The following notebook is an exercise for the Convolutional Neural Networks for Computer Vision course at Afeka College of Engineering.
It uses Kaggle's Face Mask Detection dataset for Multi-Label Object Detection.
Table of Contents:
- Imports
- Configurations
- Data Load
- Data Exploration
- Spliting the Data
- Data Samples
- Training
- Evaluation
- References
Submitted By:
- Tal Goldengoren
- Guy Kabiri
import os
import sys
from datetime import datetime
import numpy as np
import pandas as pd
import cv2
import torch
from torch.utils.data import DataLoader
import torch.optim as optim
import torchvision
from sklearn import metrics
from iterstrat.ml_stratifiers import MultilabelStratifiedKFold
from matplotlib import pyplot as plt
import seaborn as snsfrom data_handler.FaceMaskData import FaceMaskData
from data_handler.FaceMaskDataset import FaceMaskDataset
from trainer import *
from metrics.metrics import *
from criterion.criterion import *assert torch.cuda.is_available()DEBUG = False
class CFG:
seed = 42
model_name = 'faster_rcnn'
pretrained = True
img_width = 480
img_height = 480
batch_size = 10
n_epochs = 20
n_folds = 5
drop_rate = 0.
train_size = 0.90
nms_thresh = 0.2
score_threshold = 0.5
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
num_workers = 4
num_classes = 4
imgs_path = 'images' # images dir
msks_path = 'annotations' # annotations dir
multilabelKFold = True
optimizer = optim.SGD
optimizer_dict = { 'lr': 0.005,
'momentum': 0.9,
'weight_decay': 0.0005 }
scheduler = optim.lr_scheduler.StepLR
scheduler_dict = { 'step_size': 5,
'gamma': 0.1 }
def save(path):
save_path = path + '/model_dict.json'
with open(save_path, 'w') as f:
for key, val in CFG.__dict__.items():
f.write('{}\t\t= {}\n'.format(key, val))
# optimizer = optim.Adam
# optimizer_dict = None
# scheduler = optim.lr_scheduler.StepLR
# scheduler_dict = { 'step_size': 5,
# 'gamma': 0.1 }
# optimizer = optim.ASGD
# optimizer_dict = None
# scheduler = optim.lr_scheduler.ExponentialLR
# scheduler_dict = { gamma: 0.9 }
np.random.seed(CFG.seed)Load the data into a dataset
df = pd.read_csv(os.path.join(sys.path[0], 'annotation.csv'))
if CFG.multilabelKFold and CFG.n_folds > 1:
faceMasksData = FaceMaskData(CFG.imgs_path, CFG.msks_path, multilabelKFold=True, df_file=df)
(x_train, y_train, l_train), (x_test, y_test, l_test) = faceMasksData.load_data(
train_size=CFG.train_size,
drop_rate=CFG.drop_rate,
seed=CFG.seed)
else:
faceMasksData = FaceMaskData(CFG.imgs_path, CFG.msks_path)
(x_train, y_train), (x_test, y_test) = faceMasksData.load_data(
train_size=CFG.train_size,
drop_rate=CFG.drop_rate,
seed=CFG.seed)
print('Training contains {} samples which is {:g}% of the data'.format(len(x_train), len(x_train) * 100 / (len(x_train) + len(x_test))))
print('Testing contains {} samples which is {:g}% of the data'.format(len(x_test), len(x_test) * 100 / (len(x_train) + len(x_test))))Training contains 767 samples which is 89.9179% of the data
Testing contains 86 samples which is 10.0821% of the data
testset = FaceMaskDataset(x_test, y_test, CFG.imgs_path, CFG.msks_path, CFG.img_width, CFG.img_height, transforms=get_transformer('test'))
test_loader = DataLoader(dataset=testset, batch_size=CFG.batch_size, shuffle=False, num_workers=CFG.num_workers, collate_fn=collate_fn)df.head()| xmin | ymin | xmax | ymax | name | file | width | height | class | Xcent | Ycent | boxW | boxH | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0 | 28 | 55 | 46 | 71 | with_mask | maksssksksss737 | 400 | 226 | 0 | 0.09250 | 0.278761 | 0.0450 | 0.070796 |
| 1 | 98 | 62 | 111 | 78 | with_mask | maksssksksss737 | 400 | 226 | 0 | 0.26125 | 0.309735 | 0.0325 | 0.070796 |
| 2 | 159 | 50 | 193 | 90 | mask_weared_incorrect | maksssksksss737 | 400 | 226 | 1 | 0.44000 | 0.309735 | 0.0850 | 0.176991 |
| 3 | 293 | 59 | 313 | 80 | with_mask | maksssksksss737 | 400 | 226 | 0 | 0.75750 | 0.307522 | 0.0500 | 0.092920 |
| 4 | 352 | 51 | 372 | 72 | with_mask | maksssksksss737 | 400 | 226 | 0 | 0.90500 | 0.272124 | 0.0500 | 0.092920 |
faces = len(df['file'].unique())
annotates = len(df)
print('There are total {} images in the data'.format(faces))
print('There are total {} annotated faces in the data'.format(annotates))
print('Average of {:.5f} per image'.format(annotates/faces))There are total 853 images in the data
There are total 4072 annotated faces in the data
Average of 4.77374 per image
label_df = pd.DataFrame(columns=['label', 'count'])
for k, v in df['name'].value_counts().to_dict().items():
label_df = label_df.append({'label':k, 'count':v}, ignore_index=True)
display(label_df)
fig, ax = plt.subplots()
ax.bar(label_df['label'], label_df['count'], width=0.4)
plt.title('Label Count')
for index, data in enumerate(label_df['count']):
plt.text(x=index-0.1, y=data+1, s=data , fontdict=dict(fontsize=20))
plt.tight_layout()
plt.show()| label | count | |
|---|---|---|
| 0 | with_mask | 3232 |
| 1 | without_mask | 717 |
| 2 | mask_weared_incorrect | 123 |
def get_class_distribute(df):
classes = ["with_mask","without_mask","mask_weared_incorrect"]
classes_bit_shifter_amount = {
"with_mask":0,
"without_mask":1,
"mask_weared_incorrect":2,
}
dists = np.zeros(8)
for name in faceMasksData.images:
img_name = name.split('.')[0]
img_classes = df[df['file']==img_name]['name'].unique()
index = 0
for cls in img_classes:
index |= 1 << classes_bit_shifter_amount[cls]
dists[index] += 1
if dists[0] == 0:
return dists[1:]
return dists
dist_names = ['None', 'with_mask', 'without_mask', 'with_mask\nwithout_mask',
'mask_weared_incorrect', 'with_mask\nmask_weared_incorrect',
'without_mask\nmask_weared_incorrect',
'with_mask\nwithout_mask\nmask_weared_incorrect']
dist_values = get_class_distribute(df)
if len(dist_values) == 7:
dist_names = dist_names[1:]
plt.figure(figsize=(15, 5))
axs = plt.bar(dist_names, dist_values)
for ax in axs:
plt.text(ax.get_xy()[0] + 0.3, ax.get_height(), int(ax.get_height()), fontsize='x-large')
plt.xticks(rotation=30, ha='right')
plt.title('Classes per Image')
plt.show()
Clearly, the data in these charts show a very imbalanced trend.
Only 123 faces have an incorrectly worn mask, appearing in (21 + 30 + 4 + 42) = 97 images, which is (97 / 853) = 11% of all the images.
However, the masks appeared on 768 / 853 = 90% of all images.
In later stages of training, it will be necessary to divide the train and validation sets evenly to preserve the different labels at the same percentage.
def get_num_faces(df, image_name_col):
max_faces = find_max_faces(df,image_name_col)
arr = [0] * (max_faces + 1)
faces_count =1
current_img =df.at[0,image_name_col]
for i in range(1,len(df[1:])):
if current_img == df.at[i,image_name_col]:
faces_count = faces_count + 1
else:
arr[faces_count] = arr[faces_count] +1
faces_count =1
current_img =df.at[i,image_name_col]
df = pd.DataFrame(columns=['faces', 'count'])
for i, val in enumerate(arr):
if val > 0:
df = df.append({'faces': i, 'count': val}, ignore_index=True)
return df
def find_max_faces(df,image_name_col):
max_faces=1
faces_count =1
current_img =df.at[0,image_name_col]
for i in range(1,len(df[1:])):
if current_img == df.at[i,image_name_col]:
faces_count = faces_count +1
else:
if faces_count > max_faces:
max_faces = faces_count
current_img = df.at[i,image_name_col]
faces_count =1
return max_facesface_dist_df = get_num_faces(df, 'file')
ax = face_dist_df.plot.bar(x='faces', y='count', figsize=(15,5), title='Faces per Image')
for idx, label in enumerate(list(face_dist_df.faces)):
val = int(face_dist_df[face_dist_df['faces']==label]['count'])
ax.annotate(val,
(idx-0.2, val),
xytext=(0, 15),
textcoords='offset points')
print('STD of faces in the images: {:.5}'.format(face_dist_df['count'].std()))STD of faces in the images: 58.229
The data is very imbalanced, as it has been shown above.
The number of appearances of each class varies greatly, as well as the amount of faces per image.
We would like to use Cross-Validation for better training.
Due to our data being imbalanced, using standard splitting methods can cause some classes to not appear at all in some folds; Moreover, our task involves Multi-label Object Detection, for which the well-known splitting algorithm cannot be used.
Stratified KFold splits the folds by preserving the percentages of samples for each label.
Scikit-learn's algorithm cannot be applied to the second problem, so we will utilize the MultilabelStratifiedKFold implementation, which can be found here.
# self.classes = [None, 'without_mask','with_mask','mask_weared_incorrect']
# MultilabelStratifiedKFold needs to get the Y as binary for each class
x_names = [ name.split('.')[0] for name in x_train ]
y_classes = []
for name in x_names:
classes = list(df[df['file']==name]['name'].unique())
indice = np.zeros(4, dtype=np.uint8)
for c in classes:
index = testset.classes.index(c)
indice[index] = 1
y_classes.append(list(indice))
print(y_classes[:5])[[0, 0, 1, 0], [0, 0, 1, 0], [0, 0, 1, 0], [0, 0, 1, 0], [0, 0, 1, 0]]
df_folds = pd.DataFrame()
mskf = MultilabelStratifiedKFold(CFG.n_folds)
for fold_id, (train_idx, test_idx) in enumerate(mskf.split(x_names, y_classes), start=1):
classes = np.zeros(len(testset.classes))
for idx in train_idx:
name = x_names[idx]
for cl, val in df[df['file']==name]['name'].value_counts().to_dict().items():
class_index = testset.classes.index(cl)
classes[class_index] += val
for cl, val in df['name'].value_counts().to_dict().items():
class_index = testset.classes.index(cl)
df_folds = df_folds.append({ 'fold': int(fold_id),
'class':cl,
'count': classes[class_index] ,
}, ignore_index=True)
print('fold {}: train_size={}, test_size={}'.format(fold_id, len(train_idx), len(test_idx)))
# df_folds = df_folds.append({ 'fold': fold_id,
# 'train': len(train_idx),
# 'valid': len(test_idx),
# 'without_mask': classes[testset.classes.index('without_mask')],
# 'with_mask': classes[testset.classes.index('with_mask')],
# 'mask_weared_incorrect': classes[testset.classes.index('mask_weared_incorrect')]
# }, ignore_index=True)
display(df_folds)
g = ax = sns.catplot(x='fold', y='count', hue='class', data=df_folds, kind='bar', legend=True)
g.fig.set_size_inches(15, 8)
g.fig.subplots_adjust(top=0.81, right=0.86)
# extract the matplotlib axes_subplot objects from the FacetGrid
ax = g.facet_axis(0, 0)
# iterate through the axes containers
for c in ax.containers:
labels = [int(v.get_height()) for v in c]
ax.bar_label(c, labels=labels, label_type='edge')fold 1: train_size=616, test_size=151
fold 2: train_size=613, test_size=154
fold 3: train_size=615, test_size=152
fold 4: train_size=612, test_size=155
fold 5: train_size=612, test_size=155
| fold | class | count | |
|---|---|---|---|
| 0 | 1.0 | with_mask | 2348.0 |
| 1 | 1.0 | without_mask | 567.0 |
| 2 | 1.0 | mask_weared_incorrect | 92.0 |
| 3 | 2.0 | with_mask | 2255.0 |
| 4 | 2.0 | without_mask | 510.0 |
| 5 | 2.0 | mask_weared_incorrect | 90.0 |
| 6 | 3.0 | with_mask | 2383.0 |
| 7 | 3.0 | without_mask | 483.0 |
| 8 | 3.0 | mask_weared_incorrect | 87.0 |
| 9 | 4.0 | with_mask | 2369.0 |
| 10 | 4.0 | without_mask | 551.0 |
| 11 | 4.0 | mask_weared_incorrect | 89.0 |
| 12 | 5.0 | with_mask | 2205.0 |
| 13 | 5.0 | without_mask | 557.0 |
| 14 | 5.0 | mask_weared_incorrect | 90.0 |
From the chart above, it can be seen that MultilabelStratifiedKFold split the images into almost identical train-test sizes in each fold, while preserving the number of labels from each class within each fold.
def norm(img):
img = np.array(img, dtype=np.float32)
img -= img.min()
img /= img.max()
return img
def get_annotated_img(img, annt, is_pred=False):
img = norm(np.array(np.transpose(img, (1, 2, 0)))) * 255.0 # multiple by 255 as in the dataset we divide it
# needed here because the image come from the dataset in values of [0, 1]
# and the annotations are in values of [0, 255] and plt should get images
# with values of [0, 1], therefore a normalizing of the images is needed
num_faces = len(annt['boxes'])
for i in range(num_faces):
box, cat = annt['boxes'][i], annt['labels'][i]
if is_pred:
score = annt['scores'][i]
xmin, ymin, xmax, ymax = np.array(box, dtype=np.int32)
start_point = (xmin, ymin)
end_point = (xmax, ymax)
# [None, 'without_mask','with_mask','mask_weared_incorrect']
color = (0, 0, 0)
if cat == 2:
color = (0, 255, 0) # green
elif cat == 3:
color = (0, 0, 255) # blue
elif cat == 1:
color = (255, 0, 0) # red
thickness = 1
img = cv2.rectangle(img, start_point, end_point, color, thickness)
font = cv2.FONT_HERSHEY_SIMPLEX
fontScale = 1 / 3
if not is_pred:
img = cv2.putText(img, '{}'.format(faceMasksData.classes[cat]), start_point, font, fontScale, color = (0, 0, 0), thickness=2)
img = cv2.putText(img, '{}'.format(faceMasksData.classes[cat]), start_point, font, fontScale, color, thickness)
else:
img = cv2.putText(img, '{:.2f}'.format(score), start_point, font, fontScale, color = (0, 0, 0), thickness=2)
img = cv2.putText(img, '{:.2f}'.format(score), start_point, font, fontScale, color, thickness)
return img
def show_augmented_samples(dataloader, model=None):
samples, annotations = next(iter(dataloader)) # get the first batch
amount_samples = min(12, len(samples)) # number of examples will be at most 12
for i, (img, ant) in enumerate(zip(samples, annotations)):
if i >= amount_samples:
break
plt.figure(figsize=(10, 10))
# plt.subplot(amount_samples, 1, i + 1)
img = get_annotated_img(img, ant)
plt.imshow(norm(img))
plt.axis('off')
plt.tight_layout()
plt.show()if DEBUG == False:
show_augmented_samples(test_loader)
testset = FaceMaskDataset(x_test, y_test, CFG.imgs_path, CFG.msks_path, CFG.img_width, CFG.img_height, transforms=get_transformer('train'))
test_loader = DataLoader(dataset=testset, batch_size=CFG.batch_size, shuffle=False, num_workers=CFG.num_workers, collate_fn=collate_fn)
if DEBUG == False:
show_augmented_samples(test_loader)
def get_model(num_classes, pretrained=True):
model = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=pretrained) # get model
in_features = model.roi_heads.box_predictor.cls_score.in_features # get input size of last layer
model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes) # regenerate the last layer
return modelmodel = get_model(num_classes=CFG.num_classes, pretrained=CFG.pretrained)datetime_srt = datetime.today().strftime("%d-%m-%y_%H:%M")
run_path = os.path.join(sys.path[0], 'runs', datetime_srt)if DEBUG == False:
if CFG.multilabelKFold and CFG.n_folds > 1:
train(model, x_train, (y_train, l_train), run_path, CFG)
else:
train(model, x_train, y_train, run_path, CFG)This running path is: `/home/linuxgpu/anaconda3/envs/lab1/Face-Mask-Detection/runs/09-11-21_20:49`
�[1m�[4mFold 1 of 5�[0m
Epoch 1 of 20:
train: 0%| | 0/62 [00:00<?, ?it/s]/home/linuxgpu/anaconda3/envs/lab1/lib/python3.9/site-packages/torch/functional.py:445: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at /opt/conda/conda-bld/pytorch_1634272204863/work/aten/src/ATen/native/TensorShape.cpp:2157.)
return _VF.meshgrid(tensors, **kwargs) # type: ignore[attr-defined]
train: 100%|██████████| 62/62 [01:29<00:00, 1.44s/it]
valid: 100%|██████████| 16/16 [00:11<00:00, 1.43it/s]
train: loss_classifier=14.55984 loss_box_reg=14.94884 loss_objectness=7.57883 loss_rpn_box_reg=1.70882 total loss=38.79634
valid: loss_classifier=2.52951 loss_box_reg=3.70048 loss_objectness=0.43461 loss_rpn_box_reg=0.31982 total loss=6.98442
Model saved. Loss < PrevLoss (6.98442 < inf)
Epoch 2 of 20:
train: 100%|██████████| 62/62 [01:29<00:00, 1.44s/it]
valid: 100%|██████████| 16/16 [00:11<00:00, 1.40it/s]
train: loss_classifier=10.31991 loss_box_reg=15.21334 loss_objectness=1.83208 loss_rpn_box_reg=1.13317 total loss=28.49851
valid: loss_classifier=1.88260 loss_box_reg=3.19624 loss_objectness=0.39928 loss_rpn_box_reg=0.24210 total loss=5.72021
Model saved. Loss < PrevLoss (5.72021 < 6.98442)
Epoch 3 of 20:
train: 100%|██████████| 62/62 [01:33<00:00, 1.50s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.49it/s]
train: loss_classifier=8.37889 loss_box_reg=13.87635 loss_objectness=1.31881 loss_rpn_box_reg=1.00075 total loss=24.57480
valid: loss_classifier=1.86610 loss_box_reg=3.14533 loss_objectness=0.28374 loss_rpn_box_reg=0.22568 total loss=5.52086
Model saved. Loss < PrevLoss (5.52086 < 5.72021)
Epoch 4 of 20:
train: 100%|██████████| 62/62 [01:29<00:00, 1.44s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.51it/s]
train: loss_classifier=7.99943 loss_box_reg=13.56459 loss_objectness=1.27645 loss_rpn_box_reg=0.86038 total loss=23.70084
valid: loss_classifier=1.57474 loss_box_reg=2.77011 loss_objectness=0.26650 loss_rpn_box_reg=0.22109 total loss=4.83244
Model saved. Loss < PrevLoss (4.83244 < 5.52086)
Epoch 5 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=7.42718 loss_box_reg=12.46731 loss_objectness=0.97449 loss_rpn_box_reg=0.82367 total loss=21.69265
valid: loss_classifier=1.61933 loss_box_reg=2.96249 loss_objectness=0.20516 loss_rpn_box_reg=0.22687 total loss=5.01384
Epoch 6 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.52it/s]
train: loss_classifier=6.60729 loss_box_reg=11.55502 loss_objectness=0.77213 loss_rpn_box_reg=0.75366 total loss=19.68809
valid: loss_classifier=1.52544 loss_box_reg=2.66514 loss_objectness=0.18426 loss_rpn_box_reg=0.19806 total loss=4.57290
Model saved. Loss < PrevLoss (4.57290 < 4.83244)
Epoch 7 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.53it/s]
train: loss_classifier=6.48084 loss_box_reg=11.18433 loss_objectness=0.67169 loss_rpn_box_reg=0.70574 total loss=19.04260
valid: loss_classifier=1.51454 loss_box_reg=2.66865 loss_objectness=0.18496 loss_rpn_box_reg=0.19349 total loss=4.56163
Model saved. Loss < PrevLoss (4.56163 < 4.57290)
Epoch 8 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.53it/s]
train: loss_classifier=6.36256 loss_box_reg=11.12686 loss_objectness=0.60876 loss_rpn_box_reg=0.64691 total loss=18.74509
valid: loss_classifier=1.49312 loss_box_reg=2.64090 loss_objectness=0.18642 loss_rpn_box_reg=0.19516 total loss=4.51560
Model saved. Loss < PrevLoss (4.51560 < 4.56163)
Epoch 9 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=6.49383 loss_box_reg=11.26893 loss_objectness=0.64274 loss_rpn_box_reg=0.63443 total loss=19.03992
valid: loss_classifier=1.52493 loss_box_reg=2.67580 loss_objectness=0.22352 loss_rpn_box_reg=0.19233 total loss=4.61657
Epoch 10 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=6.42897 loss_box_reg=11.21191 loss_objectness=0.68523 loss_rpn_box_reg=0.70564 total loss=19.03175
valid: loss_classifier=1.55092 loss_box_reg=2.81906 loss_objectness=0.18899 loss_rpn_box_reg=0.20040 total loss=4.75938
Epoch 11 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.51it/s]
train: loss_classifier=6.46314 loss_box_reg=11.24431 loss_objectness=0.63966 loss_rpn_box_reg=0.66912 total loss=19.01622
valid: loss_classifier=1.50922 loss_box_reg=2.66300 loss_objectness=0.20616 loss_rpn_box_reg=0.19196 total loss=4.57034
Epoch 12 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.52it/s]
train: loss_classifier=6.36753 loss_box_reg=10.99467 loss_objectness=0.59974 loss_rpn_box_reg=0.67159 total loss=18.63354
valid: loss_classifier=1.48818 loss_box_reg=2.68145 loss_objectness=0.20885 loss_rpn_box_reg=0.19247 total loss=4.57095
Epoch 13 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=6.16429 loss_box_reg=11.06495 loss_objectness=0.64548 loss_rpn_box_reg=0.66106 total loss=18.53579
valid: loss_classifier=1.50100 loss_box_reg=2.68242 loss_objectness=0.19746 loss_rpn_box_reg=0.19333 total loss=4.57421
Epoch 14 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=6.30050 loss_box_reg=11.04009 loss_objectness=0.61786 loss_rpn_box_reg=0.63727 total loss=18.59573
valid: loss_classifier=1.49688 loss_box_reg=2.70135 loss_objectness=0.19851 loss_rpn_box_reg=0.19258 total loss=4.58932
Epoch 15 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.52it/s]
train: loss_classifier=6.16774 loss_box_reg=10.89944 loss_objectness=0.53491 loss_rpn_box_reg=0.65406 total loss=18.25616
valid: loss_classifier=1.49507 loss_box_reg=2.63868 loss_objectness=0.21180 loss_rpn_box_reg=0.18905 total loss=4.53459
Epoch 16 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.53it/s]
train: loss_classifier=6.41260 loss_box_reg=11.02292 loss_objectness=0.63147 loss_rpn_box_reg=0.63893 total loss=18.70592
valid: loss_classifier=1.48150 loss_box_reg=2.63900 loss_objectness=0.20336 loss_rpn_box_reg=0.19214 total loss=4.51600
Epoch 17 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.53it/s]
train: loss_classifier=6.33803 loss_box_reg=11.31839 loss_objectness=0.59926 loss_rpn_box_reg=0.64637 total loss=18.90205
valid: loss_classifier=1.48851 loss_box_reg=2.64537 loss_objectness=0.19744 loss_rpn_box_reg=0.19368 total loss=4.52499
Epoch 18 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=6.17073 loss_box_reg=11.04405 loss_objectness=0.59753 loss_rpn_box_reg=0.67411 total loss=18.48641
valid: loss_classifier=1.49577 loss_box_reg=2.64654 loss_objectness=0.20144 loss_rpn_box_reg=0.19270 total loss=4.53645
Epoch 19 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.57it/s]
train: loss_classifier=6.31196 loss_box_reg=11.08092 loss_objectness=0.64887 loss_rpn_box_reg=0.64354 total loss=18.68528
valid: loss_classifier=1.50889 loss_box_reg=2.65763 loss_objectness=0.20364 loss_rpn_box_reg=0.18930 total loss=4.55947
Epoch 20 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=6.28305 loss_box_reg=10.92692 loss_objectness=0.59834 loss_rpn_box_reg=0.65377 total loss=18.46208
valid: loss_classifier=1.49762 loss_box_reg=2.64607 loss_objectness=0.21274 loss_rpn_box_reg=0.19294 total loss=4.54937
�[1m�[4mFold 2 of 5�[0m
Epoch 1 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=6.81814 loss_box_reg=11.67234 loss_objectness=0.71450 loss_rpn_box_reg=0.69237 total loss=19.89735
valid: loss_classifier=1.52259 loss_box_reg=2.84186 loss_objectness=0.21646 loss_rpn_box_reg=0.21484 total loss=4.79574
Epoch 2 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=6.42684 loss_box_reg=11.18179 loss_objectness=0.69653 loss_rpn_box_reg=0.61717 total loss=18.92233
valid: loss_classifier=1.58965 loss_box_reg=3.08537 loss_objectness=0.13274 loss_rpn_box_reg=0.21907 total loss=5.02683
Epoch 3 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=6.16551 loss_box_reg=11.32156 loss_objectness=0.61722 loss_rpn_box_reg=0.62915 total loss=18.73344
valid: loss_classifier=1.50763 loss_box_reg=3.20170 loss_objectness=0.11712 loss_rpn_box_reg=0.26398 total loss=5.09043
Epoch 4 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.50it/s]
train: loss_classifier=6.29596 loss_box_reg=11.27375 loss_objectness=0.61891 loss_rpn_box_reg=0.60353 total loss=18.79215
valid: loss_classifier=1.44440 loss_box_reg=2.73941 loss_objectness=0.12655 loss_rpn_box_reg=0.22657 total loss=4.53692
Epoch 5 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.57it/s]
train: loss_classifier=6.06435 loss_box_reg=10.73738 loss_objectness=0.56822 loss_rpn_box_reg=0.58929 total loss=17.95924
valid: loss_classifier=1.54106 loss_box_reg=3.43084 loss_objectness=0.12239 loss_rpn_box_reg=0.24453 total loss=5.33882
Epoch 6 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.52it/s]
train: loss_classifier=5.36201 loss_box_reg=9.56199 loss_objectness=0.41353 loss_rpn_box_reg=0.53422 total loss=15.87174
valid: loss_classifier=1.38374 loss_box_reg=2.71180 loss_objectness=0.12191 loss_rpn_box_reg=0.21607 total loss=4.43352
Model saved. Loss < PrevLoss (4.43352 < 4.51560)
Epoch 7 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=5.35620 loss_box_reg=9.62900 loss_objectness=0.41985 loss_rpn_box_reg=0.50166 total loss=15.90671
valid: loss_classifier=1.36137 loss_box_reg=2.64834 loss_objectness=0.11545 loss_rpn_box_reg=0.21180 total loss=4.33696
Model saved. Loss < PrevLoss (4.33696 < 4.43352)
Epoch 8 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=5.15339 loss_box_reg=9.48356 loss_objectness=0.49098 loss_rpn_box_reg=0.54373 total loss=15.67166
valid: loss_classifier=1.35071 loss_box_reg=2.66201 loss_objectness=0.12730 loss_rpn_box_reg=0.21328 total loss=4.35331
Epoch 9 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=5.22334 loss_box_reg=9.46209 loss_objectness=0.40369 loss_rpn_box_reg=0.48905 total loss=15.57816
valid: loss_classifier=1.34566 loss_box_reg=2.72925 loss_objectness=0.11519 loss_rpn_box_reg=0.22071 total loss=4.41081
Epoch 10 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.53it/s]
train: loss_classifier=5.25994 loss_box_reg=9.45002 loss_objectness=0.47393 loss_rpn_box_reg=0.50827 total loss=15.69216
valid: loss_classifier=1.38197 loss_box_reg=2.76685 loss_objectness=0.12503 loss_rpn_box_reg=0.22647 total loss=4.50033
Epoch 11 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=5.05451 loss_box_reg=9.28040 loss_objectness=0.37193 loss_rpn_box_reg=0.49666 total loss=15.20350
valid: loss_classifier=1.36208 loss_box_reg=2.67029 loss_objectness=0.12392 loss_rpn_box_reg=0.21212 total loss=4.36841
Epoch 12 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=5.17243 loss_box_reg=9.45861 loss_objectness=0.42570 loss_rpn_box_reg=0.49943 total loss=15.55616
valid: loss_classifier=1.34324 loss_box_reg=2.67248 loss_objectness=0.10810 loss_rpn_box_reg=0.21374 total loss=4.33756
Epoch 13 of 20:
train: 100%|██████████| 62/62 [01:29<00:00, 1.44s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=5.09986 loss_box_reg=9.46279 loss_objectness=0.45588 loss_rpn_box_reg=0.48156 total loss=15.50010
valid: loss_classifier=1.34182 loss_box_reg=2.67871 loss_objectness=0.11912 loss_rpn_box_reg=0.21502 total loss=4.35467
Epoch 14 of 20:
train: 100%|██████████| 62/62 [01:29<00:00, 1.44s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=4.96293 loss_box_reg=9.29894 loss_objectness=0.45479 loss_rpn_box_reg=0.51750 total loss=15.23416
valid: loss_classifier=1.38179 loss_box_reg=2.72603 loss_objectness=0.12204 loss_rpn_box_reg=0.22164 total loss=4.45150
Epoch 15 of 20:
train: 100%|██████████| 62/62 [01:29<00:00, 1.44s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=4.93508 loss_box_reg=9.12708 loss_objectness=0.38756 loss_rpn_box_reg=0.47872 total loss=14.92844
valid: loss_classifier=1.37108 loss_box_reg=2.67638 loss_objectness=0.11827 loss_rpn_box_reg=0.21879 total loss=4.38453
Epoch 16 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.51it/s]
train: loss_classifier=4.92323 loss_box_reg=9.08934 loss_objectness=0.39607 loss_rpn_box_reg=0.48447 total loss=14.89312
valid: loss_classifier=1.36178 loss_box_reg=2.68203 loss_objectness=0.10965 loss_rpn_box_reg=0.21472 total loss=4.36817
Epoch 17 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=5.16750 loss_box_reg=9.39665 loss_objectness=0.42634 loss_rpn_box_reg=0.46196 total loss=15.45245
valid: loss_classifier=1.35160 loss_box_reg=2.68737 loss_objectness=0.12157 loss_rpn_box_reg=0.21558 total loss=4.37611
Epoch 18 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.53it/s]
train: loss_classifier=5.11507 loss_box_reg=9.24839 loss_objectness=0.48113 loss_rpn_box_reg=0.50711 total loss=15.35169
valid: loss_classifier=1.37616 loss_box_reg=2.68389 loss_objectness=0.11311 loss_rpn_box_reg=0.21346 total loss=4.38661
Epoch 19 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=5.02759 loss_box_reg=9.02574 loss_objectness=0.38890 loss_rpn_box_reg=0.47150 total loss=14.91373
valid: loss_classifier=1.34483 loss_box_reg=2.67535 loss_objectness=0.12299 loss_rpn_box_reg=0.22083 total loss=4.36401
Epoch 20 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=5.10216 loss_box_reg=9.46653 loss_objectness=0.40348 loss_rpn_box_reg=0.53256 total loss=15.50472
valid: loss_classifier=1.34559 loss_box_reg=2.69263 loss_objectness=0.12516 loss_rpn_box_reg=0.21452 total loss=4.37791
�[1m�[4mFold 3 of 5�[0m
Epoch 1 of 20:
train: 100%|██████████| 62/62 [01:29<00:00, 1.44s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=5.29630 loss_box_reg=10.05864 loss_objectness=0.48215 loss_rpn_box_reg=0.60266 total loss=16.43974
valid: loss_classifier=1.21952 loss_box_reg=2.43896 loss_objectness=0.09098 loss_rpn_box_reg=0.20272 total loss=3.95218
Model saved. Loss < PrevLoss (3.95218 < 4.33696)
Epoch 2 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=5.17768 loss_box_reg=10.32614 loss_objectness=0.42926 loss_rpn_box_reg=0.56119 total loss=16.49428
valid: loss_classifier=1.31547 loss_box_reg=3.04749 loss_objectness=0.08677 loss_rpn_box_reg=0.20489 total loss=4.65461
Epoch 3 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=5.17384 loss_box_reg=9.98724 loss_objectness=0.42308 loss_rpn_box_reg=0.55610 total loss=16.14026
valid: loss_classifier=1.33611 loss_box_reg=2.72164 loss_objectness=0.09014 loss_rpn_box_reg=0.18419 total loss=4.33209
Epoch 4 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.53it/s]
train: loss_classifier=5.12655 loss_box_reg=10.06167 loss_objectness=0.38585 loss_rpn_box_reg=0.54894 total loss=16.12301
valid: loss_classifier=1.40101 loss_box_reg=2.74668 loss_objectness=0.09026 loss_rpn_box_reg=0.22735 total loss=4.46530
Epoch 5 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=5.02072 loss_box_reg=9.84159 loss_objectness=0.46280 loss_rpn_box_reg=0.54545 total loss=15.87056
valid: loss_classifier=1.28565 loss_box_reg=2.50110 loss_objectness=0.10377 loss_rpn_box_reg=0.17978 total loss=4.07030
Epoch 6 of 20:
train: 100%|██████████| 62/62 [01:29<00:00, 1.44s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=4.56205 loss_box_reg=8.83081 loss_objectness=0.41739 loss_rpn_box_reg=0.47118 total loss=14.28143
valid: loss_classifier=1.22286 loss_box_reg=2.39532 loss_objectness=0.09660 loss_rpn_box_reg=0.16503 total loss=3.87982
Model saved. Loss < PrevLoss (3.87982 < 3.95218)
Epoch 7 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=4.31234 loss_box_reg=8.68379 loss_objectness=0.35495 loss_rpn_box_reg=0.47466 total loss=13.82574
valid: loss_classifier=1.21976 loss_box_reg=2.40189 loss_objectness=0.08253 loss_rpn_box_reg=0.16665 total loss=3.87083
Model saved. Loss < PrevLoss (3.87083 < 3.87982)
Epoch 8 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=4.38207 loss_box_reg=8.63387 loss_objectness=0.29640 loss_rpn_box_reg=0.45453 total loss=13.76688
valid: loss_classifier=1.23222 loss_box_reg=2.42774 loss_objectness=0.08264 loss_rpn_box_reg=0.17077 total loss=3.91337
Epoch 9 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=4.17644 loss_box_reg=8.62865 loss_objectness=0.30108 loss_rpn_box_reg=0.44456 total loss=13.55073
valid: loss_classifier=1.24046 loss_box_reg=2.45912 loss_objectness=0.09591 loss_rpn_box_reg=0.16995 total loss=3.96544
Epoch 10 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=4.24148 loss_box_reg=8.56676 loss_objectness=0.34705 loss_rpn_box_reg=0.47793 total loss=13.63322
valid: loss_classifier=1.20491 loss_box_reg=2.44213 loss_objectness=0.08674 loss_rpn_box_reg=0.16969 total loss=3.90347
Epoch 11 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.50it/s]
train: loss_classifier=4.11093 loss_box_reg=8.44191 loss_objectness=0.34525 loss_rpn_box_reg=0.43974 total loss=13.33783
valid: loss_classifier=1.20295 loss_box_reg=2.39838 loss_objectness=0.08311 loss_rpn_box_reg=0.17149 total loss=3.85593
Model saved. Loss < PrevLoss (3.85593 < 3.87083)
Epoch 12 of 20:
train: 100%|██████████| 62/62 [01:29<00:00, 1.44s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=4.18013 loss_box_reg=8.41082 loss_objectness=0.31653 loss_rpn_box_reg=0.41583 total loss=13.32331
valid: loss_classifier=1.18960 loss_box_reg=2.36876 loss_objectness=0.08126 loss_rpn_box_reg=0.16771 total loss=3.80733
Model saved. Loss < PrevLoss (3.80733 < 3.85593)
Epoch 13 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=4.32926 loss_box_reg=8.49101 loss_objectness=0.35807 loss_rpn_box_reg=0.42594 total loss=13.60427
valid: loss_classifier=1.18774 loss_box_reg=2.36607 loss_objectness=0.07944 loss_rpn_box_reg=0.16925 total loss=3.80249
Model saved. Loss < PrevLoss (3.80249 < 3.80733)
Epoch 14 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=4.40295 loss_box_reg=8.53910 loss_objectness=0.28315 loss_rpn_box_reg=0.45611 total loss=13.68131
valid: loss_classifier=1.20161 loss_box_reg=2.40475 loss_objectness=0.08713 loss_rpn_box_reg=0.16986 total loss=3.86335
Epoch 15 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=4.24205 loss_box_reg=8.61536 loss_objectness=0.35154 loss_rpn_box_reg=0.45784 total loss=13.66680
valid: loss_classifier=1.20190 loss_box_reg=2.36636 loss_objectness=0.08611 loss_rpn_box_reg=0.16832 total loss=3.82269
Epoch 16 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=4.19732 loss_box_reg=8.44667 loss_objectness=0.29993 loss_rpn_box_reg=0.45241 total loss=13.39634
valid: loss_classifier=1.19025 loss_box_reg=2.38040 loss_objectness=0.07774 loss_rpn_box_reg=0.17157 total loss=3.81996
Epoch 17 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=4.18506 loss_box_reg=8.37908 loss_objectness=0.26500 loss_rpn_box_reg=0.44673 total loss=13.27588
valid: loss_classifier=1.22286 loss_box_reg=2.37353 loss_objectness=0.08683 loss_rpn_box_reg=0.16709 total loss=3.85030
Epoch 18 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=4.15906 loss_box_reg=8.27909 loss_objectness=0.29277 loss_rpn_box_reg=0.41898 total loss=13.14991
valid: loss_classifier=1.22118 loss_box_reg=2.37414 loss_objectness=0.08969 loss_rpn_box_reg=0.16601 total loss=3.85102
Epoch 19 of 20:
train: 100%|██████████| 62/62 [01:29<00:00, 1.44s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.57it/s]
train: loss_classifier=4.40226 loss_box_reg=8.40437 loss_objectness=0.30971 loss_rpn_box_reg=0.43458 total loss=13.55091
valid: loss_classifier=1.19229 loss_box_reg=2.38534 loss_objectness=0.08976 loss_rpn_box_reg=0.16778 total loss=3.83517
Epoch 20 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=4.27569 loss_box_reg=8.51250 loss_objectness=0.38033 loss_rpn_box_reg=0.44488 total loss=13.61340
valid: loss_classifier=1.21717 loss_box_reg=2.38471 loss_objectness=0.09314 loss_rpn_box_reg=0.16964 total loss=3.86466
�[1m�[4mFold 4 of 5�[0m
Epoch 1 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.52it/s]
train: loss_classifier=5.01045 loss_box_reg=9.78068 loss_objectness=0.44959 loss_rpn_box_reg=0.55807 total loss=15.79880
valid: loss_classifier=0.93051 loss_box_reg=1.85362 loss_objectness=0.04986 loss_rpn_box_reg=0.09816 total loss=2.93216
Model saved. Loss < PrevLoss (2.93216 < 3.80249)
Epoch 2 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.51it/s]
train: loss_classifier=4.99632 loss_box_reg=9.90219 loss_objectness=0.44928 loss_rpn_box_reg=0.56473 total loss=15.91252
valid: loss_classifier=0.85634 loss_box_reg=2.03114 loss_objectness=0.05460 loss_rpn_box_reg=0.11332 total loss=3.05539
Epoch 3 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.53it/s]
train: loss_classifier=4.93038 loss_box_reg=10.05001 loss_objectness=0.47636 loss_rpn_box_reg=0.61745 total loss=16.07420
valid: loss_classifier=0.93626 loss_box_reg=2.03956 loss_objectness=0.05904 loss_rpn_box_reg=0.11790 total loss=3.15275
Epoch 4 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.53it/s]
train: loss_classifier=4.84155 loss_box_reg=9.43750 loss_objectness=0.36683 loss_rpn_box_reg=0.56536 total loss=15.21124
valid: loss_classifier=0.88835 loss_box_reg=1.94999 loss_objectness=0.05383 loss_rpn_box_reg=0.12534 total loss=3.01752
Epoch 5 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.52it/s]
train: loss_classifier=4.74613 loss_box_reg=9.86788 loss_objectness=0.38734 loss_rpn_box_reg=0.59772 total loss=15.59908
valid: loss_classifier=0.92606 loss_box_reg=1.96283 loss_objectness=0.06114 loss_rpn_box_reg=0.10738 total loss=3.05741
Epoch 6 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.52it/s]
train: loss_classifier=4.34204 loss_box_reg=8.62916 loss_objectness=0.32403 loss_rpn_box_reg=0.49322 total loss=13.78844
valid: loss_classifier=0.86444 loss_box_reg=1.88869 loss_objectness=0.05894 loss_rpn_box_reg=0.09708 total loss=2.90915
Model saved. Loss < PrevLoss (2.90915 < 2.93216)
Epoch 7 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.51it/s]
train: loss_classifier=4.29596 loss_box_reg=8.85147 loss_objectness=0.37958 loss_rpn_box_reg=0.51758 total loss=14.04461
valid: loss_classifier=0.83501 loss_box_reg=1.77094 loss_objectness=0.04967 loss_rpn_box_reg=0.09526 total loss=2.75088
Model saved. Loss < PrevLoss (2.75088 < 2.90915)
Epoch 8 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.53it/s]
train: loss_classifier=4.27485 loss_box_reg=8.44950 loss_objectness=0.31232 loss_rpn_box_reg=0.45110 total loss=13.48777
valid: loss_classifier=0.81911 loss_box_reg=1.76916 loss_objectness=0.04457 loss_rpn_box_reg=0.09199 total loss=2.72483
Model saved. Loss < PrevLoss (2.72483 < 2.75088)
Epoch 9 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.50it/s]
train: loss_classifier=4.08536 loss_box_reg=8.48752 loss_objectness=0.30272 loss_rpn_box_reg=0.45049 total loss=13.32610
valid: loss_classifier=0.81484 loss_box_reg=1.81223 loss_objectness=0.04067 loss_rpn_box_reg=0.09072 total loss=2.75846
Epoch 10 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.52it/s]
train: loss_classifier=4.30853 loss_box_reg=8.45487 loss_objectness=0.31926 loss_rpn_box_reg=0.47008 total loss=13.55274
valid: loss_classifier=0.83762 loss_box_reg=1.80864 loss_objectness=0.04118 loss_rpn_box_reg=0.09237 total loss=2.77981
Epoch 11 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.50it/s]
train: loss_classifier=4.26765 loss_box_reg=8.53749 loss_objectness=0.28034 loss_rpn_box_reg=0.45764 total loss=13.54312
valid: loss_classifier=0.82711 loss_box_reg=1.78607 loss_objectness=0.04394 loss_rpn_box_reg=0.09273 total loss=2.74985
Epoch 12 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.51it/s]
train: loss_classifier=3.98940 loss_box_reg=8.27722 loss_objectness=0.29300 loss_rpn_box_reg=0.43633 total loss=12.99593
valid: loss_classifier=0.80736 loss_box_reg=1.78185 loss_objectness=0.04675 loss_rpn_box_reg=0.09183 total loss=2.72779
Epoch 13 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.52it/s]
train: loss_classifier=4.33019 loss_box_reg=8.51491 loss_objectness=0.28869 loss_rpn_box_reg=0.47466 total loss=13.60845
valid: loss_classifier=0.80979 loss_box_reg=1.79134 loss_objectness=0.04956 loss_rpn_box_reg=0.09283 total loss=2.74352
Epoch 14 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.51it/s]
train: loss_classifier=4.01974 loss_box_reg=8.22312 loss_objectness=0.29367 loss_rpn_box_reg=0.45659 total loss=12.99311
valid: loss_classifier=0.82477 loss_box_reg=1.78684 loss_objectness=0.04330 loss_rpn_box_reg=0.09092 total loss=2.74583
Epoch 15 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.50it/s]
train: loss_classifier=4.10931 loss_box_reg=8.23071 loss_objectness=0.31868 loss_rpn_box_reg=0.46261 total loss=13.12131
valid: loss_classifier=0.82687 loss_box_reg=1.77465 loss_objectness=0.04676 loss_rpn_box_reg=0.09016 total loss=2.73843
Epoch 16 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.51it/s]
train: loss_classifier=4.11051 loss_box_reg=8.19198 loss_objectness=0.30949 loss_rpn_box_reg=0.43948 total loss=13.05147
valid: loss_classifier=0.81724 loss_box_reg=1.77781 loss_objectness=0.04473 loss_rpn_box_reg=0.09044 total loss=2.73023
Epoch 17 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.53it/s]
train: loss_classifier=4.28086 loss_box_reg=8.38024 loss_objectness=0.30397 loss_rpn_box_reg=0.44733 total loss=13.41239
valid: loss_classifier=0.80994 loss_box_reg=1.77486 loss_objectness=0.05051 loss_rpn_box_reg=0.09075 total loss=2.72606
Epoch 18 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.51it/s]
train: loss_classifier=4.33978 loss_box_reg=8.58205 loss_objectness=0.35799 loss_rpn_box_reg=0.48077 total loss=13.76059
valid: loss_classifier=0.80758 loss_box_reg=1.77697 loss_objectness=0.04526 loss_rpn_box_reg=0.09157 total loss=2.72138
Model saved. Loss < PrevLoss (2.72138 < 2.72483)
Epoch 19 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.51it/s]
train: loss_classifier=4.11305 loss_box_reg=8.32378 loss_objectness=0.31888 loss_rpn_box_reg=0.47125 total loss=13.22696
valid: loss_classifier=0.82873 loss_box_reg=1.77459 loss_objectness=0.04346 loss_rpn_box_reg=0.09185 total loss=2.73863
Epoch 20 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.51it/s]
train: loss_classifier=3.98643 loss_box_reg=8.23501 loss_objectness=0.29916 loss_rpn_box_reg=0.44082 total loss=12.96142
valid: loss_classifier=0.81920 loss_box_reg=1.77603 loss_objectness=0.04569 loss_rpn_box_reg=0.09165 total loss=2.73257
�[1m�[4mFold 5 of 5�[0m
Epoch 1 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=4.75750 loss_box_reg=9.90629 loss_objectness=0.40213 loss_rpn_box_reg=0.56773 total loss=15.63364
valid: loss_classifier=0.81628 loss_box_reg=1.94861 loss_objectness=0.02717 loss_rpn_box_reg=0.08378 total loss=2.87585
Epoch 2 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.58it/s]
train: loss_classifier=4.64836 loss_box_reg=9.66521 loss_objectness=0.45440 loss_rpn_box_reg=0.59994 total loss=15.36791
valid: loss_classifier=0.73029 loss_box_reg=1.65242 loss_objectness=0.03310 loss_rpn_box_reg=0.07725 total loss=2.49306
Model saved. Loss < PrevLoss (2.49306 < 2.72138)
Epoch 3 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.41s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.58it/s]
train: loss_classifier=4.66250 loss_box_reg=9.71517 loss_objectness=0.43311 loss_rpn_box_reg=0.56149 total loss=15.37227
valid: loss_classifier=0.67239 loss_box_reg=1.54756 loss_objectness=0.02647 loss_rpn_box_reg=0.06760 total loss=2.31401
Model saved. Loss < PrevLoss (2.31401 < 2.49306)
Epoch 4 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=4.36495 loss_box_reg=9.33008 loss_objectness=0.36826 loss_rpn_box_reg=0.54079 total loss=14.60408
valid: loss_classifier=0.81560 loss_box_reg=1.83050 loss_objectness=0.02303 loss_rpn_box_reg=0.07325 total loss=2.74238
Epoch 5 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=4.55119 loss_box_reg=9.33605 loss_objectness=0.36680 loss_rpn_box_reg=0.52222 total loss=14.77625
valid: loss_classifier=0.78369 loss_box_reg=1.76101 loss_objectness=0.02369 loss_rpn_box_reg=0.07994 total loss=2.64833
Epoch 6 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=4.09056 loss_box_reg=8.47766 loss_objectness=0.32095 loss_rpn_box_reg=0.45212 total loss=13.34130
valid: loss_classifier=0.71258 loss_box_reg=1.52779 loss_objectness=0.02085 loss_rpn_box_reg=0.06566 total loss=2.32688
Epoch 7 of 20:
train: 100%|██████████| 62/62 [01:29<00:00, 1.44s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.52it/s]
train: loss_classifier=4.00072 loss_box_reg=8.36098 loss_objectness=0.28402 loss_rpn_box_reg=0.45292 total loss=13.09863
valid: loss_classifier=0.68208 loss_box_reg=1.49344 loss_objectness=0.01769 loss_rpn_box_reg=0.06418 total loss=2.25738
Model saved. Loss < PrevLoss (2.25738 < 2.31401)
Epoch 8 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=3.87103 loss_box_reg=8.23573 loss_objectness=0.33236 loss_rpn_box_reg=0.48052 total loss=12.91963
valid: loss_classifier=0.71531 loss_box_reg=1.45590 loss_objectness=0.01974 loss_rpn_box_reg=0.06138 total loss=2.25232
Model saved. Loss < PrevLoss (2.25232 < 2.25738)
Epoch 9 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=3.87284 loss_box_reg=8.03685 loss_objectness=0.30199 loss_rpn_box_reg=0.43473 total loss=12.64642
valid: loss_classifier=0.68127 loss_box_reg=1.44604 loss_objectness=0.01348 loss_rpn_box_reg=0.06241 total loss=2.20320
Model saved. Loss < PrevLoss (2.20320 < 2.25232)
Epoch 10 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=3.92226 loss_box_reg=8.09959 loss_objectness=0.28075 loss_rpn_box_reg=0.44109 total loss=12.74369
valid: loss_classifier=0.69403 loss_box_reg=1.41649 loss_objectness=0.01881 loss_rpn_box_reg=0.06138 total loss=2.19071
Model saved. Loss < PrevLoss (2.19071 < 2.20320)
Epoch 11 of 20:
train: 100%|██████████| 62/62 [01:27<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=3.85472 loss_box_reg=8.04091 loss_objectness=0.30514 loss_rpn_box_reg=0.43802 total loss=12.63879
valid: loss_classifier=0.65885 loss_box_reg=1.42345 loss_objectness=0.01693 loss_rpn_box_reg=0.06193 total loss=2.16117
Model saved. Loss < PrevLoss (2.16117 < 2.19071)
Epoch 12 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=3.81633 loss_box_reg=7.82985 loss_objectness=0.27837 loss_rpn_box_reg=0.44096 total loss=12.36551
valid: loss_classifier=0.67133 loss_box_reg=1.42733 loss_objectness=0.01882 loss_rpn_box_reg=0.06191 total loss=2.17939
Epoch 13 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.54it/s]
train: loss_classifier=4.03471 loss_box_reg=8.03164 loss_objectness=0.27359 loss_rpn_box_reg=0.43124 total loss=12.77118
valid: loss_classifier=0.67612 loss_box_reg=1.42304 loss_objectness=0.01682 loss_rpn_box_reg=0.06171 total loss=2.17769
Epoch 14 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.53it/s]
train: loss_classifier=3.90455 loss_box_reg=8.13206 loss_objectness=0.28638 loss_rpn_box_reg=0.43309 total loss=12.75608
valid: loss_classifier=0.67050 loss_box_reg=1.41271 loss_objectness=0.02222 loss_rpn_box_reg=0.06152 total loss=2.16695
Epoch 15 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=3.82431 loss_box_reg=7.82751 loss_objectness=0.29736 loss_rpn_box_reg=0.44886 total loss=12.39804
valid: loss_classifier=0.67411 loss_box_reg=1.42022 loss_objectness=0.02304 loss_rpn_box_reg=0.06168 total loss=2.17905
Epoch 16 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.55it/s]
train: loss_classifier=3.70396 loss_box_reg=7.89814 loss_objectness=0.28545 loss_rpn_box_reg=0.44193 total loss=12.32948
valid: loss_classifier=0.66730 loss_box_reg=1.41717 loss_objectness=0.02378 loss_rpn_box_reg=0.06164 total loss=2.16989
Epoch 17 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.56it/s]
train: loss_classifier=3.69627 loss_box_reg=7.86215 loss_objectness=0.26493 loss_rpn_box_reg=0.43404 total loss=12.25740
valid: loss_classifier=0.68288 loss_box_reg=1.42481 loss_objectness=0.02075 loss_rpn_box_reg=0.06152 total loss=2.18996
Epoch 18 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.43s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.53it/s]
train: loss_classifier=3.78514 loss_box_reg=7.67920 loss_objectness=0.25909 loss_rpn_box_reg=0.42823 total loss=12.15166
valid: loss_classifier=0.67166 loss_box_reg=1.42271 loss_objectness=0.01591 loss_rpn_box_reg=0.06153 total loss=2.17180
Epoch 19 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.57it/s]
train: loss_classifier=3.94499 loss_box_reg=8.24162 loss_objectness=0.31013 loss_rpn_box_reg=0.47798 total loss=12.97473
valid: loss_classifier=0.68002 loss_box_reg=1.42221 loss_objectness=0.01116 loss_rpn_box_reg=0.06154 total loss=2.17493
Epoch 20 of 20:
train: 100%|██████████| 62/62 [01:28<00:00, 1.42s/it]
valid: 100%|██████████| 16/16 [00:10<00:00, 1.58it/s]
train: loss_classifier=3.75961 loss_box_reg=7.80162 loss_objectness=0.26578 loss_rpn_box_reg=0.43370 total loss=12.26071
valid: loss_classifier=0.69305 loss_box_reg=1.43105 loss_objectness=0.02125 loss_rpn_box_reg=0.06152 total loss=2.20687
It's best to look at Tensorboard, but for a quick glance, here are a few examples.
To better understand the learning progress of the model over time, it would be useful to analyze the graph of the folds as a series.
print(DEBUG)
run_path = os.path.join(sys.path[0], 'runs', '09-11-21_20:49')
print(run_path)
model_path = os.path.join(run_path, 'models', 'faster_rcnn_model.pth')
model = get_model(4)
model.load_state_dict(torch.load(model_path))False
/home/linuxgpu/anaconda3/envs/lab1/Face-Mask-Detection/runs/09-11-21_20:49
<All keys matched successfully>
testset = FaceMaskDataset(x_test, y_test, CFG.imgs_path, CFG.msks_path, CFG.img_width, CFG.img_height, transforms=get_transformer('test'))
test_loader = DataLoader(dataset=testset, batch_size=CFG.batch_size, shuffle=False, num_workers=CFG.num_workers, collate_fn=collate_fn)def plot_test_examples(model, loader, nms_threshold, score_threshold):
device = CFG.device
model = model.to(device).eval()
imgs, annts = next(iter(loader))
imgs = list(img.to(device) for img in imgs)
output = model(imgs)
amount_samples = min(12, len(imgs)) # number of examples will be at most 12
for i, (img, ant, pred) in enumerate(zip(imgs, annts, output)):
if i >= amount_samples:
break
fig, axes = plt.subplots(nrows=1, ncols=2, figsize=(14, 60))
axes[0].set_title('Ground Truth')
axes[1].set_title('Prediction')
axes[0].axis('off')
axes[1].axis('off')
annot_img = get_annotated_img(img.cpu().detach().numpy(), ant)
axes[0].imshow(norm(annot_img))
pred_thrs = get_pred_by_threshold(pred, nms_threshold, score_threshold)
pred_cpu = {k: v.cpu().detach().numpy() for k, v in pred_thrs.items()}
annot_img = get_annotated_img(img.cpu().detach().numpy(), pred_cpu, is_pred=True)
axes[1].imshow(norm(annot_img))
fig.tight_layout()
plt.show()
del imgs, pred, pred_thrs, pred_cpu
torch.cuda.empty_cache()
plot_test_examples(model, test_loader, nms_threshold=CFG.nms_thresh, score_threshold=CFG.score_threshold)
mAP@[0.5:0.05:0.95]
def plot_precision_recall(df, iou_thresh, save_path=None):
auc_dict = dict()
fig, axes = plt.subplots(nrows=1, ncols=3, figsize=(25, 7))
plt.title('IoU Threshold {:.3f}'.format(iou_thresh))
for idx, c in enumerate(df['real_class'].unique()):
prec = df[df['real_class']==c]['precision']
recall = df[df['real_class']==c]['recall']
prec = pd.concat([ pd.Series([prec.iloc[0]]), prec ])
recall = pd.concat([ pd.Series([0]), recall ])
prec = pd.concat([ prec, pd.Series([0]) ])
recall = pd.concat([ recall, pd.Series([ recall.iloc[len(recall)-1] ]) ])
auc = metrics.auc(x=recall, y=prec)
auc_dict[c] = auc
axes[idx].plot(recall, prec)
axes[idx].set_title('Class {} AP@{:.2f}={:.5f}'.format(faceMasksData.classes[c], iou_thresh, auc))
axes[idx].grid()
axes[idx].set_ylabel('Precision')
axes[idx].set_xlabel('Recall')
axes[idx].axis([-0.01, 1.01, -0.01, 1.01])
fig.tight_layout()
plt.show()
if save_path:
fig.savefig('{}/AP@{:.3}.png'.format(save_path, iou_thresh))
return auc_dictdef plot_mAP_graph(iou_dict, save_path=None, iou_start=0.5, iou_step=0.05, iou_end=0.95):
x, y = zip(*sorted(iou_dict.items()))
fig = plt.figure(figsize=(25, 7))
plt.title('mAP Vs. IoU Threshold')
plt.plot(x, y, marker='o')
plt.tight_layout()
plt.grid()
plt.show()
save_path = '{}/mAP@[{:.3}:{:.3}:{:.3}].png'.format(save_path, iou_start, iou_step, iou_end)
fig.savefig(save_path)def evaluate(model, loader, nms_thresh, score_threshold=0, iou_start=0.5, iou_step=0.05, iou_end=0.95, eps=1e-6):
metrics_path = '{}/metrics'.format(run_path)
df = get_iou_as_df(model, loader, nms_thresh, score_threshold)
auc_dict = dict()
for iou_trsh in np.arange(iou_start, iou_end + iou_step, iou_step):
df = calc_precision_recall(df, iou_trsh, metrics_path)
# display(df)
auc_dict[iou_trsh] = plot_precision_recall(df, iou_trsh, metrics_path)
iou_dict = calc_mAP_from_auc_dict(auc_dict)
classes_mAP = calc_mAP_per_class(auc_dict, iou_dict)
plot_mAP_graph(iou_dict, metrics_path, iou_start, iou_step, iou_end)
total_mAP = .0
print()
for iou in iou_dict:
print('mAP@{:.3} =\t{:.5}'.format(iou, iou_dict[iou]))
print('\nmAP Per Class:')
for c, val in classes_mAP.items():
total_mAP += val
print('{}\t\tmAP@[0.5:0.05:0.95] =\t{:.5f}'.format(faceMasksData.classes[c], classes_mAP[c]))
total_mAP /= len(classes_mAP)
print('Total mAP@[0.5:0.05:0.95] =\t{:.5f}'.format(total_mAP))
evaluate(model, test_loader, CFG.nms_thresh, CFG.score_threshold)evaluating IoU: 100%|██████████| 9/9 [00:08<00:00, 1.11it/s]
[email protected] = 0.77244
[email protected] = 0.77223
[email protected] = 0.76981
[email protected] = 0.7628
[email protected] = 0.74783
[email protected] = 0.71212
[email protected] = 0.65929
[email protected] = 0.43441
[email protected] = 0.21814
[email protected] = 0.019173
mAP Per Class:
with_mask mAP@[0.5:0.05:0.95] = 0.75002
without_mask mAP@[0.5:0.05:0.95] = 0.77128
mask_weared_incorrect mAP@[0.5:0.05:0.95] = 0.23917
Total mAP@[0.5:0.05:0.95] = 0.58682