126 lines
6.1 KiB
Python
126 lines
6.1 KiB
Python
import torch
|
|
import numpy as np
|
|
from PIL import Image
|
|
from transformers import SegformerImageProcessor, SegformerForSemanticSegmentation
|
|
import cv2 # Imported for BGR to RGB conversion, though PIL can also do it.
|
|
|
|
# Global variables for caching
|
|
HAIR_SEGMENTER_PROCESSOR = None
|
|
HAIR_SEGMENTER_MODEL = None
|
|
MODEL_NAME = "isjackwild/segformer-b0-finetuned-segments-skin-hair-clothing"
|
|
|
|
def segment_hair(image_np: np.ndarray) -> np.ndarray:
|
|
"""
|
|
Segments hair from an image.
|
|
|
|
Args:
|
|
image_np: NumPy array representing the image (BGR format from OpenCV).
|
|
|
|
Returns:
|
|
NumPy array representing the binary hair mask.
|
|
"""
|
|
global HAIR_SEGMENTER_PROCESSOR, HAIR_SEGMENTER_MODEL
|
|
|
|
if HAIR_SEGMENTER_PROCESSOR is None or HAIR_SEGMENTER_MODEL is None:
|
|
print(f"Loading hair segmentation model and processor ({MODEL_NAME}) for the first time...")
|
|
try:
|
|
HAIR_SEGMENTER_PROCESSOR = SegformerImageProcessor.from_pretrained(MODEL_NAME)
|
|
HAIR_SEGMENTER_MODEL = SegformerForSemanticSegmentation.from_pretrained(MODEL_NAME)
|
|
|
|
if torch.cuda.is_available():
|
|
try:
|
|
HAIR_SEGMENTER_MODEL = HAIR_SEGMENTER_MODEL.to('cuda')
|
|
print("INFO: Hair segmentation model moved to CUDA (GPU).")
|
|
except Exception as e_cuda:
|
|
print(f"ERROR: Failed to move hair segmentation model to CUDA: {e_cuda}. Using CPU instead.")
|
|
# Fallback to CPU if .to('cuda') fails
|
|
HAIR_SEGMENTER_MODEL = HAIR_SEGMENTER_MODEL.to('cpu')
|
|
else:
|
|
print("INFO: CUDA not available. Hair segmentation model will use CPU.")
|
|
|
|
print("INFO: Hair segmentation model and processor loaded successfully (device: {}).".format(HAIR_SEGMENTER_MODEL.device))
|
|
except Exception as e:
|
|
print(f"ERROR: Failed to load hair segmentation model/processor: {e}")
|
|
# Return an empty mask compatible with expected output shape (H, W)
|
|
return np.zeros((image_np.shape[0], image_np.shape[1]), dtype=np.uint8)
|
|
|
|
# Convert BGR (OpenCV) to RGB (PIL)
|
|
image_rgb = cv2.cvtColor(image_np, cv2.COLOR_BGR2RGB)
|
|
image_pil = Image.fromarray(image_rgb)
|
|
|
|
inputs = HAIR_SEGMENTER_PROCESSOR(images=image_pil, return_tensors="pt")
|
|
|
|
if HAIR_SEGMENTER_MODEL.device.type == 'cuda':
|
|
try:
|
|
# SegformerImageProcessor output (BatchEncoding) is a dict-like object.
|
|
# We need to move its tensor components, commonly 'pixel_values'.
|
|
if 'pixel_values' in inputs:
|
|
inputs['pixel_values'] = inputs['pixel_values'].to('cuda')
|
|
else: # Fallback if the structure is different than expected
|
|
inputs = inputs.to('cuda')
|
|
# If inputs has other tensor components that need to be moved, they'd need similar handling.
|
|
except Exception as e_inputs_cuda:
|
|
print(f"ERROR: Failed to move inputs to CUDA: {e_inputs_cuda}. Attempting inference on CPU.")
|
|
# If moving inputs to CUDA fails, we should ensure model is also on CPU for this inference pass
|
|
# This is a tricky situation; ideally, this failure shouldn't happen if model moved successfully.
|
|
# For simplicity, we'll assume if model is on CUDA, inputs should also be.
|
|
# A more robust solution might involve moving model back to CPU if inputs can't be moved.
|
|
|
|
with torch.no_grad(): # Important for inference
|
|
outputs = HAIR_SEGMENTER_MODEL(**inputs)
|
|
|
|
logits = outputs.logits # Shape: batch_size, num_labels, height, width
|
|
|
|
# Upsample logits to original image size
|
|
upsampled_logits = torch.nn.functional.interpolate(
|
|
logits,
|
|
size=(image_np.shape[0], image_np.shape[1]), # H, W
|
|
mode='bilinear',
|
|
align_corners=False
|
|
)
|
|
|
|
segmentation_map = upsampled_logits.argmax(dim=1).squeeze().cpu().numpy().astype(np.uint8)
|
|
|
|
# Label 2 is for hair in this model
|
|
return np.where(segmentation_map == 2, 255, 0).astype(np.uint8)
|
|
|
|
if __name__ == '__main__':
|
|
# This is a conceptual test.
|
|
# In a real scenario, you would load an image using OpenCV or Pillow.
|
|
# For example:
|
|
# sample_image_np = cv2.imread("path/to/your/image.jpg")
|
|
# if sample_image_np is not None:
|
|
# hair_mask_output = segment_hair(sample_image_np)
|
|
# cv2.imwrite("hair_mask_output.png", hair_mask_output)
|
|
# print("Hair mask saved to hair_mask_output.png")
|
|
# else:
|
|
# print("Failed to load sample image.")
|
|
|
|
print("Conceptual test: Hair segmenter module created.")
|
|
# Create a dummy image for a basic test run if no image is available.
|
|
dummy_image_np = np.zeros((100, 100, 3), dtype=np.uint8) # 100x100 BGR image
|
|
dummy_image_np[:, :, 1] = 255 # Make it green to distinguish from black mask
|
|
|
|
try:
|
|
print("Running segment_hair with a dummy image...")
|
|
hair_mask_output = segment_hair(dummy_image_np)
|
|
print(f"segment_hair returned a mask of shape: {hair_mask_output.shape}")
|
|
# Check if the output is a 2D array (mask) and has the same H, W as input
|
|
assert hair_mask_output.shape == (dummy_image_np.shape[0], dummy_image_np.shape[1])
|
|
# Check if the mask is binary (0 or 255)
|
|
assert np.all(np.isin(hair_mask_output, [0, 255]))
|
|
print("Dummy image test successful. Hair mask seems to be generated correctly.")
|
|
|
|
# Attempt to save the dummy mask (optional, just for visual confirmation if needed)
|
|
# cv2.imwrite("dummy_hair_mask_output.png", hair_mask_output)
|
|
# print("Dummy hair mask saved to dummy_hair_mask_output.png")
|
|
|
|
except ImportError as e:
|
|
print(f"An ImportError occurred: {e}. This might be due to missing dependencies like transformers, torch, or Pillow.")
|
|
print("Please ensure all required packages are installed by updating requirements.txt and installing them.")
|
|
except Exception as e:
|
|
print(f"An error occurred during the dummy image test: {e}")
|
|
print("This could be due to issues with model loading, processing, or other runtime errors.")
|
|
|
|
print("To perform a full test, replace the dummy image with a real image path.")
|