matsant01's picture
Minor changes
56c5ad3
raw
history blame
16.2 kB
import gradio as gr
import os
import random
import csv
from pathlib import Path
from datetime import datetime, timedelta
import tempfile
from huggingface_hub import HfApi, hf_hub_download, login
from huggingface_hub.utils import RepositoryNotFoundError, EntryNotFoundError
from apscheduler.schedulers.background import BackgroundScheduler
import atexit
import threading
import time
import shutil
# --- Configuration ---
DATASET_REPO_ID = os.getenv("DATASET_REPO_ID", "matsant01/user-study-collected-preferences")
HF_TOKEN = os.getenv("HF_TOKEN")
RESULTS_FILENAME_IN_REPO = "preferences.csv"
TEMP_DIR = tempfile.mkdtemp()
LOCAL_RESULTS_FILE = Path(TEMP_DIR) / RESULTS_FILENAME_IN_REPO
UPLOAD_INTERVAL_HOURS = 0.1
DATA_DIR = Path("data")
IMAGE_EXTENSIONS = [".png", ".jpg", ".jpeg", ".webp"]
# --- Global State for Upload Logic ---
hf_api = None
scheduler = BackgroundScheduler(daemon=True)
upload_lock = threading.Lock()
new_preferences_recorded_since_last_upload = threading.Event()
# --- Hugging Face Hub Login & Initialization ---
def initialize_hub_and_results():
global hf_api
if HF_TOKEN:
print("Logging into Hugging Face Hub...")
try:
login(token=HF_TOKEN)
hf_api = HfApi()
print(f"Attempting initial download of {RESULTS_FILENAME_IN_REPO} from {DATASET_REPO_ID}")
hf_hub_download(
repo_id=DATASET_REPO_ID,
filename=RESULTS_FILENAME_IN_REPO,
repo_type="dataset",
token=HF_TOKEN,
local_dir=TEMP_DIR,
local_dir_use_symlinks=False
)
print(f"Successfully downloaded existing {RESULTS_FILENAME_IN_REPO} to {LOCAL_RESULTS_FILE}")
except EntryNotFoundError:
print(f"{RESULTS_FILENAME_IN_REPO} not found in repo. Will create locally.")
except RepositoryNotFoundError:
print(f"Error: Dataset repository {DATASET_REPO_ID} not found or token lacks permissions.")
print("Results saving will be disabled.")
hf_api = None
except Exception as e:
print(f"Error during initial download/login: {e}")
print("Proceeding without initial download. File will be created locally.")
else:
print("Warning: HF_TOKEN secret not found. Results will not be saved to the Hub.")
hf_api = None
# --- Data Loading ---
def find_image(folder_path: Path, base_name: str) -> Path | None:
for ext in IMAGE_EXTENSIONS:
file_path = folder_path / f"{base_name}{ext}"
if file_path.exists():
return file_path
return None
def get_sample_ids() -> list[str]:
sample_ids = []
if DATA_DIR.is_dir():
for item in DATA_DIR.iterdir():
if item.is_dir():
prompt_file = item / "prompt.txt"
input_bg = find_image(item, "input_bg")
input_fg = find_image(item, "input_fg")
output_baseline = find_image(item, "baseline")
output_tficon = find_image(item, "tf-icon")
if prompt_file.exists() and input_bg and input_fg and output_baseline and output_tficon:
sample_ids.append(item.name)
return sample_ids
def load_sample_data(sample_id: str) -> dict | None:
sample_path = DATA_DIR / sample_id
if not sample_path.is_dir():
return None
prompt_file = sample_path / "prompt.txt"
input_bg_path = find_image(sample_path, "input_bg")
input_fg_path = find_image(sample_path, "input_fg")
output_baseline_path = find_image(sample_path, "baseline")
output_tficon_path = find_image(sample_path, "tf-icon")
if not all([prompt_file.exists(), input_bg_path, input_fg_path, output_baseline_path, output_tficon_path]):
print(f"Warning: Missing files in sample {sample_id}")
return None
try:
prompt = prompt_file.read_text().strip()
except Exception as e:
print(f"Error reading prompt for {sample_id}: {e}")
return None
return {
"id": sample_id,
"prompt": prompt,
"input_bg": str(input_bg_path),
"input_fg": str(input_fg_path),
"output_baseline": str(output_baseline_path),
"output_tficon": str(output_tficon_path),
}
# --- State and UI Logic ---
INITIAL_SAMPLE_IDS = get_sample_ids()
def get_next_sample(available_ids: list[str]) -> tuple[dict | None, list[str]]:
if not available_ids:
return None, []
chosen_id = random.choice(available_ids)
remaining_ids = [id for id in available_ids if id != chosen_id]
sample_data = load_sample_data(chosen_id)
return sample_data, remaining_ids
def display_new_sample(state: dict, available_ids: list[str]):
sample_data, remaining_ids = get_next_sample(available_ids)
if not sample_data:
return {
prompt_display: gr.update(value="**Prompt:** No more samples available. Thank you!"),
input_bg_display: gr.update(value=None, visible=False),
input_fg_display: gr.update(value=None, visible=False),
output_a_display: gr.update(value=None, visible=False),
output_b_display: gr.update(value=None, visible=False),
choice_button_a: gr.update(visible=False),
choice_button_b: gr.update(visible=False),
next_button: gr.update(visible=False),
status_display: gr.update(value="**Status:** Completed!"),
app_state: state,
available_samples_state: remaining_ids
}
outputs = [
{"model_name": "baseline", "path": sample_data["output_baseline"]},
{"model_name": "tf-icon", "path": sample_data["output_tficon"]},
]
random.shuffle(outputs)
output_a = outputs[0]
output_b = outputs[1]
state = {
"current_sample_id": sample_data["id"],
"output_a_model_name": output_a["model_name"],
"output_b_model_name": output_b["model_name"],
}
return {
prompt_display: gr.update(value=f"**Prompt:** {sample_data['prompt']}"),
input_bg_display: gr.update(value=sample_data["input_bg"], visible=True),
input_fg_display: gr.update(value=sample_data["input_fg"], visible=True),
output_a_display: gr.update(value=output_a["path"], visible=True),
output_b_display: gr.update(value=output_b["path"], visible=True),
choice_button_a: gr.update(visible=True, interactive=True),
choice_button_b: gr.update(visible=True, interactive=True),
next_button: gr.update(visible=False),
status_display: gr.update(value="**Status:** Please choose the image you prefer."),
app_state: state,
available_samples_state: remaining_ids
}
def record_preference(choice: str, state: dict, request: gr.Request):
if not request:
print("Error: Request object is None. Cannot get session ID.")
session_id = "unknown_session"
else:
try:
session_id = request.client.host
except AttributeError:
print("Error: request.client is None or has no 'host' attribute.")
session_id = "unknown_client"
if not state or "current_sample_id" not in state:
print("Warning: State missing, cannot record preference.")
return {
choice_button_a: gr.update(interactive=False),
choice_button_b: gr.update(interactive=False),
next_button: gr.update(visible=True, interactive=True),
status_display: gr.update(value="**Status:** Error: Session state lost. Click Next Sample."),
app_state: state
}
chosen_model_name = state["output_a_model_name"] if choice == "A" else state["output_b_model_name"]
baseline_display = "A" if state["output_a_model_name"] == "baseline" else "B"
tficon_display = "B" if state["output_a_model_name"] == "baseline" else "A"
new_row = {
"timestamp": datetime.now().isoformat(),
"session_id": session_id,
"sample_id": state["current_sample_id"],
"baseline_displayed_as": baseline_display,
"tficon_displayed_as": tficon_display,
"chosen_display": choice,
"chosen_model_name": chosen_model_name
}
header = list(new_row.keys())
try:
with upload_lock:
file_exists = LOCAL_RESULTS_FILE.exists()
mode = 'a' if file_exists else 'w'
with open(LOCAL_RESULTS_FILE, mode, newline='', encoding='utf-8') as f:
writer = csv.DictWriter(f, fieldnames=header)
if not file_exists or os.path.getsize(LOCAL_RESULTS_FILE) == 0:
writer.writeheader()
print(f"Created or wrote header to {LOCAL_RESULTS_FILE}")
writer.writerow(new_row)
print(f"Appended preference for {state['current_sample_id']} to local file.")
new_preferences_recorded_since_last_upload.set()
except Exception as e:
print(f"Error writing local results file {LOCAL_RESULTS_FILE}: {e}")
return {
choice_button_a: gr.update(interactive=False),
choice_button_b: gr.update(interactive=False),
next_button: gr.update(visible=True, interactive=True),
status_display: gr.update(value=f"**Status:** Error saving preference locally: {e}. Click Next."),
app_state: state
}
return {
choice_button_a: gr.update(interactive=False),
choice_button_b: gr.update(interactive=False),
next_button: gr.update(visible=True, interactive=True),
status_display: gr.update(value=f"**Status:** Preference recorded (Chose {choice}). Click Next Sample."),
app_state: state
}
def upload_preferences_to_hub():
print("Periodic upload check triggered.")
if not hf_api:
print("Upload check skipped: Hugging Face API not available.")
return
if not new_preferences_recorded_since_last_upload.is_set():
print("Upload check skipped: No new preferences recorded since last upload.")
return
with upload_lock:
if not new_preferences_recorded_since_last_upload.is_set():
print("Upload check skipped (race condition avoided): No new preferences.")
return
if not LOCAL_RESULTS_FILE.exists() or os.path.getsize(LOCAL_RESULTS_FILE) == 0:
print("Upload check skipped: Local results file is missing or empty.")
new_preferences_recorded_since_last_upload.clear()
return
try:
print(f"Attempting to upload {LOCAL_RESULTS_FILE} to {DATASET_REPO_ID}/{RESULTS_FILENAME_IN_REPO}")
start_time = time.time()
hf_api.upload_file(
path_or_fileobj=str(LOCAL_RESULTS_FILE),
path_in_repo=RESULTS_FILENAME_IN_REPO,
repo_id=DATASET_REPO_ID,
repo_type="dataset",
commit_message=f"Periodic upload of preferences - {datetime.now().isoformat()}"
)
end_time = time.time()
print(f"Successfully uploaded preferences. Took {end_time - start_time:.2f} seconds.")
new_preferences_recorded_since_last_upload.clear()
except Exception as e:
print(f"Error uploading results file: {e}")
def handle_choice_a(state: dict, request: gr.Request):
return record_preference("A", state, request)
def handle_choice_b(state: dict, request: gr.Request):
return record_preference("B", state, request)
with gr.Blocks(title="Image Composition User Study") as demo:
gr.Markdown("# Image Composition User Study")
gr.Markdown(
"> Please look at the input images and the prompt below. "
"Then, compare the two output images (Output A and Output B) and click the button below the one you prefer."
)
app_state = gr.State({})
available_samples_state = gr.State(INITIAL_SAMPLE_IDS)
status_display = gr.Markdown("**Status:** Loading first sample...")
gr.Markdown("## Inputs")
with gr.Row():
prompt_display = gr.Markdown("**Prompt:** Loading...")
with gr.Row():
with gr.Column():
gr.Markdown("<div style='text-align: center;'>Input Background</div>")
input_bg_display = gr.Image(type="filepath", height=250, width=250, interactive=False, show_label=False)
with gr.Column():
gr.Markdown("<div style='text-align: center;'>Input Foreground</div>")
input_fg_display = gr.Image(type="filepath", height=250, width=250, interactive=False, show_label=False)
gr.Markdown("---")
gr.Markdown("## Choose your preferred output")
with gr.Row():
with gr.Column():
output_a_display = gr.Image(label="Output A", type="filepath", height=400, width=400, interactive=False)
choice_button_a = gr.Button("Choose Output A", variant="primary")
with gr.Column():
output_b_display = gr.Image(label="Output B", type="filepath", height=400, width=400, interactive=False)
choice_button_b = gr.Button("Choose Output B", variant="primary")
next_button = gr.Button("🔁 Next Sample 🔁", visible=False)
demo.load(
fn=display_new_sample,
inputs=[app_state, available_samples_state],
outputs=[
prompt_display, input_bg_display, input_fg_display,
output_a_display, output_b_display,
choice_button_a, choice_button_b, next_button, status_display,
app_state, available_samples_state
]
)
choice_button_a.click(
fn=handle_choice_a,
inputs=[app_state],
outputs=[choice_button_a, choice_button_b, next_button, status_display, app_state],
api_name=False,
)
choice_button_b.click(
fn=handle_choice_b,
inputs=[app_state],
outputs=[choice_button_a, choice_button_b, next_button, status_display, app_state],
api_name=False,
)
next_button.click(
fn=display_new_sample,
inputs=[app_state, available_samples_state],
outputs=[
prompt_display, input_bg_display, input_fg_display,
output_a_display, output_b_display,
choice_button_a, choice_button_b, next_button, status_display,
app_state, available_samples_state
],
api_name=False,
)
def cleanup_temp_dir():
if Path(TEMP_DIR).exists():
print(f"Cleaning up temporary directory: {TEMP_DIR}")
shutil.rmtree(TEMP_DIR, ignore_errors=True)
def shutdown_hook():
print("Application shutting down. Performing final upload check...")
upload_preferences_to_hub()
if scheduler.running:
print("Shutting down scheduler...")
scheduler.shutdown(wait=False)
cleanup_temp_dir()
print("Shutdown complete.")
atexit.register(shutdown_hook)
if __name__ == "__main__":
initialize_hub_and_results()
if not INITIAL_SAMPLE_IDS:
print("Error: No valid samples found in the 'data' directory.")
print("Please ensure the 'data' directory exists and contains subdirectories")
print("named like 'sample_id', each with 'prompt.txt', 'input_bg.*',")
print("'input_fg.*', 'baseline.*', and 'tf-icon.*' files.")
elif not DATASET_REPO_ID:
print("Error: DATASET_REPO_ID environment variable is not set or is set to the default placeholder.")
print("Please set the DATASET_REPO_ID environment variable or update the script.")
elif hf_api:
print(f"Starting periodic upload scheduler (every {UPLOAD_INTERVAL_HOURS} hours)...")
scheduler.add_job(upload_preferences_to_hub, 'interval', hours=UPLOAD_INTERVAL_HOURS)
scheduler.start()
print(f"Found {len(INITIAL_SAMPLE_IDS)} samples.")
print(f"Configured to save results periodically to Hugging Face Dataset: {DATASET_REPO_ID}")
print("Starting Gradio app...")
demo.launch(server_name="0.0.0.0")
else:
print("Warning: Running without Hugging Face Hub integration (HF_TOKEN or DATASET_REPO_ID missing/invalid).")
print(f"Found {len(INITIAL_SAMPLE_IDS)} samples.")
print("Starting Gradio app...")
demo.launch(server_name="0.0.0.0")