utils.py

from PIL import Image, ImageOps
from torchvision import transforms

class ScaleAndPadTransform:
    def __init__(self, target_size):
        self.target_size = target_size

    def transform(self, img):
        width, height = img.size
        if width > height:
            scale = self.target_size / width
            new_height = int(height * scale)
            img = img.resize((self.target_size, new_height))
            padding = (self.target_size - new_height) // 2
            img = ImageOps.expand(img, (0, padding, 0, self.target_size - new_height - padding))
        else:
            scale = self.target_size / height
            new_width = int(width * scale)
            img = img.resize((new_width, self.target_size))
            padding = (self.target_size - new_width) // 2
            img = ImageOps.expand(img, (padding, 0, self.target_size - new_width - padding, 0))

        
        IMG_MEAN = [0.485, 0.456, 0.406]
        IMG_STD = [0.229, 0.224, 0.225]

        transform = transforms.Compose([
            transforms.CenterCrop(self.target_size),
            transforms.ToTensor(),
            transforms.Normalize(IMG_MEAN, IMG_STD)
        ])
        img = transform(img)

        return img


class Body_Figure(object):
    def __str__(self):
        return f"Body Figure Information:\n"\
                f" - Waist-to-Shoulder Ratio (WSR): {self.WSR}\n"\
               f" - Waist-to-Thigh Ratio (WTR): {self.WTR}\n"\
               f" - Waist-to-Hip Ratio (WHpR): {self.WHpR}\n"\
               f" - Waist-to-Head Ratio (WHdR): {self.WHdR}\n"\
               f" - Hip-to-Head Ratio (HpHdR): {self.HpHdR}\n"\
               f" - Area: {self.Area}\n"\
               f" - Height-to-Waist Ratio (H2W): {self.H2W}\n"

               
    def __init__(self, waist_width, thigh_width, hip_width, head_width, Area, height, shoulder_width):
        self._waist_width = waist_width
        self._thigh_width = thigh_width
        self._hip_width = hip_width
        self._head_width = head_width
        self._Area = Area
        self._height = height
        self._shoulder_width = shoulder_width
        if self._head_width == 0:
            self._head_width = self._hip_width/3

    @property
    def WSR(self):
        return (self._waist_width) / (self._shoulder_width)

    @property
    def WTR(self):
        return (self._waist_width / self._thigh_width)  # **2

    @property
    def WHpR(self):
        return (self._waist_width / self._hip_width)  # **2

    @property
    def WHdR(self):
        return (self._waist_width / self._head_width)  # **2

    @property
    def HpHdR(self):
        return (self._hip_width / self._head_width)  # **2

    @property
    def Area(self):
        return self._Area

    @property
    def H2W(self):
        return self._height / self._waist_width

import torch
import torch.nn as nn
import torch.optim as optim

def custom_resnet():
    # resnet101
    resnet_model = torch.hub.load('pytorch/vision:v0.10.0', 'resnet101', pretrained=True)

    resnet_model._modules.pop('fc') #1000 fc

    resnet_model.fc1 = nn.Linear(2048, 15)
    # resnet_model.fc1 = nn.Linear(2048, 15)
    resnet_model.fc2 = nn.Sequential(
        nn.ReLU(inplace=True),
        nn.Linear(15, 1)
    )
    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)  # 2048*7*7

        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        x = self.fc1(x)
        x = self.fc2(x)
        return x

    # add new_forward function to the resnet instance as a class method
    bound_method = forward.__get__(resnet_model, resnet_model.__class__)
    setattr(resnet_model, 'forward', bound_method)

    return resnet_model


def custom_resnet_optimizer(resnet_model):
    optimizer = optim.Adam(resnet_model.parameters(), lr=0.0001, betas=(0.9, 0.999), weight_decay=0.001)
    return optimizer


# scaling the longer side of image to 224 and pad the shorter size with zeroes to match 224x224
from PIL import Image, ImageOps

def scale_and_pad(img):
    width, height = img.size
    if width > height:
        scale = 224 / width
        new_height = int(height * scale)
        img = img.resize((224, new_height))
        padding = (224 - new_height) // 2
        img = ImageOps.expand(img, (0, padding, 0, 224 - new_height - padding))
    else:
        scale = 224 / height
        new_width = int(width * scale)
        img = img.resize((new_width, 224))
        padding = (224 - new_width) // 2
        img = ImageOps.expand(img, (padding, 0, 224 - new_width - padding, 0))
    return img


from torchvision import transforms
IMG_SIZE = 224
IMG_MEAN = [0.485, 0.456, 0.406]
IMG_STD = [0.229, 0.224, 0.225]
transform = transforms.Compose([
    transforms.CenterCrop(IMG_SIZE),
    transforms.ToTensor(),
    transforms.Normalize(IMG_MEAN, IMG_STD)
])


from torch.utils.data import Dataset, DataLoader
from PIL import Image
import re
class CustomDataset(Dataset):
    def __init__(self, dataset, transform=None):
        self.data = dataset
        self.transform = transform

    def __len__(self):
        return len(self.data.index)

    def __getitem__(self, idx):
        img_name = self.data.iloc[idx, 0] 
        img_path = 'datasets/Images/' + img_name  # adjust the path to your actual image directory
        image = Image.open(img_path)
        image = scale_and_pad(image)
        ret = re.match(r"\d+?_([FMfm])_(\d+?)_(\d+?)_(\d+).+", img_name)
        BMI = (int(ret.group(4)) / 100000) / (int(ret.group(3)) / 100000) ** 2
        
        if self.transform:
            image = self.transform(image)

        return (image,img_name), BMI


# train the resnet model on the train_img_tensors and train_labels
import torch
import torch.nn as nn
import torch.optim as optim
import numpy as np

from sklearn.metrics import mean_absolute_error
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print(device)

# from detectron2 import detectron2
import numpy as np
import cv2
from detectron2 import model_zoo
from detectron2.engine import DefaultPredictor
from detectron2.config import get_cfg
from detectron2.utils.visualizer import Visualizer
from detectron2.data import MetadataCatalog, DatasetCatalog

# from Human_Parse import HumanParser

# "COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"
# "COCO-Keypoints/keypoint_rcnn_R_50_FPN_3x.yaml"



class Image_Processor(object):

    def __init__(self, masks_file, key_file, key_thresh=0.7):
        
        self._KeypointCfg = self.__init_key(key_file, key_thresh)
        self._KeypointsPredictor = DefaultPredictor(self._KeypointCfg)
        
        self._Contourcfg=self.__init_mask(masks_file,key_thresh)
        self._ContourPredictor = DefaultPredictor(self._Contourcfg)
        
        # self._HumanParser = HumanParser()
        
    def __init_key(self, key_file, key_thresh):
        
        cfg = get_cfg()
        cfg.merge_from_file(model_zoo.get_config_file(key_file))
        cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = key_thresh  # set threshold for this model
        cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url(key_file)
        
        return cfg
    
    def __init_mask(self, mask_file, key_thresh):
        
        cfg = get_cfg()
        cfg.merge_from_file(model_zoo.get_config_file(mask_file))
        cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = key_thresh  # set threshold for this model
        cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url(mask_file)
        return cfg


    def get_keyandcontour_output(self, img):
        
        Keypoints=self._Keypoints_detected(img)
        
        ContourOutput=self._Contour_detected(img)

        # """ Detect Arms Mask by Human parser """
        # Arms_mask = self._HumanParser.Arms_detect(img)
        # ContourOutput = ContourOutput ^ Arms_mask

        return Keypoints, ContourOutput

    def _Contour_detected(self,img):
    
        ContourOutput=self._ContourPredictor(img)
        sorted_idxs = np.argsort(-ContourOutput["instances"].scores.cpu().numpy())
        ContourMasks = None
        for sorted_idx in sorted_idxs:
            if ContourOutput["instances"].pred_classes[sorted_idx] == 0:
                ContourMasks = ContourOutput["instances"].pred_masks[sorted_idx].cpu().numpy()
                
        ContourOutput = ContourMasks
        return ContourOutput
        

    def _Keypoints_detected(self,img):
        
        KeypointsOutput = self._KeypointsPredictor(img)
        sorted_idxs = np.argsort(-KeypointsOutput["instances"].scores.cpu().numpy())
        Keypoints = KeypointsOutput["instances"].pred_keypoints[sorted_idxs[0]].cpu().numpy()
        
        return Keypoints

    # def Process(self, img_RGB):
    def get_figure(self, img):
        Keypoints, ContourOutput = self.get_keyandcontour_output(img)
        
        nose,left_ear,right_ear,left_shoulder,right_shoulder = Keypoints[0],Keypoints[4],Keypoints[3],Keypoints[6], Keypoints[5]
        
        left_hip, right_hip, left_knee, right_knee = Keypoints[12], Keypoints[11], Keypoints[14],Keypoints[13]
        
        y_hip = (left_hip[1] + right_hip[1]) / 2
        y_knee = (left_knee[1] + right_knee[1]) / 2

        center_shoulder = (left_shoulder + right_shoulder) / 2
        
        y_waist = y_hip * 2 / 3 + (nose[1] + center_shoulder[1]) / 6
        
        left_thigh = (left_knee + left_hip) / 2
        right_thigh = (right_knee + right_hip) / 2

        # estimate the waist width
        waist_width = self.waist_width_estimate(center_shoulder, y_waist, ContourOutput)
        
        # estimate the thigh width
        thigh_width = self.thigh_width_estimate(left_thigh, right_thigh, ContourOutput)
        
        # estimate the hip width
        hip_width = self.hip_width_estimate(center_shoulder, y_hip, ContourOutput)
        
        # estimate the head_width
        head_width = self.head_width_estimate(left_ear, right_ear)
        
        # estimate the Area
        Area = self.Area_estimate(y_waist, y_hip, waist_width, hip_width, ContourOutput)
        
        # estimate the height2waist
        height = self.Height_estimate(y_knee, nose[1])
        
        # estimate tht shoulder_width
        shoulder_width = self.shoulder_width_estimate(left_shoulder, right_shoulder)

        figure = Body_Figure(waist_width, thigh_width, hip_width, head_width, Area, height, shoulder_width)
        
#         outputs = self._KeypointsPredictor(img)
#         v = Visualizer(img[:,:,::-1], MetadataCatalog.get( self._KeypointCfg.DATASETS.TRAIN[0]), scale=1.2)
#         out = v.draw_instance_predictions(outputs["instances"].to("cpu"))
#         # cv2_imshow(out.get_image()[:, :, ::-1])
#         cv2.imwrite('random.jpg', out.get_image()[:, :, ::-1])
        
#         outputs = self._ContourPredictor(img)
#         v = Visualizer(img[:,:,::-1], MetadataCatalog.get( self._Contourcfg.DATASETS.TRAIN[0]), scale=1.2)
#         out = v.draw_instance_predictions(outputs["instances"].to("cpu"))
#         # cv2_imshow(out.get_image()[:, :, ::-1])
#         cv2.imwrite('random1.jpg', out.get_image()[:, :, ::-1])

        return figure

    def Height_estimate(self, y_k, y_n):
        Height = np.abs(y_n - y_k)
        return Height

    def Area_estimate(self, y_w, y_h, W_w, H_w, mask):
        # '''
        #     Area is expressed as thenumber of
        #     pixels per unit area between waist and hip
        # '''
        try:
            pixels = np.sum(mask[int(y_w):int(y_h)][:])
        except:
            pixels=100
        
        area = (y_h - y_w) * 0.5 * (W_w + H_w)
        Area = pixels / area
        return Area

    def shoulder_width_estimate(self, left_shoulder, right_shoulder):
        shoulder_width = np.sqrt((right_shoulder[0] - left_shoulder[0]) ** 2 + (right_shoulder[1] - left_shoulder[1]) ** 2)
        return shoulder_width

    def head_width_estimate(self, left_ear, right_eat):
        head_width = np.sqrt((right_eat[0] - left_ear[0]) ** 2 + (right_eat[1] - left_ear[1]) ** 2)
        return head_width

    def hip_width_estimate(self, center_shoulder, y_hip, ContourOutput):
        x_hip_center = int(center_shoulder[0])
        try:
            x_lhb = np.where(ContourOutput[int(y_hip)][:x_hip_center] == 0)[0]
            x_lhb = x_lhb[-1] if len(x_lhb) else 0
        except:
            x_lhb = 10
        try:
            x_rhb = np.where(ContourOutput[int(y_hip)][x_hip_center:] == 0)[0]
            x_rhb = x_rhb[0] + x_hip_center if len(x_rhb) else len(ContourOutput[0])
        except:
            x_rhb = 5
        hip_width = x_rhb - x_lhb
        return hip_width

    def thigh_width_estimate(self, left_thigh, right_thigh, mask):
        lx, ly = int(left_thigh[0]), int(left_thigh[1])
        rx, ry = int(right_thigh[0]), int(right_thigh[1])
        try:
            x_ltb = np.where(mask[ly][:lx] == 0)[0]
            x_ltb = x_ltb[-1] if len(x_ltb) else 0
        except:
            x_ltb = 10
        try:
            
            x_rtb = np.where(mask[ry][rx:] == 0)[0]
            x_rtb = x_rtb[0] + rx if len(x_rtb) else len(mask[0])
        except:
            x_rtb = 0
        l_width = (lx - x_ltb) * 2
        r_width = (x_rtb - rx) * 2

        thigh_width = (l_width + r_width) / 2
        return thigh_width

    def waist_width_estimate(self, center_shoulder, y_waist, ContourOutput):
        x_waist_center = int(center_shoulder[0])
        # plt.imshow(ContourOutput)
        # plt.show()
        try:
            x_lwb = np.where(ContourOutput[int(y_waist)][:x_waist_center] == 0)[0]
            x_lwb = x_lwb[-1] if len(x_lwb) else 0
        except:
            x_lwb = 10
            print("err waist width")
        try:
            x_rwb = np.where(ContourOutput[int(y_waist)][x_waist_center:] == 0)[0]
            x_rwb = x_rwb[0] + x_waist_center if len(x_rwb) else len(ContourOutput[0])
        except:
            x_rwb=0
            print("err waist width")
        # print(x_rwb)
        waist_width = x_rwb - x_lwb
        return waist_width

import numpy as np
import pandas
import cv2
from PIL import Image
import torchvision.models.detection
from torchvision.models.detection import maskrcnn_resnet50_fpn, MaskRCNN_ResNet50_FPN_Weights



class Data_Processor(object):
    
    def __init__(self,mask_model="COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml",
        keypoints_model = "COCO-Keypoints/keypoint_rcnn_R_50_FPN_3x.yaml"):
        
        
        self._img_pro = Image_Processor(mask_model,keypoints_model)
        

    def get_image_info(self,df):
        return df
    
    def test(self,img):
        # img = cv2.imread(img_path)
        img = np.array(img)
        figure = self._img_pro.get_figure(img)
        return figure

class LayerActivations:
    features = None

    def __init__(self, model, layer_num):
        self.hook = model.register_forward_hook(self.hook_fn)

    def hook_fn(self, module, input, output):
        self.features = output

    def remove(self):
        self.hook.remove()