def generate_mean_pixel_file():
C = Config()
all_imgs, _, _ = get_data(ROI_BBOX_FILE)
avg = [0, 0, 0]
for img_data in all_imgs:
print(img_data['filepath'])
img_data_aug, x_img = augment(img_data, C, augment=False)
(width, height) = (img_data_aug['width'], img_data_aug['height'])
(rows, cols, _) = x_img.shape
# get image dimensions for resizing
(resized_width,
resized_height) = get_new_img_size(width, height, C.im_size)
# resize the image so that smalles side is length = 600px
x_img = cv2.resize(x_img, (resized_width, resized_height),
interpolation=cv2.INTER_CUBIC)
pixels = (resized_width * resized_height)
avg[0] += np.sum(x_img[:, :, 0]) / pixels
avg[1] += np.sum(x_img[:, :, 1]) / pixels
avg[2] += np.sum(x_img[:, :, 2]) / pixels
avg = [a / len(all_imgs) for a in list(avg)]
np.savetxt(MEAN_PIXEL_FILE, avg, delimiter=',')
def train(model_name, epochs=60, batch_size=1, lr=0.0001, decay=0.001):
t = time.time()
all_imgs, classes_count, class_mapping = get_data(ROI_BBOX_FILE)
print("Parsing annotation files took " + str((time.time() - t) / 1000) + "s")
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
C.model_name = model_name
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
if not os.path.isfile(ROI_CLASSES_FILE):
with open(ROI_CLASSES_FILE, 'w') as class_data_json:
json.dump(class_mapping, class_data_json)
print('Num classes (including bg) = {}'.format(len(classes_count)))
random.shuffle(all_imgs)
train_imgs = [s for s in all_imgs if s['imageset'] == 'trainval']
val_imgs = [s for s in all_imgs if s['imageset'] == 'test']
print('Num train samples {}'.format(len(train_imgs)))
print('Num val samples {}'.format(len(val_imgs)))
model = build_model(classes_count, num_anchors)
optimizer = Adam(lr=lr, decay=decay)
model.compile(optimizer=optimizer,
loss=[losses.rpn_loss_cls(num_anchors), losses.rpn_loss_regr(num_anchors),
losses.class_loss_cls,
losses.class_loss_regr(C.num_rois, len(classes_count) - 1)],
metrics={'dense_class_{}_loss'.format(len(classes_count)): 'accuracy'})
data_gen_train = data_generators.get_anchor_gt(train_imgs, class_mapping, classes_count,
C, K.image_dim_ordering(), mode='train')
data_gen_val = data_generators.get_anchor_gt(val_imgs, class_mapping, classes_count,
C, K.image_dim_ordering(), mode='val')
callbacks = [EarlyStopping(monitor='val_loss', patience=20, verbose=0),
ModelCheckpoint(C.get_model_path(), monitor='val_loss', save_best_only=True, verbose=0),
ReduceLROnPlateau(monitor='loss', factor=0.1, patience=5, min_lr=1e-7, verbose=1),
LoggingCallback(C)]
print('Starting training')
model.fit_generator(data_gen_train, steps_per_epoch=ceil(len(train_imgs) / batch_size),
epochs=epochs, validation_data=data_gen_val,
validation_steps=ceil(len(train_imgs) / batch_size),
callbacks=callbacks, max_q_size=1, workers=1)
def train_kitti():
# config for data argument
cfg = config.Config()
cfg.use_horizontal_flips = True
cfg.use_vertical_flips = True
cfg.rot_90 = True
cfg.num_rois = 32
cfg.base_net_weights = os.path.join('./model/', nn.get_weight_path())
# cfg.base_net_weights=r''
# TODO: the only file should to be change for other data to train
cfg.model_path = '/media/private/Ci/log/plane/frcnn/vgg-adam'
now = datetime.datetime.now()
day = now.strftime('%y-%m-%d')
for i in range(10000):
if not os.path.exists('%s-%s-%d' % (cfg.model_path, day, i)):
cfg.model_path = '%s-%s-%d' % (cfg.model_path, day, i)
break
make_dir(cfg.model_path)
make_dir(cfg.model_path + '/loss')
make_dir(cfg.model_path + '/loss_rpn_cls')
make_dir(cfg.model_path + '/loss_rpn_regr')
make_dir(cfg.model_path + '/loss_class_cls')
make_dir(cfg.model_path + '/loss_class_regr')
cfg.simple_label_file = '/media/public/GEOWAY/plane/plane0817.csv'
all_images, classes_count, class_mapping = get_data(cfg.simple_label_file)
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
cfg.class_mapping = class_mapping
cfg.config_save_file = os.path.join(cfg.model_path, 'config.pickle')
with open(cfg.config_save_file, 'wb') as config_f:
pickle.dump(cfg, config_f)
print(
'Config has been written to {}, and can be loaded when testing to ensure correct results'
.format(cfg.config_save_file))
inv_map = {v: k for k, v in class_mapping.items()}
print('Training images per class:')
pprint.pprint(classes_count)
print('Num classes (including bg) = {}'.format(len(classes_count)))
random.shuffle(all_images)
num_imgs = len(all_images)
train_imgs = [s for s in all_images if s['imageset'] == 'trainval']
val_imgs = [s for s in all_images if s['imageset'] == 'test']
print('Num train samples {}'.format(len(train_imgs)))
print('Num val samples {}'.format(len(val_imgs)))
data_gen_train = data_generators.get_anchor_gt(train_imgs,
classes_count,
cfg,
nn.get_img_output_length,
K.image_dim_ordering(),
mode='train')
data_gen_val = data_generators.get_anchor_gt(val_imgs,
classes_count,
cfg,
nn.get_img_output_length,
K.image_dim_ordering(),
mode='val')
Q = multiprocessing.Manager().Queue(maxsize=30)
def fill_Q(n):
while True:
if not Q.full():
Q.put(next(data_gen_train))
#print(Q.qsize(),'put',n)
else:
time.sleep(0.00001)
threads = []
for i in range(4):
thread = multiprocessing.Process(target=fill_Q, args=(i, ))
threads.append(thread)
thread.start()
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=(None, 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(cfg.anchor_box_scales) * len(cfg.anchor_box_ratios)
rpn = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(shared_layers,
roi_input,
cfg.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)
# model_all.summary()
from keras.utils import plot_model
# os.environ['PATH'] = os.environ['PATH'] + r';C:\Program Files (x86)\Graphviz2.38\bin;'
plot_model(model_all,
'model_all.png',
show_layer_names=True,
show_shapes=True)
plot_model(model_classifier,
'model_classifier.png',
show_layer_names=True,
show_shapes=True)
plot_model(model_rpn,
'model_rpn.png',
show_layer_names=True,
show_shapes=True)
'''
try:
print('loading weights from {}'.format(cfg.base_net_weights))
model_rpn.load_weights(cfg.model_path, by_name=True)
model_classifier.load_weights(cfg.model_path, by_name=True)
except Exception as e:
print(e)
print('Could not load pretrained model weights. Weights can be found in the keras application folder '
'https://github.com/fchollet/keras/tree/master/keras/applications')
'''
optimizer = adadelta()
optimizer_classifier = adadelta()
model_rpn.compile(optimizer=optimizer,
loss=[
losses_fn.rpn_loss_cls(num_anchors),
losses_fn.rpn_loss_regr(num_anchors)
])
model_classifier.compile(
optimizer=optimizer_classifier,
loss=[
losses_fn.class_loss_cls,
losses_fn.class_loss_regr(len(classes_count) - 1)
],
metrics={'dense_class_{}'.format(len(classes_count)): 'accuracy'})
model_all.compile(optimizer='sgd', loss='mae')
epoch_length = 10
num_epochs = int(cfg.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
best_rpn_cls = np.Inf
best_rpn_regr = np.Inf
best_class_cls = np.Inf
best_class_regr = np.Inf
class_mapping_inv = {v: k for k, v in class_mapping.items()}
print('Starting training')
vis = True
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 cfg.verbose:
mean_overlapping_bboxes = float(
sum(rpn_accuracy_rpn_monitor)) / len(
rpn_accuracy_rpn_monitor)
rpn_accuracy_rpn_monitor = []
print(
'Average number of overlapping bounding boxes from RPN = {} for {} previous iterations'
.format(mean_overlapping_bboxes, epoch_length))
if mean_overlapping_bboxes == 0:
print(
'RPN is not producing bounding boxes that overlap'
' the ground truth boxes. Check RPN settings or keep training.'
)
# X, Y, img_data = next(data_gen_train)
while True:
if Q.empty():
time.sleep(0.00001)
continue
X, Y, img_data = Q.get()
# print(Q.qsize(),'get')
break
# print(X.shape,Y.shape)
loss_rpn = model_rpn.train_on_batch(X, Y)
P_rpn = model_rpn.predict_on_batch(X)
result = roi_helpers.rpn_to_roi(P_rpn[0],
P_rpn[1],
cfg,
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, IouS = roi_helpers.calc_iou(
result, img_data, cfg, 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 cfg.num_rois > 1:
if len(pos_samples) < cfg.num_rois // 2:
selected_pos_samples = pos_samples.tolist()
else:
selected_pos_samples = np.random.choice(
pos_samples, cfg.num_rois // 2,
replace=False).tolist()
try:
selected_neg_samples = np.random.choice(
neg_samples,
cfg.num_rois - len(selected_pos_samples),
replace=False).tolist()
except:
selected_neg_samples = np.random.choice(
neg_samples,
cfg.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, 0] = loss_rpn[1]
losses[iter_num, 1] = loss_rpn[2]
losses[iter_num, 2] = loss_class[1]
losses[iter_num, 3] = loss_class[2]
losses[iter_num, 4] = loss_class[3]
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])),
('detector_regr', np.mean(losses[:iter_num, 3]))])
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 cfg.verbose:
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 cfg.verbose:
print(
'Total loss decreased from {} to {}, saving weights'
.format(best_loss, curr_loss))
best_loss = curr_loss
model_all.save_weights(
'%s/%s/E-%d-loss-%.4f-rpnc-%.4f-rpnr-%.4f-cls-%.4f-cr-%.4f.hdf5'
% (cfg.model_path, 'loss', epoch_num, curr_loss,
loss_rpn_cls, loss_rpn_regr, loss_class_cls,
loss_class_regr))
if loss_rpn_cls < best_rpn_cls:
if cfg.verbose:
print(
'loss_rpn_cls decreased from {} to {}, saving weights'
.format(best_rpn_cls, loss_rpn_cls))
best_rpn_cls = loss_rpn_cls
model_all.save_weights(
'%s/%s/E-%d-loss-%.4f-rpnc-%.4f-rpnr-%.4f-cls-%.4f-cr-%.4f.hdf5'
% (cfg.model_path, 'loss_rpn_cls', epoch_num,
curr_loss, loss_rpn_cls, loss_rpn_regr,
loss_class_cls, loss_class_regr))
if loss_rpn_regr < best_rpn_regr:
if cfg.verbose:
print(
'loss_rpn_regr decreased from {} to {}, saving weights'
.format(best_rpn_regr, loss_rpn_regr))
best_rpn_regr = loss_rpn_regr
model_all.save_weights(
'%s/%s/E-%d-loss-%.4f-rpnc-%.4f-rpnr-%.4f-cls-%.4f-cr-%.4f.hdf5'
% (cfg.model_path, 'loss_rpn_regr', epoch_num,
curr_loss, loss_rpn_cls, loss_rpn_regr,
loss_class_cls, loss_class_regr))
if loss_class_cls < best_class_cls:
if cfg.verbose:
print(
'loss_class_cls decreased from {} to {}, saving weights'
.format(best_loss, loss_class_cls))
best_class_cls = loss_class_cls
model_all.save_weights(
'%s/%s/E-%d-loss-%.4f-rpnc-%.4f-rpnr-%.4f-cls-%.4f-cr-%.4f.hdf5'
% (cfg.model_path, 'loss_class_cls', epoch_num,
curr_loss, loss_rpn_cls, loss_rpn_regr,
loss_class_cls, loss_class_regr))
if loss_class_regr < best_class_regr:
if cfg.verbose:
print(
'loss_class_regr decreased from {} to {}, saving weights'
.format(best_loss, loss_class_regr))
best_class_regr = loss_class_regr
model_all.save_weights(
'%s/%s/E-%d-loss-%.4f-rpnc-%.4f-rpnr-%.4f-cls-%.4f-cr-%.4f.hdf5'
% (cfg.model_path, 'loss_class_regr', epoch_num,
curr_loss, loss_rpn_cls, loss_rpn_regr,
loss_class_cls, loss_class_regr))
break
except Exception as e:
# print('Exception: {}'.format(e))
# save model
# model_all.save_weights(cfg.model_path)
continue
print('Training complete, exiting.')
def Measure_map(test_path,
network_arch,
config_filename,
preprocessing_function=None,
mAP_threshold=0.5):
"""Function to measure Mean Average prediction metric for object detection
Keyword Arguments
test_path --str: Path to the .txt file of testing train (No default)
network_arc --object: the full faster rcnn network .py file passed as an object (no default)
config_filename --str: Path to config file (No default)
preprocessing_function --function: optional image preprocessing function (Default None)
mAP threshold --float: (0,1) The min threshold to consider as a correct prediction (default 0.5)
Output:
prints the Map on the test dataset and returns a list of all Maps
"""
nn = network_arch
def get_map(pred, gt, f, threshold):
T = {}
P = {}
fx, fy = f
for bbox in gt:
bbox['bbox_matched'] = False
pred_probs = np.array([s['prob'] for s in pred])
box_idx_sorted_by_prob = np.argsort(pred_probs)[::-1]
for box_idx in box_idx_sorted_by_prob:
pred_box = pred[box_idx]
pred_class = pred_box['class']
pred_x1 = pred_box['x1']
pred_x2 = pred_box['x2']
pred_y1 = pred_box['y1']
pred_y2 = pred_box['y2']
pred_prob = pred_box['prob']
if pred_class not in P:
P[pred_class] = []
T[pred_class] = []
P[pred_class].append(pred_prob)
found_match = False
for gt_box in gt:
gt_class = gt_box['class']
gt_x1 = gt_box['x1'] / fx
gt_x2 = gt_box['x2'] / fx
gt_y1 = gt_box['y1'] / fy
gt_y2 = gt_box['y2'] / fy
gt_seen = gt_box['bbox_matched']
if gt_class != pred_class:
continue
if gt_seen:
continue
iou = data_generators.iou((pred_x1, pred_y1, pred_x2, pred_y2),
(gt_x1, gt_y1, gt_x2, gt_y2))
if iou >= threshold: #0.5 default
found_match = True
gt_box['bbox_matched'] = True
break
else:
continue
T[pred_class].append(int(found_match))
for gt_box in gt:
if not gt_box['bbox_matched']:
if gt_box['class'] not in P:
P[gt_box['class']] = []
T[gt_box['class']] = []
T[gt_box['class']].append(1)
P[gt_box['class']].append(0)
#import pdb
#pdb.set_trace()
return T, P
with open(config_filename, 'rb') as f_in:
C = pickle.load(f_in)
# turn off any train augmentation at test time
C.use_horizontal_flips = False
C.use_vertical_flips = False
C.rot_90 = False
def format_img(img, C, preprocessing_function):
img_min_side = float(C.im_size)
(height, width, _) = img.shape
if width <= height:
f = img_min_side / width
new_height = int(f * height)
new_width = int(img_min_side)
else:
f = img_min_side / height
new_width = int(f * width)
new_height = int(img_min_side)
fx = width / float(new_width)
fy = height / float(new_height)
img = cv2.resize(img, (new_width, new_height),
interpolation=cv2.INTER_CUBIC)
img = img[:, :, (2, 1, 0)] #bgr to RGB
if preprocessing_function:
img = preprocessing_function(img)
# img = np.transpose(img, (2, 0, 1)) theano format
img = np.expand_dims(img, axis=0)
return img, fx, fy
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.items()}
print(class_mapping)
# load the models
input_shape_img = (None, None, 3)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(None, 4))
shared_layers = nn.nn_base(img_input)
num_features = shared_layers.get_shape().as_list()[3] #512 for vgg-16
feature_map_input = Input(shape=(None, None, num_features))
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(feature_map_input, roi_input, C.num_rois,
len(class_mapping))
# create a keras model
model_rpn = Model(img_input, rpn)
model_classifier = Model([feature_map_input, roi_input], classifier)
#Note: The model_classifier in training and testing are different.
# In training model_classifier and model_rpn both have the base_nn.
# while testing only model_rpn has the base_nn it returns the FM of base_nn
# Thus the model_classifier has the FM and ROI as input
# This is done to increase the testing speed
print('Loading weights from {}'.format(C.weights_all_path))
model_rpn.load_weights(C.weights_all_path, by_name=True)
model_classifier.load_weights(C.weights_all_path, by_name=True)
test_imgs, _, _ = get_data(test_path)
T = {}
P = {}
ALL_MAP_LIST = []
for idx, img_data in enumerate(test_imgs):
print('{}/{}'.format(idx, len(test_imgs)))
st = time.time()
filepath = img_data['filepath']
img = cv2.imread(filepath)
X, fx, fy = format_img(img, C, preprocessing_function)
# 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,
K.image_dim_ordering(),
overlap_thresh=C.rpn_nms_threshold,
flag="test")
# 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] // 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:
# pad R
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.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 /= 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(
[16 * x, 16 * y, 16 * (x + w), 16 * (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=C.test_roi_nms_threshold,
max_boxes=C.TEST_RPN_POST_NMS_TOP_N)
for jk in range(new_boxes.shape[0]):
(x1, y1, x2, y2) = new_boxes[jk, :]
det = {
'x1': x1,
'x2': x2,
'y1': y1,
'y2': y2,
'class': key,
'prob': new_probs[jk]
}
all_dets.append(det)
print('Elapsed time = {}'.format(time.time() - st))
t, p = get_map(all_dets, img_data['bboxes'], (fx, fy), mAP_threshold)
for key in t.keys():
if key not in T:
T[key] = []
P[key] = []
T[key].extend(t[key])
P[key].extend(p[key])
all_aps = []
for key in T.keys():
ap = average_precision_score(T[key], P[key])
print('{} AP: {}'.format(key, ap))
all_aps.append(ap)
print('mAP = {}'.format(np.mean(np.array(all_aps))))
ALL_MAP_LIST.append(np.mean(np.array(all_aps)))
return (ALL_MAP_LIST)
with open(config_output_filename, 'rb') as f_in:
C = pickle.load(f_in)
if C.network == 'resnet50':
import keras_frcnn.resnet as nn
elif C.network == 'vgg':
import keras_frcnn.vgg as nn
# turn off any data augmentation at test time
C.use_horizontal_flips = False
C.use_vertical_flips = False
C.rot_90 = False
#img_path = options.test_path
img_path, _, _ = get_data(options.test_path)
def format_img_size(img, C):
""" formats the image size based on config """
img_min_side = float(C.im_size)
(height, width, _) = img.shape
if width <= height:
ratio = img_min_side / width
new_height = int(ratio * height)
new_width = int(img_min_side)
else:
ratio = img_min_side / height
new_width = int(ratio * width)
new_height = int(img_min_side)
with open(config_output_filename, 'rb') as f_in:
C = pickle.load(f_in)
if C.network == 'resnet50':
import keras_frcnn.resnet as nn
elif C.network == 'vgg':
import keras_frcnn.vgg as nn
# turn off any data augmentation at test time
C.use_horizontal_flips = False
C.use_vertical_flips = False
C.rot_90 = False
class_mapping = C.class_mapping
all_imgs, classes_count, class_mapping = get_data(options.test_path)
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
test_imgs = [
s for s in all_imgs
if s['imageset'] == 'trainval' or s['imageset'] == 'test'
]
print('Num train samples {}'.format(len(test_imgs)))
data_gen_test = data_generators.get_anchor_gt(test_imgs,
classes_count,
C,
if not os.path.isdir("models"):
os.mkdir("models")
C.num_rois = int(options.num_rois)
# we will use resnet. may change to others
from keras_frcnn import vgg16 as nn
# check if weight path was passed via command line
if options.input_weight_path:
C.model_path = options.input_weight_path
else:
# set the path to weights based on backend and model
C.model_path = nn.get_weight_path()
all_imgs, classes_count, class_mapping = get_data(options.train_path, options.cat)
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
C.class_mapping = class_mapping
inv_map = {v: k for k, v in class_mapping.items()}
print('Training images per class:')
pprint.pprint(classes_count)
print('Num classes (including bg) = {}'.format(len(classes_count)))
config_output_filename = options.config_filename
# load model
nn: Any = None
num_features = 1024
if C.network == "vgg":
from keras_frcnn import vgg
nn = vgg
num_features = 512
elif C.network == "resnet":
from keras_frcnn import resnet
nn = resnet
else:
print(f"Not a valid model: {C.network}")
raise ValueError
all_imgs, classes_count, class_mapping = get_data(C.train_path)
train_imgs, val_imgs = train_test_split(all_imgs,
test_size=0.2,
random_state=C.seed)
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
C.class_mapping = class_mapping
print(class_mapping)
inv_map = {v: k for k, v in class_mapping.items()}
print('Training images per class:')
def train_fasterrcnn():
# config for data argument
cfg = config.Config()
cfg.balanced_classes = True
cfg.use_horizontal_flips = True
cfg.use_vertical_flips = True
cfg.rot_90 = True
cfg.num_rois = 50 #50# 对于星图杯的光学遥感飞机检测，应该改为50+
cfg.anchor_box_scales = [10, 30, 50, 80, 100] #[41,70,120,20,90]
cfg.anchor_box_ratios = [[1, 1.2], [1, 1],
[1.2,
1]] #[[1,1.4],[1,0.84],[1,1.17],[1,0.64],[1,1]]
#cfg.rpn_stride = 8
cfg.im_size = 512
cfg.num_epochs = 100
cfg.epoch_length = 150 #1462
cfg.base_net_weights = os.path.join('./model/', nn.get_weight_path())
# TODO: the only file should to be change for other data to train
cfg.model_path = './model/kitti_frcnn_last.hdf5'
cfg.simple_label_file = 'DOTA2018_OpticalAircraft_bboxes.txt' #'kitti_simple_label.txt'#'E:/Xingtubei/official_datas/OpticalAircraft/laptop_Chreoc_OpticalAircraft_bboxes.txt' # '/media/liuhuaqing/Elements/Xingtubei/official_datas/OpticalAircraft/Chreoc_OpticalAircraft_bboxes.txt'#'F:/Xingtubei/official_datas/OpticalAircraft/Chreoc_OpticalAircraft_bboxes.txt' # 'kitti_simple_label.txt'
all_images, classes_count, class_mapping = get_data(
cfg.simple_label_file) #读取数据集，cv2.imread()要求数据里不能有中文路径
if 'bg' not in classes_count: #'bg'应该是代表背景
classes_count['bg'] = 0 # =0表示训练数据中没有“背景”这一类别
class_mapping['bg'] = len(class_mapping)
cfg.class_mapping = class_mapping
with open(cfg.config_save_file, 'wb') as config_f:
pickle.dump(cfg, config_f)
print(
'Config has been written to {}, and can be loaded when testing to ensure correct results'
.format(cfg.config_save_file))
inv_map = {v: k for k, v in class_mapping.items()} #class_mapping的逆向map
print('Training images per class:')
pprint.pprint(classes_count)
print('Num classes (including bg) = {}'.format(len(classes_count)))
random.shuffle(all_images)
num_imgs = len(all_images)
train_imgs = [s for s in all_images
if s['imageset'] == 'trainval'] #训练集，列表形式，列表中的元素是字典
val_imgs = [s for s in all_images
if s['imageset'] == 'test'] #验证集，列表形式，列表中的元素是字典
print('Num train samples {}'.format(len(train_imgs)))
print('Num val samples {}'.format(len(val_imgs)))
data_gen_train = data_generators.get_anchor_gt(
train_imgs,
classes_count,
cfg,
nn.get_img_output_length,
K.image_dim_ordering(),
mode='train') #数据扩增，然后生成frcnn所需的训练数据（如：图片、rpn的梯度等等）
data_gen_val = data_generators.get_anchor_gt(
val_imgs,
classes_count,
cfg,
nn.get_img_output_length,
K.image_dim_ordering(),
mode='val') #数据扩增，然后生成frcnn所需的验证数据（如：图片、rpn的梯度等等）
# 根据keras实际用的后端，定义相应的输入数据维度，因为两类后端的维度顺序不一样
if K.image_dim_ordering() == 'th':
input_shape_img = (3, None, None) #当后端是thaneo
else:
input_shape_img = (None, None, 3) #(None, None, 3)#当后端是tensorflow
img_input = Input(shape=input_shape_img) # 输入图片
roi_input = Input(shape=(None, 4)) # 输入人工标注的roi坐标，4表示x1,y1,x2,y2
# define the base network (resnet here, can be VGG, Inception, etc)
shared_layer, shared_layers_stage3, shared_layers_stage4 = nn.nn_base(
img_input,
trainable=True) # shared_layers是frcnn网络底部那些共享的层,在这里是ResNet。由nn定义好
# define the RPN, built on the base layers
num_anchors = len(cfg.anchor_box_scales) * len(cfg.anchor_box_ratios)
rpn_stage3 = nn.rpn(shared_layers_stage3, num_anchors)
print(rpn_stage3[1].shape)
rpn_stage4 = nn.rpn(shared_layers_stage4,
num_anchors) # [x_class, x_regr, base_layers]
print(rpn_stage4[1].shape)
# x_class的shape是(?,sharelayer的w/2,sharelayer的h/2,scale数*ratio数),x_regr的shape是(?,sharelayer的w/2,sharelayer的h/2,4*scale数*ratio数)
rpn = nn.rpn(shared_layer, num_anchors)
print(rpn[1].shape)
# 在这里合并两个rpn分支
classifier = nn.classifier(shared_layer,
roi_input,
cfg.num_rois,
nb_classes=len(classes_count),
trainable=True)
model_rpn = Model(
inputs=img_input, outputs=rpn[:2]
) #rpn网络由keras_frcnn/resnet定义。rpn[:2]的前两个元素分别表示rpn网络的分类输出和回归输出
model_classifier = Model(
inputs=[img_input, roi_input],
outputs=classifier) #Keras的函数式模型为Model，即广义的拥有输入和输出的模型
# this is a model that holds both the RPN and the classifier, used to load/save weights for the models
model_all = Model(inputs=[img_input, roi_input],
outputs=rpn[:2] +
classifier) #rpn[:2]+classifier的含义是？？？？？？
try:
# 尝试载入已训练网络权值
print('loading weights from {}'.format(cfg.base_net_weights))
model_rpn.load_weights(cfg.model_path, by_name=True)
model_classifier.load_weights(cfg.model_path, by_name=True)
except Exception as e:
print(e)
print(
'Could not load pretrained model weights. Weights can be found in the keras application folder '
'https://github.com/fchollet/keras/tree/master/keras/applications')
optimizer = Adam(lr=1e-5) # 定义一个Adam求解器，学习率lr
optimizer_classifier = Adam(lr=1e-5) # 定义一个Adam求解器，学习率lr
# num_anchors等于9
model_rpn.compile(optimizer=optimizer,
loss=[
losses_fn.rpn_loss_cls(num_anchors),
losses_fn.rpn_loss_regr(num_anchors)
])
model_classifier.compile(
optimizer=optimizer_classifier,
loss=[
losses_fn.class_loss_cls,
losses_fn.class_loss_regr(len(classes_count) - 1)
],
metrics={'dense_class_{}'.format(len(classes_count)): 'accuracy'})
model_all.compile(optimizer='sgd', loss='mae') #mae表示绝对值均差
epoch_length = cfg.epoch_length # epoch_length是一个周期的迭代次数（也等于训练数据量）。每迭代epoch_length次就检查一次是否要保存网络权值，然后重置iter_num = 0
num_epochs = int(cfg.num_epochs)
iter_num = 0 # 迭代次数的初值
losses = np.zeros((epoch_length, 5)) # 初始化loss数组，记录每个周期的loss
rpn_accuracy_rpn_monitor = [] # 初始化一个数组，记录rpn的训练过程中的精度变化
rpn_accuracy_for_epoch = [] # 初始化一个数组，记录rpn的每个训练周期的的精度变化
start_time = time.time() # 开始训练的时间
best_loss = np.Inf # 训练以来最小的loss
class_mapping_inv = {v: k
for k, v in class_mapping.items()
} # class_mapping_inv是一个字典，key是目标类别编号，value是类别名称
print('Starting training')
vis = True
for epoch_num in range(num_epochs):
progbar = generic_utils.Progbar(epoch_length) # 生成一个进度条对象
print('Epoch {}/{}'.format(epoch_num + 1,
num_epochs)) # 输出当前训练周期数/总周期数
while True: # 什么时候才结束这个循环？答：第247行的break(每迭代epoch_length次)
try:
if len(
rpn_accuracy_rpn_monitor
) == epoch_length and cfg.verbose: # 每epoch_length次训练周期就在窗口显示一次RPN平均精度
mean_overlapping_bboxes = float(
sum(rpn_accuracy_rpn_monitor)) / len(
rpn_accuracy_rpn_monitor)
rpn_accuracy_rpn_monitor = []
print(
'Average number of overlapping bounding boxes from RPN = {} for {} previous iterations'
.format(mean_overlapping_bboxes, epoch_length))
if mean_overlapping_bboxes == 0:
print(
'RPN is not producing bounding boxes that overlap'
' the ground truth boxes. Check RPN settings or keep training.'
)
# X是原图，如kitti尺寸是(1,600,1987,3)。
# Y是label，是有两个元素的list，
# 其中第一个元素是类别，具体：shape是(1,share_layer_h,share_layer_w,2*scale数*ratio数),前一个元素为1（0）则表示是（不是）正或负样本，后一个为1（0）则表示是（不是）正样本
# 第二个元素是bbox，具体：shape是(1,share_layer_h,share_layer_w,8*scale数*ratio数),前四个元素表示是不是正样，后四个元素才是bbox#为什么来个repeat赋值给前面一半
# img_data是字典，包含文件名、尺寸、人工标记的roi和类别等
#
X, Y, img_data = next(data_gen_train)
#Y_1=Y[0]
#Y_1=Y_1[0,:,:,:]
loss_rpn = model_rpn.train_on_batch(
X, Y) #为什么Y的尺寸与P_rpn的尺寸不同？为什么loss_rpn的尺寸是3，含义是什么，在哪里定义的？
P_rpn = model_rpn.predict_on_batch(
X) #P_rpn的尺寸是(1, 124, 38, 9) (1, 124, 38, 36)
result = roi_helpers.rpn_to_roi(
P_rpn[0],
P_rpn[1],
cfg,
K.image_dim_ordering(),
use_regr=True,
overlap_thresh=0.7,
max_boxes=300) #result的尺寸是300*4
# note: calc_iou converts from (x1,y1,x2,y2) to (x,y,w,h) format
# X2的尺寸是100*4，Y1的尺寸是1*100*8（8=训练集中目标类别总数），IouS尺寸是100
X2, Y1, Y2, IouS = roi_helpers.calc_iou(
result, img_data, cfg, class_mapping
) #Y2的尺寸是1*1*56,56=28*2，（28=4*7）前28是coords，后28是labels(是该类别则标1)
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) #Y1的尺寸是1*1*8表示分类预测结果，最后一个元素为1表示是背景
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 cfg.num_rois > 1:
if len(pos_samples) < cfg.num_rois // 2:
selected_pos_samples = pos_samples.tolist()
else:
selected_pos_samples = np.random.choice(
pos_samples, cfg.num_rois // 2,
replace=False).tolist()
try:
selected_neg_samples = np.random.choice(
neg_samples,
cfg.num_rois - len(selected_pos_samples),
replace=False).tolist()
except:
selected_neg_samples = np.random.choice(
neg_samples,
cfg.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, :]
]) #用rpn输出的roi输入给classifier
losses[iter_num, 0] = loss_rpn[1]
losses[iter_num, 1] = loss_rpn[2]
losses[iter_num, 2] = loss_class[1]
losses[iter_num, 3] = loss_class[2]
losses[iter_num, 4] = loss_class[3]
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])),
('detector_regr', np.mean(losses[:iter_num, 3]))])
if iter_num == epoch_length: # 每迭代epoch_length次就检查一次是否要保存网络权值，然后重置iter_num = 0
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 cfg.verbose:
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 cfg.verbose:
print(
'Total loss decreased from {} to {}, saving weights'
.format(best_loss, curr_loss))
best_loss = curr_loss
model_all.save_weights(cfg.model_path)
break
except Exception as e:
print('Exception: {}'.format(e))
# save model
model_all.save_weights(cfg.model_path)
continue
print('Training complete, exiting.')
# pass the settings from the command line, and persist them in the config object
C = config.Config()
# set data augmentation
C.model_path = options.output_weight_path
C.num_rois = int(options.num_rois)
# we will use resnet. may change to vgg
if options.network == 'resnet50':
from keras_frcnn import resnet as nn
C.network = 'resnet50'
else:
print('Not a valid model')
raise ValueError
all_imgs, classes_count, _ = get_data(options.train_path)
# add background class as 21st class
if 'bg' not in classes_count:
classes_count['bg'] = 0
print('Training images per class:')
pprint.pprint(classes_count)
print('Num classes (including bg) = {}'.format(len(classes_count)))
with open(config_output_filename, 'wb') 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)
def measure_map(config_output_filename, real_model_path):
with open(config_output_filename, 'r') as f_in:
C = pickle.load(f_in)
# img_path = options.test_path
img_path = '/home/comp/e4252392/map4frcnn.txt'
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()}
print(class_mapping)
class_to_color = {
class_mapping[v]: np.random.randint(0, 255, 3)
for v in class_mapping
}
# C.num_rois = int(options.num_rois)
C.num_rois = 32
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)
# 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)
model_rpn.load_weights(real_model_path, by_name=True)
model_classifier.load_weights(real_model_path, by_name=True)
model_rpn.compile(optimizer='sgd', loss='mse')
model_classifier.compile(optimizer='sgd', loss='mse')
all_imgs, _, _ = get_data(img_path)
# test_imgs = [s for s in all_imgs if s['imageset'] == 'test']
test_imgs = [s for s in all_imgs]
T = {}
P = {}
print('Calculating mAP')
st = time.time()
for idx, img_data in enumerate(test_imgs):
# print('{}/{}'.format(idx,len(test_imgs)))
# st = time.time()
filepath = img_data['filepath']
img = cv2.imread(filepath)
X, fx, fy = format_img(img, C)
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,
C,
K.image_dim_ordering(),
overlap_thresh=0.7)
# 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] // 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:
# pad R
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.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 /= 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(
[16 * x, 16 * y, 16 * (x + w), 16 * (y + h)])
probs[cls_name].append(np.max(P_cls[0, ii, :]))
all_dets = []
album_ap = 0.0
logo_ap = 0.0
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, :]
det = {
'x1': x1,
'x2': x2,
'y1': y1,
'y2': y2,
'class': key,
'prob': new_probs[jk]
}
all_dets.append(det)
# print('Elapsed time = {}'.format(time.time() - st))
t, p = get_map(all_dets, img_data['bboxes'], (fx, fy))
for key in t.keys():
if key not in T:
T[key] = []
P[key] = []
T[key].extend(t[key])
P[key].extend(p[key])
all_aps = []
for key in T.keys():
ap = average_precision_score(T[key], P[key])
# print('{} AP: {}'.format(key, ap))
all_aps.append(ap)
if idx == len(test_imgs) - 1:
if key == 'album':
album_ap = ap
if key == 'logo':
logo_ap = ap
# print('mAP = {}'.format(np.mean(np.array(all_aps))))
if idx == len(test_imgs) - 1:
mAP = np.mean(np.array(all_aps))
print('Elapsed time = {}'.format(time.time() - st))
print('album ap = {}'.format(album_ap))
print('logo ap = {}'.format(logo_ap))
print('mAP = {}'.format(mAP))
return [album_ap, logo_ap, mAP]
C.num_rois,
nb_classes=len(class_mapping))
model_rpn = Model(img_input, rpn_layers)
model_classifier = Model([feature_map_input, roi_input], classifier)
# model loading
print('Loading weights from {}'.format(options.load))
model_rpn.load_weights(options.load, by_name=True)
model_classifier.load_weights(options.load, by_name=True)
model_rpn.compile(optimizer='adam', loss='mse')
model_classifier.compile(optimizer='adam', loss='mse')
from keras_frcnn.simple_parser import get_data
all_imgs, classes_count, class_mapping_2 = get_data(options.test_path,
test_only=True)
print(f'{len(all_imgs)} images to test.')
with open('log/frcnn/frcnn_results.csv', 'w') as result_csv:
log_writer = csv.writer(result_csv, delimiter=';')
log_writer.writerow([
'iou_threshold', 'precision', 'recall', 'f1_score', 'avg_precision',
'mean_distance', 'precision_b_small', 'precision_b_medium',
'precision_b_large', 'recall_b_small', 'recall_b_medium',
'recall_b_large', 'mean_distance_b_small', 'mean_distance_b_medium',
'mean_distance_b_high'
])
for iou_threshold_tp in [0.1, 0.2, 0.3, 0.4]:
# Confusion Matrix initialization
def testModel(config_filename='config_ui.pickle'):
st.markdown('## Starting validation of test data set')
sys.setrecursionlimit(40000)
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
config.log_device_placement = True
sess = tf.compat.v1.Session(config=config)
K.set_session(sess)
test_path = 'test'
num_rois = 4
config_filename = config_filename
network = 'resnet50'
config_output_filename = config_filename
with open(config_output_filename, 'rb') as f_in:
C = pickle.load(f_in)
if C.network == 'resnet50':
import keras_frcnn.resnet as nn
elif C.network == 'vgg':
import keras_frcnn.vgg as nn
# turn off any data augmentation at test time
C.use_horizontal_flips = False
C.use_vertical_flips = False
C.rot_90 = False
img_path = test_path
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.items()}
st.write('Class Mapping', class_mapping)
class_to_color = {
class_mapping[v]: np.random.randint(0, 255, 3)
for v in class_mapping
}
C.num_rois = int(num_rois)
if C.network == 'resnet50':
num_features = 1024
elif C.network == 'vgg':
num_features = 512
if K.image_data_format() == 'channels_first':
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=(C.num_rois, 4))
feature_map_input = Input(shape=input_shape_features)
## Defining Model
# 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)
st.write(f'Loading weights from {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')
all_imgs = []
classes = {}
## few hyper parameters to select bbox
bbox_threshold = 0.9
visualise = True
set_Overlap_threshold = 0.1
progress_bar = st.progress(0.0)
with st.spinner('Wait for sample test images...'):
for idx, img_name in enumerate(sorted(os.listdir(img_path))):
progress_bar.progress((idx + 0.1) / len(os.listdir(img_path)))
if not img_name.lower().endswith(
('.bmp', '.jpeg', '.jpg', '.png', '.tif', '.tiff')):
continue
st.write(img_name)
stTime = time.time()
filepath = os.path.join(img_path, img_name)
img = cv2.imread(filepath)
X, ratio = format_img(img, C)
if K.image_data_format() == 'channels_last':
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,
K.image_data_format(),
overlap_thresh=0.7) #0.7
# 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] // 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:
print("ROI Shape: ", ROIs.shape[1])
break
if jk == R.shape[0] // C.num_rois:
#pad R
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]):
#print("np max:",np.max(P_cls[0, ii, :]))
#print("np argmax:",np.argmax(P_cls[0, ii, :]))
#print(np.max(P_cls[0, ii, :]) < bbox_threshold)
if np.max(P_cls[0, ii, :]) < bbox_threshold or np.argmax(
P_cls[0, ii, :]) == (P_cls.shape[2] - 1):
continue
cls_name = class_mapping[np.argmax(P_cls[0, ii, :])]
#print('class name:',cls_name)
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 /= 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, :]))
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=set_Overlap_threshold)
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 = f'{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)
st.write(f'Elapsed time = {time.time() - stTime}')
st.write(all_dets)
plt.figure(figsize=(10, 10))
plt.grid()
plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
#plt.show()
st.image(img, use_column_width=True, clamp=True)
#cv2.imwrite('./results_imgs-fp-mappen-test/{}.png'.format(os.path.splitext(str(img_name))[0]),img)
from keras_frcnn.simple_parser import get_data
test_path = 'test_annotationAlt.txt' # Test data (annotation file)
startTime = time.time()
test_imgs, classes_count, class_mapping = get_data(test_path)
st.write('Spend %0.2f mins to load test data' %
((time.time() - startTime) / 60))
class_mapping = C.class_mapping
class_mapping = {v: k for k, v in class_mapping.items()}
st.write(class_mapping)
class_to_color = {
class_mapping[v]: np.random.randint(0, 255, 3)
for v in class_mapping
}
from sklearn.metrics import average_precision_score
set_Overlap_threshold = 0.1
T = {}
P = {}
mAPs = []
iou_map = []
progress_bar1 = st.progress(0.0)
with st.spinner('Wait for test set evaluation...'):
for idx, img_data in enumerate(test_imgs):
progress_bar1.progress((idx + 0.1) / len(test_imgs))
st.write('{}/{}'.format(idx, len(test_imgs)))
startTime = time.time()
filepath = img_data['filepath']
img = cv2.imread(filepath)
X, fx, fy = format_img_map(img, C)
# Change X (img) shape from (1, channel, height, width) to (1, height, width, channel)
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 = rpn_to_roi(Y1,
Y2,
C,
K.image_data_format(),
overlap_thresh=0.7)
# 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] // 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:
# pad R
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])
# Calculate all classes' bboxes coordinates on resized image (300, 400)
# Drop 'bg' classes bboxes
for ii in range(P_cls.shape[1]):
# If class name is 'bg', continue
if np.argmax(P_cls[0, ii, :]) == (P_cls.shape[2] - 1):
continue
# Get class name
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 /= 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(
[16 * x, 16 * y, 16 * (x + w), 16 * (y + h)])
probs[cls_name].append(np.max(P_cls[0, ii, :]))
all_dets = []
for key in bboxes:
bbox = np.array(bboxes[key])
# Apply non-max-suppression on final bboxes to get the output bounding boxe
new_boxes, new_probs = non_max_suppression_fast(
bbox,
np.array(probs[key]),
overlap_thresh=set_Overlap_threshold)
for jk in range(new_boxes.shape[0]):
(x1, y1, x2, y2) = new_boxes[jk, :]
det = {
'x1': x1,
'x2': x2,
'y1': y1,
'y2': y2,
'class': key,
'prob': new_probs[jk]
}
all_dets.append(det)
st.write('Elapsed time = {}'.format(time.time() - startTime))
t, p, iouVal = get_map(all_dets, img_data['bboxes'], (fx, fy))
for key in t.keys():
if key not in T:
T[key] = []
P[key] = []
T[key].extend(t[key])
P[key].extend(p[key])
all_aps = []
for key in T.keys():
ap = average_precision_score(T[key], P[key])
st.write('{} AP: {}'.format(key, ap))
all_aps.append(ap)
st.write('mAP = {}'.format(np.mean(np.array(all_aps))))
st.write('iou = {}'.format(np.mean(iouVal)))
mAPs.append(np.mean(np.array(all_aps)))
iou_map.append(iouVal)
st.markdown('## Mean IOU:', Average(iou_map))
st.markdown('## Mean average precision:', np.mean(np.array(mAPs)))
def train_mscoco():
# ===========================模型的配置和加载======================================
# config for data argument
cfg = config.Config()
cfg.use_horizontal_flips = True
cfg.use_vertical_flips = True
cfg.rot_90 = True
cfg.num_rois = 32
#resnet前四卷积部分的权值
cfg.base_net_weights = nn.get_weight_path()
#保存模型的权重值
cfg.model_path = './model/mscoco_frcnn.hdf5'
#all_images, class_mapping = get_data()
#加载训练的图片
train_imgs, class_mapping = get_data('train')
cfg.class_mapping = class_mapping
print('Num classes (including bg) = {}'.format(len(class_mapping)))
#保存所有的配置文件
with open(cfg.config_save_file, 'wb') as config_f:
pickle.dump(cfg, config_f)
print(
'Config has been written to {}, and can be loaded when testing to ensure correct results'
.format(cfg.config_save_file))
#图片随机洗牌
random.shuffle(train_imgs)
print('Num train samples {}'.format(len(train_imgs)))
data_gen_train = data_generators.get_anchor_gt(train_imgs,
class_mapping,
cfg,
nn.get_img_output_length,
K.image_dim_ordering(),
mode='train')
# ==============================================================================
# ===============================模型的定义======================================
#keras内核为tensorflow
input_shape_img = (None, None, 3)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(None, 4))
# define the base resnet50 network
shared_layers = nn.nn_base(img_input, trainable=False)
# define the RPN, built on the base layers
num_anchors = len(cfg.anchor_box_scales) * len(cfg.anchor_box_ratios)
rpn = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(shared_layers,
roi_input,
cfg.num_rois,
nb_classes=len(class_mapping),
trainable=True)
#model(input=,output=)
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)
# ==============================================================================
# ===========================基本模型加载ImageNet权值=============================
try:
print('loading base model weights from {}'.format(
cfg.base_net_weights))
model_rpn.load_weights(cfg.base_net_weights, by_name=True)
model_classifier.load_weights(cfg.base_net_weights, by_name=True)
except Exception as e:
print('基本模型加载ImageNet权值: ', e)
print('Could not load pretrained model weights on ImageNet.')
# ==============================================================================
# ===============================模型优化========================================
#在调用model.compile()之前初始化一个优化器对象，然后传入该函数
optimizer = Adam(lr=1e-5)
optimizer_classifier = Adam(lr=1e-5)
model_rpn.compile(optimizer=optimizer,
loss=[
losses_fn.rpn_loss_cls(num_anchors),
losses_fn.rpn_loss_regr(num_anchors)
])
model_classifier.compile(
optimizer=optimizer_classifier,
loss=[
losses_fn.class_loss_cls,
losses_fn.class_loss_regr(len(class_mapping) - 1)
],
metrics={'dense_class_{}'.format(len(class_mapping)): 'accuracy'})
model_all.compile(optimizer='sgd', loss='mae')
# ==============================================================================
# ================================训练、输出设置==================================
epoch_length = len(train_imgs)
num_epochs = int(cfg.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
logger = Logger(os.path.join('.', 'log.txt'))
# ==============================================================================
print('Starting training')
for epoch_num in range(num_epochs):
progbar = generic_utils.Progbar(epoch_length)
logger.write('Epoch {}/{}'.format(epoch_num + 1, num_epochs))
while True:
try:
if len(rpn_accuracy_rpn_monitor
) == epoch_length and cfg.verbose:
mean_overlapping_bboxes = float(
sum(rpn_accuracy_rpn_monitor)) / len(
rpn_accuracy_rpn_monitor)
rpn_accuracy_rpn_monitor = []
print(
'Average number of overlapping bounding boxes from RPN = {} for {} previous iterations'
.format(mean_overlapping_bboxes, epoch_length))
if mean_overlapping_bboxes == 0:
print(
'RPN is not producing bounding boxes that overlap'
' the ground truth boxes. Check RPN settings or keep training.'
)
#图片，标准的cls、rgr，盒子数据
X, Y, img_data = next(data_gen_train)
#训练rpn
loss_rpn = model_rpn.train_on_batch(X, Y)
#边训练rpn得到的区域送入roi
#x_class, x_regr, base_layers
P_rpn = model_rpn.predict_on_batch(X)
result = roi_helpers.rpn_to_roi(P_rpn[0],
P_rpn[1],
cfg,
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
#区域、cls、rgr、iou
X2, Y1, Y2, IouS = roi_helpers.calc_iou(
result, img_data, cfg, 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 cfg.num_rois > 1:
if len(pos_samples) < cfg.num_rois // 2:
selected_pos_samples = pos_samples.tolist()
else:
selected_pos_samples = np.random.choice(
pos_samples, cfg.num_rois // 2,
replace=False).tolist()
try:
selected_neg_samples = np.random.choice(
neg_samples,
cfg.num_rois - len(selected_pos_samples),
replace=False).tolist()
except:
selected_neg_samples = np.random.choice(
neg_samples,
cfg.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)
#训练classifier
loss_class = model_classifier.train_on_batch(
[X, X2[:, sel_samples, :]],
[Y1[:, sel_samples, :], Y2[:, sel_samples, :]])
losses[iter_num, 0] = loss_rpn[1]
losses[iter_num, 1] = loss_rpn[2]
losses[iter_num, 2] = loss_class[1]
losses[iter_num, 3] = loss_class[2]
losses[iter_num, 4] = loss_class[3]
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])),
('detector_regr', np.mean(losses[:iter_num, 3]))])
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 cfg.verbose:
logger.write(
'Mean number of bounding boxes from RPN overlapping ground truth boxes: {}'
.format(mean_overlapping_bboxes))
logger.write(
'Classifier accuracy for bounding boxes from RPN: {}'
.format(class_acc))
logger.write(
'Loss RPN classifier: {}'.format(loss_rpn_cls))
logger.write(
'Loss RPN regression: {}'.format(loss_rpn_regr))
logger.write('Loss Detector classifier: {}'.format(
loss_class_cls))
logger.write('Loss Detector regression: {}'.format(
loss_class_regr))
logger.write('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 cfg.verbose:
logger.write(
'Total loss decreased from {} to {}, saving weights'
.format(best_loss, curr_loss))
best_loss = curr_loss
model_all.save_weights(cfg.model_path)
break
except Exception as e:
print('Exception: {}'.format(e))
# save model
model_all.save_weights(cfg.model_path)
continue
print('Training complete, exiting.')
# model loading
C.model_path = options.load
print('Loading weights from {}'.format(C.model_path))
model_rpn.load_weights(C.model_path, by_name=True)
optimizer = Adam(lr=1e-3, clipnorm=0.001)
model_rpn.compile(optimizer=optimizer,
loss=[
loss_func.rpn_loss_cls(num_anchors),
loss_func.rpn_loss_regr(num_anchors)
])
#### load images here ####
from keras_frcnn.simple_parser import get_data
all_imgs, classes_count, class_mapping = get_data(options.test_path,
test_only=False)
train_imgs = [s for s in all_imgs if s['imageset'] == 'train']
data_gen_val = data_generators.get_anchor_gt(train_imgs,
classes_count,
C,
nn.get_img_output_length,
K.common.image_dim_ordering(),
mode='val')
with open('log/losses.csv', 'w') as log:
log_writer = csv.writer(log, delimiter=';')
log_writer.writerow(['loss', 'loss_rpn_cls', 'loss_rpn_regr', 'img'])
for img in train_imgs:
X_gen, Y_gen, imgdata = next(data_gen_val)
def train_net():
# config for data argument
cfg = config.Config()
cfg.use_horizontal_flips = False
cfg.use_vertical_flips = False
cfg.rot_90 = False
cfg.num_rois = 32 # config中设置的是4
cfg.base_net_weights = os.path.join('./model/', nn.get_weight_path())
# TODO: the only file should to be change for other data to train
cfg.model_path = 'samples.hdf5'
cfg.simple_label_file = 'annotations_train.txt' # 训练集产生的标签
all_images, classes_count, class_mapping = get_data(cfg.simple_label_file)
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
cfg.class_mapping = class_mapping
with open(cfg.config_save_file, 'wb') as config_f:
pickle.dump(cfg, config_f)
print('Config has been written to {}, and can be loaded when testing to ensure correct results'.format(
cfg.config_save_file))
inv_map = {v: k for k, v in class_mapping.items()}
print('Training images per class:')
pprint.pprint(classes_count)
print('Num classes (including bg) = {}'.format(len(classes_count)))
random.shuffle(all_images)
num_imgs = len(all_images)
train_imgs = [s for s in all_images if s['imageset'] == 'trainval']
val_imgs = [s for s in all_images if s['imageset'] == 'test']
print('Num train samples {}'.format(len(train_imgs)))
print('Num val samples {}'.format(len(val_imgs)))
# there图片
data_gen_train = data_generators.get_anchor_gt(train_imgs, classes_count, cfg, nn.get_img_output_length,
K.image_dim_ordering(), mode='train')
data_gen_val = data_generators.get_anchor_gt(val_imgs, classes_count, cfg, nn.get_img_output_length,
K.image_dim_ordering(), mode='val')
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=(None, 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(cfg.anchor_box_scales) * len(cfg.anchor_box_ratios)
rpn = nn.rpn(shared_layers, num_anchors)
# classifier是什么？
# classes_count {} 每一个类的数量：{'cow': 4, 'dog': 10, ...}
# C.num_rois每次取的感兴趣区域，默认为32
# roi_input = Input(shape=(None, 4)) 框框
# classifier是faster rcnn的两个损失函数[out_class, out_reg]
# shared_layers是vgg的输出feature map
classifier = nn.classifier(shared_layers, roi_input, cfg.num_rois, nb_classes=len(classes_count), trainable=True)
# 定义model_rpn
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:
print('loading weights from {}'.format(cfg.base_net_weights))
model_rpn.load_weights(cfg.model_path, by_name=True)
model_classifier.load_weights(cfg.model_path, by_name=True)
except Exception as e:
print(e)
print('Could not load pretrained model weights. Weights can be found in the keras application folder '
'https://github.com/fchollet/keras/tree/master/keras/applications')
optimizer = Adam(lr=1e-5)
optimizer_classifier = Adam(lr=1e-5)
model_rpn.compile(optimizer=optimizer,
loss=[losses_fn.rpn_loss_cls(num_anchors), losses_fn.rpn_loss_regr(num_anchors)])
model_classifier.compile(optimizer=optimizer_classifier,
loss=[losses_fn.class_loss_cls, losses_fn.class_loss_regr(len(classes_count) - 1)],
metrics={'dense_class_{}'.format(len(classes_count)): 'accuracy'})
model_all.compile(optimizer='sgd', loss='mae')
epoch_length = 10
num_epochs = int(cfg.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.items()}
print('Starting training')
vis = True
for epoch_num in range(num_epochs):
progbar = generic_utils.Progbar(epoch_length)
print('Epoch {}/{}'.format(epoch_num + 1, num_epochs))
while True:
try:
# 用来监督每一次epoch的平均正回归框的个数
if len(rpn_accuracy_rpn_monitor) == epoch_length and cfg.verbose:
mean_overlapping_bboxes = float(sum(rpn_accuracy_rpn_monitor)) / len(rpn_accuracy_rpn_monitor)
rpn_accuracy_rpn_monitor = []
print(
'Average number of overlapping bounding boxes from RPN = {} for {} previous iterations'.format(
mean_overlapping_bboxes, epoch_length))
if mean_overlapping_bboxes == 0:
# 每次都框不到正样本，说明rpn有问题
print('RPN is not producing bounding boxes that overlap'
' the ground truth boxes. Check RPN settings or keep training.')
# 迭代器，取数据
# 训练rpn网络，X是图片，Y是对应类别和回归梯度（不是所有的点都参加训练，符合条件才参加训练）
# next(data_gen_train)是一个迭代器。
# 返回的是 np.copy(x_img), [np.copy(y_rpn_cls), np.copy(y_rpn_regr)],
# img_data_aug（我们这里假设数据没有进行水平翻转等操作。那么，x_img = img_data_aug）,
# y_rpn_cls和y_rpn_regr是RPN的两个损失函数。
X, Y, img_data = next(data_gen_train)
# classifer和rpn网络交叉训练
loss_rpn = model_rpn.train_on_batch(X, Y)
P_rpn = model_rpn.predict_on_batch(X)
# result是得到的预选框
# 得到了region proposals，接下来另一个重要的思想就是ROI pooling，
# 可将不同shape的特征图转化为固定shape，送到全连接层进行最终的预测。
# rpn_to_roi接收的是每张图片的预测输出，返回的R = [boxes, probs]
# ---------------------
result = roi_helpers.rpn_to_roi(P_rpn[0], P_rpn[1], cfg, 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
# Y1根据预选框，得到这个预选框属于哪一类，
# Y2这个类相应的回归梯度
# X2是返回这个框
"""
# 通过calc_iou()找出剩下的不多的region对应ground truth里重合度最高的bbox，从而获得model_classifier的数据和标签。
# X2保留所有的背景和match bbox的框； Y1 是类别one-hot转码； Y2是对应类别的标签及回归要学习的坐标位置; IouS是debug用的。
"""
X2, Y1, Y2, IouS = roi_helpers.calc_iou(result, img_data, cfg, class_mapping)
if X2 is None:
# 如果没有有效的预选框则结束本次循环
rpn_accuracy_rpn_monitor.append(0)
rpn_accuracy_for_epoch.append(0)
continue
# 因为是one—hot，最后一位是1，则代表是背景
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] # 将其变为1维的数组
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 cfg.num_rois > 1:
# 选择num_rois个数的框，送入classifier网络进行训练。 分类网络一次要训练多少个框
# 思路：当num_rois大于1的时候正负样本尽量取到一半，小于1的时候正负样本随机取一个。
if len(pos_samples) < cfg.num_rois // 2:
# 挑选正样本
selected_pos_samples = pos_samples.tolist()
else:
selected_pos_samples = np.random.choice(pos_samples, cfg.num_rois // 2, replace=False).tolist()
try:
# 挑选负样本
selected_neg_samples = np.random.choice(neg_samples, cfg.num_rois - len(selected_pos_samples),
replace=False).tolist()
except:
selected_neg_samples = np.random.choice(neg_samples, cfg.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)
# 训练classifier网络
# 是从位置中挑选，
loss_class = model_classifier.train_on_batch([X, X2[:, sel_samples, :]],
[Y1[:, sel_samples, :], Y2[:, sel_samples, :]])
#
losses[iter_num, 0] = loss_rpn[1] # rpn_cls平均值
losses[iter_num, 1] = loss_rpn[2] # rpn_regr平均值
losses[iter_num, 2] = loss_class[1] # detector_cls平均值
losses[iter_num, 3] = loss_class[2] # detector_regr平均值
losses[iter_num, 4] = loss_class[3] # 4是准确率
iter_num += 1
# 进度条更新
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])),
('detector_regr', np.mean(losses[:iter_num, 3]))])
if iter_num == epoch_length:
loss_rpn_cls = np.mean(losses[:, 0]) # loss中存放了每一次训练出的losses
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 cfg.verbose:
# 打印出前n次loss的平均值
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:
# 当结束一轮的epoch时，只有当这轮epoch的loss小于最优的时候才会存储这轮的训练数据，
# 并结束这轮epoch进入下一轮epoch。
if cfg.verbose:
print('Total loss decreased from {} to {}, saving weights'.format(best_loss, curr_loss))
best_loss = curr_loss
model_all.save_weights(cfg.model_path)
break
except Exception as e:
print('Exception: {}'.format(e))
# save model
model_all.save_weights(cfg.model_path)
continue
print('Training complete, exiting.')
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_path specified in config file
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')
all_imgs = []
classes = {}
all_imgs, _, _ = get_data(options.test_path)
#test_imgs = [s for s in all_imgs if s['imageset'] == 'train'] # mAP
print("DEBUGGING 218: test imgs: ", len(test_imgs))
classification_threshold = 0.8 # threshold above which we classify as positive
# for mAP
T, P = {}, {}
counter = 0
for idx, img_data in enumerate(test_imgs):
print('{}/{}'.format(idx, len(test_imgs)))
img_name = img_data['filepath'].split('/')[-1]
print("DEBUGGING 232 img_name:", img_name)
print("img {}: {}".format(str(counter), img_name))
counter += 1
start_time = time.time()
def build_and_train(hype_space, save_best_weights=False):
train_path = '/home/comp/e4252392/retraindata4frcnn.txt'
config_output_filename = '/home/comp/e4252392/hyperopt/hyperopt_config.pickle'
num_epochs = 20
#for retrain best model only
diagnose_path = '/home/comp/e4252392/hyperopt/models/hyperopt_loss_ap_plt.npy'
real_model_path = '/home/comp/e4252392/hyperopt/models/hyperopt_model_plt_'
print("Hyperspace:")
print(hype_space)
C = config.Config()
C.num_rois = int(hype_space['num_rois']) #why int?
# C.anchor_box_scales = hype_space['anchor_box_scales']
# C.base_net_weights = '/home/comp/e4252392/second_res_more_epoch.h5'
C.base_net_weights = 'model_frcnn.hdf5'
#data
all_imgs, classes_count, class_mapping = get_data(train_path)
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
C.class_mapping = class_mapping
print('Training images per class:')
pprint.pprint(classes_count)
print('Num classes (including bg) = {}'.format(len(classes_count)))
with open(config_output_filename, 'wb') 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]
print('Num train samples {}'.format(len(train_imgs)))
data_gen_train = data_generators.get_anchor_gt(train_imgs,
classes_count,
C,
nn.get_img_output_length,
K.image_dim_ordering(),
mode='train')
#data
# build_model
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=(None, 4))
shared_layers = nn.nn_base(int(hype_space['kernel_size']),
img_input,
trainable=True)
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn = nn.rpn(int(hype_space['kernel_size']), shared_layers, num_anchors)
classifier = nn.classifier(int(hype_space['kernel_size']),
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)
model_all = Model([img_input, roi_input], rpn[:2] + classifier)
try:
print('loading weights from {}'.format(C.base_net_weights))
model_rpn.load_weights(C.base_net_weights, by_name=True)
model_classifier.load_weights(C.base_net_weights, by_name=True)
except:
print(
'Could not load pretrained model weights. Weights can be found in the keras application folder \
https://github.com/fchollet/keras/tree/master/keras/applications')
# optimizer = Adam(lr=1e-5)
# optimizer_classifier = Adam(lr=1e-5)
optimizer = Adam(lr=hype_space['optimizer_lr'],
decay=hype_space['optimizer_decay'])
optimizer_classifier = Adam(lr=hype_space['optimizer_lr'],
decay=hype_space['optimizer_decay'])
model_rpn.compile(optimizer=optimizer,
loss=[
thelosses.rpn_loss_cls(num_anchors),
thelosses.rpn_loss_regr(num_anchors)
])
model_classifier.compile(
optimizer=optimizer_classifier,
loss=[
thelosses.class_loss_cls,
thelosses.class_loss_regr(len(classes_count) - 1)
],
metrics={'dense_class_{}'.format(len(classes_count)): 'accuracy'})
sgd = SGD(lr=hype_space['sgd_lr'], decay=hype_space['sgd_decay'])
model_all.compile(optimizer=sgd, loss='mae')
# build_model
#build_and_train
epoch_length = 10
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
print('Starting training')
loss_array = []
ap_array = []
epoch_array = []
epoch_array.append(0)
result = {}
model_name = ''
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 = []
print(
'Average number of overlapping bounding boxes from RPN = {} for {} previous iterations'
.format(mean_overlapping_bboxes, epoch_length))
if mean_overlapping_bboxes == 0:
print(
'RPN is not producing bounding boxes that overlap the ground truth boxes. Check RPN settings or keep training.'
)
# train
X, Y, img_data = next(data_gen_train)
loss_rpn = model_rpn.train_on_batch(X, Y)
P_rpn = model_rpn.predict_on_batch(X)
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)
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 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:
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, :]])
# train
losses[iter_num, 0] = loss_rpn[1]
losses[iter_num, 1] = loss_rpn[2]
losses[iter_num, 2] = loss_class[1]
losses[iter_num, 3] = loss_class[2]
losses[iter_num, 4] = loss_class[3]
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])),
('detector_regr', np.mean(losses[:iter_num, 3]))])
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(
'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))
# result
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(
'Total loss decreased from {} to {}, saving weights'
.format(best_loss, curr_loss))
best_loss = curr_loss
if save_best_weights:
real_model_path = real_model_path + str(
epoch_num + 1) + '.hdf5'
model_all.save_weights(real_model_path,
overwrite=True)
print("Best weights so far saved to " +
real_model_path + ". best_loss = " +
str(best_loss))
epoch_array.append(epoch_num + 1)
loss_array.append([
loss_rpn_cls, loss_rpn_regr, loss_class_cls,
loss_class_regr, best_loss
])
album_ap, logo_ap, mAP = measure_map.measure_map(
config_output_filename, real_model_path)
ap_array.append([album_ap, logo_ap, mAP])
np.save(diagnose_path,
[epoch_array, loss_array, ap_array])
else:
album_ap = 'not applicable'
logo_ap = 'not applicable'
mAP = 'not applicable'
model_name = "model_{}_{}".format(
str(best_loss),
str(uuid.uuid4())[:5])
result = {
'loss': best_loss,
'loss_rpn_cls': loss_rpn_cls,
'loss_rpn_regr': loss_rpn_regr,
'loss_class_cls': loss_class_cls,
'loss_class_regr': loss_class_regr,
'album_ap': album_ap,
'logo_ap': logo_ap,
'mAP': mAP,
'model_name': model_name,
'space': hype_space,
'status': STATUS_OK
}
print("RESULT UPDATED.")
print("Model name: {}".format(model_name))
# result
break
except Exception as e:
print('Exception: {}'.format(e))
continue
print('Training complete, exiting.')
print("BEST MODEL: {}".format(model_name))
print("FINAL RESULT:")
print_json(result)
save_json_result(model_name, result)
try:
K.clear_session()
del model_all, model_rpn, model_classifier
except Exception as err:
try:
K.clear_session()
except:
pass
err_str = str(err)
print(err_str)
traceback_str = str(traceback.format_exc())
print(traceback_str)
return {
'status': STATUS_FAIL,
'err': err_str,
'traceback': traceback_str
}
print("\n\n")
return model_name, result
C = config.Config()
C.num_rois = 6
C.use_vertical_flips = True
C.use_horizontal_flips = True
C.scale_augment = True
C.rot_90 = True
C.balanced_classes = False
C2 = config.Config()
C2.num_rois = 6
#from keras_frcnn.pascal_voc_parser import get_data
from keras_frcnn.simple_parser import get_data
all_imgs, classes_count, class_mapping = get_data('C:/Fraps/im2txt.txt')
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
with open('classes.json', 'w') as class_data_json:
json.dump(class_mapping, class_data_json)
inv_map = {v: k for k, v in class_mapping.items()}
#pprint.pprint(classes_count)
print(('Num classes (including bg) = {}'.format(len(classes_count))))
random.shuffle(all_imgs)
num_imgs = len(all_imgs)
def Train_frcnn(
train_path='./data/flickr_logos_27_dataset_training_set_annotation.txt', # path to the text file containing the data
network_arch='vgg', # the type of the base faster rcnn network architecture
num_epochs=50, # num of epochs
output_weight_path='./models/model_frcnn.hdf5', # path to save the model_all.weights as hdf5
preprocessing_function=None,
config_filename="config.pickle",
input_weights_path='./models/vgg16_weights_tf_dim_ordering_tf_kernels.h5',
train_rpn=True,
train_final_classifier=True,
train_base_nn=True,
losses_to_watch=['rpn_cls', 'rpn_reg', 'final_cls', 'final_reg'],
tb_log_dir="log",
num_rois=32,
horizontal_flips=False,
vertical_flips=False,
rot_90=False,
anchor_box_scales=[128, 256, 512],
anchor_box_ratios=[[1, 1], [1. / math.sqrt(2), 2. / math.sqrt(2)],
[2. / math.sqrt(2), 1. / math.sqrt(2)]],
im_size=600,
rpn_stride=16, # depends on network architecture
visualize_model=None,
verify_trainable=True,
optimizer_rpn=Adam(lr=1e-5),
optimizer_classifier=Adam(lr=1e-5),
validation_interval=3,
rpn_min_overlap=0.3,
rpn_max_overlap=0.7,
classifier_min_overlap=0.1,
classifier_max_overlap=0.5,
rpn_nms_threshold=0.7, # original implementation
seed=5000):
"""
Trains a Faster RCNN for object detection in keras
NOTE: This trains 2 models namely model_rpn and model_classifer with the same shared base_nn (fixed feature extractor)
Keyword Arguments
train_path -- str: path to the text file or pascal_voc (no Default)
network_arch --object: the full faster rcnn network .py file passed as an object (no default)
num_epochs -- int: number of epochs to train (no Default)
output_weight_path --str: path to save the frcnn weights (no Default)
preprocessing_function --function: Optional preprocessing function (must be defined like given in keras docs) (Default None)
config_filename --str: Path to save the config file. Used when testing (Default "config.pickle")
input_weight_path --str: Path to hdf5 file containing weights for the model (Default None)
you can pass path to both classification and detection checkpoints as long as the names dont' change
train_rpn --bool: whether to train the rpn layer (Default True)
train_final_classifier --bool:Whether to train the final_classifier (Fast Rcnn layer) (Default True)
train_base_nn --bool:Whether to train the base_nn/fixed_feature_extractor (Default True)
losses_to_watch --list: A list of losses to watch (Default ['rpn_cls','rpn_reg','final_cls','final_reg']).
The losses in this list are added and then weights are saved wrt to that.
The list can contain any combination of the above 4 only.
tb_log_dir --str: path to log dir for tensorboard logging (Default 'log')
num_rois --int: The number of rois to use at once (Default = 32)
horizontal_flips --bool: augment training data by horizontal flips (Default False)
vertical_flips --bool: augment training data by vertical flips (Default False)
rot_90 --bool: augment training data by 90 deg rotations (Default False)
anchor_box_scales --list: The list of anchor box scales to use (Default [128,256,512])
anchor_box ratios --list of list: The list of anchorbox aspect ratios to use (Default [[1, 1], [1./math.sqrt(2), 2./math.sqrt(2)], [2./math.sqrt(2), 1./math.sqrt(2)]])
im_size --int: The size to resize the image (Default 600). This is the smallest side of Pascal VOC format
rpn_stride --int: The stride for rpn (Default = 16)
visualize_model --str: Path to save the model as .png file
verify_trainable --bool: print layer wise names and prints if it is trainable or not (Default True)
optimizer_rpn --keras.optimizer: The optimizer for rpn (Default Adam(lr=1e-5))
optimizer_classifier --keras.optimizer: The optimizer for classifier (Default Adam(lr=1e-5))
validation_interval --int: The frequency (in epochs) to do validation. supply 0 if no validation
rpn_min_overlap --float: (0,1) The Min IOU in rpn layer (Default 0.3) (original implementation)
rpn_max_overlap --float: (0,1) The max IOU in rpn layer (Default 0.7) (original implementation)
classifier_min_overlap --float: (0,1) same as above but in final classifier (Default 0.1) (original implementation)
classifier_max_overlap --float: (0,1) same as above (Default 0.5) (original implementation)
rpn_nms_threshold --float :(0,1) The threshold above which to supress the bbox using Non max supression in rpn (Default 0.7)(from original implementation)
seed --int: To seed the random shuffling of training data (Default = 5000)
Performing alternating training:
- Use the train_rpn,train_final_classifier and train_base_nn boolean arguments to accomplish
alternating training.
- While using the above arguments change the members of losses_to_watch = ['rpn_cls','rpn_reg','final_cls','final_reg']
accordingly else it will throw error
- for eg if you are training only the base_nn and the rpn set:
train_rpn = True
train_base_nn = True
train_final_classifier = False
losses_to_watch = ['rpn_cls','rpn_reg'] (do not include 'final_cls', 'final_reg')
OUTPUT:
prints the training log. Does not return anything
Save details:
1.saves the weights of the full FRCNN model as .h5
2.saves a tensorboard file
3.saves the history of weights saved in ./saving_log.txt so that it can be known at which epoch the model is saved
4.saves the model configuration as a .pickle file
5.optionally saves the full FRCNN architecture as .png
NOTE:
as of now the batch size = 1
Prints loss = 0 for losses from model which is not being trained
TODO: The training is a bit slow because of the data generation step. Generate_data in multiple threads and queue them for faster training
"""
check_list = ['rpn_cls', 'rpn_reg', 'final_cls', 'final_reg']
for n in losses_to_watch:
if n not in check_list:
raise ValueError(
"unsupported loss the supported losses are: {}".format(
check_list))
if not train_rpn:
if "rpn_cls" in losses_to_watch or "rpn_reg" in losses_to_watch:
raise ValueError(
"Cannot watch rpn_cls and rpn_reg when train_rpn == False")
if not train_final_classifier:
if "final_cls" in losses_to_watch or "final_reg" in losses_to_watch:
raise ValueError(
"cannot watch final_cls and final_reg when train_final_classifier == False"
)
if network_arch == 'vgg':
from keras_frcnn import nn_arch_vgg16 as nn
elif network_arch == 'resnet50':
from keras_frcnn import nn_arch_resnet50 as nn
else:
print('Not a valid model')
raise ValueError
random.seed(seed)
np.random.seed(seed)
# pass the settings from the function call, and persist them in the config object
C = config.Config()
C.rpn_max_overlap = rpn_max_overlap
C.rpn_min_overlap = rpn_min_overlap
C.classifier_min_overlap = classifier_min_overlap
C.classifier_max_overlap = classifier_max_overlap
C.anchor_box_scales = anchor_box_scales
C.anchor_box_ratios = anchor_box_ratios
C.im_size = im_size
C.use_horizontal_flips = bool(horizontal_flips)
C.use_vertical_flips = bool(vertical_flips)
C.rot_90 = bool(rot_90)
C.rpn_stride = rpn_stride
C.rpn_nms_threshold = rpn_nms_threshold
C.weights_all_path = output_weight_path
C.num_rois = int(num_rois)
# check if weight path was passed via command line
if input_weights_path:
C.initial_weights = input_weights_path
all_imgs, classes_count, class_mapping = get_data(train_path)
print("The class mapping is:")
print(class_mapping)
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
C.class_mapping = class_mapping
print('Training images per class:')
pprint.pprint(classes_count)
print('Num classes (including bg) = {}'.format(len(classes_count)))
with open(config_filename, 'wb') 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_filename))
np.random.shuffle(all_imgs)
train_imgs = [s for s in all_imgs if s['imageset'] == 'train']
val_imgs = [s for s in all_imgs if s['imageset'] == 'test']
print('Num train samples {}'.format(len(train_imgs)))
print('Num val samples {}'.format(len(val_imgs)))
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, can be VGG, Inception, etc)
shared_layers = nn.nn_base(img_input, trainable=train_base_nn)
# 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, trainable=train_rpn)
# define the classifier, built on base layers
classifier = nn.classifier(shared_layers,
roi_input,
C.num_rois,
len(classes_count),
trainable=train_final_classifier)
# create models
model_base = Model(img_input,
shared_layers) # for computing the output shape
model_rpn = Model(img_input, rpn[:2]) # used for training
model_classifier = Model([img_input, roi_input],
classifier) # used for training
# this is a model that holds both the RPN and the classifier, used to load/save and freeze/unfreeze weights for the models
model_all = Model([img_input, roi_input], rpn[:2] + classifier)
# tensorboard
tbCallBack = TensorBoard(log_dir=tb_log_dir,
histogram_freq=1,
write_graph=False,
write_images=False)
tbCallBack.set_model(model_all)
#NOTE: both model_rpn and model_classifer contains the base_nn
try:
print('loading weights from {}'.format(C.initial_weights))
model_all.load_weights(C.initial_weights, by_name=True)
except:
print('Could not load pretrained model weights')
# number of trainable parameters
trainable_count = int(
np.sum([K.count_params(p) for p in set(model_all.trainable_weights)]))
non_trainable_count = int(
np.sum(
[K.count_params(p) for p in set(model_all.non_trainable_weights)]))
print('Total params: {:,}'.format(trainable_count + non_trainable_count))
print('Trainable params: {:,}'.format(trainable_count))
print('Non-trainable params: {:,}'.format(non_trainable_count))
if verify_trainable:
for layer in model_all.layers:
print(layer.name, layer.trainable)
model_rpn.compile(optimizer=optimizer_rpn,
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='mse')
# save model_all as png for visualization
if visualize_model != None:
# from IPython.display import SVG
# from keras.utils.vis_utils import model_to_dot
# SVG(model_to_dot(model_all).create(prog='dot', format='svg'))
plot_model(model=model_all,
to_file=visualize_model,
show_shapes=True,
show_layer_names=True)
epoch_length = len(train_imgs)
validation_epoch_length = len(val_imgs)
num_epochs = int(num_epochs)
iter_num = 0
# train and valid data generator
data_gen_train = data_generators.get_anchor_gt(train_imgs,
classes_count,
C,
model_base,
K.image_dim_ordering(),
preprocessing_function,
mode='train')
data_gen_val = data_generators.get_anchor_gt(val_imgs,
classes_count,
C,
model_base,
K.image_dim_ordering(),
preprocessing_function,
mode='val')
losses_val = np.zeros((validation_epoch_length, 5))
losses = np.zeros((epoch_length, 5))
rpn_accuracy_rpn_monitor = []
rpn_accuracy_for_epoch = []
start_time = time.time()
best_loss = np.Inf
val_best_loss = np.Inf
val_best_loss_epoch = 0
print('Starting training')
def write_log(callback, names, logs, batch_no):
for name, value in zip(names, logs):
summary = tf.Summary()
summary_value = summary.value.add()
summary_value.simple_value = value
summary_value.tag = name
callback.writer.add_summary(summary, batch_no)
callback.writer.flush()
train_names = [
'train_loss_rpn_cls', 'train_loss_rpn_reg', 'train_loss_class_cls',
'train_loss_class_reg', 'train_total_loss', 'train_acc'
]
val_names = [
'val_loss_rpn_cls', 'val_loss_rpn_reg', 'val_loss_class_cls',
'val_loss_class_reg', 'val_total_loss', 'val_acc'
]
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 = []
print(
'Average number of overlapping bounding boxes from RPN = {} for {} previous iterations'
.format(mean_overlapping_bboxes, epoch_length))
if mean_overlapping_bboxes == 0:
print(
'RPN is not producing bounding boxes that overlap the ground truth boxes. Check RPN settings or keep training.'
)
X, Y, img_data = next(data_gen_train)
if train_rpn:
loss_rpn = model_rpn.train_on_batch(X, Y)
P_rpn = model_rpn.predict_on_batch(X)
R = roi_helpers.rpn_to_roi(P_rpn[0],
P_rpn[1],
C,
K.image_dim_ordering(),
use_regr=True,
overlap_thresh=C.rpn_nms_threshold,
flag="train")
# note: calc_iou converts from (x1,y1,x2,y2) to (x,y,w,h) format
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 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)
if train_final_classifier:
loss_class = model_classifier.train_on_batch(
[X, X2[:, sel_samples, :]],
[Y1[:, sel_samples, :], Y2[:, sel_samples, :]])
# losses
if train_rpn:
losses[iter_num, 0] = loss_rpn[1]
losses[iter_num, 1] = loss_rpn[2]
else:
losses[iter_num, 0] = 0
losses[iter_num, 1] = 0
if train_final_classifier:
losses[iter_num, 2] = loss_class[1]
losses[iter_num, 3] = loss_class[2]
losses[iter_num, 4] = loss_class[3] # accuracy
else:
losses[iter_num, 2] = 0
losses[iter_num, 3] = 0
losses[iter_num, 4] = 0
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])),
('detector_regr', np.mean(losses[:iter_num, 3]))])
if iter_num == epoch_length:
if train_rpn:
loss_rpn_cls = np.mean(losses[:, 0])
loss_rpn_regr = np.mean(losses[:, 1])
else:
loss_rpn_cls = 0
loss_rpn_regr = 0
if train_final_classifier:
loss_class_cls = np.mean(losses[:, 2])
loss_class_regr = np.mean(losses[:, 3])
class_acc = np.mean(losses[:, 4])
else:
loss_class_cls = 0
loss_class_regr = 0
class_acc = 0
mean_overlapping_bboxes = float(sum(
rpn_accuracy_for_epoch)) / len(rpn_accuracy_for_epoch)
rpn_accuracy_for_epoch = []
if C.verbose:
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))
loss_dict_train = {
"rpn_cls": loss_rpn_cls,
"rpn_reg": loss_rpn_regr,
"final_cls": loss_class_cls,
"final_reg": loss_class_regr
}
curr_loss = 0
for l in losses_to_watch:
curr_loss += loss_dict_train[l]
iter_num = 0
start_time = time.time()
write_log(tbCallBack, train_names, [
loss_rpn_cls, loss_rpn_regr, loss_class_cls,
loss_class_regr, curr_loss, class_acc
], epoch_num)
if curr_loss < best_loss:
if C.verbose:
print(
'Total loss decreased from {} to {} in training, saving weights'
.format(best_loss, curr_loss))
save_log_data = '\nTotal loss decreased from {} to {} in epoch {}/{} in training, saving weights'.format(
best_loss, curr_loss, epoch_num + 1,
num_epochs)
with open("./saving_log.txt", "a") as f:
f.write(save_log_data)
best_loss = curr_loss
model_all.save_weights(C.weights_all_path)
break
except Exception as e:
print('Exception: {}'.format(e))
continue
if validation_interval > 0:
# validation
if (epoch_num + 1) % validation_interval == 0:
progbar = generic_utils.Progbar(validation_epoch_length)
print("Validation... \n")
while True:
try:
X, Y, img_data = next(data_gen_val)
if train_rpn:
val_loss_rpn = model_rpn.test_on_batch(X, Y)
P_rpn = model_rpn.predict_on_batch(X)
R = roi_helpers.rpn_to_roi(
P_rpn[0],
P_rpn[1],
C,
K.image_dim_ordering(),
use_regr=True,
overlap_thresh=C.rpn_nms_threshold,
flag="train")
# note: calc_iou converts from (x1,y1,x2,y2) to (x,y,w,h) format
X2, Y1, Y2, IouS = roi_helpers.calc_iou(
R, img_data, C, class_mapping)
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 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)
if train_final_classifier:
val_loss_class = model_classifier.test_on_batch(
[X, X2[:, sel_samples, :]],
[Y1[:, sel_samples, :], Y2[:, sel_samples, :]])
if train_rpn:
losses_val[iter_num, 0] = val_loss_rpn[1]
losses_val[iter_num, 1] = val_loss_rpn[2]
else:
losses_val[iter_num, 0] = 0
losses_val[iter_num, 1] = 0
if train_final_classifier:
losses_val[iter_num, 2] = val_loss_class[1]
losses_val[iter_num, 3] = val_loss_class[2]
losses_val[iter_num, 4] = val_loss_class[3]
else:
losses_val[iter_num, 2] = 0
losses_val[iter_num, 3] = 0
losses_val[iter_num, 4] = 0
iter_num += 1
progbar.update(
iter_num,
[('rpn_cls', np.mean(losses_val[:iter_num, 0])),
('rpn_regr', np.mean(losses_val[:iter_num, 1])),
('detector_cls', np.mean(losses_val[:iter_num,
2])),
('detector_regr', np.mean(losses_val[:iter_num,
3]))])
if iter_num == validation_epoch_length:
if train_rpn:
val_loss_rpn_cls = np.mean(losses_val[:, 0])
val_loss_rpn_regr = np.mean(losses_val[:, 1])
else:
val_loss_rpn_cls = 0
val_loss_rpn_regr = 0
if train_final_classifier:
val_loss_class_cls = np.mean(losses_val[:, 2])
val_loss_class_regr = np.mean(losses_val[:, 3])
val_class_acc = np.mean(losses_val[:, 4])
else:
val_loss_class_cls = 0
val_loss_class_regr = 0
val_class_acc = 0
mean_overlapping_bboxes = float(
sum(rpn_accuracy_for_epoch)) / len(
rpn_accuracy_for_epoch)
rpn_accuracy_for_epoch = []
loss_dict_valid = {
"rpn_cls": val_loss_rpn_cls,
"rpn_reg": val_loss_rpn_regr,
"final_cls": val_loss_class_cls,
"final_reg": val_loss_class_regr
}
val_curr_loss = 0
for l in losses_to_watch:
val_curr_loss += loss_dict_valid[l]
write_log(tbCallBack, val_names, [
val_loss_rpn_cls, val_loss_rpn_regr,
val_loss_class_cls, val_loss_class_regr,
val_curr_loss, val_class_acc
], epoch_num)
if C.verbose:
print('[INFO VALIDATION]')
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(val_class_acc))
print('Loss RPN classifier: {}'.format(
val_loss_rpn_cls))
print('Loss RPN regression: {}'.format(
val_loss_rpn_regr))
print('Loss Detector classifier: {}'.format(
val_loss_class_cls))
print('Loss Detector regression: {}'.format(
val_loss_class_regr))
print(
"current loss: %.2f, best loss: %.2f at epoch: %d"
% (val_curr_loss, val_best_loss,
val_best_loss_epoch))
print('Elapsed time: {}'.format(time.time() -
start_time))
if val_curr_loss < val_best_loss:
if C.verbose:
print(
'Total loss decreased from {} to {}, saving weights'
.format(val_best_loss, val_curr_loss))
save_log_data = '\nTotal loss decreased from {} to {} in epoch {}/{} in validation, saving weights'.format(
val_best_loss, val_curr_loss,
epoch_num + 1, num_epochs)
with open("./saving_log.txt", "a") as f:
f.write(save_log_data)
val_best_loss = val_curr_loss
val_best_loss_epoch = epoch_num
model_all.save_weights(C.weights_all_path)
start_time = time.time()
iter_num = 0
break
except:
pass
print('Training complete, exiting.')
import pprint
import sys
import json
from keras_frcnn import config
sys.setrecursionlimit(40000)
C = config.Config()
C.num_rois = 2
C.use_vertical_flips = True
#from keras_frcnn.pascal_voc_parser import get_data
from keras_frcnn.simple_parser import get_data
all_imgs, classes_count, class_mapping = get_data(sys.argv[1])
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
with open('classes.json', 'w') as class_data_json:
json.dump(class_mapping, class_data_json)
inv_map = {v: k for k, v in class_mapping.items()}
#pprint.pprint(classes_count)
print(('Num classes (including bg) = {}'.format(len(classes_count))))
random.shuffle(all_imgs)
num_imgs = len(all_imgs)
def test_kitti():
# config for data argument
cfg = config.Config()
cfg.use_horizontal_flips = True
cfg.use_vertical_flips = True
cfg.rot_90 = True
cfg.num_rois = 32
cfg.base_net_weights = os.path.join('./model/', nn.get_weight_path())
# TODO: the only file should to be change for other data to train
cfg.model_path = './model/kitti_frcnn_last.hdf5'
cfg.simple_label_file = 'kitti_simple_label.txt'
#查看绝对路径
#t = os.path.abspath('kitti_simple_label.txt')
all_images, classes_count, class_mapping = get_data(cfg.simple_label_file)
pedestrain_num = classes_count['Pedestrian']
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
cfg.class_mapping = class_mapping
with open(cfg.config_save_file, 'wb') as config_f:
pickle.dump(cfg, config_f)
print(
'Config has been written to {}, and can be loaded when testing to ensure correct results'
.format(cfg.config_save_file))
inv_map = {v: k for k, v in class_mapping.items()}
print('Training images per class:')
pprint.pprint(classes_count)
print('Num classes (including bg) = {}'.format(len(classes_count)))
random.shuffle(all_images)
num_imgs = len(all_images)
train_imgs = [s for s in all_images if s['imageset'] == 'trainval']
val_imgs = [s for s in all_images if s['imageset'] == 'test']
print('Num train samples {}'.format(len(train_imgs)))
print('Num val samples {}'.format(len(val_imgs)))
data_gen_train = data_generators.get_anchor_gt(train_imgs,
classes_count,
cfg,
nn.get_img_output_length,
K.image_dim_ordering(),
mode='train')
data_gen_val = data_generators.get_anchor_gt(val_imgs,
classes_count,
cfg,
nn.get_img_output_length,
K.image_dim_ordering(),
mode='val')
if K.image_dim_ordering() == 'th':
input_shape_img = (3, None, None)
else:
input_shape_img = (None, None, 3)
#img_input: 三通道，为输入图片
img_input = Input(shape=input_shape_img)
#roi_input：为输入图片boudingbox的四维值
roi_input = Input(shape=(None, 4))
# define the base network (resnet here, can be VGG, Inception, etc)
#shared_layers : 基础的网络结构（例如： resnet，vgg）通过该网络来提取原始图片的featuremap特征,最后将这些特征送入RPN网络和RCNN网络
# 1.定义nn的输入层，faster-rcnn共享卷积层，
shared_layers = nn.nn_base(img_input, trainable=True)
# define the RPN, built on the base layers 2.定义RPN层
num_anchors = len(cfg.anchor_box_scales) * len(cfg.anchor_box_ratios)
#RPN网络用于生成region proposals，该层通过sigmoid函数判断anchors属于foreground或者background， 再利用bounding box regression修正anchors获得修正后的RoI。
# rpn： 在基础的网络结构使用9个bounding box产生了分类和回归的rpn网络。定义rpn层，return [x_class, x_regr, base_layers]
rpn = nn.rpn(shared_layers, num_anchors)
#定义分类器层，定义classifier的输入和输出
classifier = nn.classifier(shared_layers,
roi_input,
cfg.num_rois,
nb_classes=len(classes_count),
trainable=True)
#定义rpn模型的输入和输出一个框2分类（最后使用的sigmod而不是softmax）和框的回归
model_rpn = Model(img_input, rpn[:2])
#定义classifier的输入和输出
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:
print('loading weights from {}'.format(cfg.base_net_weights))
#TODO 第一次运行因为model_path没有hdf5文件，因此修改为cfg.base_net_weights，现在可以修改回来
model_rpn.load_weights(cfg.model_path, by_name=True)
model_classifier.load_weights(cfg.model_path, by_name=True)
# model_rpn.load_weights(cfg.base_net_weights, by_name=True)
# model_classifier.load_weights(cfg.base_net_weights, by_name=True)
except Exception as e:
print(e)
print(
'Could not load pretrained model weights. Weights can be found in the keras application folder '
'https://github.com/fchollet/keras/tree/master/keras/applications')
optimizer = Adam(lr=1e-5)
optimizer_classifier = Adam(lr=1e-5)
model_rpn.compile(optimizer=optimizer,
loss=[
losses_fn.rpn_loss_cls(num_anchors),
losses_fn.rpn_loss_regr(num_anchors)
])
model_classifier.compile(
optimizer=optimizer_classifier,
loss=[
losses_fn.class_loss_cls,
losses_fn.class_loss_regr(len(classes_count) - 1)
],
metrics={'dense_class_{}'.format(len(classes_count)): 'accuracy'})
model_all.compile(optimizer='sgd', loss='mae')
#todo 增加tensorboard日志文件
log_path = './graph'
callback = TensorBoard(log_path,
histogram_freq=0,
write_graph=True,
write_images=True)
callback.set_model(model_all)
#todo epoch的大小为训练图片的个数
# epoch_length = len(train_imgs)
epoch_length = len(val_imgs)
#epoch_length = 47182
num_epochs = int(cfg.num_epochs)
iter_num = 0
losses = np.zeros((epoch_length, 5))
#todo
losses_val = np.zeros((epoch_length, 5))
rpn_accuracy_rpn_monitor = []
rpn_accuracy_for_epoch = []
rpn_accuracy_rpn_monitor_val = []
rpn_accuracy_for_epoch_val = []
start_time = time.time()
best_loss = np.Inf
class_mapping_inv = {v: k for k, v in class_mapping.items()}
print('Starting testing')
vis = True
allbbox = 0
#只有训练的,改成只有测试的
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 cfg.verbose:
mean_overlapping_bboxes = float(
sum(rpn_accuracy_rpn_monitor)) / len(
rpn_accuracy_rpn_monitor)
rpn_accuracy_rpn_monitor = []
print(
'Average number of overlapping bounding boxes from RPN = {} for {} previous iterations'
.format(mean_overlapping_bboxes, epoch_length))
if mean_overlapping_bboxes == 0:
print(
'RPN is not producing bounding boxes that overlap'
' the ground truth boxes. Check RPN settings or keep training.'
)
#todo train修改为val
# X, Y, img_data = next(data_gen_train)
X, Y, img_data = next(data_gen_val)
# loss_rpn = model_rpn.train_on_batch(X, Y)
loss_rpn = model_rpn.test_on_batch(X, Y)
P_rpn = model_rpn.predict_on_batch(X)
result = roi_helpers.rpn_to_roi(P_rpn[0],
P_rpn[1],
cfg,
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
#todo 增加 count
# X2, Y1, Y2, IouS = roi_helpers.calc_iou(result, img_data, cfg, class_mapping)
X2, Y1, Y2, IouS, count = roi_helpers.calc_iou(
result, img_data, cfg, class_mapping)
allbbox = allbbox + count
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 cfg.num_rois > 1:
if len(pos_samples) < cfg.num_rois // 2:
selected_pos_samples = pos_samples.tolist()
else:
selected_pos_samples = np.random.choice(
pos_samples, cfg.num_rois // 2,
replace=False).tolist()
try:
selected_neg_samples = np.random.choice(
neg_samples,
cfg.num_rois - len(selected_pos_samples),
replace=False).tolist()
except:
selected_neg_samples = np.random.choice(
neg_samples,
cfg.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, 0] = loss_rpn[1]
losses[iter_num, 1] = loss_rpn[2]
losses[iter_num, 2] = loss_class[1]
losses[iter_num, 3] = loss_class[2]
losses[iter_num, 4] = loss_class[3]
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])),
('detector_regr', np.mean(losses[:iter_num, 3]))])
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 cfg.verbose:
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 cfg.verbose:
print(
'Total loss decreased from {} to {}, saving weights'
.format(best_loss, curr_loss))
best_loss = curr_loss
model_all.save_weights(cfg.model_path)
break
except Exception as e:
print('Exception: {}'.format(e))
# save model
model_all.save_weights(cfg.model_path)
continue
print("检测准确率：")
print(float(allbbox / pedestrain_num))
print('testing complete, exiting.')
from keras.utils import generic_utils
from keras_frcnn.simple_parser import get_data
C = config.Config()
now = datetime.datetime.now()
otherStyleTime = now.strftime("%Y-%m-%d %H:%M:%S")
model_path = 'model/' + str(otherStyleTime) + 'stag_model_frcnn.hdf5'
config_path = 'model/' + str(otherStyleTime) + 'stag_config.pickle'
C.model_path = model_path
C.num_rois = 32
all_imgs, classes_count, class_mapping = get_data('managedata/stag_data_with_person.txt')
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
C.class_mapping = class_mapping
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)))
config_output_filename = config_path
def train_kitti():
# config for data argument
cfg = config.Config()
cfg.use_horizontal_flips = True
cfg.use_vertical_flips = True
cfg.rot_90 = True
cfg.num_rois = 32
cfg.base_net_weights = os.path.join('./model/', nn.get_weight_path())
# TODO: the only file should to be change for other data to train
cfg.model_path = './model/kitti_frcnn_last.hdf5'
cfg.simple_label_file = 'kitti_simple_label.txt'
all_images, classes_count, class_mapping = get_data(cfg.simple_label_file)
if 'bg' not in classes_count:
classes_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
cfg.class_mapping = class_mapping
with open(cfg.config_save_file, 'wb') as config_f:
pickle.dump(cfg, config_f)
print('Config has been written to {}, and can be loaded when testing to ensure correct results'.format(
cfg.config_save_file))
inv_map = {v: k for k, v in class_mapping.items()}
print('Training images per class:')
pprint.pprint(classes_count)
print('Num classes (including bg) = {}'.format(len(classes_count)))
random.shuffle(all_images)
num_imgs = len(all_images)
train_imgs = [s for s in all_images if s['imageset'] == 'trainval']
val_imgs = [s for s in all_images if s['imageset'] == 'test']
print('Num train samples {}'.format(len(train_imgs)))
print('Num val samples {}'.format(len(val_imgs)))
data_gen_train = data_generators.get_anchor_gt(train_imgs, classes_count, cfg, nn.get_img_output_length,
K.image_dim_ordering(), mode='train')
data_gen_val = data_generators.get_anchor_gt(val_imgs, classes_count, cfg, nn.get_img_output_length,
K.image_dim_ordering(), mode='val')
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=(None, 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(cfg.anchor_box_scales) * len(cfg.anchor_box_ratios)
rpn = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(shared_layers, roi_input, cfg.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:
print('loading weights from {}'.format(cfg.base_net_weights))
model_rpn.load_weights(cfg.model_path, by_name=True)
model_classifier.load_weights(cfg.model_path, by_name=True)
except Exception as e:
print(e)
print('Could not load pretrained model weights. Weights can be found in the keras application folder '
'https://github.com/fchollet/keras/tree/master/keras/applications')
optimizer = Adam(lr=1e-5)
optimizer_classifier = Adam(lr=1e-5)
model_rpn.compile(optimizer=optimizer,
loss=[losses_fn.rpn_loss_cls(num_anchors), losses_fn.rpn_loss_regr(num_anchors)])
model_classifier.compile(optimizer=optimizer_classifier,
loss=[losses_fn.class_loss_cls, losses_fn.class_loss_regr(len(classes_count) - 1)],
metrics={'dense_class_{}'.format(len(classes_count)): 'accuracy'})
model_all.compile(optimizer='sgd', loss='mae')
epoch_length = 1000
num_epochs = int(cfg.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.items()}
print('Starting training')
vis = True
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 cfg.verbose:
mean_overlapping_bboxes = float(sum(rpn_accuracy_rpn_monitor)) / len(rpn_accuracy_rpn_monitor)
rpn_accuracy_rpn_monitor = []
print(
'Average number of overlapping bounding boxes from RPN = {} for {} previous iterations'.format(
mean_overlapping_bboxes, epoch_length))
if mean_overlapping_bboxes == 0:
print('RPN is not producing bounding boxes that overlap'
' the ground truth boxes. Check RPN settings or keep training.')
X, Y, img_data = next(data_gen_train)
loss_rpn = model_rpn.train_on_batch(X, Y)
P_rpn = model_rpn.predict_on_batch(X)
result = roi_helpers.rpn_to_roi(P_rpn[0], P_rpn[1], cfg, 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, IouS = roi_helpers.calc_iou(result, img_data, cfg, 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 cfg.num_rois > 1:
if len(pos_samples) < cfg.num_rois // 2:
selected_pos_samples = pos_samples.tolist()
else:
selected_pos_samples = np.random.choice(pos_samples, cfg.num_rois // 2, replace=False).tolist()
try:
selected_neg_samples = np.random.choice(neg_samples, cfg.num_rois - len(selected_pos_samples),
replace=False).tolist()
except:
selected_neg_samples = np.random.choice(neg_samples, cfg.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, 0] = loss_rpn[1]
losses[iter_num, 1] = loss_rpn[2]
losses[iter_num, 2] = loss_class[1]
losses[iter_num, 3] = loss_class[2]
losses[iter_num, 4] = loss_class[3]
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])),
('detector_regr', np.mean(losses[:iter_num, 3]))])
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 cfg.verbose:
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 cfg.verbose:
print('Total loss decreased from {} to {}, saving weights'.format(best_loss, curr_loss))
best_loss = curr_loss
model_all.save_weights(cfg.model_path)
break
except Exception as e:
print('Exception: {}'.format(e))
# save model
model_all.save_weights(cfg.model_path)
continue
print('Training complete, exiting.')
def meature_map(test_path,length):
test_imgs, _, _ = sp.get_data(test_path)
T = {}
P = {}
mAPs = []
for idx, img_data in enumerate(test_imgs):
print('{}/{}'.format(idx, len(test_imgs)))
st = time.time()
filepath = img_data['filepath']
img = cv2.imread(filepath)
X, fx, fy = format_img_map(img, C)
# Change X (img) shape from (1, channel, height, width) to (1, height, width, channel)
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, K.image_dim_ordering(), overlap_thresh=0.7)
# 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] // 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:
# pad R
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])
# Calculate all classes' bboxes coordinates on resized image (300, 400)
# Drop 'bg' classes bboxes
for ii in range(P_cls.shape[1]):
# If class name is 'bg', continue
if np.argmax(P_cls[0, ii, :]) == (P_cls.shape[2] - 1):
continue
# Get class name
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 /= 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([16 * x, 16 * y, 16 * (x + w), 16 * (y + h)])
probs[cls_name].append(np.max(P_cls[0, ii, :]))
all_dets = []
for key in bboxes:
bbox = np.array(bboxes[key])
# Apply non-max-suppression on final bboxes to get the output bounding boxe
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, :]
det = {'x1': x1, 'x2': x2, 'y1': y1, 'y2': y2, 'class': key, 'prob': new_probs[jk]}
all_dets.append(det)
print('Elapsed time = {}'.format(time.time() - st))
t, p = get_map(all_dets, img_data['bboxes'], (fx, fy))
for key in t.keys():
if key not in T:
T[key] = []
P[key] = []
T[key].extend(t[key])
P[key].extend(p[key])
all_aps = []
for key in T.keys():
ap = average_precision_score(T[key], P[key])
print('{} AP: {}'.format(key, ap))
all_aps.append(ap)
print('mAP = {}'.format(np.mean(np.array(all_aps))))
mAPs.append(np.mean(np.array(all_aps)))
# print(T)
# print(P)
print()
print('mean average precision:', np.mean(np.array(mAPs)))
mAP = [mAP for mAP in mAPs if str(mAP) != 'nan']
mean_average_prec = round(np.mean(np.array(mAP)), 3)
print('After training %dk batches, the mean average precision is %0.3f' % (length, mean_average_prec))