def train(im_size,
train_mode,
epoch_length=1000,
num_epochs=10,
val_steps=200,
lr_mode='adaptive'):
""" Train the model
Args:
im_size: user input. Input images are resized to this size.
train_mode: 1 to train new model with pre-trained weights on generic dataset.
2 to keep training on the WashingtonOB Race dataset.
epoch_length: number of steps per training epoch
num_epochs: maximum number of training epochs
val_steps: number of validation steps, at the end of each training epoch
lr_mode: if 'adaptive', learning rate varies with eponential decay. If 'constant', it is 1e-5.
Returns:
None - (trained model saved automatically)
"""
C = config.Config()
C.network = 'resnet50'
C.im_size = im_size
C.model_path = f'models/model_frcnn_{C.im_size}.hdf5'
all_imgs, classes_count, class_mapping = get_data(
'data/training/corners_training.txt')
# add background class
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
C.class_mapping = class_mapping
print(
f'Total number of objects per class (across all dataset): {class_mapping}'
)
config_output_filename = f'models/conf/config_frcnn_{C.im_size}.pickle'
record_path = f'models/logs/frcnn_{C.im_size}.csv' # Where to record data (used to save the losses, classification accuracy and mean average precision)
if lr_mode == 'adaptive':
record_path = f'models/logs/frcnn_{C.im_size}_adap.csv'
pickle.dump(C, open(config_output_filename, 'wb'))
train_imgs = [s for s in all_imgs if s['imageset'] == 'train']
val_imgs = [s for s in all_imgs if s['imageset'] == 'val']
# Shuffle the images with seed 1
random.seed(1)
random.shuffle(train_imgs)
random.shuffle(val_imgs)
print(
f'{len(train_imgs)} and {len(val_imgs)} training and validation samples (before augmentation), respectively.'
)
data_gen_train = regions_proposal_network.get_anchor_gt(
train_imgs, C, nn.get_img_output_length, mode='train')
data_gen_val = regions_proposal_network.get_anchor_gt(
val_imgs, C, nn.get_img_output_length, mode='val')
X, Y, image_data = next(data_gen_train)
print('Original image: height=%d width=%d' %
(image_data['height'], image_data['width']))
print('Resized image: height=%d width=%d' % (X.shape[1], X.shape[2]))
print('Feature map size: height=%d width=%d C.rpn_stride=%d' %
(Y[0].shape[1], Y[0].shape[2], C.rpn_stride))
input_shape_img = (None, None, 3)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(None, 4))
# define the base network (resnet here)
shared_layers = nn.nn_base(img_input, trainable=True)
# define the RPN, built on the base layers
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(shared_layers,
roi_input,
C.num_rois,
nb_classes=len(classes_count))
model_rpn = Model(img_input, rpn[:2])
model_classifier = Model([img_input, roi_input], classifier)
# this is a model that holds both the RPN and the classifier, used to load/save weights for the models
model_all = Model([img_input, roi_input], rpn[:2] + classifier)
model_rpn, model_classifier, record_df = load_weights(
im_size, model_rpn, model_classifier, train_mode)
optimizer = Adam(lr=1e-5)
optimizer_classifier = Adam(lr=1e-5)
model_rpn.compile(optimizer=optimizer,
loss=[
losses.rpn_loss_cls(num_anchors),
losses.rpn_loss_regr(num_anchors)
])
model_classifier.compile(
optimizer=optimizer_classifier,
loss=[
losses.class_loss_cls,
losses.class_loss_regr(len(classes_count) - 1)
],
metrics={'dense_class_{}'.format(len(classes_count)): 'accuracy'})
model_all.compile(optimizer='sgd', loss='mae')
# Training setting
total_epochs = len(record_df)
if len(record_df) == 0:
best_loss = np.Inf
else:
best_loss = np.min(record_df['curr_loss_val'])
print(
f'Resuming training. Already trained for {len(record_df)} epochs.')
validation_trend_hold = False
total_epochs += num_epochs
iter_num = 0
loss = np.zeros((epoch_length, 5))
rpn_accuracy_rpn_monitor = []
rpn_accuracy_for_epoch = []
print('Starting training')
start_time = time.time()
for epoch_num in range(num_epochs):
progbar = generic_utils.Progbar(epoch_length)
print('Epoch {}/{}'.format(epoch_num + 1, num_epochs))
while True:
try:
if len(rpn_accuracy_rpn_monitor) == epoch_length and C.verbose:
mean_overlapping_bboxes = float(
sum(rpn_accuracy_rpn_monitor)) / len(
rpn_accuracy_rpn_monitor)
rpn_accuracy_rpn_monitor = []
if mean_overlapping_bboxes == 0:
print(
'RPN is not producing bounding boxes that overlap the ground truth boxes. Check RPN settings or keep training.'
)
# use next to extract data since it is a generator
X, Y, img_data = next(data_gen_train)
# Train rpn model and get loss value [_, loss_rpn_cls, loss_rpn_regr]
current_learning_rate = calculate_learning_rate(epoch_num,
mode=lr_mode)
K.set_value(model_rpn.optimizer.lr, current_learning_rate)
loss_rpn = model_rpn.train_on_batch(X, Y)
# Get predicted rpn from rpn model [rpn_cls, rpn_regr]
P_rpn = model_rpn.predict_on_batch(X)
# R: bboxes (shape=(im_size,4))
# Convert rpn layer to roi bboxes
R = roi_helpers.rpn_to_roi(P_rpn[0],
P_rpn[1],
C,
overlap_thresh=0.6,
max_boxes=C.max_boxes)
# note: calc_iou converts from (x1,y1,x2,y2) to (x,y,w,h) format
# X2: bboxes with iou > C.classifier_min_overlap for all gt bboxes in 20 non_max_suppression bboxes
# Y2: corresponding labels and corresponding ground truth bboxes
X2, Y1, Y2, IouS = roi_helpers.calc_iou(
R, img_data, C, class_mapping)
if X2 is None:
rpn_accuracy_rpn_monitor.append(0)
rpn_accuracy_for_epoch.append(0)
continue
neg_samples = np.where(Y1[0, :, -1] == 1)
pos_samples = np.where(Y1[0, :, -1] == 0)
if len(neg_samples) > 0:
neg_samples = neg_samples[0]
else:
neg_samples = []
if len(pos_samples) > 0:
pos_samples = pos_samples[0]
else:
pos_samples = []
rpn_accuracy_rpn_monitor.append(len(pos_samples))
rpn_accuracy_for_epoch.append((len(pos_samples)))
if C.num_rois > 1:
# If number of positive anchors is larger than 4//2 = 2, randomly choose 2 pos samples
if len(pos_samples) < C.num_rois // 2:
selected_pos_samples = pos_samples.tolist()
else:
selected_pos_samples = np.random.choice(
pos_samples, C.num_rois // 2,
replace=False).tolist()
# Randomly choose (num_rois - num_pos) neg samples
try:
selected_neg_samples = np.random.choice(
neg_samples,
C.num_rois - len(selected_pos_samples),
replace=False).tolist()
except:
selected_neg_samples = np.random.choice(
neg_samples,
C.num_rois - len(selected_pos_samples),
replace=True).tolist()
# Save all the pos and neg samples in sel_samples
sel_samples = selected_pos_samples + selected_neg_samples
else:
# in the extreme case where num_rois = 1, we pick a random pos or neg sample
selected_pos_samples = pos_samples.tolist()
selected_neg_samples = neg_samples.tolist()
if np.random.randint(0, 2):
sel_samples = random.choice(neg_samples)
else:
sel_samples = random.choice(pos_samples)
# training_data: [X, X2[:, sel_samples, :]]
# labels: [Y1[:, sel_samples, :], Y2[:, sel_samples, :]]
# X => img_data resized image
# X2[:, sel_samples, :] => num_rois (4 in here) bboxes which contains selected neg and pos
# Y1[:, sel_samples, :] => one hot encode for num_rois bboxes which contains selected neg and pos
# Y2[:, sel_samples, :] => labels and gt bboxes for num_rois bboxes which contains selected neg and pos
K.set_value(model_classifier.optimizer.lr,
current_learning_rate)
loss_class = model_classifier.train_on_batch(
[X, X2[:, sel_samples, :]],
[Y1[:, sel_samples, :], Y2[:, sel_samples, :]])
loss[iter_num, 0] = loss_rpn[1]
loss[iter_num, 1] = loss_rpn[2]
loss[iter_num, 2] = loss_class[1]
loss[iter_num, 3] = loss_class[2]
loss[iter_num, 4] = loss_class[3]
progbar.update(
iter_num + 1,
[('RPN Classifier Loss', loss[iter_num, 0]),
('RPN Regression Loss', loss[iter_num, 1]),
('Detector Classifier Loss', loss[iter_num, 2]),
('Detector Regression Loss', loss[iter_num, 3])])
iter_num += 1
# end of epoch check
if iter_num == epoch_length:
loss_rpn_cls = np.mean(loss[:, 0])
loss_rpn_regr = np.mean(loss[:, 1])
loss_class_cls = np.mean(loss[:, 2])
loss_class_regr = np.mean(loss[:, 3])
class_acc = np.mean(loss[:, 4])
print("Performing validation.")
val_loss = validate(val_steps, data_gen_val, model_rpn, C,
class_mapping, model_classifier)
mean_overlapping_bboxes = float(sum(
rpn_accuracy_for_epoch)) / len(rpn_accuracy_for_epoch)
rpn_accuracy_for_epoch = []
if C.verbose:
print(
f'Classifier accuracy for bounding boxes: {class_acc}'
)
curr_loss = loss_rpn_cls + loss_rpn_regr + loss_class_cls + loss_class_regr
iter_num = 0
if val_loss['curr_loss'] <= best_loss:
if C.verbose:
print(
f'Total validation loss decreased from {best_loss} to {val_loss["curr_loss"]}, saving weights.'
)
print('')
best_loss = val_loss['curr_loss']
model_all.save_weights(C.model_path)
validation_trend_hold = False
elif not validation_trend_hold:
if C.verbose:
print(
f'Total validation loss increased for the first time, from {best_loss} to {val_loss["curr_loss"]}. Performing one more epoch to verify trend. Not saving weights for now.'
)
print('')
validation_trend_hold = True
else:
if C.verbose:
print(
f'Total validation loss increased for the second time, from {best_loss} to {val_loss["curr_loss"]}.'
)
print(
f'Terminating training now to prevent over-fitting. Keeping weights from epoch {epoch_num - 1}.'
)
exit()
new_row = {
'mean_overlapping_bboxes':
round(mean_overlapping_bboxes, 3),
'class_acc':
round(class_acc, 3),
'loss_rpn_cls':
round(loss_rpn_cls, 3),
'loss_rpn_regr':
round(loss_rpn_regr, 3),
'loss_class_cls':
round(loss_class_cls, 3),
'loss_class_regr':
round(loss_class_regr, 3),
'curr_loss':
round(curr_loss, 3),
'class_acc_val':
round(val_loss['class_acc'], 3),
'curr_loss_val':
round(val_loss['curr_loss'], 3),
'elapsed_time':
round(time.time() - start_time, 3)
}
start_time = time.time()
record_df = record_df.append(new_row, ignore_index=True)
record_df.to_csv(record_path, index=False)
break
except Exception as e:
print(f'Exception: {e}')
continue
print('Training complete.')
return
def detect_img(img_name):
use_horizontal_flips = False
use_vertical_flips = False
rot_90 = False
im_size = 600
anchor_box_scales = [64, 128, 256, 512]
anchor_box_ratios = [[1, 1], [1, 2], [2, 1]]
im_size = 600
img_channel_mean = [103.939, 116.779, 123.68]
img_scaling_factor = 1.0
num_rois = 4
rpn_stride = 16
balanced_classes = False
std_scaling = 4.0
classifier_regr_std = [8.0, 8.0, 4.0, 4.0]
rpn_min_overlap = 0.3
rpn_max_overlap = 0.7
classifier_min_overlap = 0.1
classifier_max_overlap = 0.5
class_mapping = {'MALIGNANT': 0, 'BENIGN': 1, 'bg': 2}
if 'bg' not in class_mapping:
class_mapping['bg'] = len(class_mapping)
class_mapping = {v: k for k, v in class_mapping.items()}
print(class_mapping)
class_to_color = {
class_mapping[v]: np.random.randint(0, 255, 3)
for v in class_mapping
}
num_features = 1024
if K.image_dim_ordering() == 'th':
input_shape_img = (3, None, None)
input_shape_features = (num_features, None, None)
else:
input_shape_img = (None, None, 3)
input_shape_features = (None, None, num_features)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(num_rois, 4))
feature_map_input = Input(shape=input_shape_features)
# define the base network (resnet here, can be VGG, Inception, etc)
shared_layers = nn.nn_base(img_input, trainable=True)
# define the RPN, built on the base layers
num_anchors = len(anchor_box_scales) * len(anchor_box_ratios)
rpn_layers = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(feature_map_input,
roi_input,
num_rois,
nb_classes=len(class_mapping),
trainable=True)
model_rpn = Model(img_input, rpn_layers)
model_classifier_only = Model([feature_map_input, roi_input], classifier)
model_classifier = Model([feature_map_input, roi_input], classifier)
model_path = "frcnn\\model_final_1.hdf5"
# print('Loading weights from {}'.format(model_path))
# model_rpn.load_weights(model_path, by_name=True)
# model_classifier.load_weights(model_path, by_name=True)
# model_rpn.compile(optimizer='sgd', loss='mse')
# model_classifier.compile(optimizer='sgd', loss='mse')
model_rpn, model_classifier = get_model(model_path, model_rpn,
model_classifier)
all_imgs = []
classes = {}
bbox_threshold = 0.8
print(img_name)
st = time.time()
# filepath = os.path.join(img_path,img_name)
img = cv2.imread(img_name)
X, ratio = format_img(img, im_size, img_channel_mean, img_scaling_factor)
if K.image_dim_ordering() == 'tf':
X = np.transpose(X, (0, 2, 3, 1))
# get the feature maps and output from the RPN
[Y1, Y2, F] = model_rpn.predict(X)
R = roi_helpers.rpn_to_roi(Y1,
Y2,
anchor_box_scales,
anchor_box_ratios,
std_scaling,
rpn_stride,
K.image_dim_ordering(),
overlap_thresh=0.5)
# convert from (x1,y1,x2,y2) to (x,y,w,h)
R[:, 2] -= R[:, 0]
R[:, 3] -= R[:, 1]
# apply the spatial pyramid pooling to the proposed regions
bboxes = {}
probs = {}
for jk in range(R.shape[0] // num_rois + 1):
ROIs = np.expand_dims(R[num_rois * jk:num_rois * (jk + 1), :], axis=0)
if ROIs.shape[1] == 0:
break
if jk == R.shape[0] // num_rois:
#pad R
curr_shape = ROIs.shape
target_shape = (curr_shape[0], num_rois, curr_shape[2])
ROIs_padded = np.zeros(target_shape).astype(ROIs.dtype)
ROIs_padded[:, :curr_shape[1], :] = ROIs
ROIs_padded[0, curr_shape[1]:, :] = ROIs[0, 0, :]
ROIs = ROIs_padded
[P_cls, P_regr] = model_classifier_only.predict([F, ROIs])
for ii in range(P_cls.shape[1]):
if np.argmax(P_cls[0, ii, :]) == (P_cls.shape[2] - 1):
continue
cls_name = class_mapping[np.argmax(P_cls[0, ii, :])]
if cls_name not in bboxes:
bboxes[cls_name] = []
probs[cls_name] = []
(x, y, w, h) = ROIs[0, ii, :]
cls_num = np.argmax(P_cls[0, ii, :])
try:
(tx, ty, tw, th) = P_regr[0, ii, 4 * cls_num:4 * (cls_num + 1)]
tx /= classifier_regr_std[0]
ty /= classifier_regr_std[1]
tw /= classifier_regr_std[2]
th /= classifier_regr_std[3]
x, y, w, h = roi_helpers.apply_regr(x, y, w, h, tx, ty, tw, th)
except:
pass
bboxes[cls_name].append([
rpn_stride * x, rpn_stride * y, rpn_stride * (x + w),
rpn_stride * (y + h)
])
probs[cls_name].append(np.max(P_cls[0, ii, :]))
all_dets = []
for key in bboxes:
bbox = np.array(bboxes[key])
new_boxes, new_probs = roi_helpers.non_max_suppression_fast(
bbox, np.array(probs[key]), overlap_thresh=0.5)
for jk in range(new_boxes.shape[0]):
(x1, y1, x2, y2) = new_boxes[jk, :]
(real_x1, real_y1, real_x2,
real_y2) = get_real_coordinates(ratio, x1, y1, x2, y2)
cv2.rectangle(
img, (real_x1, real_y1), (real_x2, real_y2),
(int(class_to_color[key][0]), int(
class_to_color[key][1]), int(class_to_color[key][2])), 2)
textLabel = '{}: {}'.format(key, int(100 * new_probs[jk]))
all_dets.append((key, 100 * new_probs[jk]))
(retval, baseLine) = cv2.getTextSize(textLabel,
cv2.FONT_HERSHEY_COMPLEX, 1,
1)
textOrg = (real_x1, real_y1 - 0)
cv2.rectangle(
img, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5),
(0, 0, 0), 2)
cv2.rectangle(
img, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5),
(255, 255, 255), -1)
cv2.putText(img, textLabel, textOrg, cv2.FONT_HERSHEY_DUPLEX, 1,
(0, 0, 0), 1)
# print('Elapsed time = {}'.format(time.time() - st))
# print(all_dets)
# cv2.imshow('img', img)
# cv2.waitKey(0)
img_name = img_name.split('\\')[-1]
# cv2.imwrite(f'./static/images/{img_name}.png', img)
cv2.imwrite('./predict/kq.jpg', img)
try:
a = all_dets[0]
except:
a = ("khong phat hien", "khong phat hien")
print(a)
return img_name, a
# print("tp: {} \nfp: {}".format(tp, fp))
# img_name = r"D:\Desktop\thesis\Images\mass_crop_train\P_01981_RIGHT_MLO_FULL.jpg"
# a, b = detect_img(img_name)
# print(b[0], b[1])
# print(type(b[0]), type(b[1]))
def setupFRCNN(options):
# suppressing tensorflow warnings and allowing gpu memory allocation to grow
import os
import tensorflow as tf
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
configTF = tf.ConfigProto()
configTF.gpu_options.allow_growth = True
sess = tf.Session(config=configTF)
# setting system recursion limit
import sys
sys.setrecursionlimit(40000)
# loading config file
with open(options.config_filename, 'r') as f_in:
C = pickle.load(f_in)
# turn off any data augmentation at run time
C.use_horizontal_flips = False
C.use_vertical_flips = False
C.rot_90 = False
# obtaining classes used to train classifier
class_mapping = C.class_mapping
if 'bg' not in class_mapping:
class_mapping['bg'] = len(class_mapping)
class_mapping = {v: k for k, v in class_mapping.iteritems()}
class_to_color = {class_mapping[v]: np.random.randint(0, 255, 3) for v in class_mapping}
# number of ROIs fed to classifier per batch
C.num_rois = int(options.num_rois)
# acccounting for difference in img dimesion order in theano/tensorflow
if K.image_dim_ordering() == 'th':
input_shape_img = (3, None, None)
input_shape_features = (1024, None, None)
else:
input_shape_img = (None, None, 3)
input_shape_features = (None, None, 1024)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(C.num_rois, 4))
feature_map_input = Input(shape=input_shape_features)
# setting up the layers in RPN and CNN classifier
shared_layers = nn.nn_base(img_input, trainable=True)
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn_layers = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(feature_map_input, roi_input, C.num_rois, nb_classes=len(class_mapping), trainable=True)
model_rpn = Model(img_input, rpn_layers)
model_classifier_only = Model([feature_map_input, roi_input], classifier)
model_classifier = Model([feature_map_input, roi_input], classifier)
# loading weights
model_rpn.load_weights(options.model_weight_path, by_name=True)
model_classifier.load_weights(options.model_weight_path, by_name=True)
# compiling models
model_rpn.compile(optimizer='sgd', loss='mse')
model_classifier.compile(optimizer='sgd', loss='mse')
return C,model_rpn,model_classifier,model_classifier_only
input_shape_features = (num_features, None, None)
else:
input_shape_img = (None, None, 3)
input_shape_features = (None, None, num_features)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(C.num_rois, 4))
feature_map_input = Input(shape=input_shape_features)
# define the base network (resnet here, can be VGG, Inception, etc)
shared_layers = nn.nn_base(img_input, trainable=True)
# define the RPN, built on the base layers
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn_layers = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(feature_map_input, roi_input, C.num_rois, nb_classes=len(class_mapping), trainable=True)
model_rpn = Model(img_input, rpn_layers)
model_classifier_only = Model([feature_map_input, roi_input], classifier)
model_classifier = Model([feature_map_input, roi_input], classifier)
print('Loading weights from {}'.format(C.model_path))
model_rpn.load_weights(C.model_path, by_name=True)
model_classifier.load_weights(C.model_path, by_name=True)
model_rpn.compile(optimizer='sgd', loss='mse')
model_classifier.compile(optimizer='sgd', loss='mse')
(x2 * C.rpn_stride, y2 * C.rpn_stride),
colors[class_id].tolist(), 2)
cv2.putText(a_image, '{} {:.2f}'.format(class_name,
pos_ious[idx]), (x1, y1 + 10),
cv2.FONT_HERSHEY_COMPLEX, 0.7, (0, 0, 255), 1)
cv2.namedWindow('a_image', cv2.WINDOW_NORMAL)
cv2.imshow('a_image', a_image)
cv2.waitKey(0)
image_input_shape = (None, None, 3)
image_input = Input(shape=image_input_shape)
num_anchors = 9
# define the base network (resnet here, can be VGG, Inception, etc)
base_net_output = nn.base_net(image_input)
rpn_output = nn.rpn(base_net_output, num_anchors)
with open('../config.pickle', 'rb') as f_in:
C = pickle.load(f_in)
class_name_idx_mapping = C.class_name_idx_mapping
if 'bg' not in class_name_idx_mapping:
class_name_idx_mapping['bg'] = len(class_name_idx_mapping)
class_idx_name_mapping = {v: k for k, v in class_name_idx_mapping.items()}
model_rpn = Model(image_input, rpn_output)
model_rpn.load_weights(
'../frcnn_resnet50_8_0.7191_0.1186_0.2485_0.1342_0.9118.hdf5',
by_name=True)
annotations = json.load(open('annotation_data.json'))
train_annotations = [
annotation for annotation in annotations
if annotation['imageset'] == 'train'
]
def test(im_size=600,
mode='normal',
detect_threshold=0.9,
overlap_threshold=0.5):
""" Test the model
Args:
im_size: trained model available for 300, 400 or 600. Input images are resized to this size.
mode: 'normal' means predictions will be saved. Any other mode means only a CSV corner file will be saved.
detect_threshold: minimum class belonging probability for a proposal to be accepted
overlap_threshold: maximum IoU between two proposals
Returns:
avg_time: time taken over the last 10 images. Time is measured from loading the image to saving the predictions.
"""
conf_file = f'models/conf/config_frcnn_{im_size}.pickle'
C = pickle.load(open(conf_file, 'rb'))
img_path = 'data/testing/images'
class_mapping = C.class_mapping
class_mapping['bg'] = len(class_mapping)
class_mapping = {v: k for k, v in class_mapping.items()}
class_to_color = {
class_mapping[v]: np.array([0, 128, 255])
for v in class_mapping
}
input_shape_img = (None, None, 3)
input_shape_features = (None, None, 1024)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(C.num_rois, 4))
feature_map_input = Input(shape=input_shape_features)
# define the base resnet network
shared_layers = nn.nn_base(img_input, trainable=True)
# define the RPN, built on the base layers
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn_layers = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(feature_map_input,
roi_input,
C.num_rois,
nb_classes=len(class_mapping))
model_rpn = Model(img_input, rpn_layers)
model_classifier_only = Model([feature_map_input, roi_input], classifier)
model_classifier = Model([feature_map_input, roi_input], classifier)
model_rpn, model_classifier = load_weights(im_size, model_rpn,
model_classifier)
model_rpn.compile(optimizer='sgd', loss='mse')
model_classifier.compile(optimizer='sgd', loss='mse')
output_folder = f'output/predictions/frcnn_size{int(im_size)}_p{int(detect_threshold * 100)}'
if not os.path.exists(output_folder):
os.makedirs(output_folder)
output_file = open(
f'output/predictions/frcnn_size{int(im_size)}_p{int(detect_threshold * 100)}/'
f'predicted_corners_size{int(im_size)}_p{int(detect_threshold * 100)}.csv',
'w')
writer = csv.writer(output_file)
print(
f'Predicting gates for im_size={im_size} and detection probability threshold of {int(detect_threshold * 100)}%.'
)
print(
f'Output to be saved in directory "/output/predictions/frcnn_size{int(im_size)}_p{int(detect_threshold * 100)}/"'
)
progbar = Progbar(len(os.listdir(img_path)) - 1)
for idx, img_name in enumerate(sorted(os.listdir(img_path))):
if not img_name.lower().endswith(('.png', '.jpg')):
continue
filepath = os.path.join(img_path, img_name)
start_time = time.time()
img = cv2.imread(filepath)
X, ratio = format_img(img, C)
X = np.transpose(X, (0, 2, 3, 1))
# get the feature maps and output from the RPN
[Y1, Y2, F] = model_rpn.predict(X)
R = roi_helpers.rpn_to_roi(Y1,
Y2,
C,
overlap_thresh=overlap_threshold,
max_boxes=C.max_boxes)
# convert from (x1,y1,x2,y2) to (x,y,w,h)
R[:, 2] -= R[:, 0]
R[:, 3] -= R[:, 1]
# apply the spatial pyramid pooling to the proposed regions
bboxes = {}
probs = {}
time_list = []
for jk in range(R.shape[0] // C.num_rois + 1):
ROIs = np.expand_dims(R[C.num_rois * jk:C.num_rois * (jk + 1), :],
axis=0)
if ROIs.shape[1] == 0:
break
if jk == R.shape[0] // C.num_rois:
curr_shape = ROIs.shape
target_shape = (curr_shape[0], C.num_rois, curr_shape[2])
ROIs_padded = np.zeros(target_shape).astype(ROIs.dtype)
ROIs_padded[:, :curr_shape[1], :] = ROIs
ROIs_padded[0, curr_shape[1]:, :] = ROIs[0, 0, :]
ROIs = ROIs_padded
[P_cls, P_regr] = model_classifier_only.predict([F, ROIs])
for ii in range(P_cls.shape[1]):
if np.max(P_cls[0, ii, :]) < detect_threshold or np.argmax(
P_cls[0, ii, :]) == (P_cls.shape[2] - 1):
continue
# only gate objects
cls_name = 'gate'
if cls_name not in bboxes:
bboxes[cls_name] = []
probs[cls_name] = []
(x, y, w, h) = ROIs[0, ii, :]
# only gate objects, which is index 0
cls_num = 0
try:
(tx, ty, tw, th) = P_regr[0, ii,
4 * cls_num:4 * (cls_num + 1)]
tx /= C.classifier_regr_std[0]
ty /= C.classifier_regr_std[1]
tw /= C.classifier_regr_std[2]
th /= C.classifier_regr_std[3]
x, y, w, h = roi_helpers.apply_regr(
x, y, w, h, tx, ty, tw, th)
except:
pass
bboxes[cls_name].append([
C.rpn_stride * x, C.rpn_stride * y, C.rpn_stride * (x + w),
C.rpn_stride * (y + h)
])
probs[cls_name].append(np.max(P_cls[0, ii, :]))
for key in bboxes:
bbox = np.array(bboxes[key])
new_boxes, new_probs = roi_helpers.non_max_suppression_fast(
bbox, np.array(probs[key]), overlap_thresh=overlap_threshold)
for jk in range(new_boxes.shape[0]):
(x1, y1, x2, y2) = new_boxes[jk, :]
w = (x2 - x1)
h = (y2 - y1)
scale = 0.16
x1 += w * scale
x2 -= w * scale
y1 += h * scale
y2 -= h * scale
(real_x1, real_y1, real_x2,
real_y2) = get_real_coordinates(ratio, x1, y1, x2, y2)
writer.writerow([
img_name, real_x1, real_y1, real_x2, real_y1, real_x2,
real_y2, real_x1, real_y2
])
if mode == 'normal':
cv2.rectangle(img, (real_x1, real_y1), (real_x2, real_y2),
(int(class_to_color[key][0]),
int(class_to_color[key][1]),
int(class_to_color[key][2])), 2)
textLabel = f'{key}: {int(100 * new_probs[jk])}%'
(retval,
baseLine) = cv2.getTextSize(textLabel,
cv2.FONT_HERSHEY_SIMPLEX, 0.5,
1)
textOrg = (real_x1, real_y1)
cv2.rectangle(img, (textOrg[0], textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5,
textOrg[1] - retval[1] - 5),
(255, 255, 255), 2)
cv2.rectangle(img, (textOrg[0], textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5,
textOrg[1] - retval[1] - 5), (0, 128, 255),
-1)
cv2.putText(img, textLabel, (real_x1, real_y1 - 3),
cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255),
1)
time_taken = time.time() - start_time
progbar.update(idx)
print(f' - Elapsed time: {time_taken:.3}s for {img_name}')
if idx > 20:
time_list.append(time_taken)
if mode == 'normal':
cv2.imwrite(
f'output/predictions/frcnn_size{int(im_size)}_p{int(detect_threshold * 100)}/predict_{img_name}',
img)
plt.close()
output_file.close()
# return the prediction time over the last 10 images
return sum(time_list) / len(time_list)
def trainModel(options):
import random
import pprint
import sys
import time
import numpy as np
import pickle
import os
from keras import backend as K
from keras.optimizers import Adam, SGD, RMSprop
from keras.layers import Input
from keras.models import Model
from frcnn import config, data_generators
from frcnn import losses as losses
from frcnn import resnet as nn
import frcnn.roi_helpers as roi_helpers
from keras.utils import generic_utils
from frcnn.simple_parser import get_data
from frcnn.simple_parser import load_data
# os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
import tensorflow as tf
configTF = tf.ConfigProto()
configTF.gpu_options.allow_growth = True
sess = tf.Session(config=configTF)
sys.setrecursionlimit(40000)
# Config class stores all relevant settings to be recalled during test/further training
C = config.Config()
C.im_size = options.im_size
C.anchor_box_scales = options.anchor_box_scales
C.anchor_box_ratios = options.anchor_box_ratios
C.num_rois = int(options.num_rois)
C.use_horizontal_flips = bool(options.horizontal_flips)
C.use_vertical_flips = bool(options.vertical_flips)
C.rot_90 = bool(options.rot_90)
C.rpn_max_overlap = options.rpn_max_overlap_threshold
C.model_path = options.output_weight_path
C.balanced_classes = options.balanced_classes
C.rpn_stride = options.rpn_stride
C.cutImage = options.cutImage
# loading old weights to continue training
if options.load_weights:
C.base_net_weights = options.input_weight_path
# get_data() function returns image list along with info on bbox,
# height, width in all_imgs,and class data in classes_count and class_mapping
if options.load_data:
all_imgs, classes_count, class_mapping = load_data(options.name)
else:
all_imgs, classes_count, class_mapping = get_data(
options.train_path, C, options.name, options.num_frames)
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
C.class_mapping = class_mapping
# making assigned value of class the key to index the dictionary
inv_map = {v: k for k, v in class_mapping.iteritems()}
print('Training images per class:')
pprint.pprint(classes_count)
print('Num classes (including bg) = {}'.format(len(classes_count)))
print '\nUsing RPN Stride = %d' % C.rpn_stride
config_output_filename = options.config_filename
with open(config_output_filename, 'w') as config_f:
pickle.dump(C, config_f)
print(
'Config has been written to {}, and can be loaded when testing to ensure correct results'
.format(config_output_filename))
random.shuffle(all_imgs)
num_imgs = len(all_imgs)
train_imgs = [s for s in all_imgs if s['imageset'] == 'trainval']
print('\nTraining on {} Frames'.format(len(train_imgs)))
# ground truth boxes are obtained
data_gen_train = data_generators.get_anchor_gt(train_imgs,
classes_count,
C,
K.image_dim_ordering(),
mode='train')
if K.image_dim_ordering() == 'th':
input_shape_img = (3, None, None)
else:
input_shape_img = (None, None, 3)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(C.num_rois, 4))
# define the base network (resnet here, can be VGG, Inception, etc)
shared_layers = nn.nn_base(img_input, trainable=True)
# define the RPN, built on the base layers
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(shared_layers,
roi_input,
C.num_rois,
nb_classes=len(classes_count),
trainable=True)
model_rpn = Model(img_input, rpn[:2])
model_classifier = Model([img_input, roi_input], classifier)
# this is a model that holds both the RPN and the classifier, used to load/save weights for the models
model_all = Model([img_input, roi_input], rpn[:2] + classifier)
try:
model_rpn.load_weights(C.base_net_weights, by_name=True)
model_classifier.load_weights(C.base_net_weights, by_name=True)
print('\nLoaded weights from {}'.format(C.base_net_weights))
except:
print('\nNo pretrained weights found in folder...')
print('Proceeding to train from scratch\n\n')
# setting learning rates and optimizers
optimizer = Adam(lr=1e-4)
optimizer_classifier = Adam(lr=1e-4)
model_rpn.compile(optimizer=optimizer,
loss=[
losses.rpn_loss_cls(num_anchors),
losses.rpn_loss_regr(num_anchors)
])
model_classifier.compile(
optimizer=optimizer_classifier,
loss=[
losses.class_loss_cls,
losses.class_loss_regr(len(classes_count) - 1)
],
metrics={'dense_class_{}'.format(len(classes_count)): 'accuracy'})
model_all.compile(optimizer='sgd', loss='mae')
# if epoch_length was set as default, a epoch goes over entire trainset images
if options.epoch_length == 'default':
epoch_length = len(train_imgs)
# else it uses the number of images given
else:
epoch_length = int(options.epoch_length)
num_epochs = int(options.num_epochs)
iter_num = 0
losses = np.zeros((epoch_length, 5))
rpn_accuracy_rpn_monitor = []
rpn_accuracy_for_epoch = []
start_time = time.time()
best_loss = np.Inf
class_mapping_inv = {v: k for k, v in class_mapping.iteritems()}
print('Starting training\n')
prev_pos_samples = []
prev_neg_samples = []
for epoch_num in range(num_epochs):
progbar = generic_utils.Progbar(epoch_length)
print('Epoch {}/{}'.format(epoch_num + 1, num_epochs))
while True:
try:
if iter_num == epoch_length and C.verbose:
mean_overlapping_bboxes = float(
sum(rpn_accuracy_rpn_monitor)) / len(
rpn_accuracy_rpn_monitor)
rpn_accuracy_rpn_monitor = []
if mean_overlapping_bboxes == 0:
print(
'\n\nRPN is not producing bounding boxes that overlap the ground truth boxes. Check RPN settings or keep training.\n'
)
# date_gen_train is a generator defined in data_generators.py
# it reads the the image with the filepath in image list and returns relevant data for training
X, Y, img_data = data_gen_train.next()
# train_on_batch() is the keras function defined in Sequential.
# does a single gradient update on given data (essentially, one epoch)
loss_rpn = model_rpn.train_on_batch(X, Y)
# uses trained model to make prediction
P_rpn = model_rpn.predict_on_batch(X)
# converting RPN prediction to ROI
R = roi_helpers.rpn_to_roi(P_rpn[0],
P_rpn[1],
C,
K.image_dim_ordering(),
use_regr=True,
overlap_thresh=0.7,
max_boxes=300)
# note: calc_iou converts from (x1,y1,x2,y2) to (x,y,w,h) format
X2, Y1, Y2 = roi_helpers.calc_iou(R, img_data, C,
class_mapping)
# if no ROI is detected
if X2 is None:
rpn_accuracy_rpn_monitor.append(0)
rpn_accuracy_for_epoch.append(0)
continue
neg_samples = np.where(Y1[0, :, -1] == 1)
pos_samples = np.where(Y1[0, :, -1] == 0)
if len(neg_samples) > 0:
neg_samples = neg_samples[0]
else:
neg_samples = []
if len(pos_samples) > 0:
pos_samples = pos_samples[0]
else:
pos_samples = []
rpn_accuracy_rpn_monitor.append(len(pos_samples))
rpn_accuracy_for_epoch.append((len(pos_samples)))
if C.num_rois > 1:
# Take half of positive samples and half of negative samples for classfier training
if len(pos_samples) < C.num_rois / 2:
selected_pos_samples = pos_samples.tolist()
else:
selected_pos_samples = np.random.choice(
pos_samples, C.num_rois / 2,
replace=False).tolist()
try:
selected_neg_samples = np.random.choice(
neg_samples,
C.num_rois - len(selected_pos_samples),
replace=False).tolist()
except:
selected_neg_samples = np.random.choice(
neg_samples,
C.num_rois - len(selected_pos_samples),
replace=True).tolist()
sel_samples = selected_pos_samples + selected_neg_samples
else:
# in the extreme case where num_rois = 1, we pick a random pos or neg sample
selected_pos_samples = pos_samples.tolist()
selected_neg_samples = neg_samples.tolist()
if np.random.randint(0, 2):
sel_samples = random.choice(neg_samples)
else:
sel_samples = random.choice(pos_samples)
loss_class = model_classifier.train_on_batch(
[X, X2[:, sel_samples, :]],
[Y1[:, sel_samples, :], Y2[:, sel_samples, :]])
losses[iter_num, 2] = loss_class[1]
losses[iter_num, 3] = loss_class[2]
losses[iter_num, 4] = loss_class[3]
losses[iter_num, 0] = loss_rpn[1]
losses[iter_num, 1] = loss_rpn[2]
iter_num += 1
progbar.update(
iter_num,
[('rpn_cls', np.mean(losses[:iter_num, 0])),
('rpn_regr', np.mean(losses[:iter_num, 1])),
('detector_cls', np.mean(losses[:iter_num, 2])),
('rpn_overlap', float(sum(rpn_accuracy_rpn_monitor)) /
len(rpn_accuracy_rpn_monitor))])
if iter_num == epoch_length:
loss_rpn_cls = np.mean(losses[:, 0])
loss_rpn_regr = np.mean(losses[:, 1])
loss_class_cls = np.mean(losses[:, 2])
loss_class_regr = np.mean(losses[:, 3])
class_acc = np.mean(losses[:, 4])
mean_overlapping_bboxes = float(sum(
rpn_accuracy_for_epoch)) / len(rpn_accuracy_for_epoch)
rpn_accuracy_for_epoch = []
if C.verbose:
print(
'\n---------------------------------------------------------------------------------------'
)
print(
'Mean number of bounding boxes from RPN overlapping ground truth boxes: {}'
.format(mean_overlapping_bboxes))
print(
'Classifier accuracy for bounding boxes from RPN: {}'
.format(class_acc))
print('Loss RPN classifier: {}'.format(loss_rpn_cls))
print('Loss RPN regression: {}'.format(loss_rpn_regr))
print('Loss Detector classifier: {}'.format(
loss_class_cls))
print('Loss Detector regression: {}'.format(
loss_class_regr))
print('Elapsed time: {}'.format(time.time() -
start_time))
curr_loss = loss_rpn_cls + loss_rpn_regr + loss_class_cls + loss_class_regr
iter_num = 0
start_time = time.time()
if curr_loss < best_loss:
if C.verbose:
print(
'\nTotal loss decreased from {} to {}'.format(
best_loss, curr_loss))
print(
'---------------------------------------------------------------------------------------\n'
)
best_loss = curr_loss
# saving weights with smallest loss (overwrite)
model_all.save_weights(C.model_path)
else:
if C.verbose:
print('\nLoss did not improve')
print(
'---------------------------------------------------------------------------------------\n'
)
# also saving weights for each epoch
model_all.save_weights(C.model_path[:-5] + '_%03d' %
(epoch_num + 1) +
'_%2.2f' % mean_overlapping_bboxes +
'.hdf5')
break
except Exception as e:
print('\nException: {}\n'.format(e))
continue
print('\n-----------------\nTraining complete!\n')