Spaces:
Sleeping
Sleeping
File size: 10,620 Bytes
8d3e73e 8a7a9dd 8d3e73e 8a7a9dd 8d3e73e 8a7a9dd 8d3e73e 8a7a9dd 8d3e73e 8a7a9dd 8d3e73e 8a7a9dd 8d3e73e 8a7a9dd 8d3e73e f9eab93 8d3e73e f9eab93 8d3e73e f9eab93 8d3e73e |
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 |
import sys
from pathlib import Path
import numpy as np
from joblib import Parallel, delayed
from neus_v.automaton.video_automaton import VideoAutomaton
from neus_v.model_checking.stormpy import StormModelChecker
from neus_v.veval.parse import parse_tl_formula, parse_until_to_next_frame
from neus_v.video.frame import VideoFrame
from neus_v.video.read_video import read_video
def create_frame_windows(frames: list, window_size: int) -> list[list]:
"""Create non-overlapping windows of frames, with remainder in last window.
Args:
frames: List of frames
window_size: Size of each window
Returns:
List of frame windows
"""
windows = []
for i in range(0, len(frames), window_size):
windows.append(frames[i : i + window_size])
return windows
def evaluate_video(
vision_language_model,
confidence_as_token_probability: bool,
video_path: Path | str,
proposition_set: list,
tl_spec: str,
parallel_inference: bool = False,
threshold: float = 0.1,
num_of_frame_in_sequence: int = 1,
) -> dict:
"""Evaluate a video using the given vision language model."""
output_log = {
"specification": None,
"propositions": None,
"probability": None,
"min_probability": None,
"max_probability": None,
"propositions_avg_probability": {},
}
if isinstance(video_path, str):
video_path = Path(video_path)
video = read_video(video_path=video_path)
# TODO: if there's F in the tl_spec
ltl_formula = parse_tl_formula(tl_spec)
ltl_formula = parse_until_to_next_frame(ltl_formula)
video_automaton = VideoAutomaton(include_initial_state=True)
video_automaton.set_up(proposition_set=proposition_set)
model_checker = StormModelChecker(
proposition_set=proposition_set,
ltl_formula=ltl_formula,
)
proposition_probability_record = {}
for proposition in proposition_set:
proposition_probability_record[proposition] = []
if model_checker.validate_tl_specification(ltl_formula):
frame_count = 0
all_frames: list[np.ndarray] = video.get_all_frames_of_video(
return_format="ndarray",
desired_interval_in_sec=1,
)
try:
# for frame_img in all_frames:
def process_frame(frame_img: np.ndarray, frame_count: int):
sys.stdout.write(f"\rProcessing frame: {frame_count+1}/{len(all_frames)} ")
sys.stdout.flush()
object_of_interest = {}
for proposition in proposition_set:
detected_object = vision_language_model.detect(
frame_img=frame_img,
scene_description=proposition,
confidence_as_token_probability=confidence_as_token_probability,
threshold=threshold,
)
object_of_interest[proposition] = detected_object
# proposition_probability_record.get(proposition).append(
# detected_object.probability
# )
video_frame = VideoFrame(
frame_idx=frame_count,
timestamp=None,
frame_image=frame_img,
object_of_interest=object_of_interest,
)
return video_frame, object_of_interest
if parallel_inference:
frame_windows = create_frame_windows(frames=all_frames, window_size=num_of_frame_in_sequence)
results = Parallel(n_jobs=len(all_frames))(
delayed(process_frame)(frame_img, i) for i, frame_img in enumerate(all_frames)
)
else:
frame_windows = create_frame_windows(frames=all_frames, window_size=num_of_frame_in_sequence)
results = [process_frame(frame_img, i) for i, frame_img in enumerate(all_frames)]
for video_frame, object_of_interest in results:
video_automaton.add_frame(frame=video_frame)
for proposition, detected_object in object_of_interest.items():
proposition_probability_record[proposition].append(detected_object.probability)
video_automaton.add_terminal_state(add_with_terminal_label=True)
sys.stdout.write("\n") # Move to the next line after processing all frames
result = model_checker.check_automaton(
states=video_automaton.states,
transitions=video_automaton.transitions,
model_type="dtmc",
use_filter=True,
)
output_log["specification"] = tl_spec
output_log["propositions"] = proposition_set
output_log["probability"] = round(float(str(result)), 6)
output_log["min_probability"] = round(float(str(result.min)), 6)
output_log["max_probability"] = round(float(str(result.max)), 6)
for (
proposition,
probabilities,
) in proposition_probability_record.items():
avg_probability = sum(probabilities) / len(probabilities)
output_log["propositions_avg_probability"][proposition] = round(avg_probability, 3)
except Exception as e: # noqa: BLE001
# print(f"\nError processing frame {frame_count}: {e}")
import traceback
print(f"\nError processing frame {frame_count}: {e}")
traceback.print_exc()
return output_log
def evaluate_video_with_sequence_of_images(
vision_language_model,
confidence_as_token_probability: bool,
video_path: Path | str,
proposition_set: list,
tl_spec: str,
parallel_inference: bool = False,
num_of_frame_in_sequence: int = 3,
threshold: float = 0.1,
) -> dict:
"""Evaluate a video using the given vision language model."""
output_log = {
"specification": None,
"propositions": None,
"probability": None,
"min_probability": None,
"max_probability": None,
"propositions_avg_probability": {},
}
if isinstance(video_path, str):
video_path = Path(video_path)
video = read_video(video_path=video_path)
# TODO: if there's F in the tl_spec
ltl_formula = parse_tl_formula(tl_spec)
ltl_formula = parse_until_to_next_frame(ltl_formula)
video_automaton = VideoAutomaton(include_initial_state=True)
video_automaton.set_up(proposition_set=proposition_set)
model_checker = StormModelChecker(
proposition_set=proposition_set,
ltl_formula=ltl_formula,
)
proposition_probability_record = {}
for proposition in proposition_set:
proposition_probability_record[proposition] = []
if model_checker.validate_tl_specification(ltl_formula):
frame_count = 0
all_frames: list[np.ndarray] = video.get_all_frames_of_video(
return_format="ndarray",
desired_interval_in_sec=0.5,
)
try:
# for frame_img in all_frames:
def process_frame(sequence_of_frames: list[np.ndarray], frame_count: int):
sys.stdout.write(f"\rProcessing frame window: {frame_count+1}/{len(frame_windows)} ")
sys.stdout.flush()
object_of_interest = {}
for proposition in proposition_set:
detected_object = vision_language_model.detect(
seq_of_frames=sequence_of_frames,
scene_description=proposition,
# confidence_as_token_probability=confidence_as_token_probability,
threshold=threshold,
)
object_of_interest[proposition] = detected_object
# proposition_probability_record.get(proposition).append(
# detected_object.probability
# )
print(f"{proposition}: {detected_object.probability}")
video_frame = VideoFrame(
frame_idx=frame_count,
timestamp=None,
frame_image=sequence_of_frames,
object_of_interest=object_of_interest,
)
return video_frame, object_of_interest
if parallel_inference:
frame_windows = create_frame_windows(frames=all_frames, window_size=num_of_frame_in_sequence)
results = Parallel(n_jobs=len(frame_windows))(
delayed(process_frame)(frame_img, i) for i, frame_img in enumerate(frame_windows)
)
else:
frame_windows = create_frame_windows(frames=all_frames, window_size=num_of_frame_in_sequence)
results = [process_frame(sequence_of_frames, i) for i, sequence_of_frames in enumerate(frame_windows)]
for video_frame, object_of_interest in results:
video_automaton.add_frame(frame=video_frame)
for proposition, detected_object in object_of_interest.items():
proposition_probability_record[proposition].append(detected_object.probability)
video_automaton.add_terminal_state(add_with_terminal_label=False)
sys.stdout.write("\n") # Move to the next line after processing all frames
result = model_checker.check_automaton(
states=video_automaton.states,
transitions=video_automaton.transitions,
model_type="dtmc",
use_filter=False,
)
output_log["specification"] = tl_spec
output_log["propositions"] = proposition_set
output_log["probability"] = round(float(str(result.at(0))), 6)
output_log["min_probability"] = round(float(str(result.min)), 6)
output_log["max_probability"] = round(float(str(result.max)), 6)
for (
proposition,
probabilities,
) in proposition_probability_record.items():
avg_probability = sum(probabilities) / len(probabilities)
output_log["propositions_avg_probability"][proposition] = round(avg_probability, 3)
except Exception as e: # noqa: BLE001
# print(f"\nError processing frame {frame_count}: {e}")
import traceback
print(f"\nError processing frame {frame_count}: {e}")
traceback.print_exc()
return output_log
|