kitooo's picture
initial commit
fc9bdaf verified
raw
history blame
2.08 kB
import torch
import gradio as gr
import matplotlib.pyplot as plt
from PIL import Image
from transformers import SamModel, SamProcessor
device = 'cuda' if torch.cuda.is_available() else 'cpu'
processor = SamProcessor.from_pretrained('facebook/sam-vit-base')
model = SamModel.from_pretrained('hmdliu/sidewalks-seg-base')
model.to(device)
def segment_sidewalk(image, threshold):
# init data
width, height = image.size
prompt = [0, 0, width, height]
inputs = processor(image, input_boxes=[[prompt]], return_tensors='pt')
# make prediction
outputs = model(pixel_values=inputs['pixel_values'].to(device),
input_boxes=inputs['input_boxes'].to(device),
multimask_output=False)
prob_map = torch.sigmoid(outputs.pred_masks.squeeze()).cpu().detach()
prediction = (prob_map > threshold).float()
prob_map, prediction = prob_map.numpy(), prediction.numpy()
# visualize results
save_image(image, 'image.png')
save_image(prob_map, 'prob.png', cmap='jet')
save_image(prediction, 'mask.png', cmap='gray')
return Image.open('image.png'), Image.open('mask.png'), Image.open('prob.png')
def save_image(image, path, **kwargs):
plt.figure(figsize=(8, 8))
plt.imshow(image, interpolation='nearest', **kwargs)
plt.axis('off')
plt.tight_layout()
plt.savefig(path, bbox_inches='tight', pad_inches=0)
plt.close()
with gr.Blocks() as demo:
with gr.Row():
with gr.Column():
image_input = gr.Image(type='pil', label='TIFF Image')
threshold_slider = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.5, label='Prediction Threshold')
segment_button = gr.Button('Segment')
with gr.Column():
prediction = gr.Image(type='pil', label='Segmentation Result')
prob_map = gr.Image(type='pil', label='Probability Map')
segment_button.click(
segment_image,
inputs=[image_input, threshold_slider],
outputs=[image_input, prediction, prob_map]
)
demo.launch(debug=True, show_error=True)