548 KiB
548 KiB
None
<html>
<head>
</head>
</html>
Question 7 - FasterRCNN¶
In [1]:
import os
import glob
import cv2
import torch
import torchvision
import numpy as np
from PIL import Image
import pandas as pd
from pathlib import Path
from tqdm import tqdm
from collections import Counter
from xml.etree import ElementTree as ET
from torchvision import transforms, models
from torch.utils.data import Dataset, DataLoader
In [2]:
MODEL_NAME = 'data/weights_masks.pth'
class_dict = {"with_mask": 1,
"without_mask": 2,
"mask_weared_incorrect": 3}
class_dct_inv = {1: 'with_mask',
2: "without_mask",
3: "mask_weared_incorrect"}
In [3]:
import albumentations as A #note installing this changes your version of OpenCV and you loose sift
from albumentations.pytorch import ToTensorV2
# bbox_params = A.BboxParams(
# format = 'pascal_voc',
# label_fields = ['labels'],
# min_visibility = 0.3
# )
# test_transform = A.Compose([
# A.Resize(height=480, width=480),
# # A.Normalize(
# # mean=[0.485, 0.456, 0.406],
# # std=[0.229, 0.224, 0.225]),
# ToTensorV2(p=1.0)],
# bbox_params = bbox_params
# )
bbox_params = A.BboxParams(
format='pascal_voc',
min_area=5, # 5
min_visibility=0.9, # 0.9
# check_each_transform =True,
label_fields=['labels']
)
test_transform = A.Compose([
A.Resize(height=446, width=446),
ToTensorV2()],
bbox_params=bbox_params,
is_check_shapes=False
)
Load the model¶
In [4]:
from torchvision.models.detection.faster_rcnn import FastRCNNPredictor
In [5]:
def get_model_detection(num_classes):
model = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True)
in_features = model.roi_heads.box_predictor.cls_score.in_features
model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes)
return model
In [6]:
device='cuda' if torch.cuda.is_available() else 'cpu'
model = get_model_detection(4)
model.load_state_dict(torch.load(MODEL_NAME, map_location=torch.device('cpu')), False)
model.eval()
model.to(device)
Out[6]:
Load the dataset¶
In [7]:
from bs4 import BeautifulSoup
class MaskedFaceTestDataset(torch.utils.data.Dataset):
def __init__(self, path_to_imgs, path_to_xml_files, transform=None):
self.path_to_imgs = path_to_imgs
self.path_to_xml_files = path_to_xml_files
self.transform = transform
self.file_names = [file for file in os.listdir(self.path_to_imgs) if file.endswith(".png")]
self.file_paths = [os.path.join(self.path_to_imgs, file_name) for file_name in self.file_names]
def __getitem__(self, index: int):
path_to_img = self.file_paths[index]
path = Path(path_to_img)
fname = path.stem
self.name = self.file_names[index]
image = Image.open(path_to_img).convert('RGB')
image = np.array(image) / 255.0
target = self.__generate_target(self.path_to_xml_files, self.name)
if self.transform:
augmentations = self.transform(image=image, bboxes=target['boxes'], labels=target['labels'])
image = augmentations['image']
target['boxes'] = torch.tensor(augmentations['bboxes'])
target['labels'] = torch.tensor(augmentations['labels'])
return image, target, fname
def __generate_target(self, path_to_xml, name_of_img):
path_to_annot = os.path.join(path_to_xml, name_of_img.replace('png', 'xml'))
boxes_lst = []
labels_lst = []
clases_lst = []
target = {}
with open(path_to_annot, 'r') as file:
xml_content = file.read()
soup = BeautifulSoup(xml_content, 'xml')
objects = soup.find_all('object')
for obj in objects:
name = obj.find('name').text
xmin = int(float(obj.find('xmin').text)) - 1
ymin = int(float(obj.find('ymin').text)) - 1
xmax = int(float(obj.find('xmax').text)) - 1
ymax = int(float(obj.find('ymax').text)) - 1
labels_lst.append(name)
boxes_lst.append([xmin, ymin, xmax, ymax])
clases_lst.append(class_dict[name])
target["boxes"] = torch.tensor(boxes_lst, dtype=torch.float32)
target["labels"] = torch.tensor(clases_lst, dtype=torch.int64)
return target
def __len__(self):
return len(self.file_paths)
In [8]:
import xmltodict
from torch.utils.data.sampler import SubsetRandomSampler
def collate_fn(batch):
return tuple(zip(*batch))
test_ds = MaskedFaceTestDataset(path_to_imgs = './data/MaskedFace/val/',
path_to_xml_files = './data/MaskedFace/val/',
transform = test_transform)
test_sampler = SubsetRandomSampler( list(range(0, len(test_ds))) )
test_loader = DataLoader(test_ds, batch_size=16, num_workers=0, pin_memory=True,
sampler=test_sampler, collate_fn=collate_fn)
Test an image¶
In [9]:
index = 10
In [10]:
import matplotlib.pyplot as plt
image, target, fname = test_ds[index]
cv_image = np.transpose(image, (1, 2, 0)).numpy()
for box, label in zip(target['boxes'], target['labels']):
cv2.rectangle(cv_image, (int(box[0]), int(box[1])), (int(box[2]), int(box[3])), (255, 255, 0))
cv2.putText(cv_image,
class_dct_inv[label.item()],
(int(box[0])-10, int(box[1])-10),
cv2.FONT_HERSHEY_DUPLEX,
0.5,
(255, 255, 0),
)
plt.imshow(cv_image)
plt.show()
In [11]:
model.eval()
image, target, fname = test_ds[index]
with torch.no_grad():
prediction = model(image.unsqueeze(0).to(device, dtype=torch.float))[0]
cv_image = np.transpose(image, (1, 2, 0)).numpy()
for box, label, score in zip(prediction['boxes'], prediction['labels'], prediction['scores']):
if score >= 0.5:
cv2.rectangle(cv_image, (int(box[0]), int(box[1])), (int(box[2]), int(box[3])), (255, 255, 0))
cv2.putText(cv_image,
class_dct_inv[label.item()],
(int(box[0])-10, int(box[1])-10),
cv2.FONT_HERSHEY_DUPLEX,
0.5,
(255, 255, 0),
)
plt.imshow(cv_image)
plt.show()
Evaluation on the Test set¶
In [12]:
def read_true_object(root):
for child in root:
if child.tag == "name":
label_desc = str(child.text)
if label_desc == "with_mask":
return class_dict["with_mask"]
elif label_desc == "without_mask":
return class_dict["without_mask"]
else:
return class_dict["mask_weared_incorrect"]
def get_true_result(filename):
wi_m = 0
wo_m = 0
inc_m = 0
tree = ET.parse('./data/MaskedFace/val/' + filename + ".xml")
root = tree.getroot()
result = None
for child in root:
if child.tag == "object":
result = read_true_object(child)
if result == class_dict["with_mask"]:
wi_m+=1
elif result == class_dict["without_mask"]:
wo_m+=1
elif result == class_dict["mask_weared_incorrect"]:
inc_m+=1
return [wi_m, wo_m, inc_m]
In [13]:
pred_list = []
true_list = []
model.eval()
for images, targets, fname in tqdm(test_loader):
for i in range(len(fname)):
true_list.append(get_true_result(fname[i]))
images = list(image.to(device, dtype=torch.float) for image in images)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
with torch.no_grad():
prediction = model(images)
for i in range(len(images)):
with_m = 0
with_o_m = 0
wrong_m = 0
for pred in range(len(prediction[i]['boxes'])):
box = prediction[i]['boxes'][pred].cpu().numpy()
label = prediction[i]['labels'][pred].cpu().numpy()
score = prediction[i]['scores'][pred].cpu().numpy()
if score >= 0.5:
if label.item() == class_dict["with_mask"]:
with_m += 1
elif label.item() == class_dict["without_mask"]:
with_o_m += 1
elif label.item() == class_dict["mask_weared_incorrect"]:
wrong_m += 1
pred_list.append([int(with_m), int(with_o_m), int(wrong_m)])
MAPE¶
In [14]:
def compute_mape(prediction, truth):
mape = np.mean( np.abs(truth - prediction) / np.maximum(truth, np.ones_like(truth)) ) * 100
return mape
In [15]:
MAPE = compute_mape( np.array(pred_list), np.array(true_list) )
In [16]:
print(MAPE)
Final Score¶
In [17]:
if MAPE <= 10:
print("Score: ", 25*1.0)
elif MAPE <= 15:
print("Score: ", 25*0.875)
elif MAPE <= 20:
print("Score: ", 25*0.75)
elif MAPE <= 25:
print("Score: ", 25*0.625)
elif MAPE <= 30:
print("Score: ", 25*0.5)
else:
print("Score: ", 0)
In [ ]: