File size: 11,558 Bytes
8f51ef2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 | from __future__ import annotations
from typing import Any, Dict, List, Optional, Type
from pathlib import Path
import tempfile
import shutil
import datetime
import numpy as np
import cv2
import torch
from pydantic import BaseModel, Field, field_validator
from langchain_core.tools import BaseTool
from langchain_core.callbacks import (
CallbackManagerForToolRun,
AsyncCallbackManagerForToolRun,
)
# EchoFlow demo primitives (mirror your guarded import style)
try:
import sys
import os
from pathlib import Path
# Add EchoFlow path to sys.path
current_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
echoflow_path = os.path.join(current_dir, "model_weights", "EchoFlow")
if echoflow_path not in sys.path:
sys.path.insert(0, echoflow_path)
from demo import (
load_view_mask, # (view: str) -> editor dict
generate_latent_image, # (mask, view, sampling_steps) -> latent
convert_latent_to_display, # (latent) -> np.ndarray (grayscale uint8)
decode_latent_to_pixel, # (latent) -> np.ndarray (RGB uint8)
check_privacy, # (latent, view) -> (Optional[latent], str)
generate_animation, # (latent, ef, sampling_steps, cfg_scale) -> latent_video
latent_animation_to_grayscale, # (latent_video) -> mp4_path
decode_animation, # (latent_video) -> mp4_path
)
except Exception:
load_view_mask = None # type: ignore
generate_latent_image = None # type: ignore
convert_latent_to_display = None # type: ignore
decode_latent_to_pixel = None # type: ignore
check_privacy = None # type: ignore
generate_animation = None # type: ignore
latent_animation_to_grayscale = None # type: ignore
decode_animation = None # type: ignore
# ----------------------------- Input schema -----------------------------
class EchoSynthesisInput(BaseModel):
"""Generate synthetic echo images and EF-conditioned videos via EchoFlow demo."""
views: List[str] = Field(
default_factory=lambda: ["A4C", "PLAX", "PSAX"],
description="Cardiac echo views to synthesize (e.g., A4C, PLAX, PSAX).",
)
efs: List[int] = Field(
default_factory=lambda: [35, 55, 70],
description="Ejection fraction percentages used to condition the animation.",
)
img_steps: int = Field(150, ge=1, le=2000, description="Sampling steps for latent image generation.")
vid_steps: int = Field(150, ge=1, le=2000, description="Sampling steps for latent video.")
cfg_scale: float = Field(1.0, ge=0.0, le=20.0, description="CFG scale for animation generation.")
max_privacy_retries: int = Field(3, ge=0, le=20, description="Max retries if privacy filter fails.")
outdir: Optional[str] = Field(
None,
description="Root output dir. If omitted, a timestamped folder is created under the tool temp dir.",
)
save_decoded_image: bool = Field(True, description="Save decoded RGB PNG per view.")
save_latent_preview: bool = Field(True, description="Save latent grayscale PNG per view.")
keep_failed_privacy_preview: bool = Field(
True,
description="If privacy fails after retries, save the last latent preview for diagnostics.",
)
@field_validator("views")
@classmethod
def _nonempty_views(cls, v: List[str]) -> List[str]:
if not v:
raise ValueError("At least one view must be provided.")
return v
@field_validator("efs")
@classmethod
def _valid_efs(cls, v: List[int]) -> List[int]:
if not v:
raise ValueError("At least one EF must be provided.")
for x in v:
if x < 0 or x > 100:
raise ValueError(f"EF {x} out of range [0, 100].")
return v
# ----------------------------- Tool class -------------------------------
class EchoSynthesisTool(BaseTool):
"""EchoFlow synthesis tool consistent with your EchoPrime tool suite."""
name: str = "echo_synthesis"
description: str = (
"Synthesize echocardiography images and EF-conditioned videos using EchoFlow demo primitives. "
"For each view: generate latent, save latent preview/decoded image, pass privacy filter, then render EF videos "
"(latent grayscale MP4 and decoded RGB MP4). Returns artifact paths and metadata."
)
args_schema: Type[BaseModel] = EchoSynthesisInput
device: Optional[str] = "cuda"
temp_dir: Path = Path("temp")
def __init__(self, device: Optional[str] = None, temp_dir: Optional[str] = None):
super().__init__()
# Present for symmetry with your other tools; EchoFlow demo may not use device directly
self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
self.temp_dir = Path(temp_dir or tempfile.mkdtemp())
self.temp_dir.mkdir(parents=True, exist_ok=True)
# ----------------------------- helpers -----------------------------
def _ensure_demo(self):
if any(x is None for x in [
load_view_mask, generate_latent_image, convert_latent_to_display,
decode_latent_to_pixel, check_privacy, generate_animation,
latent_animation_to_grayscale, decode_animation,
]):
raise RuntimeError(
"EchoFlow demo functions not importable. Ensure the 'demo' module and assets are in PYTHONPATH / working directory."
)
@staticmethod
def _ensure_dirs(root: Path) -> Dict[str, Path]:
d = {
"grayscale_frames": root / "grayscale_frames",
"decoded_images": root / "decoded_images",
"latent_videos": root / "latent_videos",
"decoded_videos": root / "decoded_videos",
"meta": root / "meta",
}
for p in d.values():
p.mkdir(parents=True, exist_ok=True)
return d
@staticmethod
def _save_png(path: Path, arr: np.ndarray) -> str:
if arr.dtype != np.uint8:
arr = np.clip(arr, 0, 255).astype(np.uint8)
if arr.ndim == 3 and arr.shape[2] == 3:
arr = arr[:, :, ::-1] # RGB->BGR for cv2.imwrite
if not cv2.imwrite(str(path), arr):
raise IOError(f"Failed to write image: {path}")
return str(path)
# ----------------------------- core run -----------------------------
def _run(
self,
views: List[str],
efs: List[int],
img_steps: int = 150,
vid_steps: int = 150,
cfg_scale: float = 1.0,
max_privacy_retries: int = 3,
outdir: Optional[str] = None,
save_decoded_image: bool = True,
save_latent_preview: bool = True,
keep_failed_privacy_preview: bool = True,
run_manager: Optional[CallbackManagerForToolRun] = None,
) -> Dict[str, Any]:
self._ensure_demo()
stamp = datetime.datetime.utcnow().strftime("%Y%m%d_%H%M%S")
root = Path(outdir) if outdir else (self.temp_dir / f"echoflow_run_{stamp}")
root.mkdir(parents=True, exist_ok=True)
paths = self._ensure_dirs(root)
run_meta = {
"timestamp_utc": stamp,
"device": self.device,
"views": views,
"efs": efs,
"img_steps": img_steps,
"vid_steps": vid_steps,
"cfg_scale": cfg_scale,
"max_privacy_retries": max_privacy_retries,
}
results: Dict[str, Any] = {"outdir": str(root), "meta": run_meta, "views": {}}
for view in views:
view_rec: Dict[str, Any] = {
"view": view,
"latent_preview_png": None,
"decoded_image_png": None,
"privacy_passed": False,
"privacy_message": None,
"videos": [], # [{ef, latent_grayscale_mp4, decoded_rgb_mp4}]
}
results["views"][view] = view_rec
# 1) Load mask
mask = load_view_mask(view)
# 2) Latent image
latent = generate_latent_image(mask, view, sampling_steps=img_steps)
# 3) Save previews
if save_latent_preview:
preview = convert_latent_to_display(latent)
view_rec["latent_preview_png"] = self._save_png(
paths["grayscale_frames"] / f"{view}_latent.png", preview
)
if save_decoded_image:
decoded = decode_latent_to_pixel(latent)
view_rec["decoded_image_png"] = self._save_png(
paths["decoded_images"] / f"{view}_decoded.png", decoded
)
# 4) Privacy gate
filtered_latent, msg = check_privacy(latent, view)
tries = 0
while filtered_latent is None and tries < max_privacy_retries:
latent = generate_latent_image(mask, view, sampling_steps=img_steps)
filtered_latent, msg = check_privacy(latent, view)
tries += 1
view_rec["privacy_message"] = msg
if filtered_latent is None:
if keep_failed_privacy_preview:
rejected_preview = convert_latent_to_display(latent)
self._save_png(paths["grayscale_frames"] / f"{view}_privacy_reject.png", rejected_preview)
continue # Skip videos for this view
view_rec["privacy_passed"] = True
# 5) EF-conditioned videos
for ef in efs:
lat_vid = generate_animation(filtered_latent, int(ef), sampling_steps=vid_steps, cfg_scale=cfg_scale)
gray_tmp = latent_animation_to_grayscale(lat_vid)
gray_target = paths["latent_videos"] / f"{view}_EF{ef}.mp4"
shutil.move(gray_tmp, gray_target)
dec_tmp = decode_animation(lat_vid)
dec_target = paths["decoded_videos"] / f"{view}_EF{ef}.mp4"
shutil.move(dec_tmp, dec_target)
view_rec["videos"].append({
"ef": int(ef),
"latent_grayscale_mp4": str(gray_target),
"decoded_rgb_mp4": str(dec_target),
})
return results
async def _arun( # pragma: no cover
self,
views: List[str],
efs: List[int],
img_steps: int = 150,
vid_steps: int = 150,
cfg_scale: float = 1.0,
max_privacy_retries: int = 3,
outdir: Optional[str] = None,
save_decoded_image: bool = True,
save_latent_preview: bool = True,
keep_failed_privacy_preview: bool = True,
run_manager: Optional[AsyncCallbackManagerForToolRun] = None,
) -> Dict[str, Any]:
return self._run(
views=views,
efs=efs,
img_steps=img_steps,
vid_steps=vid_steps,
cfg_scale=cfg_scale,
max_privacy_retries=max_privacy_retries,
outdir=outdir,
save_decoded_image=save_decoded_image,
save_latent_preview=save_latent_preview,
keep_failed_privacy_preview=keep_failed_privacy_preview,
)
################## How to run ##################
# tool = EchoSynthesisTool()
# result = tool.run({
# "views": ["A4C", "PLAX"],
# "efs": [35, 60],
# "img_steps": 150,
# "vid_steps": 150,
# "cfg_scale": 1.0,
# "max_privacy_retries": 3,
# "outdir": "out/agent_echoflow_run",
# }) |