Fine-tuning-data / analyze_data.py
ch-min's picture
Upload folder using huggingface_hub
849ca03 verified
import json
import os
import re
import pandas as pd
from collections import Counter
import argparse
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np
import colorsys # For creating color variations
import random
# --- 1. Configuration Settings ---
BASE_DATA_DIR = '.'
BASE_METADATA_DIR = 'metadata'
# UPDATED: Output directory structure
OUTPUT_DIR = 'analyzed_results' # Base output directory
ERQA_OUTPUT_DIR = os.path.join(OUTPUT_DIR, 'ERQA') # Subdirectory for ERQA plots
DATASET_CATEGORIES = {
"SYNTHETIC": "synthetic_SAT_Spatial457_PRISM_80.0k",
"REAL": "real_SPAR-7M_RoboSpatial_80.0k",
"STATIC": "static_RoboSpatial_PRISM_Spatial457_SAT_SPAR-7M_80.0k",
"DYNAMIC": "dynamic_Spatial457_SAT_SPAR-7M_80k",
"PERCEPTION": "perception_Spatial457_SAT_SPAR-7M_80k",
"REASONING": "reasoning_RoboSpatial_PRISM_Spatial457_SAT_SPAR-7M_80.0k",
"_2D": "2d_Spatial457_SAT_SPAR-7M_80k",
"_3D": "3d_RoboSpatial_PRISM_Spatial457_SAT_SPAR-7M_80.0k",
}
# --- 2. ERQA Task Mapping & Colors ---
ERQA_TASK_MAPPING_FROM_FILENAME = {
"Spatial Reasoning": [
'relative_spatial', # SAT: Asks about left/right/above/below relationships.
'relative_depth', # SAT: Asks about closer/further, higher/lower relationships.
'obj_spatial_relation', # SPAR-7M: Asks about relative object locations (e.g., left, below, farther).
'robospatial_configuration',# RoboSpatial: Asks simple positional questions (e.g., A in front of B?).
'Spatial457_all', # Spatial457: Likely contains general spatial relationship questions.
'static', # Spatial457: Focuses on static scenes, implies spatial relationships.
'dynamic',
'2D', # Spatial457: 2D spatial relationships.
'3D', # Spatial457: 3D spatial relationships.
],
"Action Reasoning": [
'action_consequence', # SAT: Asks about the outcome of an action (e.g., facing away after turning?).
'robospatial_compatibility',# RoboSpatial: Asks if an action is possible (e.g., can X fit Y?).
'point_prediction', # PRISM: Asks for the grasp point *to accomplish a specific task*.
],
"Trajectory Reasoning": [
'egocentric_movement', # SAT: Asks how the camera moved/rotated.
'object_movement', # SAT: Asks if/how objects moved.
'goal_aiming', # SAT: Asks which way to turn to face a target.
'dynamic', # Spatial457: Likely involves predicting outcomes of movement (e.g., collision).
],
"State Estimation": [
'count', # SAT, SPAR-7M: Asks for the number of objects.
'perception', # Spatial457: Asks about object properties (color, shape, size).
'relative_spatial'
# Note: Spatial457's 2D/3D/static/reasoning also contain property questions, reinforcing this link.
],
"Task Reasoning": [
'reasoning', # Spatial457: Contains broader reasoning Qs, some task-related (e.g., size comparison for collision). Fits ERQA's broad definition.
'robospatial_compatibility'
# Note: action_consequence could also weakly fit here.
],
"Multi-view Reasoning": [
'allocentric_perspective', # SAT: Asks about spatial relationships from a different imagined viewpoint.
'spatial_imagination', # SPAR-7M: Asks how relationships change after observer movement.
'obj_spatial_relation'
# Note: obj_spatial_relation might include multi-view variants based on original SPAR-7M file names,
# but the grouped name 'obj_spatial_relation' doesn't specify view count.
],
"Pointing": [
'point_prediction', # RoboSpatial: Asks to pinpoint multiple points in a specified vacant area.
],
"Other": [
# Fine-tuning tasks that don't clearly fit above.
# Currently, all major tasks seem reasonably mapped.
]
}
# ERQA_TASK_MAPPING_FROM_FILENAME = {
# "Spatial Reasoning": ['relative_spatial', 'relative_depth', 'obj_spatial_relation', 'robospatial_configuration', 'Spatial457_all', 'static', '2D', '3D'],
# "Action Reasoning": ['action_consequence', 'robospatial_compatibility', 'point_prediction'], # Includes PRISM
# "Trajectory Reasoning": ['egocentric_movement', 'object_movement', 'goal_aiming', 'dynamic'],
# "State Estimation": ['count', 'perception'],
# "Task Reasoning": ['reasoning'],
# "Multi-view Reasoning": ['allocentric_perspective', 'spatial_imagination'],
# "Pointing": ['point_prediction'], # Includes RoboSpatial context
# "Other": []
# }
# Define NEW representative colors for ERQA tasks
ERQA_COLORS = {
"Spatial Reasoning": "#1f77b4", # ์„ ๋ช…ํ•œ ํŒŒ๋ž‘ (Blue)
"Action Reasoning": "#ff7f0e", # ์„ ๋ช…ํ•œ ์ฃผํ™ฉ (Orange)
"Trajectory Reasoning": "#2ca02c", # ์„ ๋ช…ํ•œ ์ดˆ๋ก (Green)
"State Estimation": "#d62728", # ์„ ๋ช…ํ•œ ๋นจ๊ฐ• (Red)
"Task Reasoning": "#9467bd", # ์„ ๋ช…ํ•œ ๋ณด๋ผ (Purple)
"Multi-view Reasoning": "#8c564b", # ๊ฐˆ์ƒ‰ (Brown)
"Pointing": "#e377c2", # ๋ถ„ํ™ (Pink)
"Other": "#17becf" # ์ฒญ๋ก (Teal/Cyan)
}
# --- 3. Helper Functions ---
def categorize_answer(answer_str):
# (Existing function - unchanged)
if not isinstance(answer_str, str): return "not_a_string"
parsed_answer = answer_str
if answer_str.startswith('{"Reasoning":'):
try: data = json.loads(answer_str); parsed_answer = str(data.get('Answer', answer_str)).strip()
except json.JSONDecodeError: parsed_answer = answer_str
parsed_answer = parsed_answer.strip()
if "<point x=" in parsed_answer: return "point_prediction_prism"
if parsed_answer.startswith('[(') and parsed_answer.endswith(')]'): return "point_list_robospatial"
if parsed_answer.lower() in ['yes', 'no']: return "yes_no"
if parsed_answer.upper() in ['A', 'B', 'C', 'D'] and len(parsed_answer) == 1: return "multiple_choice_letter_spar7m"
if re.fullmatch(r"^\d+$", parsed_answer) and len(parsed_answer)<4 : return "number_only"
if parsed_answer.lower() in ['left', 'right', 'above', 'below', 'front', 'behind', 'back']: return "simple_direction_sat"
if 'no objects moved' in parsed_answer or 'was moved' in parsed_answer: return "object_movement_sat"
if "From the observer's perspective" in parsed_answer or "the Object" in parsed_answer: return "descriptive_sentence_spar7m"
if len(parsed_answer) < 50: return "short_answer_other"
return "long_descriptive_other"
def map_filename_to_task(file_name, dataset_name):
# (Existing function - unchanged)
base_name, _ = os.path.splitext(file_name)
if 'obj_count' in base_name or base_name == 'count': return 'count'
elif base_name == 'robospatial_context': return 'point_prediction'
elif base_name == 'train_data' and dataset_name == 'PRISM': return 'point_prediction'
elif 'obj_spatial_relation' in base_name: return 'obj_spatial_relation'
elif 'spatial_imagination' in base_name: return 'spatial_imagination'
elif base_name.endswith('_tasks'): return base_name.replace('_tasks', '')
elif base_name == 'Spatial457_all': return 'Spatial457_all'
else: return base_name
def generate_color_variations(base_hex, num_variations, min_sat=0.3, max_sat=0.9, min_val=0.45, max_val=0.9):
# (Updated function - unchanged from previous version with color adjustments)
is_gray = base_hex.lower() in ["#808080", "#7f7f7f", "gray", "grey"]
if is_gray: min_sat=0.0; max_sat=0.05; min_val=0.35; max_val=0.75
base_rgb = tuple(int(base_hex.lstrip('#')[i:i+2], 16) / 255.0 for i in (0, 2, 4))
base_h, base_l, base_s = colorsys.rgb_to_hls(*base_rgb)
variations = []
for i in range(num_variations):
l_ratio = i / max(1, num_variations - 1) if num_variations > 1 else 0.5
s_ratio = i / max(1, num_variations - 1) if num_variations > 1 else 0.5
l = min_val + (max_val - min_val) * l_ratio
s = min_sat + (max_sat - min_sat) * s_ratio
l = max(0.0, min(1.0, l)); s = max(0.0, min(1.0, s))
r, g, b = colorsys.hls_to_rgb(base_h, l, s)
variations.append(f"#{int(r*255):02x}{int(g*255):02x}{int(b*255):02x}")
if num_variations > 1 : random.shuffle(variations)
return variations
# --- 4. Analysis Functions ---
def analyze_answer_formats():
# (Existing function - unchanged)
print("--- Starting Answer Format Analysis ---")
all_results = {}
for name, file_prefix in DATASET_CATEGORIES.items():
file_path = os.path.join(BASE_DATA_DIR, f"{file_prefix}.json")
if not os.path.exists(file_path): all_results[name] = {"status": "File Not Found"}; continue
gpt_answers = []
try:
with open(file_path, 'r', encoding='utf-8') as f: data = json.load(f)
for item in data:
convs = item.get('conversations', [])
for conv in convs:
if conv.get('from') == 'gpt': gpt_answers.append(conv.get('value')); break
if not gpt_answers: all_results[name] = {"status": "No 'gpt' answers"}; continue
total = len(gpt_answers)
cats = [categorize_answer(ans) for ans in gpt_answers]
counts = Counter(cats)
percentages = {cat: round(count / total * 100, 2) for cat, count in counts.items()}
all_results[name] = {"status": "Success", "category_counts_percentages": percentages}
except Exception as e: all_results[name] = {"status": f"Error: {e}"}
return all_results
def analyze_question_tasks():
# (Existing function - unchanged)
print("--- Starting Question Task Analysis ---")
all_results = {}
for name, file_prefix in DATASET_CATEGORIES.items():
metadata_path = os.path.join(BASE_METADATA_DIR, f"{file_prefix}_metadata.json")
if not os.path.exists(metadata_path): all_results[name] = {"status": "Metadata Not Found"}; continue
try:
with open(metadata_path, 'r', encoding='utf-8') as f: metadata = json.load(f)
task_counts = Counter(); total_entries = 0
if 'source_files' not in metadata: all_results[name] = {"status": "Invalid Metadata"}; continue
for src in metadata['source_files']:
fname, dname = src.get('file_name'), src.get('dataset_name')
entries = src.get('sampled_entries', 0)
if fname and dname:
task = map_filename_to_task(fname, dname)
task_counts[task] += entries; total_entries += entries
if total_entries == 0: all_results[name] = {"status": "No Sampled Entries"}; continue
percentages = {task: round(count / total_entries * 100, 2) for task, count in task_counts.items()}
all_results[name] = {"status": "Success", "task_counts_percentages": percentages}
except Exception as e: all_results[name] = {"status": f"Error: {e}"}
return all_results
# --- 5. Plotting Function ---
def plot_distribution(analysis_results, analysis_type, highlight_tasks=None, base_color=None, save_path=None):
"""Generates and saves a stacked bar plot, optionally highlighting specific tasks."""
print(f"\n--- Generating Plot for {analysis_type} Distribution ---")
plot_data = {}; all_categories = set()
data_key = "category_counts_percentages" if analysis_type == "answer_format" else "task_counts_percentages"
for dataset, info in analysis_results.items():
if info['status'] == 'Success' and data_key in info:
plot_data[dataset] = info[data_key]
all_categories.update(info[data_key].keys())
else: plot_data[dataset] = {}
df = pd.DataFrame(plot_data).fillna(0).T
if df.empty or len(all_categories) == 0: print("!!! ERROR: No data available for plotting."); return
plot_df = df[sorted(list(all_categories))] # Include all categories initially
num_total_categories = len(plot_df.columns)
# --- Generate Colors (Highlighting or Default) ---
color_map = {}
if highlight_tasks and base_color:
target_tasks = [task for task in highlight_tasks if task in plot_df.columns]
other_tasks = [task for task in plot_df.columns if task not in target_tasks]
target_colors = generate_color_variations(base_color, len(target_tasks))
gray_variations = generate_color_variations("#808080", len(other_tasks)) # Use the updated gray generator
for i, task in enumerate(target_tasks): color_map[task] = target_colors[i % len(target_colors)]
for i, task in enumerate(other_tasks): color_map[task] = gray_variations[i % len(gray_variations)]
# Order columns: highlighted first (sorted), then others (sorted)
plot_df = plot_df[sorted(target_tasks) + sorted(other_tasks)]
else: # Default plotting (no highlighting)
# Apply Top N + Other logic ONLY for default plots
category_totals = plot_df.sum().sort_values(ascending=False)
top_n = 20
top_categories = category_totals.head(top_n).index.tolist()
other_categories = category_totals.iloc[top_n:].index.tolist()
if other_categories:
plot_df['Other'] = plot_df[other_categories].sum(axis=1)
plot_df = plot_df[top_categories + ['Other']]
else: plot_df = plot_df[top_categories]
plot_df = plot_df[sorted(plot_df.columns)] # Sort final columns
num_colors = len(plot_df.columns)
default_colors = sns.color_palette("tab20", n_colors=num_colors) if num_colors <= 20 else sns.color_palette("husl", n_colors=num_colors)
color_map = {col: default_colors[i % len(default_colors)] for i, col in enumerate(plot_df.columns)}
plot_colors = [color_map[col] for col in plot_df.columns]
# --- Plotting ---
plt.style.use('seaborn-v0_8-whitegrid')
fig, ax = plt.subplots(figsize=(14, 8))
plot_df.plot(kind='bar', stacked=True, ax=ax, color=plot_colors, width=0.8)
# --- Formatting ---
ax.set_xlabel("Dataset Category", fontsize=12, fontweight='bold')
ax.set_ylabel("Percentage Distribution (%)", fontsize=12, fontweight='bold')
# Use basename of save_path for title suffix if highlighting
title_suffix = f" (Highlighting: {os.path.basename(save_path).replace('.png','')})" if highlight_tasks else ""
title = f"Distribution of {analysis_type.replace('_', ' ').title()} Across Datasets{title_suffix}"
ax.set_title(title, fontsize=16, fontweight='bold', pad=20)
ax.tick_params(axis='x', rotation=45, labelsize=11); ax.tick_params(axis='y', labelsize=11)
ax.set_ylim(0, 100); ax.yaxis.grid(True, linestyle='--', alpha=0.7); ax.xaxis.grid(False)
# --- Legend ---
legend_title = f"{analysis_type.replace('_', ' ').title()} Categories"
ax.legend(title=legend_title, bbox_to_anchor=(1.02, 1), loc='upper left', borderaxespad=0., fontsize='small', title_fontsize='medium')
plt.tight_layout(rect=[0, 0, 0.83, 1])
# --- Save the plot ---
plt.savefig(save_path, dpi=300, bbox_inches='tight')
print(f"Plot saved as {save_path}")
plt.close(fig)
# --- 6. Main Execution Block ---
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Analyze fine-tuning dataset distributions.")
group = parser.add_mutually_exclusive_group(required=True)
group.add_argument(
"--analysis_type", type=str, choices=['answer', 'question'],
help="Type of analysis: 'answer' (formats) or 'question' (tasks)."
)
group.add_argument(
"--ERQA", action='store_true',
help="Generate individual plots highlighting tasks relevant to each ERQA category."
)
args = parser.parse_args()
# Create output directories recursively
os.makedirs(OUTPUT_DIR, exist_ok=True)
os.makedirs(ERQA_OUTPUT_DIR, exist_ok=True)
if args.analysis_type == 'answer':
results = analyze_answer_formats()
# UPDATED: Save path is now under analyzed_results/
save_path = os.path.join(OUTPUT_DIR, "answer_format_distribution.png")
plot_distribution(results, 'answer_format', save_path=save_path)
elif args.analysis_type == 'question':
results = analyze_question_tasks()
# UPDATED: Save path is now under analyzed_results/
save_path = os.path.join(OUTPUT_DIR, "question_task_distribution.png")
plot_distribution(results, 'question_task', save_path=save_path)
elif args.ERQA:
print("--- Starting ERQA Task Highlighting Analysis ---")
question_task_results = analyze_question_tasks()
if not question_task_results or all(v['status'] != 'Success' for v in question_task_results.values()):
print("!!! ERROR: Could not get valid question task data for ERQA plotting.")
else:
for erqa_task_name, relevant_ft_tasks in ERQA_TASK_MAPPING_FROM_FILENAME.items():
if not relevant_ft_tasks and erqa_task_name != "Other": continue # Skip ERQA Other
print(f"\n--- Generating plot for ERQA Task: {erqa_task_name} ---")
base_color = ERQA_COLORS.get(erqa_task_name, "#7f7f7f") # Default to gray
# Define save path within ERQA subdirectory
erqa_save_filename = f"{erqa_task_name.replace(' ', '_')}.png"
# UPDATED: Save path uses ERQA_OUTPUT_DIR
erqa_save_path = os.path.join(ERQA_OUTPUT_DIR, erqa_save_filename)
# Call plot function with highlighting
plot_distribution(
analysis_results=question_task_results,
analysis_type='question_task',
highlight_tasks=relevant_ft_tasks,
base_color=base_color,
save_path=erqa_save_path
)
print("\n--- Analysis finished ---")