cam_high", "cam_left_wrist", "cam_right_wrist"]
per_view_metrics = {}
for v_i in range(num_views):
per_view_metrics[view_names[v_i]] = compute_metrics(concat_pred[v_i], concat_gt[v_i])
metrics = {
"psnr": float(np.mean([m["psnr"] for m in per_view_metrics.values()])),
"ssim": float(np.mean([m["ssim"] for m in per_view_metrics.values()])),
"lpips": float(np.mean([m["lpips"] for m in per_view_metrics.values()])),
"per_view": per_view_metrics,
"episode_id": episode_id,
"instruction": instruction,
"num_frames_pred": min_len,
"num_frames_gt": episode_length,
"interact_num": interact_num,
"mode": "replay",
}
with open(ep_save_dir / "metrics.json", "w") as f:
json.dump(metrics, f, indent=2)
return metrics
def _import_cache_module(backend, module_name):
project_root = os.path.normpath(os.path.join(os.path.dirname(__file__), ".."))
if project_root not in sys.path:
sys.path.insert(0, project_root)
return importlib.import_module(f"methods.cache_strategy.{backend}.{module_name}")
def _import_pruning_module(backend, module_name):
project_root = os.path.normpath(os.path.join(os.path.dirname(__file__), ".."))
if project_root not in sys.path:
sys.path.insert(0, project_root)
return importlib.import_module(f"methods.prunning.{backend}.{module_name}")
def _enable_backend(agent, args, num_inference_steps):
"""Enable the (single, validated) acceleration backend on agent.model.unet.
Mirrors scripts/infer_single_arm_multiview_ctrlworld.py so bimanual/multiview
supports the exact same WorldCache / DiCache / FasterCache / SiTo / ITM flags
(same CrtlWorld UNet). Returns a possibly-adjusted guidance_scale.
"""
guidance_scale = agent.args.guidance_scale
unet = agent.model.unet
if getattr(args, "use_worldcache", False):
adapter = _import_cache_module("WorldCache", "adapter")
adapter.enable_worldcache(
unet,
num_steps=num_inference_steps,
rel_l1_thresh=args.worldcache_rel_l1_thresh,
ret_ratio=args.worldcache_ret_ratio,
probe_depth=args.worldcache_probe_depth,
motion_sensitivity=args.worldcache_motion_sensitivity,
hf_enabled=args.worldcache_hf_enabled,
hf_thresh=args.worldcache_hf_thresh,
saliency_enabled=args.worldcache_saliency_enabled,
saliency_weight=args.worldcache_saliency_weight,
osi_enabled=args.worldcache_osi_enabled,
dynamic_decay=args.worldcache_dynamic_decay,
)
print(f"[Cache] WorldCache enabled: thresh={args.worldcache_rel_l1_thresh}, "
f"ret_ratio={args.worldcache_ret_ratio}, probe_depth={args.worldcache_probe_depth}")
if getattr(args, "use_dicache", False):
adapter = _import_cache_module("DiCache", "adapter")
adapter.enable_dicache(
unet,
num_steps=num_inference_steps,
rel_l1_thresh=args.dicache_rel_l1_thresh,
ret_ratio=args.dicache_ret_ratio,
probe_depth=args.dicache_probe_depth,
)
print(f"[Cache] DiCache enabled: thresh={args.dicache_rel_l1_thresh}, "
f"ret_ratio={args.dicache_ret_ratio}, probe_depth={args.dicache_probe_depth}")
if getattr(args, "use_fastercache", False):
adapter = _import_cache_module("FasterCache", "adapter")
adapter.enable_fastercache(
unet,
start_step=args.fastercache_start_step,
model_interval=args.fastercache_model_interval,
block_interval=args.fastercache_block_interval,
first_layers_fp=2,
)
print(f"[Cache] FasterCache enabled: start_step={args.fastercache_start_step}, "
f"model_interval={args.fastercache_model_interval}, block_interval={args.fastercache_block_interval}")
if getattr(args, "use_sito", False):
if getattr(args, "sito_spatiotemporal_hold", False):
st_hold = _import_pruning_module("SiTo", "spatiotemporal_hold")
st_hold.enable_spatiotemporal_hold(
unet,
keep_ratio=args.sito_st_keep_ratio,
max_downsample_ratio=args.sito_max_downsample_ratio,
recompute_every=(args.sito_plan_recompute_every or 999999),
)
print(f"[Prune] SiTo ST-hold enabled: keep_ratio={args.sito_st_keep_ratio}, "
f"max_downsample_ratio={args.sito_max_downsample_ratio}")
else:
adapter = _import_pruning_module("SiTo", "adapter")
adapter.enable_sito(
unet,
start_layer_idx=args.sito_start_layer_idx or 0,
prune_ratio=args.sito_prune_ratio,
patch_h=args.sito_patch_h,
patch_w=args.sito_patch_w,
noise_alpha=args.sito_noise_alpha,
sim_beta=args.sito_sim_beta,
max_downsample_ratio=args.sito_max_downsample_ratio,
plan_recompute_every=getattr(args, "sito_plan_recompute_every", 0),
)
print(f"[Prune] SiTo enabled: prune_ratio={args.sito_prune_ratio}")
if getattr(args, "use_itm", False):
itm_state = {}
if getattr(args, "itm_spatiotemporal_hold", False):
st_hold = _import_pruning_module("SiTo", "spatiotemporal_hold")
st_hold.enable_spatiotemporal_hold(
unet,
keep_ratio=args.itm_st_keep_ratio,
max_downsample_ratio=args.itm_max_downsample_ratio,
recompute_every=(args.itm_plan_recompute_every or 999999),
similarity_recover=True,
)
print(f"[Prune] ITM ST-hold enabled: keep_ratio={args.itm_st_keep_ratio}, "
f"max_downsample_ratio={args.itm_max_downsample_ratio}")
elif getattr(args, "itm_block_hold", False):
block_hold = _import_pruning_module("importance_token_merge", "block_hold")
itm_state["prune_from_step"] = args.itm_prune_from_step
itm_state["merge_from_step"] = args.itm_merge_from_step
block_hold.enable_itm_block_hold(
unet,
itm_state=itm_state,
start_layer_idx=args.itm_start_layer_idx or 0,
compress_ratio=args.itm_compress_ratio,
max_downsample_ratio=args.itm_max_downsample_ratio,
self_importance=True,
plan_recompute_every=getattr(args, "itm_plan_recompute_every", 0),
)
else:
adapter = _import_pruning_module("importance_token_merge", "adapter")
adapter.enable_itm(
unet,
itm_state=itm_state,
start_layer_idx=args.itm_start_layer_idx or 0,
compress_ratio=args.itm_compress_ratio,
prune_from_step=args.itm_prune_from_step,
merge_from_step=args.itm_merge_from_step,
merge_attn=args.itm_merge_attn,
merge_crossattn=args.itm_merge_crossattn,
merge_mlp=args.itm_merge_mlp,
max_downsample_ratio=args.itm_max_downsample_ratio,
)
if guidance_scale <= 1.0:
guidance_scale = 2.0
print(f"[ITM] guidance_scale overridden to {guidance_scale} (ITM requires > 1.0)")
return guidance_scale
def main():
parser = ArgumentParser()
parser.add_argument("--svd_model_path", type=str, required=True)
parser.add_argument("--clip_model_path", type=str, required=True)
parser.add_argument("--ckpt_path", type=str, required=True)
parser.add_argument("--dataset_dir", type=str, required=True,
help="Prepared bimanual_multiview dir (has annotation/{train,val}, videos/).")
parser.add_argument("--dataset_meta_info_path", type=str,
default="./models/Ctrl-World/dataset_meta_info")
parser.add_argument("--dataset_name", type=str, default="bimanual_multiview")
parser.add_argument("--save_dir", type=str, required=True)
parser.add_argument("--input_save_dir", type=str, default=None)
parser.add_argument("--num_episodes", type=int, default=None)
parser.add_argument("--num_inference_steps", type=int, default=None)
parser.add_argument("--guidance_scale", type=float, default=None,
help="CFG scale for action conditioning. >1.0 amplifies "
"action-following (config default 1.0 = CFG off).")
parser.add_argument("--infer_down_sample", type=int, default=1,
help="Must match the down_sample used at prep/train (for output fps).")
parser.add_argument("--split", choices=["val", "train", "all"], default="all")
parser.add_argument("--num_shards", type=int, default=1,
help="Split the episode list into N shards for parallel multi-GPU runs.")
parser.add_argument("--shard_id", type=int, default=0,
help="Which shard (0-based) this process handles. Requires --num_shards.")
parser.add_argument("--config", choices=["bimanual", "bimanual_grid", "humanoid", "humanoid_grid"],
default="bimanual",
help="Which model config (sets action_dim, view grid, etc.).")
# Cache backend args (mutually exclusive)
_import_cache_module("WorldCache", "config").add_worldcache_args(parser)
_import_cache_module("DiCache", "config").add_dicache_args(parser)
_import_cache_module("FasterCache", "config").add_fastercache_args(parser)
# Pruning/sparse attention backend args
_import_pruning_module("SiTo", "config").add_sito_args(parser)
_import_pruning_module("importance_token_merge", "config").add_itm_args(parser)
args = parser.parse_args()
from methods.cache_strategy.ctrl_world_utils import validate_ctrl_world_backend_args
validate_ctrl_world_backend_args(args)
sys.path.insert(0, ctrl_world_dir)
if args.config == "humanoid":
from config_humanoid import wm_args_humanoid as _wm_args
elif args.config == "humanoid_grid":
from config_humanoid_grid import wm_args_humanoid_grid as _wm_args
elif args.config == "bimanual_grid":
from config_bimanual_grid import wm_args_bimanual_grid as _wm_args
else:
from config_bimanual import wm_args_bimanual as _wm_args
model_args = _wm_args()
model_args.svd_model_path = args.svd_model_path
model_args.clip_model_path = args.clip_model_path
model_args.ckpt_path = args.ckpt_path
model_args.val_model_path = args.ckpt_path
model_args.val_dataset_dir = args.dataset_dir
model_args.dataset_meta_info_path = args.dataset_meta_info_path
model_args.data_stat_path = os.path.join(args.dataset_meta_info_path, args.dataset_name, "stat.json")
model_args.infer_down_sample = args.infer_down_sample
if args.num_inference_steps is not None:
model_args.num_inference_steps = args.num_inference_steps
if args.guidance_scale is not None:
model_args.guidance_scale = args.guidance_scale
print(f"[override] guidance_scale = {args.guidance_scale}")
agent = CtrlWorldAgent(model_args)
# Enable acceleration backend (validated mutually-exclusive) on the UNet.
guidance_scale = _enable_backend(agent, args, model_args.num_inference_steps)
if guidance_scale != model_args.guidance_scale:
model_args.guidance_scale = guidance_scale
splits = ["val", "train"] if args.split == "all" else [args.split]
episode_ids = []
for split in splits:
anno_dir = Path(args.dataset_dir) / "annotation" / split
if anno_dir.exists():
episode_ids += sorted([f.stem for f in anno_dir.glob("*.json")])
episode_ids = sorted(set(episode_ids))
if args.num_episodes:
episode_ids = episode_ids[:args.num_episodes]
if args.num_shards > 1:
# Interleaved sharding keeps each shard's episode-length mix balanced.
total_eps = len(episode_ids)
episode_ids = episode_ids[args.shard_id::args.num_shards]
print(f"[shard {args.shard_id}/{args.num_shards}] {len(episode_ids)}/{total_eps} episodes")
print(f"Dataset dir: {args.dataset_dir}")
print(f"Output dir: {args.save_dir}")
print(f"Episodes to process: {len(episode_ids)}")
all_metrics = []
for ep_idx, episode_id in enumerate(episode_ids):
ep_save_path = episode_output_dir(args.save_dir, episode_id) / "full_pred.mp4"
if ep_save_path.exists():
print(f"[{ep_idx+1}/{len(episode_ids)}] {episode_id} done, skipping")
continue
print(f"[{ep_idx+1}/{len(episode_ids)}] Episode {episode_id}")
_t0 = time.time()
metrics = run_episode(agent, episode_id, args.save_dir, input_save_dir=args.input_save_dir)
_dt = time.time() - _t0
if metrics:
metrics["wall_time_s"] = _dt
all_metrics.append(metrics)
print(f" -> {metrics['num_frames_pred']} frames, "
f"PSNR={metrics['psnr']:.2f}, SSIM={metrics['ssim']:.4f}, LPIPS={metrics['lpips']:.4f} "
f"| {_dt:.1f}s")
if all_metrics:
def _write_summary(metrics_list, out_dir, label):
summary = {
"mean_psnr": sum(m["psnr"] for m in metrics_list) / len(metrics_list),
"mean_ssim": sum(m["ssim"] for m in metrics_list) / len(metrics_list),
"mean_lpips": sum(m["lpips"] for m in metrics_list) / len(metrics_list),
"num_episodes": len(metrics_list),
}
_times = [m["wall_time_s"] for m in metrics_list if "wall_time_s" in m]
if _times:
summary["mean_wall_time_s"] = sum(_times) / len(_times)
out_dir.mkdir(parents=True, exist_ok=True)
# Sharded runs each write their own file so they don't clobber each
# other; merge_ctrlworld_shard_summaries.py aggregates them after.
fname = ("all_summary.json" if args.num_shards <= 1
else f"all_summary.shard{args.shard_id}of{args.num_shards}.json")
with open(out_dir / fname, "w") as f:
json.dump(summary, f, indent=2)
print(f"=== Summary [{label}] ({len(metrics_list)} episodes) === "
f"PSNR: {summary['mean_psnr']:.3f} "
f"SSIM: {summary['mean_ssim']:.4f} "
f"LPIPS: {summary['mean_lpips']:.4f}"
+ (f" time: {summary['mean_wall_time_s']:.1f}s" if "mean_wall_time_s" in summary else ""))
return summary
# Per-category summaries (mirrors single_arm: one all_summary.json per
# {category}/ dir) so bimanual/humanoid match the layout convention.
by_category = {}
for m in all_metrics:
category, _ = split_category(m["episode_id"])
by_category.setdefault(category, []).append(m)
print()
for category, metrics_list in by_category.items():
out_dir = Path(args.save_dir)
if category is not None:
out_dir = out_dir / category
_write_summary(metrics_list, out_dir, category or "all")
if __name__ == "__main__":
main()