Li-zhangsheng's picture
Upload model code and weights
83ab733 verified
Raw History Blame Contribute Delete
20.7 kB
import torch
import torch.nn as nn
from PIL import Image
import os
import numpy as np
from typing import Tuple, List, Optional
import logging
# Import the model architecture used for prediction; it must match the checkpoint.
from proposed_uncertainty import UNet
from scipy import io
import matplotlib.pyplot as plt
from scipy.ndimage import zoom
import cv2
import time
class Predictor:
"""
Predictor for M and D matrices.
"""
def __init__(self, model_path: str, crop_size=None, device: Optional[str] = None):
"""
Initialize the predictor.
Args:
model_path: Path to the model weights.
crop_size: Crop dimensions.
device: Device to use ('cuda' or 'cpu'); select automatically when None.
"""
# Configure logging.
self._setup_logger()
# Configure the device.
self.device = self._setup_device(device)
self.logger.info(f"Using device: {self.device}")
# Process the crop dimensions.
if crop_size is not None:
if isinstance(crop_size, (int, float)):
self.crop_height = self.crop_width = int(crop_size)
elif isinstance(crop_size, (tuple, list)) and len(crop_size) == 2:
self.crop_height, self.crop_width = crop_size
else:
raise ValueError("crop_size should be an integer or a tuple of (height, width)")
self.logger.info(f"Using crop size: ({self.crop_height}, {self.crop_width})")
else:
self.crop_height = self.crop_width = None
self.logger.info("Cropping disabled")
# Load the model.
try:
self.model = self._load_model(model_path)
self.logger.info("Model loaded successfully")
except Exception as e:
self.logger.error(f"Failed to load model: {str(e)}")
raise
def _setup_logger(self):
"""Configure logging."""
self.logger = logging.getLogger('Predictor')
self.logger.setLevel(logging.INFO)
if not self.logger.handlers:
handler = logging.StreamHandler()
formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
handler.setFormatter(formatter)
self.logger.addHandler(handler)
def _setup_device(self, device: Optional[str]) -> torch.device:
"""Configure the compute device."""
if device is None:
device = 'cuda' if torch.cuda.is_available() else 'cpu'
return torch.device(device)
def _load_model(self, model_path: str) -> nn.Module:
"""Load the model."""
if not os.path.exists(model_path):
raise FileNotFoundError(f"Model file not found: {model_path}")
# Create the model.
model = UNet(n_channels=3, bilinear=False)
checkpoint = torch.load(model_path, map_location=self.device)
model.load_state_dict(checkpoint['model_state_dict'])
model = model.to(self.device)
model.eval()
return model
def _process_images(self, img1_path: str, img2_path: str, img3_path: str) -> Tuple[torch.Tensor, Tuple[int, int]]:
"""
Convert three images into network input.
Returns the processed tensor and original dimensions:
(tensor, (orig_height, orig_width)).
"""
# Read the three images.
img1 = np.array(Image.open(img1_path), dtype=np.float32)
img2 = np.array(Image.open(img2_path), dtype=np.float32)
img3 = np.array(Image.open(img3_path), dtype=np.float32)
# Preserve the original dimensions.
orig_height, orig_width = img1.shape
# Calculate sina and cosa.
S = 3
sina = np.zeros_like(img1, dtype=np.float32)
cosa = np.zeros_like(img1, dtype=np.float32)
for n in range(1, S + 1):
sina -= [img1, img2, img3][n - 1] * np.sin(2 * np.pi * (n - 1) / S)
cosa += [img1, img2, img3][n - 1] * np.cos(2 * np.pi * (n - 1) / S)
# Normalize by dividing directly by 255.
# sina = sina / 1000
# cosa = cosa / 1000
img1 = img1 / 1000
img2 = img2 / 1000
img3 = img3 / 1000
# Crop when requested.
if self.crop_height is not None and self.crop_width is not None:
H, W = img1.shape
start_h = (H - self.crop_height) // 2
start_w = (W - self.crop_width) // 2
img1 = img1[start_h:start_h + self.crop_height, start_w:start_w + self.crop_width]
img2 = img2[start_h:start_h + self.crop_height, start_w:start_w + self.crop_width]
img3 = img3[start_h:start_h + self.crop_height, start_w:start_w + self.crop_width]
img1_tensor = torch.from_numpy(img1).float().unsqueeze(0) # [1, H, W]
img2_tensor = torch.from_numpy(img2).float().unsqueeze(0) # [1, H, W]
img3_tensor = torch.from_numpy(img3).float().unsqueeze(0) # [1, H, W]
input_tensor = torch.cat([img1_tensor, img2_tensor, img3_tensor], dim=0)
return input_tensor, (orig_height, orig_width)
def create_mask(self, img1_path: str, img2_path: str, img3_path: str) -> Tuple[np.ndarray, Tuple[int, int]]:
"""
Create a mask that is zero where all three images have zero-valued pixels.
Returns the mask and original dimensions: (mask, (orig_height, orig_width)).
"""
img1 = np.array(Image.open(img1_path))
img2 = np.array(Image.open(img2_path))
img3 = np.array(Image.open(img3_path))
# Preserve the original dimensions.
orig_height, orig_width = img1.shape
# Create the mask.
mask = np.ones_like(img1, dtype=np.float32)
mask[(img1 == 0) & (img2 == 0) & (img3 == 0)] = 0
# Crop when requested.
if self.crop_height is not None and self.crop_width is not None:
H, W = mask.shape
start_h = (H - self.crop_height) // 2
start_w = (W - self.crop_width) // 2
mask = mask[start_h:start_h + self.crop_height, start_w:start_w + self.crop_width]
return mask, (orig_height, orig_width)
def _restore_to_original_size(self, data: np.ndarray, orig_size: Tuple[int, int]) -> np.ndarray:
"""
Restore cropped data to the center of the original dimensions.
Args:
data: Cropped data.
orig_size: Original dimensions as (height, width).
Returns:
Data restored to the original dimensions.
"""
if self.crop_height is None or self.crop_width is None:
return data
orig_height, orig_width = orig_size
result = np.zeros((orig_height, orig_width), dtype=data.dtype)
# Calculate the placement coordinates.
start_h = (orig_height - self.crop_height) // 2
start_w = (orig_width - self.crop_width) // 2
# Place the cropped data in the center.
result[start_h:start_h + self.crop_height, start_w:start_w + self.crop_width] = data
return result
def predict_single(self, img1_path: str, img2_path: str, img3_path: str) -> Tuple[np.ndarray, np.ndarray]:
"""
Predict the M and D matrices once.
"""
self.logger.info("Starting single prediction...")
try:
# Process the input images.
inputs, orig_size = self._process_images(img1_path, img2_path, img3_path)
mask, _ = self.create_mask(img1_path, img2_path, img3_path)
mask = torch.from_numpy(mask).unsqueeze(0) # [1, H, W]
# Use evaluation mode during prediction to disable dropout.
self.model.eval()
with torch.no_grad():
inputs = inputs.unsqueeze(0).to(self.device) # [1, 3, H, W]
outputs= self.model(inputs)
# Process the outputs.
M = outputs['M_mean'].squeeze().cpu().numpy() # [H, W]
D = outputs['D_mean'].squeeze().cpu().numpy() # [H, W]
# Apply the mask.
M *= mask.squeeze().numpy()
D *= mask.squeeze().numpy()
# Restore the original dimensions.
M = self._restore_to_original_size(M, orig_size)
D = self._restore_to_original_size(D, orig_size)
# Create a mask with the original dimensions.
full_mask, _ = self.create_mask(img1_path, img2_path, img3_path)
if self.crop_height is not None and self.crop_width is not None:
# Reconstruct the full mask.
orig_height, orig_width = orig_size
full_mask = np.ones((orig_height, orig_width), dtype=np.float32)
full_mask[(np.array(Image.open(img1_path)) == 0) &
(np.array(Image.open(img2_path)) == 0) &
(np.array(Image.open(img3_path)) == 0)] = 0
# Apply the full mask.
M *= full_mask
D *= full_mask
self.logger.info("Single prediction completed")
return {'M': M, 'D': D}
except Exception as e:
self.logger.error(f"An error occurred during prediction: {str(e)}")
raise
def predict_with_uncertainty(self, img1_path: str, img2_path: str, img3_path: str, n_samples=50) -> dict:
"""
Predict with Monte Carlo sampling, including uncertainty estimation.
"""
self.logger.info(f"Starting Monte Carlo prediction (samples: {n_samples})...")
try:
# Process the input images.
inputs, orig_size = self._process_images(img1_path, img2_path, img3_path)
mask, _ = self.create_mask(img1_path, img2_path, img3_path)
mask = torch.from_numpy(mask).unsqueeze(0) # [1, H, W]
# Store Monte Carlo samples.
M_samples = []
D_samples = []
var_M_samples = []
var_D_samples = []
# Monte Carlo sampling uses training mode to enable dropout.
self.model.train()
inputs = inputs.unsqueeze(0).to(self.device) # [1, 3, H, W]
with torch.no_grad():
for _ in range(n_samples):
outputs = self.model(inputs)
M_samples.append(outputs['M_mean'].cpu().numpy())
D_samples.append(outputs['D_mean'].cpu().numpy())
var_M_samples.append(outputs['M_var'].cpu().numpy())
var_D_samples.append(outputs['D_var'].cpu().numpy())
# Convert to NumPy arrays.
M_samples = np.concatenate(M_samples, axis=0)
D_samples = np.concatenate(D_samples, axis=0)
var_M_samples = np.concatenate(var_M_samples, axis=0)
var_D_samples = np.concatenate(var_D_samples, axis=0)
# Calculate means and uncertainties.
M_mean = np.mean(M_samples, axis=0).squeeze()
D_mean = np.mean(D_samples, axis=0).squeeze()
var_M_mean = np.mean(var_M_samples, axis=0).squeeze()
var_D_mean = np.mean(var_D_samples, axis=0).squeeze()
# Calculate model and data uncertainties.
M_model_uncertainty = np.sqrt(np.mean((M_samples - M_mean) ** 2, axis=0)).squeeze()
D_model_uncertainty = np.sqrt(np.mean((D_samples - D_mean) ** 2, axis=0)).squeeze()
M_data_uncertainty = np.sqrt(np.mean(var_M_samples, axis=0)).squeeze()
D_data_uncertainty = np.sqrt(np.mean(var_D_samples, axis=0)).squeeze()
M_model_uncertainty = M_model_uncertainty
D_model_uncertainty = D_model_uncertainty
M_data_uncertainty = M_data_uncertainty
D_data_uncertainty = D_data_uncertainty
M_total_uncertainty = (M_model_uncertainty ** 2 + M_data_uncertainty ** 2)
D_total_uncertainty = (D_model_uncertainty ** 2 + D_data_uncertainty ** 2)
# Apply the mask.
mask = mask.squeeze().numpy()
M_mean *= mask
D_mean *= mask
M_model_uncertainty *= mask
D_model_uncertainty *= mask
M_data_uncertainty *= mask
D_data_uncertainty *= mask
M_total_uncertainty *= mask
D_total_uncertainty *= mask
var_M_mean *= mask
var_D_mean *= mask
# Create a mask with the original dimensions.
full_mask = np.ones(orig_size, dtype=np.float32)
full_mask[(np.array(Image.open(img1_path)) == 0) &
(np.array(Image.open(img2_path)) == 0) &
(np.array(Image.open(img3_path)) == 0)] = 0
# Restore the original dimensions.
M_mean = self._restore_to_original_size(M_mean, orig_size)
D_mean = self._restore_to_original_size(D_mean, orig_size)
M_model_uncertainty = self._restore_to_original_size(M_model_uncertainty, orig_size)
D_model_uncertainty = self._restore_to_original_size(D_model_uncertainty, orig_size)
M_data_uncertainty = self._restore_to_original_size(M_data_uncertainty, orig_size)
D_data_uncertainty = self._restore_to_original_size(D_data_uncertainty, orig_size)
M_total_uncertainty = self._restore_to_original_size(M_total_uncertainty, orig_size)
D_total_uncertainty = self._restore_to_original_size(D_total_uncertainty, orig_size)
var_M_mean = self._restore_to_original_size(var_M_mean, orig_size)
var_D_mean = self._restore_to_original_size(var_D_mean, orig_size)
# Apply the full mask to the restored results.
M_mean *= full_mask
D_mean *= full_mask
M_model_uncertainty *= full_mask
D_model_uncertainty *= full_mask
M_data_uncertainty *= full_mask
D_data_uncertainty *= full_mask
M_total_uncertainty *= full_mask
D_total_uncertainty *= full_mask
var_M_mean *= full_mask
var_D_mean *= full_mask
# Visualize uncertainty.
self._visualize_uncertainties({
'M': M_mean,
'D': D_mean,
'M_model_uncertainty': M_model_uncertainty,
'D_model_uncertainty': D_model_uncertainty,
'M_data_uncertainty': M_data_uncertainty,
'D_data_uncertainty': D_data_uncertainty,
'M_total_uncertainty': M_total_uncertainty,
'D_total_uncertainty': D_total_uncertainty
})
# Do not process the 3D sample arrays further.
self.logger.info("Monte Carlo prediction completed")
return {
'M': M_mean,
'D': D_mean,
'M_var': var_M_mean,
'D_var': var_D_mean,
'M_model_uncertainty': M_model_uncertainty,
'D_model_uncertainty': D_model_uncertainty,
'M_data_uncertainty': M_data_uncertainty,
'D_data_uncertainty': D_data_uncertainty,
'M_total_uncertainty':M_total_uncertainty,
'D_total_uncertainty':D_total_uncertainty
}
except Exception as e:
self.logger.error(f"An error occurred during prediction: {str(e)}")
raise
finally:
# Return to evaluation mode after sampling.
self.model.eval()
def _visualize_uncertainties(self, results: dict, save_dir: str = 'uncertainty_analysis'):
"""
Visualize the uncertainty-analysis results.
Args:
results: Value returned by predict_with_uncertainty.
save_dir: Output directory.
"""
os.makedirs(save_dir, exist_ok=True)
plt.figure(figsize=(20, 10))
# M prediction and uncertainty.
plt.subplot(2, 4, 1)
plt.imshow(results['M'])
plt.colorbar()
plt.title('Predicted M')
plt.subplot(2, 4, 2)
plt.imshow(results['M_model_uncertainty'])
plt.colorbar()
plt.title('Model Uncertainty (M)')
plt.subplot(2, 4, 3)
plt.imshow(results['M_data_uncertainty'])
plt.colorbar()
plt.title('Data Uncertainty (M)')
plt.subplot(2, 4, 4)
total_M = results['M_total_uncertainty']
plt.imshow(total_M)
plt.colorbar()
plt.title('Total Uncertainty (M)')
# D prediction and uncertainty.
plt.subplot(2, 4, 5)
plt.imshow(results['D'])
plt.colorbar()
plt.title('Predicted D')
plt.subplot(2, 4, 6)
plt.imshow(results['D_model_uncertainty'])
plt.colorbar()
plt.title('Model Uncertainty (D)')
plt.subplot(2, 4, 7)
plt.imshow(results['D_data_uncertainty'])
plt.colorbar()
plt.title('Data Uncertainty (D)')
plt.subplot(2, 4, 8)
total_D = results['D_total_uncertainty']
plt.imshow(total_D)
plt.colorbar()
plt.title('Total Uncertainty (D)')
plt.tight_layout()
plt.savefig(os.path.join(save_dir, 'uncertainty_analysis.png'))
plt.close()
self.logger.info(f"Uncertainty analysis plot saved to: {save_dir}")
def save_prediction(self, results, save_path: str, mode='single'):
"""
Save prediction results to a .mat file.
Args:
results: Prediction results as a tuple or dictionary.
save_path: Output path.
mode: Prediction mode, either 'single' or 'mc'.
"""
try:
os.makedirs(os.path.dirname(save_path), exist_ok=True)
if mode == 'single':
io.savemat(save_path, results)
else: # mode == 'mc'
io.savemat(save_path, results)
self.logger.info(f"Prediction results saved to: {save_path}")
except Exception as e:
self.logger.error(f"Failed to save results: {str(e)}")
raise
if __name__ == "__main__":
# Create the predictor and set the checkpoint used for inference.
predictor = Predictor(
model_path="path/to/your/model_weights.pth",
crop_size=(1360,672),
device='cuda'
)
# Set paths to the three input frames.
img1_path = 'path/to/your/input/image1.png'
img2_path = 'path/to/your/input/image2.png'
img3_path = 'path/to/your/input/image3.png'
# Single-prediction example.
try:
result = predictor.predict_single(img1_path, img2_path, img3_path)
predictor.save_prediction(result, 'pre_uncertainty/single_result.mat', mode='single')
except Exception as e:
print(f"Single prediction failed: {str(e)}")
# Display the mask.
mask, orig_size = predictor.create_mask(img1_path, img2_path, img3_path)
# Restore the mask to its original dimensions when cropping is enabled.
if predictor.crop_height is not None and predictor.crop_width is not None:
full_mask = np.ones(orig_size, dtype=np.float32)
full_mask[(np.array(Image.open(img1_path)) == 0) &
(np.array(Image.open(img2_path)) == 0) &
(np.array(Image.open(img3_path)) == 0)] = 0
mask_to_save = full_mask
else:
mask_to_save = mask
# Save the mask as an image.
cv2.imwrite('pre_uncertainty/mask.png', mask_to_save*255)
# Save the mask as a .mat file.
io.savemat('pre_uncertainty/mask.mat', {'mask':mask_to_save})
plt.imshow(mask_to_save)
# Monte Carlo sampling example.
try:
results = predictor.predict_with_uncertainty(img1_path, img2_path, img3_path, n_samples=50)
print("\nMonte Carlo prediction result shapes:")
for key, value in results.items():
print(f"{key}: {value.shape}")
predictor.save_prediction(results, 'pre_uncertainty/mc_result.mat', mode='mc')
predictor._visualize_uncertainties(results)
except Exception as e:
print(f"Monte Carlo prediction failed: {str(e)}")