utils/utils.py

import types

import numpy as np
import streamlit as st
import torch
from distinctipy import distinctipy
from segment_anything import (SamAutomaticMaskGenerator, SamPredictor,
                              sam_model_registry)
from torch.nn import functional as F


def get_color():
    return distinctipy.get_colors(200)


def medsam_preprocess(self, x: torch.Tensor) -> torch.Tensor:
    """Normalize pixel values and pad to a square input."""
    # Normalize colors
    x = (x - x.min()) / torch.clip(
        x.max() - x.min(), min=1e-8, max=None)  # normalize to [0, 1], (H, W, 3)

    # Pad
    h, w = x.shape[-2:]
    padh = self.image_encoder.img_size - h
    padw = self.image_encoder.img_size - w
    x = F.pad(x, (0, padw, 0, padh))
    return x


@st.cache_resource
def get_model(checkpoint='checkpoint/sam_vit_b_01ec64.pth'):
    device         = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model          = sam_model_registry['vit_b'](checkpoint=checkpoint)

    # Replace preprocess function
    funcType = types.MethodType
    model.preprocess = funcType(medsam_preprocess, model)
    model.mask_threshold = 0.5

    model = model.to(device)
    if torch.cuda.is_available():
        torch.cuda.empty_cache()

    predictor      = SamPredictor(model)
    mask_generator = SamAutomaticMaskGenerator(model)
    return predictor, mask_generator


def show_everything(sorted_anns):
    if len(sorted_anns) == 0:
        return np.array([])
    #sorted_anns = sorted(anns, key=(lambda x: x['stability_score']), reverse=True)
    h, w        = sorted_anns[0]['segmentation'].shape[-2:]
    #sorted_anns = sorted_anns[:int(len(sorted_anns) * stability_score/100)]
    mask = np.zeros((h,w,4))
    for ann in sorted_anns:
        m      = ann['segmentation']
        color  = np.concatenate([np.random.random(3), np.array([0.6])], axis=0)
        mask  += m.reshape(h,w,1) * color.reshape(1, 1, -1)
    mask = mask * 255
    return mask.astype(np.uint8)


def show_click(masks, colors):
    h, w         = masks[0].shape[-2:]
    masks_total = np.zeros((h,w,4)).astype(np.uint8)

    for mask, color in zip(masks, colors):
        if np.array_equal(mask,np.array([])):continue
        masks = np.zeros((h,w,4)).astype(np.uint8)
        masks = masks + mask.reshape(h,w,1).astype(np.uint8)
        masks = masks.astype(bool).astype(np.uint8)
        masks = masks * 255 * color.reshape(1, 1, -1)
        masks_total += masks.astype(np.uint8)

    return masks_total

def model_predict_masks_click(model,input_points,input_labels):
    if input_points == []:return np.array([])
    input_labels = np.array(input_labels)
    input_points = np.array(input_points)
    masks, _, _ = model.predict(
        point_coords=input_points,
        point_labels=input_labels,
        multimask_output=False,
    )

    if torch.cuda.is_available():
        torch.cuda.empty_cache()

    return masks

def model_predict_masks_box(model,center_point,center_label,input_box):
    masks = np.array([])
    for i in range(len(center_label)):
        if center_point[i] == []:continue
        center_point_1 = np.array([center_point[i]])
        center_label_1 = np.array(center_label[i])
        input_box_1 = np.array(input_box[i])
        mask, _, _ = model.predict(
            point_coords=center_point_1,
            point_labels=center_label_1,
            box=input_box_1,
            multimask_output=False,
        )
        try:
            masks = masks + mask
        except:
            masks = mask

    if torch.cuda.is_available():
        torch.cuda.empty_cache()

    return masks


def model_predict_masks_everything(mask_generator, image):
    masks = mask_generator.generate(image)

    if torch.cuda.is_available():
        torch.cuda.empty_cache()

    return masks