| """StatePlay action utilities — profile-driven parquet -> tensor pipeline. |
| |
| Generic replacement for the per-game `data/action_utils.py` files in the |
| StatePlay data preparation. The button schema and fixed-prompt configuration live |
| on a `GameProfile` (see `profiles.py`); this module only owns the shape / |
| upsampling behavior, which is identical across every fighting-game profile we |
| have seen so far. |
| """ |
| from __future__ import annotations |
|
|
| import os |
|
|
| import numpy as np |
| import pandas as pd |
| import torch |
|
|
| from diffsynth.core.data.operators import RouteByType |
|
|
| from .profiles import GameProfile |
|
|
|
|
| def hold_last_upsample(btn: torch.Tensor, window: int = 10) -> torch.Tensor: |
| """Forward-broadcast each window-sized block's first row across the block. |
| |
| Parquets are sampled at 10 Hz but stored 1:1 with the 20 fps video, so 9/10 |
| rows are zero-filled. window=10 collapses that back to a dense per-frame |
| signal. window=1 disables the upsample. |
| |
| Args: |
| btn: [T, num_buttons] float tensor. |
| window: block size in frames. |
| Returns: [T, num_buttons] tensor with the same dtype/device as `btn`. |
| """ |
| if window <= 1: |
| return btn |
| T = btn.shape[0] |
| idx = (torch.arange(T, device=btn.device) // window) * window |
| idx = idx.clamp(max=T - 1) |
| return btn[idx] |
|
|
|
|
| def get_action_op( |
| profile: GameProfile, |
| base_path: str, |
| num_frames: int, |
| hold_window: int | None = None, |
| ) -> RouteByType: |
| """UnifiedDataset operator: parquet path -> [1, T, num_buttons] action tensor. |
| |
| Strips the parquet to `profile.button_cols`, truncates to `num_frames`, |
| casts to float32, then hold-last upsamples. |
| """ |
| cols = list(profile.button_cols) |
| win = profile.default_action_hold_window if hold_window is None else hold_window |
|
|
| def process_parquet(rel_path: str) -> torch.Tensor: |
| full = os.path.join(base_path, rel_path) |
| df = pd.read_parquet(full) |
| arr = df[cols].values[:num_frames].astype(np.float32) |
| keyboard = torch.tensor(arr) |
| keyboard = hold_last_upsample(keyboard, window=win) |
| return keyboard.unsqueeze(0) |
|
|
| return RouteByType(operator_map=[(str, process_parquet)]) |
|
|
