nikraf's picture
Upload folder using huggingface_hub
714cf46 verified
#!/usr/bin/env python3
from __future__ import annotations
import argparse
import json
import math
from pathlib import Path
from typing import Dict, List, Tuple
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns
from visualization.utils import CLS_PREFS, DATASET_NAMES, MODEL_NAMES, REG_PREFS
def is_regression(metrics: Dict[str, float]) -> bool:
"""Heuristic based on key names."""
reg = ("spearman", "pearson", "r_squared", "rmse", "mse")
cls = ("accuracy", "f1", "mcc", "auc", "precision", "recall")
# Filter out time-related metrics
filtered_metrics = {k: v for k, v in metrics.items()
if 'training_time' not in k.lower() and 'time_seconds' not in k.lower()}
keys = {k.lower() for k in filtered_metrics}
if any(k for k in keys if any(r in k for r in reg)):
return True
if any(k for k in keys if any(c in k for c in cls)):
return False
return False # default to classification
def pick_metric(metrics: Dict[str, float], prefs: List[Tuple[str, str]]) -> Tuple[str, str]:
"""Return (key, pretty_name) for the first preference present in metrics."""
for k, nice in prefs:
for mk in metrics:
# Skip time-related metrics
if 'training_time' in mk.lower() or 'time_seconds' in mk.lower():
continue
if mk.lower().endswith(k):
return k, nice
raise KeyError("No preferred metric found.")
def parse_metric_value(value) -> Tuple[float, float]:
"""
Parse a metric value that may be in 'mean±std' format or a plain number.
Returns (mean, std) where std is 0.0 if not present.
"""
if isinstance(value, str) and '±' in value:
parts = value.split('±')
try:
mean_val = float(parts[0])
std_val = float(parts[1]) if len(parts) > 1 else 0.0
return mean_val, std_val
except ValueError:
return math.nan, 0.0
elif isinstance(value, (int, float)):
return float(value), 0.0
return math.nan, 0.0
def get_metric_value(metrics: Dict[str, float], key_suffix: str) -> float:
"""Fetch metric value case-/prefix-insensitively; NaN if absent.
For mean±std format, returns only the mean value."""
for k, v in metrics.items():
# Skip time-related metrics and _mean/_std suffixed keys
if 'training_time' in k.lower() or 'time_seconds' in k.lower():
continue
if k.lower().endswith('_mean') or k.lower().endswith('_std'):
continue
if k.lower().endswith(key_suffix):
mean_val, _ = parse_metric_value(v)
return mean_val
return math.nan
def get_metric_value_with_std(metrics: Dict[str, float], key_suffix: str) -> Tuple[float, float, str]:
"""
Fetch metric value with std case-/prefix-insensitively.
Returns (mean, std, display_string) where display_string is formatted for heatmap display.
"""
for k, v in metrics.items():
# Skip time-related metrics and _mean/_std suffixed keys
if 'training_time' in k.lower() or 'time_seconds' in k.lower():
continue
if k.lower().endswith('_mean') or k.lower().endswith('_std'):
continue
if k.lower().endswith(key_suffix):
mean_val, std_val = parse_metric_value(v)
if std_val > 0:
display_str = f"{mean_val:.2f}±{std_val:.2f}"
else:
display_str = f"{mean_val:.2f}"
return mean_val, std_val, display_str
return math.nan, 0.0, ""
def radar_factory(n_axes: int):
theta = np.linspace(0, 2 * np.pi, n_axes, endpoint=False)
fig, ax = plt.subplots(figsize=(10, 10), subplot_kw={"polar": True})
ax.set_theta_offset(np.pi / 2)
ax.set_theta_direction(-1)
return fig, ax, theta
def plot_radar(*,
categories: List[str],
models: List[str],
data: List[List[float]],
title: str,
outfile: Path,
normalize: bool = False):
# Use pretty names for categories (datasets) and models
pretty_categories = [DATASET_NAMES.get(cat, cat) for cat in categories]
pretty_models = [MODEL_NAMES.get(m, m) for m in models]
if normalize:
arr = np.asarray(data)
rng = np.where(np.ptp(arr, axis=0) == 0, 1, np.ptp(arr, axis=0))
data = (arr - arr.min(0)) / rng
# Convert back to list of lists for consistency
data = data.tolist()
# append mean column (do this after normalization if normalize=True)
pretty_categories = pretty_categories + ["Avg"]
data = [row + [np.nanmean(row)] for row in data]
fig, ax, theta = radar_factory(len(pretty_categories))
ax.set_thetagrids(np.degrees(theta), pretty_categories, fontsize=11)
ax.set_ylim(0, 1.0)
ax.set_yticks(np.linspace(0, 1, 11))
palette = [plt.cm.tab20(i / len(pretty_models)) for i in range(len(pretty_models))]
for i, (m, vals) in enumerate(zip(pretty_models, data)):
ang = np.concatenate([theta, [theta[0]]])
val = np.concatenate([vals, [vals[0]]])
ax.plot(ang, val, lw=2, label=m, color=palette[i])
ax.fill(ang, val, alpha=.25, color=palette[i])
ax.grid(True)
plt.title(title, pad=20)
plt.legend(bbox_to_anchor=(1.25, 1.05))
plt.tight_layout()
plt.savefig(outfile, dpi=450, bbox_inches="tight")
plt.close(fig)
def bar_plot(datasets: List[str],
models: List[str],
data: List[List[float]],
metric_name: str,
outfile: Path):
rows = [
{"Dataset": DATASET_NAMES.get(d, d), "Model": MODEL_NAMES.get(m, m), "Score": s}
for m, col in zip(models, data)
for d, s in zip(datasets, col)
]
dfp = pd.DataFrame(rows)
plt.figure(figsize=(max(12, .8 * len(datasets)), 8))
sns.barplot(dfp, x="Dataset", y="Score", hue="Model")
plt.title(f"{metric_name} across datasets (Cls→F1, Reg→Spearman)")
plt.xticks(rotation=45, ha="right")
plt.tight_layout()
plt.savefig(outfile, dpi=450, bbox_inches="tight")
plt.close()
def normalize_per_dataset(arr: np.ndarray) -> np.ndarray:
"""
Normalize array per dataset (row-wise).
Args:
arr: Array of shape (num_datasets, num_models)
Returns:
Normalized array of same shape, with each row normalized to [0, 1]
"""
normalized_data = np.zeros_like(arr)
for i in range(arr.shape[0]):
lowest_performance = np.nanmin(arr[i, :])
best_performance = np.nanmax(arr[i, :])
denom = best_performance - lowest_performance
denom = 1 if denom == 0 else denom
normalized_data[i, :] = (arr[i, :] - lowest_performance) / denom
return normalized_data
def heatmap_plot(datasets: List[str],
models: List[str],
data: List[List[float]],
metric_name: str,
outfile: Path,
normalize: bool = False,
display_strings: List[List[str]] = None,
no_std: bool = False):
"""
Create a heatmap plot.
Args:
datasets: List of dataset names
models: List of model names
data: List of lists of mean values (for coloring)
metric_name: Name of the metric being plotted
outfile: Output file path
normalize: Whether to normalize display values
display_strings: Optional list of lists of display strings (e.g., "0.85±0.01").
If provided, these are used for annotations instead of raw values.
"""
arr = np.array(data).T # shape: (num_datasets, num_models)
# Compute average row (mean across datasets for each model)
avg_row = np.nanmean(arr, axis=0, keepdims=True)
arr_with_avg = np.vstack([arr, avg_row])
datasets_plus_avg = datasets + ['Average']
# Clean display names
clean_model_names = [MODEL_NAMES.get(m, m) for m in models]
clean_dataset_names = [DATASET_NAMES.get(d, d) for d in datasets_plus_avg]
print(clean_dataset_names)
print(datasets_plus_avg)
# Build display string matrix if provided
if display_strings is not None and not no_std:
# Transpose to match arr shape: (num_datasets, num_models)
display_arr = np.array(display_strings).T.tolist()
# Add average row display strings
avg_display = []
for j in range(len(models)):
model_vals = [arr[i, j] for i in range(arr.shape[0]) if not math.isnan(arr[i, j])]
if model_vals:
avg_display.append(f"{np.mean(model_vals):.4f}")
else:
avg_display.append("")
display_arr.append(avg_display)
else:
display_arr = None
# For annotations: use normalized or original values based on normalize parameter
if normalize:
# Normalize values for display in annotations
normalized_data = normalize_per_dataset(arr)
# Add average row to normalized data
avg_row_norm = np.nanmean(normalized_data, axis=0, keepdims=True)
annot_arr = np.vstack([normalized_data, avg_row_norm])
annot_label = 'Normalized Performance (0-1)'
# Don't use display_strings for normalized view
display_arr = None
else:
annot_arr = arr_with_avg
annot_label = metric_name
# Always normalize colors per row (dataset) for visualization
# This creates a color array where each row is scaled 0-1
color_arr = np.zeros_like(arr_with_avg)
for i in range(arr_with_avg.shape[0]):
row_min = np.nanmin(arr_with_avg[i, :])
row_max = np.nanmax(arr_with_avg[i, :])
denom = row_max - row_min
if denom == 0 or np.isnan(denom):
color_arr[i, :] = 0.5 # neutral color if all values are the same
else:
color_arr[i, :] = (arr_with_avg[i, :] - row_min) / denom
# Calculate figure size based on content
# Increase cell width if we have mean±std strings or normalized values
has_std = display_arr is not None
#if normalize:
# cell_width = 1.1
#elif has_std:
# cell_width = 1.1
#else:
# cell_width = 0.85 # standard width for .2f format
cell_width = 1.3
cell_height = 0.7 # height per cell in inches
fig_width = max(8, cell_width * len(clean_model_names))
fig_height = max(6, cell_height * len(clean_dataset_names) + 1)
fig, ax = plt.subplots(figsize=(fig_width, fig_height))
# Create heatmap with per-row normalized colors (blue = bad, orange = good)
from matplotlib.colors import LinearSegmentedColormap
colors = ['#3498db', '#85c1e9', '#FFD700'] # Blue -> Light Blue -> Yellow
n_bins = 100
cmap = LinearSegmentedColormap.from_list('blue_yellow', colors, N=n_bins)
im = ax.imshow(color_arr, cmap=cmap, aspect='auto', vmin=0, vmax=1)
# Add colorbar with "Worst to Best" label
cbar = plt.colorbar(im, ax=ax)
cbar.set_label('Worst to Best', fontsize=16)
cbar.set_ticks([0, 0.5, 1])
cbar.set_ticklabels(['Worst', 'Mid', 'Best'], fontsize=11)
# Set ticks and labels
ax.set_xticks(np.arange(len(clean_model_names)))
ax.set_yticks(np.arange(len(clean_dataset_names)))
ax.set_xticklabels(clean_model_names, rotation=45, ha='right', fontsize=16)
ax.set_yticklabels(clean_dataset_names, rotation=0, fontsize=16)
# Add value annotations
# Use smaller font if displaying mean±std
font_size = 10 if has_std else 16
for i in range(annot_arr.shape[0]):
for j in range(annot_arr.shape[1]):
if display_arr is not None and i < len(display_arr) and j < len(display_arr[i]):
text_str = display_arr[i][j]
else:
if i == annot_arr.shape[0] - 1: # Average row
text_str = f'{annot_arr[i, j]:.4f}'
else:
text_str = f'{annot_arr[i, j]:.2f}'
text = ax.text(j, i, text_str,
ha="center", va="center", color="black", fontsize=font_size)
# Add black boxes around best performers in each row
for i in range(color_arr.shape[0]):
if not np.all(np.isnan(color_arr[i, :])):
best_idx = np.nanargmax(color_arr[i, :])
ax.add_patch(plt.Rectangle((best_idx - 0.5, i - 0.5), 1, 1,
fill=False, edgecolor='black', lw=3))
# Set appropriate title
if normalize:
title = f'{annot_label} Heatmap (Cls→F1, Reg→Spearman)\nColors normalized per dataset'
else:
title = f'{annot_label} Heatmap (Cls→F1, Reg→Spearman)\nColors normalized per dataset'
plt.title(title, pad=20, fontsize=21)
plt.ylabel('Dataset', fontsize=17)
plt.xlabel('Model', fontsize=17)
plt.tight_layout()
plt.savefig(outfile, dpi=450, bbox_inches='tight')
plt.close()
def load_tsv(tsv: Path) -> pd.DataFrame:
df = pd.read_csv(tsv, sep="\t")
for c in df.columns:
if c != "dataset":
df[c] = df[c].apply(json.loads)
return df
def create_plots(tsv: str, outdir: str, no_std: bool = False):
tsv, outdir = Path(tsv), Path(outdir)
df = load_tsv(tsv)
models = [c for c in df.columns if c != "dataset"]
# Resolve metric per-dataset (MCC or R², w/ fallbacks).
datasets, scores_by_model = [], {m: [] for m in models}
display_by_model = {m: [] for m in models} # For mean±std display strings
dataset_types = [] # Track which type each dataset is
for _, row in df.iterrows():
name = row["dataset"]
metrics0 = row[models[0]]
task = "regression" if is_regression(metrics0) else "classification"
dataset_types.append(task)
prefs = REG_PREFS if task == "regression" else CLS_PREFS
try:
suffix, pretty = pick_metric(metrics0, prefs)
except KeyError:
print(f"[WARN] {name}: no suitable metric – skipped.")
continue
datasets.append(name)
for m in models:
if no_std:
mean_val, std_val, display_str = get_metric_value_with_std(row[m], suffix)
display_str = f"{mean_val:.2f}"
else:
mean_val, std_val, display_str = get_metric_value_with_std(row[m], suffix)
scores_by_model[m].append(mean_val)
display_by_model[m].append(display_str)
if not datasets:
raise RuntimeError("No plottable datasets found.")
# Check if we have only one type of dataset
only_classification = all(t == "classification" for t in dataset_types)
only_regression = all(t == "regression" for t in dataset_types)
# Order datasets according to DATASET_NAMES keys
ordered_datasets = []
ordered_scores = {m: [] for m in models}
ordered_display = {m: [] for m in models} # For mean±std display strings
ordered_types = [] # Keep track of ordered dataset types
# First add datasets that are in DATASET_NAMES in their defined order
for ds in DATASET_NAMES.keys():
if ds in datasets:
idx = datasets.index(ds)
ordered_datasets.append(ds)
ordered_types.append(dataset_types[idx])
for m in models:
ordered_scores[m].append(scores_by_model[m][idx])
ordered_display[m].append(display_by_model[m][idx])
# Then add any remaining datasets that weren't in DATASET_NAMES
for ds in datasets:
if ds not in ordered_datasets:
ordered_datasets.append(ds)
idx = datasets.index(ds)
ordered_types.append(dataset_types[idx])
for m in models:
ordered_scores[m].append(scores_by_model[m][idx])
ordered_display[m].append(display_by_model[m][idx])
# Replace original lists with ordered ones
datasets = ordered_datasets
scores_by_model = ordered_scores
display_by_model = ordered_display
dataset_types = ordered_types
# assemble lists in model order
plot_matrix = [scores_by_model[m] for m in models]
display_matrix = [display_by_model[m] for m in models]
# Sort models by average score (ascending: worst to best)
model_avgs = [np.nanmean(scores) for scores in plot_matrix]
sorted_indices = np.argsort(model_avgs)
sorted_models = [models[i] for i in sorted_indices]
sorted_plot_matrix = [plot_matrix[i] for i in sorted_indices]
sorted_display_matrix = [display_matrix[i] for i in sorted_indices]
# For normalized heatmap, sort based on normalized averages
arr_for_norm = np.array(plot_matrix).T # shape: (num_datasets, num_models)
normalized_data = normalize_per_dataset(arr_for_norm)
# Calculate normalized averages for each model (column-wise mean)
normalized_model_avgs = np.nanmean(normalized_data, axis=0)
sorted_indices_norm = np.argsort(normalized_model_avgs)
sorted_models_norm = [models[i] for i in sorted_indices_norm]
sorted_plot_matrix_norm = [plot_matrix[i] for i in sorted_indices_norm]
sorted_display_matrix_norm = [display_matrix[i] for i in sorted_indices_norm]
fig_tag = tsv.stem
outdir = outdir / fig_tag
outdir.mkdir(parents=True, exist_ok=True)
# File paths for all plot types
radar_path = outdir / f"{fig_tag}_radar_all.png"
radar_path_norm = outdir / f"{fig_tag}_radar_all_normalized.png"
bar_path = outdir / f"{fig_tag}_bar_all.png"
bar_path_norm = outdir / f"{fig_tag}_bar_all_normalized.png"
heatmap_path = outdir / f"{fig_tag}_heatmap_all.png"
heatmap_path_norm = outdir / f"{fig_tag}_heatmap_all_normalized.png"
# Set subtitle and metric name based on dataset types
if only_classification:
subtitle = "Classification datasets plot F1"
metric_name = "F1"
elif only_regression:
subtitle = "Regression datasets plot Spearman rho"
metric_name = "Spearman rho"
else:
subtitle = "Classification datasets plot F1; Regression datasets plot Spearman rho"
metric_name = "F1 / Spearman rho"
# Radar plot keeps original order
plot_radar(categories=datasets,
models=models,
data=plot_matrix,
title=subtitle,
outfile=radar_path,
normalize=False)
plot_radar(categories=datasets,
models=models,
data=plot_matrix,
title=subtitle + " (Normalized)",
outfile=radar_path_norm,
normalize=True)
# Bar and heatmap use sorted order
bar_plot(datasets, sorted_models, sorted_plot_matrix, metric_name, bar_path)
# Normalized bar plot
# For bar plot normalization, use min-max per dataset (column-wise normalization)
arr = np.asarray(sorted_plot_matrix)
rng = np.where(np.ptp(arr, axis=0) == 0, 1, np.ptp(arr, axis=0))
arr_norm = (arr - arr.min(0)) / rng
bar_plot(datasets, sorted_models, arr_norm.tolist(), metric_name + " (Normalized)", bar_path_norm)
# Heatmap - pass display strings for mean±std annotation
heatmap_plot(datasets, sorted_models, sorted_plot_matrix, metric_name, heatmap_path,
normalize=False, display_strings=sorted_display_matrix, no_std=no_std)
# Normalized heatmap uses sorting based on normalized averages
heatmap_plot(datasets, sorted_models_norm, sorted_plot_matrix_norm, metric_name, heatmap_path_norm,
normalize=True, display_strings=sorted_display_matrix_norm, no_std=no_std)
print(f"Radar saved to {radar_path}")
print(f"Radar (normalized) saved to {radar_path_norm}")
print(f"Bar saved to {bar_path}")
print(f"Bar (normalized) saved to {bar_path_norm}")
print(f"Heatmap saved to {heatmap_path}")
print(f"Heatmap (normalized) saved to {heatmap_path_norm}")
def main() -> None:
ap = argparse.ArgumentParser(description="Generate radar, bar, and heatmap plots for all datasets. Always saves both normalized and unnormalized versions.")
ap.add_argument("--input", required=True, help="TSV file with metrics")
ap.add_argument("--output_dir", default="plots", help="Directory for plots")
ap.add_argument("--no_std", action="store_true", help="Do not display standard deviation in heatmap plots")
args = ap.parse_args()
create_plots(Path(args.input), Path(args.output_dir), no_std=args.no_std)
print("Finished.")
if __name__ == "__main__":
# py -m visualization.plot_result
# Check if input file is provided and run main to generate plots
import sys
if "--input" in sys.argv:
main()
else:
# --- TESTS FOR PLOTTING FUNCTIONS ---
print("\nRunning plot function tests...")
from pathlib import Path
tmpdir = Path("plots/test_plots")
tmpdir.mkdir(parents=True, exist_ok=True)
# Dummy data
categories = ["A", "B", "C"]
models = ["Model1", "Model2"]
data = [
[0.8, 0.6, 0.7],
[0.5, 0.9, 0.4],
]
# Radar plot
radar_path = tmpdir / "test_radar.png"
plot_radar(categories=categories, models=models, data=data, title="Test Radar", outfile=radar_path)
assert radar_path.exists(), "Radar plot not created!"
print(f"Radar plot test passed: {radar_path}")
# Normalized radar plot
radar_path_norm = tmpdir / "test_radar_normalized.png"
plot_radar(categories=categories, models=models, data=data, title="Test Radar (Normalized)", outfile=radar_path_norm, normalize=True)
assert radar_path_norm.exists(), "Normalized radar plot not created!"
print(f"Normalized radar plot test passed: {radar_path_norm}")
# Bar plot
bar_path = tmpdir / "test_bar.png"
bar_plot(categories, models, data, "Test Metric", bar_path)
assert bar_path.exists(), "Bar plot not created!"
print(f"Bar plot test passed: {bar_path}")
# Normalized bar plot
arr = np.asarray(data)
rng = np.where(np.ptp(arr, axis=0) == 0, 1, np.ptp(arr, axis=0))
arr_norm = (arr - arr.min(0)) / rng
bar_path_norm = tmpdir / "test_bar_normalized.png"
bar_plot(categories, models, arr_norm.tolist(), "Test Metric (Normalized)", bar_path_norm)
assert bar_path_norm.exists(), "Normalized bar plot not created!"
print(f"Normalized bar plot test passed: {bar_path_norm}")
# Heatmap plot
heatmap_path = tmpdir / "test_heatmap.png"
heatmap_plot(categories, models, data, "Test Metric", heatmap_path)
assert heatmap_path.exists(), "Heatmap plot not created!"
print(f"Heatmap plot test passed: {heatmap_path}")
# Normalized heatmap plot
heatmap_path_norm = tmpdir / "test_heatmap_normalized.png"
heatmap_plot(categories, models, data, "Test Metric", heatmap_path_norm, normalize=True)
assert heatmap_path_norm.exists(), "Normalized heatmap plot not created!"
print(f"Normalized heatmap plot test passed: {heatmap_path_norm}")
# Heatmap plot with mean±std display strings
display_strings = [
["0.80±0.02", "0.60±0.01", "0.70±0.03"],
["0.50±0.05", "0.90±0.02", "0.40±0.01"],
]
heatmap_path_std = tmpdir / "test_heatmap_with_std.png"
heatmap_plot(categories, models, data, "Test Metric", heatmap_path_std, display_strings=display_strings)
assert heatmap_path_std.exists(), "Heatmap with std not created!"
print(f"Heatmap with std test passed: {heatmap_path_std}")
print("All plot function tests passed!\n")