Yunhao Fang
initialize space.
c5c2e39
import gradio as gr
import os
import json
import threading
from pathlib import Path
from moviepy.editor import VideoFileClip
import hashlib
import random
import string
from PIL import Image
PHYSICAL_LAWS = [
"Violation of Newton's Law: Objects move without any external force.",
"Violation of the Law of Conservation of Mass or Solid Constitutive Law: Objects deform or distort irregularly.",
"Violation of Fluid Constitutive Law: Liquids flow in an unnatural or irregular manner.",
"Violation of Non-physical Penetration: Objects unnaturally pass through each other.",
"Violation of Gravity: Objects behave inconsistently with gravity, such as floating in the air.",
"No violation!"
]
# List of commonsense violations
COMMON_SENSE = [
"Poor Aesthetics: Visually unappealing or low-quality content.",
"Temporal Inconsistency: Flickering, choppiness, or sudden appearance/disappearance of irrelevant objects.",
"No violation!"
]
# Example images for physical law violations
EXAMPLE_IMAGES = {
"newtons_law": "test_images/law_violation1.jpg",
"mass_conservation": "test_images/law_violation2.jpg",
"fluid.": "test_images/law_violation3.jpg",
"penetration": "test_images/law_violation4.jpg",
"gravity": "test_images/law_violation5.jpg"
}
def string_to_md5(input_string, max_digits=12):
return hashlib.md5(input_string.encode()).hexdigest()[:max_digits]
def generate_random_id(length=6):
return ''.join(random.choices(string.ascii_lowercase + string.digits, k=length))
class VideoAnnotator:
def __init__(self, videos, annotation_base_dir, max_resolution=(640, 480)):
self.annotation_base_dir = Path(annotation_base_dir)
self.max_resolution = max_resolution
self.videos = videos
self.current_index = 0
self.file_locks = {}
self.current_labeler = None
self.current_labeler_file = None
def get_annotation_file_path(self, labeler_email):
md5_email = string_to_md5(labeler_email, max_digits=12)
# random_id = generate_random_id()
# file_name = f"md5-{md5_email}.{random_id}.json"
file_name = f"md5-{md5_email}.json"
return self.annotation_base_dir / file_name
def load_annotations(self, labeler_email):
file_path = self.get_annotation_file_path(labeler_email)
if file_path.exists():
with open(file_path, 'r') as f:
return json.load(f)
return {}
def save_annotations(self, labeler_email, annotations):
file_path = self.get_annotation_file_path(labeler_email)
self.annotation_base_dir.mkdir(parents=True, exist_ok=True)
if file_path not in self.file_locks:
self.file_locks[file_path] = threading.Lock()
with self.file_locks[file_path]:
with open(file_path, 'w') as f:
json.dump(annotations, f, indent=2)
def get_current_video(self):
if self.videos:
video_path = self.videos[self.current_index]
resized_path = self.resize_video_if_needed(video_path)
return str(resized_path.resolve())
return None
def resize_video_if_needed(self, video_path):
from moviepy.video.io.ffmpeg_writer import ffmpeg_write_video
clip = VideoFileClip(str(video_path))
width, height = clip.size
if width > self.max_resolution[0] or height > self.max_resolution[1]:
resized_clip = clip.resize(height=self.max_resolution[1])
cleaned_name = video_path.name.replace(" ", "_")
resized_path = video_path.with_name(f"resized_{cleaned_name}")
fps = clip.fps if clip.fps else 8.0
ffmpeg_write_video(resized_clip, str(resized_path), fps, codec="libx264")
return resized_path
return video_path
def update_annotation(self, video_name, labeler_email, instruction_check, law_annotations, commonsense):
video_name = postprocess_name_for_gradio(video_name)
annotations = self.load_annotations(labeler_email)
if instruction_check and video_name not in annotations:
annotations[video_name] = {
"labeler": labeler_email,
"law_details": law_annotations,
"commonsense": commonsense,
"instruction": instruction_check
}
self.save_annotations(labeler_email, annotations)
def next_video(self):
if self.videos:
self.current_index = min(self.current_index + 1, len(self.videos) - 1)
return self.get_current_video()
def prev_video(self):
if self.videos:
self.current_index = max(self.current_index - 1, 0)
return self.get_current_video()
def jump_to_video(self, index):
if self.videos:
self.current_index = max(0, min(index, len(self.videos) - 1))
return self.get_current_video()
def set_current_labeler(self, labeler_email):
self.current_labeler = labeler_email
self.current_labeler_file = self.get_annotation_file_path(labeler_email)
def postprocess_name_for_gradio(name):
return name.replace("–","").replace("+","").replace("-","").replace("t2v","").replace("(", "").replace(")","").replace(",","").replace("_","").replace(".","")
def get_cur_data(instruction_data, video_name):
video_name = postprocess_name_for_gradio(video_name)
if "resized_" in video_name:
clean_name = video_name.replace("resized_", "")
clean_name = "_".join(clean_name.split("_")[2:])
else:
clean_name = video_name
# print(clean_name, instruction_data.keys())
for k in instruction_data.keys():
if k in clean_name:
real_name = k
cur_data = instruction_data[real_name]
return cur_data
def create_interface(instruction_data, videos, annotation_base_dir):
annotator = VideoAnnotator(videos, annotation_base_dir)
def update_video():
video_path = annotator.get_current_video()
if video_path is None:
return (None, annotator.current_labeler or "", "[system] Video not in benchmark", "[system] Video not in benchmark", *[False for _ in PHYSICAL_LAWS], *[False for _ in COMMON_SENSE])
video_name = Path(video_path).name
cur_data = get_cur_data(instruction_data, video_name)
current_annotations = {}
if annotator.current_labeler:
annotations = annotator.load_annotations(annotator.current_labeler)
current_annotations = annotations.get(
postprocess_name_for_gradio(video_name),
{"labeler": annotator.current_labeler, "law_details": {law: False for law in PHYSICAL_LAWS}, "commonsense": {cs: False for cs in COMMON_SENSE}, "instruction": None}
)
else:
current_annotations = {"labeler": "", "law_details": {law: False for law in PHYSICAL_LAWS}, "commonsense": {cs: False for cs in COMMON_SENSE},"instruction": None}
first_frame = cur_data["text_first_frame"]
num_annotations = str(len(annotations)) if 'annotations' in locals() else "0"
text_instruction = cur_data["text_instruction"]
# Flatten the outputs
outputs = [
video_path,
current_annotations["labeler"] or "",
num_annotations,
current_annotations["instruction"],
text_instruction
]
# Add individual law checkbox values
outputs.extend([current_annotations["law_details"].get(law, False) for law in PHYSICAL_LAWS])
# Add individual commonsense checkbox values
outputs.extend([current_annotations["commonsense"].get(cs, False) for cs in COMMON_SENSE])
return outputs
def save_current_annotation(video_path, labeler_email, instruction_check, law_values, commonsense_values, skipped: bool=False):
if not skipped:
if video_path is None:
return "No video loaded to save annotations."
if not labeler_email:
return "Please enter a valid labeler email before saving annotations."
video_name = Path(video_path).name
law_annotations = {law: bool(value) for law, value in zip(PHYSICAL_LAWS, law_values)}
commonsense_annotations = {cs: bool(value) for cs, value in zip(COMMON_SENSE, commonsense_values)}
annotator.set_current_labeler(labeler_email)
annotator.update_annotation(video_name, labeler_email, instruction_check, law_annotations, commonsense_annotations)
return f"Annotation saved successfully for {labeler_email}!"
else:
video_name = Path(video_path).name
law_annotations = {law: bool(value) for law, value in zip(PHYSICAL_LAWS, law_values)}
commonsense_annotations = {cs: bool(value) for cs, value in zip(COMMON_SENSE, commonsense_values)}
annotator.set_current_labeler(labeler_email)
annotator.update_annotation(video_name, labeler_email, instruction_check, law_annotations, commonsense_annotations)
return f"Annotation saved successfully for {labeler_email}!"
def load_anns_callback(labeler_email):
"""
Load annotations for the given labeler email and jump to the next unlabeled video.
Returns the updated interface state.
"""
if not labeler_email:
return update_video()
# Set the current labeler
annotator.set_current_labeler(labeler_email)
# Load existing annotations
annotations = annotator.load_annotations(labeler_email)
# Find the first video that hasn't been annotated
next_unannotated_index = None
for i, video in enumerate(annotator.videos):
video_name = postprocess_name_for_gradio("resized_" + Path(video).name)
if video_name not in annotations:
next_unannotated_index = i
break
# If we found an unannotated video, jump to it
if next_unannotated_index is not None:
annotator.jump_to_video(next_unannotated_index)
video_path = annotator.get_current_video()
video_name = Path(video_path).name
cur_data = get_cur_data(instruction_data, video_name)
# Prepare default state for the new video
return [
video_path, # video
labeler_email, # labeler
str(len(annotations)), # num_annotations
None, # instruction_check (default value)
cur_data["text_instruction"], # text_instruction
*[False for _ in PHYSICAL_LAWS], # law checkboxes
*[False for _ in COMMON_SENSE] # commonsense checkboxes
]
else:
# If all videos are annotated, stay at current video but update the interface
current_video = annotator.get_current_video()
if current_video:
video_name = Path(current_video).name
current_annotations = annotations.get(
postprocess_name_for_gradio(video_name),
{
"labeler": labeler_email,
"law_details": {law: False for law in PHYSICAL_LAWS},
"commonsense": {cs: False for cs in COMMON_SENSE},
"instruction": "3"
}
)
cur_data = get_cur_data(instruction_data, video_name)
return [
current_video,
labeler_email,
str(len(annotations)),
current_annotations["instruction"],
cur_data["text_instruction"],
*[current_annotations["law_details"].get(law, False) for law in PHYSICAL_LAWS],
*[current_annotations["commonsense"].get(cs, False) for cs in COMMON_SENSE]
]
else:
# Fallback for empty video list
return [
None,
labeler_email,
"0",
None,
"[system] No videos available",
*[False for _ in PHYSICAL_LAWS],
*[False for _ in COMMON_SENSE]
]
def check_inputs(labeler_email, instruction_check):
"""Helper function to check input validity"""
if not labeler_email:
return False, "Please enter your email before proceeding."
if not instruction_check:
return False, "Please select whether the video follows the instruction before proceeding."
return True, ""
def confirm_callback(video_path, labeler_email, instruction_check, *checkbox_values):
pass
def skip_callback(video_path, labeler_email, instruction_check, *checkbox_values):
## save annotations with a flag skipped
num_laws = len(PHYSICAL_LAWS)
law_values = checkbox_values[:num_laws]
commonsense_values = checkbox_values[num_laws:]
breakpoint()
save_current_annotation(video_path, labeler_email, instruction_check, law_values, commonsense_values, skipped=True)
annotator.next_video()
return update_video()
def next_video_callback(video_path, labeler_email, instruction_check, *checkbox_values):
breakpoint()
# First check inputs
is_valid, message = check_inputs(labeler_email, instruction_check)
if not is_valid:
# Return current state with error message
gr.Warning(message)
return update_video()
# Split checkbox values into law and commonsense values
num_laws = len(PHYSICAL_LAWS)
law_values = checkbox_values[:num_laws]
commonsense_values = checkbox_values[num_laws:]
save_current_annotation(video_path, labeler_email, instruction_check, law_values, commonsense_values)
annotator.next_video()
return update_video()
def prev_video_callback(video_path, labeler_email, instruction_check, *checkbox_values):
# First check inputs
is_valid, message = check_inputs(labeler_email, instruction_check)
if not is_valid:
# Return current state with error message
gr.Warning(message)
return update_video()
# Split checkbox values into law and commonsense values
num_laws = len(PHYSICAL_LAWS)
law_values = checkbox_values[:num_laws]
commonsense_values = checkbox_values[num_laws:]
save_current_annotation(video_path, labeler_email, instruction_check, law_values, commonsense_values)
annotator.prev_video()
return update_video()
with gr.Blocks() as interface:
# gr.Markdown("# Video Annotation Interface")
with gr.Row():
with gr.Column(scale=1):
video = gr.Video(label="Current Video", format="mp4", height=450, width=800)
with gr.Row():
with gr.Column(scale=2):
labeler = gr.Textbox(
label="Labeler ID (your email)",
placeholder="Enter your email",
interactive=True,
)
with gr.Column(scale=1):
num_annotations = gr.Textbox(
label="Annotations Count",
placeholder="0",
interactive=False,
)
text_instruction = gr.Textbox(label="Text prompt", interactive=False)
instruction_check = gr.Radio(
label="Task1: Does this video follow the instruction?",
choices=[
"0: Not at all!!!",
"1: Correct object, wrong motion (or vice versa).",
"2: Follow instruction, fail task.",
"3: Follow instruction, complete task."
],
type="value",
value="3"
)
with gr.Row():
with gr.Column(scale=1):
skip_btn = gr.Button("Skip! Video Corrupted")
with gr.Column(scale=1):
confirm_btn = gr.Button("Confirm!")
with gr.Row():
with gr.Column(scale=1):
prev_btn = gr.Button("Previous Video")
with gr.Column(scale=1):
next_btn = gr.Button("Next Video")
load_btn = gr.Button("Load Annotations")
with gr.Column(scale=1):
gr.Markdown("Task2: [Based on your first impression] Select the major <span style='color: blue;'>commonsense violations</span> in the video: <span style='color: red;'>[multiple (0-2) choices]</span>")
commonsense_checkboxes = []
for cs in COMMON_SENSE:
commonsense_checkboxes.append(gr.Checkbox(label=cs))
gr.Markdown("Task3: Please select all physics laws the video <span style='color: blue;'>violates</span>: <span style='color: red;'>[multiple (0-5) choices]</span>")
law_checkboxes = []
for i, law in enumerate(PHYSICAL_LAWS):
checkbox = gr.Checkbox(label=law, interactive=True)
law_checkboxes.append(checkbox)
# if i < len(PHYSICAL_LAWS) - 1:
# image_path = os.path.join(os.path.abspath(__file__).rsplit("/", 1)[0], list(EXAMPLE_IMAGES.values())[i])
if i != len(PHYSICAL_LAWS) - 1:
image_path = list(EXAMPLE_IMAGES.values())[i]
image = Image.open(image_path).convert("RGB")
gr.Image(value=image, label=f"Example {i+1}", show_label=True, height=68, width=700)
# Create a flat list of all inputs
all_inputs = [video, labeler, instruction_check] + law_checkboxes + commonsense_checkboxes
# Create a flat list of all outputs
all_outputs = [video, labeler, num_annotations, instruction_check, text_instruction] + law_checkboxes + commonsense_checkboxes
# Set up event handlers with flattened inputs and outputs
skip_btn.click(
skip_callback,
inputs=all_inputs,
outputs=all_outputs
)
load_btn.click(
load_anns_callback,
inputs=[labeler],
outputs=all_outputs
)
next_btn.click(
next_video_callback,
inputs=all_inputs,
outputs=all_outputs
)
prev_btn.click(
prev_video_callback,
inputs=all_inputs,
outputs=all_outputs
)
interface.load(
fn=update_video,
inputs=None,
outputs=all_outputs
)
return interface
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description="Annotation")
parser.add_argument("--domain", type=str, default="robotics", help="")
parser.add_argument("--src", type=str, default="CogVideo-T2V", help="")
# Parse the arguments
args = parser.parse_args()
domains = ["robotics", "humans", "general", "av", "game"]
src = ["CogVideo-I2V", "CogVideo-T2V", "Open-Sora-I2V", "Open-Sora-T2V", "Pandora", "TurboT2V", "Open-Sora-Plan-I2V", "Open-Sora-Plan-T2V"]
assert args.domain in domains, f"{args.domain} not in available domain."
assert args.src in src, f"{args.src} not in available model src."
instruction_base_path = "domains"
src_video_map = {
"CogVideo-I2V": "/home/yunhaof/workspace/datasets/outputs_v2",
"CogVideo-T2V": "/home/yunhaof/workspace/datasets/outputs_v2",
"Pandora": "/lustre/fsw/portfolios/nvr/users/dachengl/VILA-EWM/outputs",
"Open-Sora-I2V": "/lustre/fsw/portfolios/nvr/users/dachengl/Open-Sora/outputs",
"Open-Sora-T2V": "/lustre/fsw/portfolios/nvr/users/dachengl/Open-Sora/outputs",
"TurboT2V": "",
"Open-Sora-Plan-I2V": "/home/yunhaof/workspace/projects/Open-Sora-Plan/ewm_benchmark/gradio_videos",
"Open-Sora-Plan-T2V": "/home/yunhaof/workspace/projects/Open-Sora-Plan/ewm_benchmark/gradio_videos"
}
# Adhoc solution to naming mismatch
domain_name_map = {
"humans": "humans",
"game": "game",
"general": "general",
"av": "av",
"robotics": "robotics"
}
cur_domain = domain_name_map[args.domain]
# video_folder = "/lustre/fsw/portfolios/nvr/users/dachengl/CogVideo/outputs"
video_folder = Path(src_video_map[args.src])
# print("Processing the 100 videos for the current annotation.")
videos = []
if args.src == "CogVideo-I2V":
for v in video_folder.glob("*.mp4"):
if "t2v" not in v.stem and "resized_" not in v.stem and f"{cur_domain}_" in v.stem:
videos.append(v)
elif args.src == "CogVideo-T2V":
for v in video_folder.glob("*.mp4"):
if "t2v" in v.stem and "resized_" not in v.stem and f"{cur_domain}_" in v.stem:
videos.append(v)
elif args.src == "Pandora":
for v in video_folder.glob("*.mp4"):
if "resized_" not in v.stem and f"{cur_domain}_" in v.stem:
videos.append(v)
elif args.src == "Open-Sora-I2V":
for v in video_folder.glob("*.mp4"):
if "t2v" not in v.stem and "resized_" not in v.stem and f"{cur_domain}_" in v.stem:
videos.append(v)
elif args.src == "Open-Sora-T2V":
for v in video_folder.glob("*.mp4"):
if "t2v" in v.stem and "resized_" not in v.stem and f"{cur_domain}_" in v.stem:
videos.append(v)
elif args.src == "Open-Sora-Plan-I2V":
for v in video_folder.glob("*.mp4"):
if "t2v" not in v.stem and "resized_" not in v.stem and f"{cur_domain}_" in v.stem:
videos.append(v)
elif args.src == "Open-Sora-Plan-T2V":
for v in video_folder.glob("*.mp4"):
if "t2v" in v.stem and "resized_" not in v.stem and f"{cur_domain}_" in v.stem:
videos.append(v)
elif args.src == "TurboT2V":
for v in video_folder.glob("*.mp4"):
if "t2v" in v.stem and "resized_" not in v.stem and f"{cur_domain}_" in v.stem:
videos.append(v)
videos = sorted(videos)
print(f"Number of videos: {len(videos)}")
instruction_file = f"domains/{args.domain}/dataset_v2/instruction_ewm.json"
annotation_base = "annotations"
os.makedirs(annotation_base, exist_ok=True)
annotation_dir = os.path.join(annotation_base, f"{args.domain}_{args.src}")
instruction_data = {}
with open(instruction_file, "r") as f:
instructions = json.load(f)
for instruction in instructions:
file_name = os.path.basename(instruction["video_path"])
# gradio will eliminate -
file_name = postprocess_name_for_gradio(file_name)#.replace("-", "").replace("_t2v","")
instruction_data[file_name] = instruction
# perform a check that these videos will appear on the instruction, with or without the resized_
for _video in videos:
try:
_ = get_cur_data(instruction_data, postprocess_name_for_gradio(Path(_video).name))#.replace("-", "").replace("_t2v",""))
except:
print(f"parsing name {_video} fails, you may want to look at the name in instruction_ewm.json")
assert False
try:
_ = get_cur_data(instruction_data, "resized_" + postprocess_name_for_gradio(Path(_video).name))# .replace("-", "").replace("_t2v",""))
except:
print(f"parsing name resized_{_video} fails, you may want to look at the name in instruction_ewm.json")
assert False
iface = create_interface(instruction_data, videos, annotation_dir)
iface.launch(share=True, allowed_paths=[src_video_map[args.src]])