Got it, let's walk through exactly how to build this plot step by step. I'll include a full, adaptable code example that covers all your requirements: using a background image, plotting your curve, and placing a rotated heatmap (saved via plt.imsave) that only covers part of the background.

Step-by-Step Breakdown & Code #

First, let's import the tools we'll need:

import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
import io  # For in-memory file handling (avoids temp files)

1. Prepare Your Rotated Heatmap

Since you mentioned needing to rotate the heatmap before use, we'll handle that and save it (here's a clean in-memory approach to avoid temp files):

# Replace this with your actual heatmap data (numpy array)
model_prediction_i = np.random.rand(100, 100)  # Example random heatmap

# Rotate the heatmap (adjust k=1 for 90°, k=2 for 180°, etc.)
rotated_heatmap = np.rot90(model_prediction_i, k=1)

# Save rotated heatmap to memory (no temp file needed!)
buf = io.BytesIO()
plt.imsave(buf, rotated_heatmap, cmap='viridis', format='png')  # Pick your preferred colormap
buf.seek(0)
heatmap_img = Image.open(buf)

2. Load Background & Set Up Canvas

We'll create a canvas that matches the background image's dimensions for perfect alignment:

# Load your background image
background = Image.open("reference.jpg")

# Create figure matching background size (dpi=100 ensures 1px = 0.01in)
fig, ax = plt.subplots(
    figsize=(background.width/100, background.height/100),
    dpi=100
)

# Draw the background as the base layer
ax.imshow(background)

3. Plot Your X_test_i Curve

Add your curve on top of the background. Replace the example data with your actual X_test_i values:

# Replace this with your actual X_test_i data (x and y arrays)
x = np.linspace(0, background.width, 200)  # Example x-values spanning background width
y = np.random.randint(50, background.height-50, 200)  # Example y-values

# Plot the curve (customize color, linewidth, label as needed)
ax.plot(x, y, color='#ff3333', linewidth=2.5, label='X_test_i')
ax.legend(fontsize=12, loc='upper right')

4. Position & Overlay the Heatmap

Now place the rotated heatmap on a specific part of the plot. We'll convert pixel coordinates to axis coordinates for precise placement:

# Define where to place the heatmap (pixel coordinates on the background)
# Adjust these values to your desired position/size
heatmap_x_start = 150  # Left edge pixel position
heatmap_y_start = 100  # Bottom edge pixel position
heatmap_width = heatmap_img.width // 2  # Resize to 50% original width
heatmap_height = heatmap_img.height // 2  # Resize to 50% original height

# Convert pixel positions to normalized axis coordinates (matches the background)
x0, y0 = ax.transData.inverted().transform((heatmap_x_start, heatmap_y_start))
x1, y1 = ax.transData.inverted().transform((
    heatmap_x_start + heatmap_width,
    heatmap_y_start + heatmap_height
))

# Add a new axis for the heatmap (no ticks/borders to keep it clean)
heatmap_ax = fig.add_axes([x0, y0, x1 - x0, y1 - y0])
heatmap_ax.axis('off')

# Draw the resized heatmap
heatmap_ax.imshow(heatmap_img.resize((heatmap_width, heatmap_height)))

5. Finalize & Display/Save

Clean up the plot and show or save your final result:

# Optional: Hide the main axis ticks if they're unnecessary
ax.axis('off')

# Show the plot
plt.show()

# Or save it to a file (high DPI for crispness)
# plt.savefig("final_overlay_plot.png", dpi=300, bbox_inches='tight')

Key Notes #

• Why plt.imsave? Saving the rotated heatmap (even in memory) lets us treat it as a raster image, making it easy to resize and position precisely without worrying about array alignment issues.
• Precise Placement: Using ax.transData.inverted().transform() ensures your heatmap aligns perfectly with the background image's pixel grid.
• No Temp Files: The io.BytesIO approach keeps everything in memory, so you won't have leftover files cluttering your workspace.

内容的提问来源于stack exchange，提问作者Noltibus