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.')
# data generator
X, Y, img_data = next(data_gen_train)
# print("加载一个batch的数据")
# print(X,Y,img_data)
loss_rpn = model_rpn.train_on_batch(X, Y)
write_log(callback, ['rpn_cls_loss', 'rpn_reg_loss'], loss_rpn, train_step)
P_rpn = model_rpn.predict_on_batch(X)
R = roi_helpers.rpn_to_roi(P_rpn[0], P_rpn[1], C, 'tf', 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
# print(R)
X2, Y1, Y2, IouS = roi_helpers.calc_iou(R, img_data, C, class_mapping)
# print("debug1")
# print(X2,Y1,Y2,IouS)
if X2 is None:
rpn_accuracy_rpn_monitor.append(0)
rpn_accuracy_for_epoch.append(0)
continue
# sampling positive/negative samples
neg_samples = np.where(Y1[0, :, -1] == 1)
pos_samples = np.where(Y1[0, :, -1] == 0)
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
def test_view_func_NN(model_classifier, model_rpn, model_inner, C):
test_cls = 'aeroplane'
input_train_file = 'pickle_data/train_data_Wflip_all.pickle'
## read the training data from pickle file or from annotations
test_pickle = 'pickle_data/test_data_{}.pickle'.format(test_cls)
if os.path.exists(test_pickle):
with open(test_pickle) as f:
all_imgs, classes_count, _ = pickle.load(f)
class_mapping = C.class_mapping
inv_class_mapping = {v: k for k, v in class_mapping.iteritems()}
backend = K.image_dim_ordering()
gt_cls_num = class_mapping[test_cls]
print('work on class {}'.format(test_cls))
base_path = os.getcwd()
# turn off any data augmentation at test time
C.use_horizontal_flips = False
C.use_vertical_flips = False
C.rot_90 = False
count = 0
good_img = 0
not_good = 0
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)
img = cv2.resize(img, (new_width, new_height),
interpolation=cv2.INTER_CUBIC)
return img, ratio
def format_img_channels(img, C):
""" formats the image channels based on config """
img = img[:, :, (2, 1, 0)]
img = img.astype(np.float32)
img[:, :, 0] -= C.img_channel_mean[0]
img[:, :, 1] -= C.img_channel_mean[1]
img[:, :, 2] -= C.img_channel_mean[2]
img /= C.img_scaling_factor
img = np.transpose(img, (2, 0, 1))
img = np.expand_dims(img, axis=0)
return img
def format_img(img, C):
""" formats an image for model prediction based on config """
img, ratio = format_img_size(img, C)
img = format_img_channels(img, C)
return img, ratio
def display_image(img):
img1 = img[:, :, (2, 1, 0)]
# img1=img
im = Image.fromarray(img1.astype('uint8'), 'RGB')
im.show()
# Method to transform the coordinates of the bounding box to its original size
def get_real_coordinates(ratio, x1, y1, x2, y2):
## read the training data from pickle file or from annotations
real_x1 = int(round(x1 // ratio))
real_y1 = int(round(y1 // ratio))
real_x2 = int(round(x2 // ratio))
real_y2 = int(round(y2 // ratio))
return (real_x1, real_y1, real_x2, real_y2)
vnum_test = 24
azimuth_vec = np.concatenate(
([0],
np.linspace((360. / (vnum_test * 2)), 360. -
(360. / (vnum_test * 2)), vnum_test)),
axis=0)
def find_interval(azimuth, azimuth_vec):
for i in range(len(azimuth_vec)):
if azimuth < azimuth_vec[i]:
break
ind = i
if azimuth > azimuth_vec[-1]:
ind = 1
return ind
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)
class_to_color = {
class_mapping[v]: np.random.randint(0, 255, 3)
for v in class_mapping
}
C.num_rois = 32
obj_num = 0
bbox_threshold_orig = 0.6
th_bbox = 0.4
## get GT for all az for single cls
feature_az = []
sorted_path = input_train_file
tmp_ind = sorted_path.index('.pickle')
sorted_path = sorted_path[:tmp_ind] + "_sorted_Angles" + sorted_path[
tmp_ind:]
if os.path.exists(sorted_path):
print("loading sorted data")
with open(sorted_path) as f:
trip_data = pickle.load(f)
im_file = []
ind = []
for ii in range(360):
for jj in range(3):
try:
im_file.append(trip_data[test_cls][ii][jj])
ind.append(ii)
except:
if jj == 0:
print('no azimuth {}'.format(ii))
data_gen_train = data_generators.get_anchor_gt(im_file, [],
C,
K.image_dim_ordering(),
mode='test')
azimuth_dict = []
inner_NN = []
azimuths = []
for tt in range(len(ind)):
try:
if tt % 100 == 0:
print('worked on {}/{}'.format(tt, len(ind)))
# print ('im num {}'.format(good_img))
X, Y, img_data = next(data_gen_train)
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, Y_view = roi_helpers.calc_iou_new(
R, img_data, C, C.class_mapping)
pos_samples = np.where(Y1[0, :, -1] == 0)
sel_samples = pos_samples[0].tolist()
R = X2[0, sel_samples, :]
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, P_view] = model_classifier.predict([X, ROIs])
iner_f = model_inner.predict([X, ROIs])
# oo = model_classifier_only.predict([F, ROIs])
for ii in range(len(sel_samples)):
if np.max(P_cls[0,
ii, :]) < bbox_threshold_orig or np.argmax(
P_cls[0,
ii, :]) == (P_cls.shape[2] - 1):
continue
## get class from the net
# cls_num = np.argmax(P_cls[0, ii, :])
## use gt class
cls_num = gt_cls_num
cls_name = inv_class_mapping[cls_num]
cls_view = P_view[0, ii, 360 * cls_num:360 * (cls_num + 1)]
# azimuths[cls_name].append(np.argmax(cls_view, axis=0))
inner_NN.append(iner_f[0, ii, :])
azimuth_dict.append(img_data['bboxes'][0]['azimuth'])
except:
print('failed on az {}'.format(img_data['bboxes'][0]['azimuth']))
## calculating some mean feature map for every az
with open('pickle_data/{}_NN.pickle'.format(C.weight_name), 'w') as f:
pickle.dump([inner_NN, azimuth_dict], f)
print('saved PICKLE')
with open('pickle_data/{}_NN.pickle'.format(C.weight_name)) as f:
inner_NN, azimuth_dict = pickle.load(f)
neigh = KNeighborsClassifier(n_neighbors=1)
neigh.fit(inner_NN, azimuth_dict)
jj = 0
for im_file in all_imgs:
jj += 1
if jj % 50 == 0:
print(jj)
filepath = im_file['filepath']
img = cv2.imread(filepath)
img_gt = np.copy(img)
if img is None:
not_good += 1
continue
else:
good_img += 1
# print ('im num {}'.format(good_img))
X, ratio = format_img(img, C)
if backend == 'tf':
X = np.transpose(X, (0, 2, 3, 1))
# get the feature maps and output from the RPN
Y1, Y2 = 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]
width, height = int(im_file["width"]), int(im_file["height"])
resized_width, resized_height = data_generators.get_new_img_size(
width, height, C.im_size)
# [_,_, F] = model_rpn.predict(X)
ROIs = []
## pass on all the labels in the image, some of them are not equal to test_cls
for bbox_gt in im_file['bboxes']:
no_bbox_flag = 1
bbox_threshold = bbox_threshold_orig
if not bbox_gt['class'] == test_cls:
continue
if bbox_gt[
'class'] == test_cls and bbox_threshold == bbox_threshold_orig:
obj_num += 1
while no_bbox_flag and bbox_threshold > th_bbox:
cls_gt = bbox_gt['class']
az_gt = bbox_gt['azimuth']
el_gt = bbox_gt['elevation']
t_gt = bbox_gt['tilt']
if len(ROIs) == 0:
# apply the spatial pyramid pooling to the proposed regions
bboxes = {}
probs = {}
azimuths = {}
inner_res = {}
# print ('obj num {}'.format(obj_num))
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,
P_view] = model_classifier.predict([X, ROIs])
inner_out = model_inner.predict([X, ROIs])
# oo = model_classifier_only.predict([F, ROIs])
for ii in range(P_cls.shape[1]):
if np.max(P_cls[
0, ii, :]) < bbox_threshold or np.argmax(
P_cls[0,
ii, :]) == (P_cls.shape[2] - 1):
continue
## get class from the net
# cls_num = np.argmax(P_cls[0, ii, :])
## use gt class
cls_num = gt_cls_num
cls_name = inv_class_mapping[cls_num]
cls_view = P_view[0, ii, 360 * cls_num:360 *
(cls_num + 1)]
if cls_name not in bboxes:
bboxes[cls_name] = []
probs[cls_name] = []
azimuths[cls_name] = []
inner_res[cls_name] = []
(x, y, w, h) = ROIs[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, :]))
azimuths[cls_name].append(
np.argmax(cls_view, axis=0))
inner_res[cls_name].append(inner_out[0, ii, :])
# cv2.rectangle(img_gt, (bbox_gt['x1'], bbox_gt['y1']), (bbox_gt['x2'], bbox_gt['y2']), (int(class_to_color[test_cls][0]), int(class_to_color[test_cls][1]), int(class_to_color[test_cls][2])), 2)
for key in bboxes:
# if 1:
if key == test_cls and bbox_gt['class'] == test_cls:
bbox = np.array(bboxes[key])
prob = np.array(probs[key])
azimuth = np.array(azimuths[key])
inner_result = np.array(inner_res[key])
# img = draw_bbox(img,bbox, prob, azimuth, ratio)
azimuth = neigh.predict(inner_result)
## get the azimuth from bbox that have more than 'overlap_thresh' overlap with gt_bbox
az = []
overlap_thresh = 0.5
try:
while np.size(az) == 0 and overlap_thresh > 0:
_, prob_bbox, az = roi_helpers.overlap_with_gt(
bbox,
prob,
azimuth,
bbox_gt,
ratio=ratio,
overlap_thresh=overlap_thresh,
max_boxes=300,
use_az=True)
overlap_thresh -= 0.1
if overlap_thresh == 0:
print("No good Bbox was found")
counts = np.bincount(az)
except:
az = []
counts = []
try:
az_fin = np.argmax(counts)
true_bin = find_interval(az_gt, azimuth_vec)
prob_bin = find_interval(az_fin, azimuth_vec)
no_bbox_flag = 0
if true_bin == prob_bin:
count += 1
break
except:
# print('here')
no_bbox_flag = 1
bbox_threshold -= 0.1
## azimuth calculations
## display
bbox_threshold -= 0.1
succ = float(count) / float(obj_num) * 100.
print(
'for class {} -true count is {} out of {} from {} images . {} success'.
format(test_cls, count, obj_num, good_img, succ))
return succ
def Test_frcnn(test_path,
config_filename,
num_rois=32,
network="vgg",
terminal_flag=False):
"""
Test the object detection network
test_path --str: Full Path to the folder containing the test images (No default)
config_filename --str: Full path to the config_file.pickle, generated while training (No default)
num_rois --int: number of ROIs to process at once (Default 32)
network --str: The base network to use (One of 'vgg','resnet50') (Default 'vgg')
terminal_flag --bool: Flag to test if accessing from terminal do not pass anything to it while calling this function
OUTPUT:
When the script is called from terminal the images are displayed using opencv (images are in BGR format)
When called as a function returns the images, dets as 2 lists (images are in RGB format)
"""
config_output_filename = config_filename
with open(config_output_filename, 'rb') as f_in:
C = pickle.load(f_in)
if network == 'resnet50':
import keras_frcnn.resnet as nn
elif network == 'vgg':
import keras_frcnn.vgg as nn
elif network == "mobilenet":
import keras_frcnn.mobilenet as nn
C.model_path = 'epoch-176.hdf5'
# 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
def format_img_size(img, C): # utility function 1
""" 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)
img = cv2.resize(img, (new_width, new_height),
interpolation=cv2.INTER_CUBIC)
return img, ratio
def format_img_channels(img, C): #utility function 2
""" formats the image channels based on config """
img = img[:, :, (2, 1, 0)]
img = img.astype(np.float32)
img[:, :, 0] -= C.img_channel_mean[0]
img[:, :, 1] -= C.img_channel_mean[1]
img[:, :, 2] -= C.img_channel_mean[2]
img /= C.img_scaling_factor
img = np.transpose(img, (2, 0, 1))
img = np.expand_dims(img, axis=0)
return img
def format_img(img, C): # utility function 3
""" formats an image for model prediction based on config """
img, ratio = format_img_size(img, C)
img = format_img_channels(img, C)
return img, ratio
# Method to transform the coordinates of the bounding box to its original size
def get_real_coordinates(ratio, x1, y1, x2, y2): #utility function 4
real_x1 = int(round(x1 // ratio))
real_y1 = int(round(y1 // ratio))
real_x2 = int(round(x2 // ratio))
real_y2 = int(round(y2 // ratio))
return (real_x1, real_y1, real_x2, real_y2)
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)
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
elif C.network == 'mobilenet':
num_features = 512
if K.image_dim_ordering() == 'th':
input_shape_img = (3, None, None)
input_shape_features = (num_features, None, None)
else:
input_shape_img = (None, None, 3)
input_shape_features = (None, None, num_features)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(C.num_rois, 4))
feature_map_input = Input(shape=input_shape_features)
# define the base network (resnet here, can be VGG, Inception, etc)
shared_layers = nn.nn_base(img_input, trainable=True)
# define the RPN, built on the base layers
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn_layers = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(feature_map_input,
roi_input,
C.num_rois,
nb_classes=len(class_mapping),
trainable=True)
model_rpn = Model(img_input, rpn_layers)
model_classifier_only = Model([feature_map_input, roi_input], classifier)
model_classifier = Model([feature_map_input, roi_input], classifier)
print('Loading weights from {}'.format(C.model_path))
model_rpn.load_weights(C.model_path, by_name=True)
model_classifier.load_weights(C.model_path, by_name=True)
model_rpn.compile(optimizer='sgd', loss='mse')
model_classifier.compile(optimizer='sgd', loss='mse')
bbox_threshold = 0.8
list_of_all_images = []
list_of_all_dets = []
for idx, img_name in enumerate(sorted(os.listdir(img_path))):
if not img_name.lower().endswith(
('.bmp', '.jpeg', '.jpg', '.png', '.tif', '.tiff')):
continue
print(img_name)
st = time.time()
filepath = os.path.join(img_path, img_name)
img = cv2.imread(filepath)
X, ratio = 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.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, :])]
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=0.5)
for jk in range(new_boxes.shape[0]):
(x1, y1, x2, y2) = new_boxes[jk, :]
(real_x1, real_y1, real_x2,
real_y2) = get_real_coordinates(ratio, x1, y1, x2, y2)
cv2.rectangle(
img, (real_x1, real_y1), (real_x2, real_y2),
(int(class_to_color[key][0]), int(class_to_color[key][1]),
int(class_to_color[key][2])), 2)
textLabel = '{}: {}'.format(key, int(100 * new_probs[jk]))
all_dets.append((key, 100 * new_probs[jk]))
(retval, baseLine) = cv2.getTextSize(textLabel,
cv2.FONT_HERSHEY_COMPLEX,
1, 1)
textOrg = (real_x1, real_y1 - 0)
cv2.rectangle(
img, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5),
(0, 0, 0), 2)
cv2.rectangle(
img, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5),
(255, 255, 255), -1)
cv2.putText(img, textLabel, textOrg, cv2.FONT_HERSHEY_DUPLEX,
1, (0, 0, 0), 1)
if terminal_flag:
print('Elapsed time = {}'.format(time.time() - st))
print(all_dets)
if len(all_dets) > 0:
cv2.imwrite(img_name + '_new.png', img)
cv2.waitKey(0)
else:
list_of_all_images.append(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
list_of_all_dets.append(all_dets)
if not terminal_flag:
return (list_of_all_images, list_of_all_dets)
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)
R = roi_helpers.rpn_to_roi(P_rpn[0],
P_rpn[1],
C,
K.image_data_format(),
use_regr=True,
overlap_thresh=0.4,
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(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)
st = 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)
# 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:
def work(input, output, textlabel, piclabel, primpiclabel):
textlabel.append("Detecting now：")
test_path = input + '/'
output_path = output + '/'
#test_path = "../big_test_input/"
sys.setrecursionlimit(40000)
parser = OptionParser()
parser.add_option("-p",
"--path",
dest="test_path",
help="Path to test data.",
default=test_path)
parser.add_option(
"-n",
"--num_rois",
type="int",
dest="num_rois",
help="Number of ROIs per iteration. Higher means more memory use.",
default=32)
parser.add_option(
"--config_filename",
dest="config_filename",
help=
"Location to read the metadata related to the training (generated when training).",
default="./config.pickle")
parser.add_option("--network",
dest="network",
help="Base network to use. Supports vgg or resnet50.",
default='resnet50')
(options, args) = parser.parse_args()
if not options.test_path: # if filename is not given
parser.error(
'Error: path to test data must be specified. Pass --path to command line'
)
config_output_filename = options.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 = options.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()}
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)
if C.network == 'resnet50':
num_features = 1024
elif C.network == 'vgg':
num_features = 512
if K.image_dim_ordering() == 'th':
input_shape_img = (3, None, None)
input_shape_features = (num_features, None, None)
else:
input_shape_img = (None, None, 3)
input_shape_features = (None, None, num_features)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(C.num_rois, 4))
feature_map_input = Input(shape=input_shape_features)
# define the base network (resnet here, can be VGG, Inception, etc)
shared_layers = nn.nn_base(img_input, trainable=True)
# define the RPN, built on the base layers
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn_layers = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(feature_map_input,
roi_input,
C.num_rois,
nb_classes=len(class_mapping),
trainable=True)
model_rpn = Model(img_input, rpn_layers)
model_classifier_only = Model([feature_map_input, roi_input], classifier)
model_classifier = Model([feature_map_input, roi_input], classifier)
print('Loading weights from {}'.format(C.model_path))
textlabel.append('Loading weights from {}'.format(C.model_path))
########################这两句以后要加上#################
#model_rpn.load_weights(C.model_path, by_name=True)
#model_classifier.load_weights(C.model_path, by_name=True)
#########################################################
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 = {}
bbox_threshold = 0.8
visualise = True
for idx, img_name in enumerate(sorted(os.listdir(img_path))):
if not img_name.lower().endswith(
('.bmp', '.jpeg', '.jpg', '.png', '.tif', '.tiff')):
continue
print(img_name)
##################更新窗口################
textlabel.append(img_name)
oldpath = test_path + img_name
beforeimage = QtGui.QPixmap(oldpath)
primpiclabel.setPixmap(beforeimage)
primpiclabel.setScaledContents(True)
##########################################
st = time.time()
filepath = os.path.join(img_path, img_name)
img = cv2.imread(filepath)
X, ratio = 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.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, :])]
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=0.5)
for jk in range(new_boxes.shape[0]):
(x1, y1, x2, y2) = new_boxes[jk, :]
(real_x1, real_y1, real_x2,
real_y2) = get_real_coordinates(ratio, x1, y1, x2, y2)
color = [0, 0, 0]
if key == "airbase": color = [0, 0, 255]
if key == "harbour": color = [21, 159, 235]
if key == "island": color = [59, 197, 184]
cv2.rectangle(img, (real_x1, real_y1), (real_x2, real_y2),
color, 2)
#textLabel = '{}: {}'.format(key,int(100*new_probs[jk]))
#all_dets.append((key,100*new_probs[jk]))
#textlabel.append('{}: {}'.format(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, 0.6, (0, 0, 0), 1)
print('Elapsed time = {}'.format(time.time() - st))
textlabel.append('Elapsed time = {}'.format(time.time() - st))
print(all_dets)
#cv2.imshow('img', img)
#cv2.waitKey(0)
cv2.imwrite(output_path + '{}.png'.format(idx), img)
afterimage = QtGui.QPixmap(output_path + '{}.png'.format(idx))
piclabel.setPixmap(afterimage)
piclabel.setScaledContents(True)
def testing(test_path,
num_rois=32,
config_filename="config.pickle",
network="resnet50",
weigths_file='./model_frcnn.hdf5',
output_folder='./results_imgs/'):
'''
test_path = Path to testing data
num_rois = Number of ROIs per iteration. Higher means more memory use
config_filename = Location to read the metadata related to the training (generated when training)
network = Base network to use. Supports vgg or resnet50
weigths_file = Select the path and location of .hdf5 file
output_folder = Select the output folder
'''
if not test_path: # if filename is not given
parser.error(
'Error: path to test data must be specified. Pass --path to command line'
)
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
C.model_path = weigths_file
# 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
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)
img = cv2.resize(img, (new_width, new_height),
interpolation=cv2.INTER_CUBIC)
return img, ratio
def format_img_channels(img, C):
""" formats the image channels based on config """
img = img[:, :, (2, 1, 0)]
img = img.astype(np.float32)
img[:, :, 0] -= C.img_channel_mean[0]
img[:, :, 1] -= C.img_channel_mean[1]
img[:, :, 2] -= C.img_channel_mean[2]
img /= C.img_scaling_factor
img = np.transpose(img, (2, 0, 1))
img = np.expand_dims(img, axis=0)
return img
def format_img(img, C):
""" formats an image for model prediction based on config """
img, ratio = format_img_size(img, C)
img = format_img_channels(img, C)
return img, ratio
# Method to transform the coordinates of the bounding box to its original size
def get_real_coordinates(ratio, x1, y1, x2, y2):
real_x1 = int(round(x1 // ratio))
real_y1 = int(round(y1 // ratio))
real_x2 = int(round(x2 // ratio))
real_y2 = int(round(y2 // ratio))
return (real_x1, real_y1, real_x2, real_y2)
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)
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_dim_ordering() == 'th':
input_shape_img = (3, None, None)
input_shape_features = (num_features, None, None)
else:
input_shape_img = (None, None, 3)
input_shape_features = (None, None, num_features)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(C.num_rois, 4))
feature_map_input = Input(shape=input_shape_features)
# define the base network (resnet here, can be VGG, Inception, etc)
shared_layers = nn.nn_base(img_input, trainable=True)
# define the RPN, built on the base layers
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn_layers = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(feature_map_input,
roi_input,
C.num_rois,
nb_classes=len(class_mapping),
trainable=True)
model_rpn = Model(img_input, rpn_layers)
model_classifier_only = Model([feature_map_input, roi_input], classifier)
model_classifier = Model([feature_map_input, roi_input], classifier)
print('Loading weights from {}'.format(C.model_path))
model_rpn.load_weights(C.model_path, by_name=True)
model_classifier.load_weights(C.model_path, by_name=True)
model_rpn.compile(optimizer='sgd', loss='mse')
model_classifier.compile(optimizer='sgd', loss='mse')
all_imgs = []
classes = {}
bbox_threshold = 0.8
visualise = True
for idx, img_name in enumerate(sorted(os.listdir(img_path))):
if not img_name.lower().endswith(
('.bmp', '.jpeg', '.jpg', '.png', '.tif', '.tiff')):
continue
print(img_name)
st = time.time()
filepath = os.path.join(img_path, img_name)
img = cv2.imread(filepath)
X, ratio = 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.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, :])]
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=0.5)
for jk in range(new_boxes.shape[0]):
(x1, y1, x2, y2) = new_boxes[jk, :]
(real_x1, real_y1, real_x2,
real_y2) = get_real_coordinates(ratio, x1, y1, x2, y2)
cv2.rectangle(
img, (real_x1, real_y1), (real_x2, real_y2),
(int(class_to_color[key][0]), int(class_to_color[key][1]),
int(class_to_color[key][2])), 2)
textLabel = '{}: {}'.format(key, int(100 * new_probs[jk]))
all_dets.append((key, 100 * new_probs[jk]))
(retval, baseLine) = cv2.getTextSize(textLabel,
cv2.FONT_HERSHEY_COMPLEX,
1, 1)
textOrg = (real_x1, real_y1 - 0)
cv2.rectangle(
img, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5),
(0, 0, 0), 2)
cv2.rectangle(
img, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5),
(255, 255, 255), -1)
cv2.putText(img, textLabel, textOrg, cv2.FONT_HERSHEY_DUPLEX,
1, (0, 0, 0), 1)
print('Elapsed time = {}'.format(time.time() - st))
print(all_dets)
#cv2.imshow('img', img)
#cv2.waitKey(0)
output_file = output_folder + '{}.png'
cv2.imwrite(output_file.format(idx), img)
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.')
def train_model(dataset_directory: str,
model_name: str,
delete_and_recreate_dataset_directory: bool,
configuration_name: str,
output_weight_path: str,
configuration_filename: str,
number_of_epochs: int,
early_stopping: int,
learning_rate_reduction_patience: int,
learning_rate_reduction_factor: float,
non_max_suppression_overlap_threshold: float,
non_max_suppression_max_boxes: int,
input_weight_path: str = None):
muscima_pp_raw_dataset_directory = os.path.join(dataset_directory,
"muscima_pp_raw")
muscima_image_directory = os.path.join(dataset_directory,
"cvcmuscima_staff_removal")
muscima_cropped_directory = os.path.join(dataset_directory,
"muscima_pp_cropped_images")
if not dataset_directory: # if filename is not given
parser.error(
'Error: path to training data must be specified. Pass --path to command line'
)
network = NetworkFactory.get_network_by_name(model_name)
try:
all_images, classes_count, class_mapping = get_data(
muscima_cropped_directory)
data_loaded = True
except:
print(
"Could not load dataset. Automatically downloading and recreating dataset."
)
data_loaded = False
delete_and_recreate_dataset_directory = True
if delete_and_recreate_dataset_directory:
print("Deleting dataset directory {0}".format(dataset_directory))
if os.path.exists(dataset_directory):
shutil.rmtree(dataset_directory)
downloader = MuscimaPlusPlusDatasetDownloader(
muscima_pp_raw_dataset_directory)
downloader.download_and_extract_dataset()
downloader = CvcMuscimaDatasetDownloader(
muscima_image_directory, CvcMuscimaDataset.StaffRemoval)
downloader.download_and_extract_dataset()
delete_unused_images(muscima_image_directory)
inverter = ImageInverter()
# We would like to work with black-on-white images instead of white-on-black images
inverter.invert_images(muscima_image_directory, "*.png")
shutil.copy("Staff-Vertical-Positions.txt", dataset_directory)
cut_images(
muscima_image_directory,
os.path.join(dataset_directory, "Staff-Vertical-Positions.txt"),
muscima_cropped_directory, muscima_pp_raw_dataset_directory)
# pass the settings from the command line, and persist them in the config object
C = ConfigurationFactory.get_configuration_by_name(configuration_name)
C.model_path = output_weight_path
start_time = time.time()
if not data_loaded:
all_images, classes_count, class_mapping = get_data(
muscima_cropped_directory)
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)))
print(
'Hyperparameters: {0} RoIs generated per run with {1} boxes remaining from non-max suppression and using '
'non-max suppression threshold of {2:.2f}'.format(
C.num_rois, non_max_suppression_max_boxes,
non_max_suppression_overlap_threshold))
config_output_filename = configuration_filename
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.seed(1)
random.shuffle(all_images)
# num_imgs = len(all_images)
train_imgs = [s for s in all_images if s['imageset'] == 'train']
val_imgs = [s for s in all_images if s['imageset'] == 'val']
print('Num train samples {}'.format(len(train_imgs)))
print('Num val samples {}'.format(len(val_imgs)))
if not use_fast_data_generators:
print("Using standard data_generator")
data_gen_train = data_generators.get_anchor_gt(
train_imgs,
classes_count,
C,
network.get_img_output_length,
mode='train')
data_gen_val = data_generators.get_anchor_gt(
val_imgs,
classes_count,
C,
network.get_img_output_length,
mode='val')
else:
print("Using fast data_generator")
data_gen_train = data_generators_fast.get_anchor_gt(
train_imgs,
classes_count,
C,
network.get_img_output_length,
mode='train')
data_gen_val = data_generators_fast.get_anchor_gt(
val_imgs,
classes_count,
C,
network.get_img_output_length,
mode='val')
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 = network.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 = network.rpn(shared_layers, num_anchors)
classifier = network.classifier(shared_layers,
roi_input,
C.num_rois,
nb_classes=len(classes_count),
trainable=True)
model_rpn = Model(img_input, rpn[:2])
model_classifier = Model([img_input, roi_input], classifier)
# this is a model that holds both the RPN and the classifier, used to load/save weights for the models
model_all = Model([img_input, roi_input], rpn[:2] + classifier)
start_of_training = datetime.date.today()
tensorboard_callback = TensorBoard(log_dir="./logs/{0}_{1}/".format(
start_of_training, configuration_name))
tensorboard_callback.set_model(model_all)
try:
print('Loading weights from {0}'.format(input_weight_path))
model_rpn.load_weights(input_weight_path, by_name=True)
model_classifier.load_weights(input_weight_path, 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 = Adadelta()
optimizer_classifier = Adadelta()
model_rpn.compile(optimizer=optimizer,
loss=[
faster_rcnn_losses.rpn_loss_cls(num_anchors),
faster_rcnn_losses.rpn_loss_regr(num_anchors)
])
model_classifier.compile(
optimizer=optimizer_classifier,
loss=[
faster_rcnn_losses.class_loss_cls,
faster_rcnn_losses.class_loss_regr(len(classes_count) - 1)
],
metrics={'dense_class_{}'.format(len(classes_count)): 'accuracy'})
model_all.compile(optimizer=Adadelta(), loss='mae')
epoch_length = 1000
validation_epoch_length = len(val_imgs)
validation_interval = 1
losses = np.zeros((epoch_length, 5))
losses_val = np.zeros((validation_epoch_length, 5))
rpn_accuracy_rpn_monitor = []
rpn_accuracy_for_epoch = []
best_loss_training = np.inf
best_loss_epoch = 0
best_total_loss_validation = np.Inf
best_loss_rpn_cls = np.inf
best_loss_rpn_regr = np.inf
best_loss_class_cls = np.inf
best_loss_class_regr = np.inf
best_class_acc = 0.0
model_classifier.summary()
print(C.summary())
print('Starting training')
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'
]
epochs_without_improvement = 0
for epoch_num in range(number_of_epochs):
progbar = generic_utils.Progbar(epoch_length)
print('Epoch {}/{}'.format(epoch_num + 1, number_of_epochs))
for iter_num in range(epoch_length):
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(
'\nAverage 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)
R = roi_helpers.rpn_to_roi(
P_rpn[0],
P_rpn[1],
C,
use_regr=True,
overlap_thresh=non_max_suppression_overlap_threshold,
max_boxes=non_max_suppression_max_boxes)
# 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
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]
progbar.update(
iter_num + 1,
[('rpn_cls', np.mean(losses[:iter_num + 1, 0])),
('rpn_regr', np.mean(losses[:iter_num + 1, 1])),
('detector_cls', np.mean(losses[:iter_num + 1, 2])),
('detector_regr', np.mean(losses[:iter_num + 1, 3]))])
except Exception as e:
print('Exception during training: {}'.format(e))
continue
# Calculate losses after the specified number of iterations
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('[INFO TRAINING]')
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))
print("Best loss for training: {0:.3f}".format(best_loss_training))
curr_total_loss = loss_rpn_cls + loss_rpn_regr + loss_class_cls + loss_class_regr
val_start_time = time.time()
write_log(tensorboard_callback, train_names, [
loss_rpn_cls, loss_rpn_regr, loss_class_cls, loss_class_regr,
curr_total_loss, class_acc
], epoch_num)
if curr_total_loss < best_loss_training:
model_path = C.model_path[:-5] + "_training.hdf5"
if C.verbose:
print(
'Total training loss decreased from {0:.3f} to {1:.3f}, saving weights to {2}'
.format(best_loss_training, curr_total_loss, model_path))
best_loss_training = curr_total_loss
model_all.save_weights(model_path)
#############
# VALIDATION
#############
if (epoch_num + 1) % validation_interval != 0:
continue
progbar = generic_utils.Progbar(validation_epoch_length)
for iter_num in range(validation_epoch_length):
try:
X, Y, img_data = next(data_gen_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,
use_regr=True,
overlap_thresh=non_max_suppression_overlap_threshold,
max_boxes=non_max_suppression_max_boxes)
# 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)
loss_class = model_classifier.test_on_batch(
[X, X2[:, sel_samples, :]],
[Y1[:, sel_samples, :], Y2[:, sel_samples, :]])
losses_val[iter_num, 0] = loss_rpn[1]
losses_val[iter_num, 1] = loss_rpn[2]
losses_val[iter_num, 2] = loss_class[1]
losses_val[iter_num, 3] = loss_class[2]
losses_val[iter_num, 4] = loss_class[3]
progbar.update(
iter_num + 1,
[('rpn_cls', np.mean(losses_val[:iter_num + 1, 0])),
('rpn_regr', np.mean(losses_val[:iter_num + 1, 1])),
('detector_cls', np.mean(losses_val[:iter_num + 1, 2])),
('detector_regr', np.mean(losses_val[:iter_num + 1, 3]))])
except Exception as e:
print('Exception during validation: {}'.format(e))
continue
# Computer aggregated losses
loss_rpn_cls = np.mean(losses_val[:, 0])
loss_rpn_regr = np.mean(losses_val[:, 1])
loss_class_cls = np.mean(losses_val[:, 2])
loss_class_regr = np.mean(losses_val[:, 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 = []
curr_total_loss = loss_rpn_cls + loss_rpn_regr + loss_class_cls + loss_class_regr
write_log(tensorboard_callback, val_names, [
loss_rpn_cls, loss_rpn_regr, loss_class_cls, loss_class_regr,
curr_total_loss, 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(
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(
"Current validation loss: {0:.3f}, Best validation loss: {1:.3f} at epoch: {2}"
.format(curr_total_loss, best_total_loss_validation,
best_loss_epoch))
print('Elapsed time: {}'.format(time.time() - val_start_time))
if curr_total_loss < best_total_loss_validation:
if C.verbose:
print(
'Total validation loss decreased from {0:.3f} to {1:.3f}, saving weights to {2}'
.format(best_total_loss_validation, curr_total_loss,
C.model_path))
best_total_loss_validation = curr_total_loss
best_loss_rpn_cls = loss_rpn_cls
best_loss_rpn_regr = loss_rpn_regr
best_loss_class_cls = loss_class_cls
best_loss_class_regr = loss_class_regr
best_class_acc = class_acc
best_loss_epoch = epoch_num
model_all.save_weights(C.model_path)
epochs_without_improvement = 0
else:
epochs_without_improvement += validation_interval
if epochs_without_improvement > early_stopping:
print(
"Early stopping training after {0} epochs without improvement on validation set"
.format(epochs_without_improvement))
break
if epochs_without_improvement > learning_rate_reduction_patience:
current_learning_rate = K.get_value(model_classifier.optimizer.lr)
new_learning_rate = current_learning_rate * learning_rate_reduction_factor
print(
"Not improved validation accuracy for {0} epochs. Reducing learning rate from {1} to {2}"
.format(learning_rate_reduction_patience,
current_learning_rate, new_learning_rate))
K.set_value(model_classifier.optimizer.lr, new_learning_rate)
K.set_value(model_rpn.optimizer.lr, new_learning_rate)
K.set_value(model_all.optimizer.lr, new_learning_rate)
end_time = time.time()
execution_time_in_seconds = round(end_time - start_time)
print("Execution time: {0:.1f}s".format(end_time - start_time))
notification_message = "Training on {0} dataset with model {1} and configuration {2} finished. " \
"Val. accuracy: {3:0.5f}%".format("muscima_pp",
model_name,
configuration_name,
best_class_acc * 100)
TelegramNotifier.send_message_via_telegram(notification_message)
today = "{0:02d}.{1:02d}.{2}".format(start_of_training.day,
start_of_training.month,
start_of_training.year)
GoogleSpreadsheetReporter.append_result_to_spreadsheet(
dataset_size=len(all_images),
model_name=model_name,
configuration_name=configuration_name,
data_augmentation="",
early_stopping=early_stopping,
reduction_patience=learning_rate_reduction_patience,
learning_rate_reduction_factor=learning_rate_reduction_factor,
optimizer="Adadelta",
initial_learning_rate=1.0,
non_max_suppression_overlap_threshold=
non_max_suppression_overlap_threshold,
non_max_suppression_max_boxes=non_max_suppression_max_boxes,
validation_accuracy=best_class_acc,
validation_total_loss=best_total_loss_validation,
best_loss_rpn_cls=best_loss_rpn_cls,
best_loss_rpn_regr=best_loss_rpn_regr,
best_loss_class_cls=best_loss_class_cls,
best_loss_class_regr=best_loss_class_regr,
date=today,
datasets="muscima_pp",
execution_time_in_seconds=execution_time_in_seconds)
'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)
R = roi_helpers.rpn_to_roi(
P_rpn[0],
P_rpn[1],
C,
K.
image_dim_ordering( #could try out different overlap_threshold
),
use_regr=True,
overlap_thresh=0.8,
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(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)
def main():
cleanup()
sys.setrecursionlimit(40000)
config_output_filename = 'keras_frcnn/config.pickle'
with open(config_output_filename, 'r') as f_in:
C = pickle.load(f_in)
# 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
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 = num_rois
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(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 = {}
bbox_threshold = 0.8
visualise = True
print("Converting video to images..")
convert_to_images()
print("anotating...")
list_files = sorted(get_file_names(img_path),
key=lambda var: [
int(x) if x.isdigit() else x
for x in re.findall(r'[^0-9]|[0-9]+', var)
])
for img_name in list_files:
if not img_name.lower().endswith(
('.bmp', '.jpeg', '.jpg', '.png', '.tif', '.tiff')):
continue
print(img_name)
st = time.time()
filepath = os.path.join(img_path, img_name)
img = cv2.imread(filepath)
X = format_img(img, C)
img_scaled = np.transpose(X.copy()[0, (2, 1, 0), :, :],
(1, 2, 0)).copy()
img_scaled[:, :, 0] += 123.68
img_scaled[:, :, 1] += 116.779
img_scaled[:, :, 2] += 103.939
img_scaled = img_scaled.astype(np.uint8)
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.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, :])]
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 = []
all_objects = []
people_count = 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, :]
cv2.rectangle(img_scaled, (x1, y1), (x2, y2),
class_to_color[key], 2)
textLabel = '{}: {}'.format(key, int(100 * new_probs[jk]))
all_dets.append((key, 100 * new_probs[jk]))
if key == "person":
people_count = people_count + 1
all_objects.append((key, 1))
(retval, baseLine) = cv2.getTextSize(textLabel,
cv2.FONT_HERSHEY_COMPLEX,
1, 1)
textOrg = (x1, y1 - 0)
cv2.rectangle(
img_scaled, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5),
(0, 0, 0), 2)
cv2.rectangle(
img_scaled, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5),
(255, 255, 255), -1)
cv2.putText(img_scaled, textLabel, textOrg,
cv2.FONT_HERSHEY_DUPLEX, 1, (0, 0, 0), 1)
print('Elapsed time = {}'.format(time.time() - st))
height, width, channels = img_scaled.shape
cv2.rectangle(img_scaled, (0, 0), (width, 30), (0, 0, 0), -1)
cv2.putText(img_scaled,
"Obj count: " + str(list(accumulate(all_objects))),
(5, 19), cv2.FONT_HERSHEY_TRIPLEX, 0.5, (255, 255, 255), 1)
cv2.imwrite(os.path.join(output_path, img_name), img_scaled)
print "Nr of people in this room: %d" % people_count
now = datetime.datetime.now()
f = open('people-count.log', 'a+')
f.write("Nr of people in this room: %d\n" % people_count)
f.close()
print(all_dets)
def predict_image(file_path):
global graph
print(file_path)
st = time.time()
img = cv2.imread(file_path)
image_name = os.path.split(file_path)[-1]
X = format_img(img, C)
img_scaled = np.transpose(X.copy()[0, (2, 1, 0), :, :], (1, 2, 0)).copy()
img_scaled[:, :, 0] += 123.68
img_scaled[:, :, 1] += 116.779
img_scaled[:, :, 2] += 103.939
img_scaled = img_scaled.astype(np.uint8)
if K.image_dim_ordering() == 'tf':
X = np.transpose(X, (0, 2, 3, 1))
with graph.as_default():
# 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.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, :])]
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=0.5)
for jk in range(new_boxes.shape[0]):
(x1, y1, x2, y2) = new_boxes[jk, :]
cv2.rectangle(img_scaled, (x1, y1), (x2, y2),
class_to_color[key], 2)
textLabel = '{}: {}'.format(key, int(100 * new_probs[jk]))
all_dets.append((key, 100 * new_probs[jk]))
(retval, baseLine) = cv2.getTextSize(textLabel,
cv2.FONT_HERSHEY_COMPLEX,
1, 1)
textOrg = (x1, y1 - 0)
cv2.rectangle(
img_scaled, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5),
(0, 0, 0), 2)
cv2.rectangle(
img_scaled, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5),
(255, 255, 255), -1)
cv2.putText(img_scaled, textLabel, textOrg,
cv2.FONT_HERSHEY_DUPLEX, 1, (0, 0, 0), 1)
print('Elapsed time = {}'.format(time.time() - st))
cv2.imwrite('./static/uploadImage/{}process.jpg'.format(image_name),
img_scaled)
print(all_dets)
print(img_name)
st = time.time()
filepath = os.path.join(img_path,img_name)
img = cv2.imread(filepath)
X, ratio = 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:
def map_main(config_output_filename,
img_path,
weights_filename,
num_rois=32,
parser='pascal_voc'):
if parser == 'pascal_voc':
from keras_frcnn.pascal_voc_parser import get_data
elif parser == 'simple':
from keras_frcnn.simple_parser import get_data
else:
raise ValueError(
"Command line option parser must be one of 'pascal_voc' or 'simple'"
)
with open(config_output_filename, 'r') as f_in:
C = pickle.load(f_in)
C.model_path = weights_filename
class_mapping = C.class_mapping
# turn off any data augmentation at test time
C.use_horizontal_flips = False
C.use_vertical_flips = False
C.rot_90 = False
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 = num_rois
input_shape_img = (None, None, 3)
input_shape_features = (None, None, 512)
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(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, _, _ = get_data(img_path)
test_imgs = [s for s in all_imgs if s['imageset'] == 'test']
T = {}
P = {}
final_map = None
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 = {}
else:
pass
st = time.time()
filepath = os.path.join(img_path, img_name)
img = cv2.imread(filepath)
X, ratio = format_img(img, C)
X = np.transpose(X, (0, 2, 3, 1))
# get the feature maps and output from the RPN
[Y1, Y2, F] = model_rpn.predict(X)
R = roi_helpers.rpn_to_roi(Y1, Y2, C, overlap_thresh=0.7, max_boxes=300)
# 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
def uploadtest(request):
if request.FILES:
str_info = ''
pic = request.FILES.get('pic')
sava_path = '%s/test/%s' % (settings.MEDIA_ROOT, pic.name)
with open(sava_path, 'wb') as f:
for content in pic.chunks():
f.write(content)
print('-->图片上传成功...')
str_info += '-->图片上传成功...\n'
# 以下是测试过程：
sys.setrecursionlimit(40000)
config_output_filename = os.path.join(settings.CONFIG_BISHE, 'bishe/config.pickle')
# print(config_output_filename)
print('-->正在检测...')
str_info += '-->正在检测...\n'
with open(config_output_filename, 'rb') as f_in:
C = pickle.load(f_in)
print('-->找到配置文件...')
C.model_path = os.path.join(settings.CONFIG_BISHE, C.model_path)
print('-->找到模型信息...')
str_info += '-->找到模型信息...\n'
print('-->模型路径地址：' + C.model_path)
str_info += '-->模型路径地址：' + C.model_path + '\n'
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
C.num_rois = 10
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)
img = cv2.resize(img, (new_width, new_height), interpolation=cv2.INTER_CUBIC)
return img, ratio
def format_img_channels(img, C):
""" formats the image channels based on config """
img = img[:, :, (2, 1, 0)]
img = img.astype(np.float32)
img[:, :, 0] -= C.img_channel_mean[0]
img[:, :, 1] -= C.img_channel_mean[1]
img[:, :, 2] -= C.img_channel_mean[2]
img /= C.img_scaling_factor
img = np.transpose(img, (2, 0, 1))
img = np.expand_dims(img, axis=0)
return img
def format_img(img, C):
""" formats an image for model prediction based on config """
img, ratio = format_img_size(img, C)
img = format_img_channels(img, C)
return img, ratio
# Method to transform the coordinates of the bounding box to its original size
def get_real_coordinates(ratio, x1, y1, x2, y2):
real_x1 = int(round(x1 // ratio))
real_y1 = int(round(y1 // ratio))
real_x2 = int(round(x2 // ratio))
real_y2 = int(round(y2 // ratio))
return (real_x1, real_y1, real_x2, real_y2)
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('-->乳腺癌病症种类：' + str(class_mapping))
str_info += '-->乳腺癌病症种类：' + str(class_mapping) + '\n'
class_to_color = {class_mapping[v]: np.random.randint(0, 255, 3) for v in class_mapping}
if C.network == 'resnet50':
num_features = 1024
elif C.network == 'vgg':
num_features = 512
if K.image_dim_ordering() == 'th':
input_shape_img = (3, None, None)
input_shape_features = (num_features, None, None)
else:
input_shape_img = (None, None, 3)
input_shape_features = (None, None, num_features)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(C.num_rois, 4))
feature_map_input = Input(shape=input_shape_features)
# define the base network (resnet here, can be VGG, Inception, etc)
shared_layers = nn.nn_base(img_input, trainable=True)
# define the RPN, built on the base layers
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn_layers = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(feature_map_input, roi_input, C.num_rois, nb_classes=len(class_mapping),
trainable=True)
model_rpn = Model(img_input, rpn_layers)
model_classifier_only = Model([feature_map_input, roi_input], classifier)
model_classifier = Model([feature_map_input, roi_input], classifier)
print('-->由 {} 加载权重信息...'.format(C.model_path))
str_info += '-->由 {} 加载权重信息...'.format(C.model_path) + '\n'
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 = {}
bbox_threshold = 0.8
st = time.time()
filepath = sava_path
img = cv2.imread(filepath)
X, ratio = 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.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, :])]
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=0.5)
for jk in range(new_boxes.shape[0]):
(x1, y1, x2, y2) = new_boxes[jk, :]
(real_x1, real_y1, real_x2, real_y2) = get_real_coordinates(ratio, x1, y1, x2, y2)
cv2.rectangle(img, (real_x1, real_y1), (real_x2, real_y2), (
int(class_to_color[key][0]), int(class_to_color[key][1]), int(class_to_color[key][2])), 2)
textLabel = '{}: {}'.format(key, int(100 * new_probs[jk]))
all_dets.append((key, 100 * new_probs[jk]))
(retval, baseLine) = cv2.getTextSize(textLabel, cv2.FONT_HERSHEY_COMPLEX, 1, 1)
textOrg = (real_x1, real_y1 - 0)
cv2.rectangle(img, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5), (0, 0, 0), 2)
cv2.rectangle(img, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5), (255, 255, 255), -1)
cv2.putText(img, textLabel, textOrg, cv2.FONT_HERSHEY_DUPLEX, 1, (0, 0, 0), 1)
timeused = (time.time() - st)
print('-->检测完成，用时： {}...'.format(timeused))
str_info += '-->检测完成，用时： {}'.format(timeused) + '\n'
aaa = str(all_dets)
print('-->检测结果：' + aaa)
str_info += '-->检测结果：' + aaa + '\n'
result_path = '%s/result/%s' % (settings.MEDIA_ROOT, pic.name)
cv2.imwrite(result_path, img)
print(str_info)
# 将图片路径回传，为上传数据库做准备
test_pic = '/static/media/test/%s' % (pic.name)
result_pic = '/static/media/result/%s' % (pic.name)
user_id = UserInfo.objects.get(username=request.session.get('username')).id
return JsonResponse(
{'res': 1, 'result_pic': result_pic, 'test_pic': test_pic, 'user_id': user_id, 'str_info': str_info})
return JsonResponse({'res': 0})
def predict_single_image(img_path, model_rpn, model_classifier_only, cfg,
class_mapping):
st = time.time()
img = cv2.imread(img_path)
if img is None:
print('reading image failed.')
exit(0)
X, ratio = format_img(img, cfg)
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)
# this is result contains all boxes, which is [x1, y1, x2, y2]
result = roi_helpers.rpn_to_roi(Y1,
Y2,
cfg,
K.image_dim_ordering(),
overlap_thresh=0.7)
# convert from (x1,y1,x2,y2) to (x,y,w,h)
result[:, 2] -= result[:, 0]
result[:, 3] -= result[:, 1]
bbox_threshold = 0.8
# apply the spatial pyramid pooling to the proposed regions
boxes = dict()
for jk in range(result.shape[0] // cfg.num_rois + 1):
rois = np.expand_dims(result[cfg.num_rois * jk:cfg.num_rois *
(jk + 1), :],
axis=0)
if rois.shape[1] == 0:
break
if jk == result.shape[0] // cfg.num_rois:
# pad R
curr_shape = rois.shape
target_shape = (curr_shape[0], cfg.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])
import pdb
pdb.set_trace()
for ii in range(p_cls.shape[1]):
if 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):
# if predicted class was "bg", skip to append to boxes.
continue
cls_num = np.argmax(p_cls[0, ii, :])
if cls_num not in boxes.keys():
boxes[cls_num] = []
(x, y, w, h) = rois[0, ii, :]
try:
(tx, ty, tw, th) = p_regr[0, ii, 4 * cls_num:4 * (cls_num + 1)]
tx /= cfg.classifier_regr_std[0]
ty /= cfg.classifier_regr_std[1]
tw /= cfg.classifier_regr_std[2]
th /= cfg.classifier_regr_std[3]
x, y, w, h = roi_helpers.apply_regr(x, y, w, h, tx, ty, tw, th)
except Exception as e:
print(e)
pass
boxes[cls_num].append([
cfg.rpn_stride * x, cfg.rpn_stride * y,
cfg.rpn_stride * (x + w), cfg.rpn_stride * (y + h),
np.max(p_cls[0, ii, :])
])
# add some nms to reduce many boxes
for cls_num, box in boxes.items():
boxes_nms = roi_helpers.non_max_suppression_fast(box,
overlap_thresh=0.5)
boxes[cls_num] = boxes_nms
print(class_mapping[cls_num] + ":")
for b in boxes_nms:
b[0], b[1], b[2], b[3] = get_real_coordinates(
ratio, b[0], b[1], b[2], b[3])
print('{} prob: {}'.format(b[0:4], b[-1]))
img = draw_boxes_and_label_on_image_cv2(img, class_mapping, boxes)
print('Elapsed time = {}'.format(time.time() - st))
# cv2.imshow('image', img)
result_path = './results_images/{}.png'.format(
os.path.basename(img_path).split('.')[0])
print('result saved into ', result_path)
cv2.imwrite(result_path, img)
cv2.waitKey(0)
def predict(self, img_path):
all_imgs = []
classes = {}
bbox_threshold = 0.8
st = time.time()
img = cv2.imread(img_path)
X, ratio = format_img(img, self.C)
print(K.image_data_format())
if K.image_data_format() == 'tf' or 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] = self.model_rpn.predict(X)
R = roi_helpers.rpn_to_roi(Y1,
Y2,
self.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] // self.C.num_rois + 1):
ROIs = np.expand_dims(R[self.C.num_rois * jk:self.C.num_rois *
(jk + 1), :],
axis=0)
if ROIs.shape[1] == 0:
break
if jk == R.shape[0] // self.C.num_rois:
# pad R
curr_shape = ROIs.shape
target_shape = (curr_shape[0], self.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] = self.model_classifier_only.predict([F, ROIs])
for ii in range(P_cls.shape[1]):
if np.max(P_cls[0, ii, :]) < bbox_threshold or np.argmax(
P_cls[0, ii, :]) == (P_cls.shape[2] - 1):
continue
cls_name = self.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 /= self.C.classifier_regr_std[0]
ty /= self.C.classifier_regr_std[1]
tw /= self.C.classifier_regr_std[2]
th /= self.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([
self.C.rpn_stride * x, self.C.rpn_stride * y,
self.C.rpn_stride * (x + w), self.C.rpn_stride * (y + h)
])
probs[cls_name].append(np.max(P_cls[0, ii, :]))
all_dets = []
candidates = [] # 모든 객체 후보 영역이 담길 list
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=0.6)
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)
real_points = TwoPoint(real_x1 + 10, real_y1 + 10,
real_x2 - 10, real_y2 - 10)
candidates.append(real_points)
cv2.rectangle(img, (real_x1 + 10, real_y1 + 10),
(real_x2 - 10, real_y2 - 10), (255, 0, 0), 2)
# cv2.rectangle(img,(x1, y1), (x2, y2), (int(class_to_color[key][0]), int(class_to_color[key][1]), int(class_to_color[key][2])), 3)
textLabel = '{}: {}'.format(key, int(100 * new_probs[jk]))
all_dets.append((key, 100 * new_probs[jk]))
(retval, baseLine) = cv2.getTextSize(textLabel,
cv2.FONT_HERSHEY_COMPLEX,
1, 1)
textOrg = (real_x1, real_y1 - 0)
cv2.rectangle(img, (textOrg[0], textOrg[1]),
(textOrg[0] + retval[0], textOrg[1] - retval[1]),
(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)
min_x1 = 10
min_y1 = 10
max_x2 = 0
max_y2 = 0
for c in candidates: # 허용된 범위를 넘어가는 bbox에 대한 최대 좌표 수정
if c.x1 <= 0:
c.x1 = 10
if c.y1 <= 0:
c.y1 = 10
if c.x2 >= IMAGE_MAX_X:
c.x2 = IMAGE_MAX_X - 10
if c.y2 >= IMAGE_MAX_Y:
c.y2 = IMAGE_MAX_Y - 10
if min_x1 > c.x1: # 객체들의 영역을 하나의 영역으로 합치는 과정
min_x1 = c.x1
if min_y1 > c.y1:
min_y1 = c.y1
if max_x2 < c.x2:
max_x2 = c.x2
if max_y2 < c.y2:
max_y2 = c.y2 # 각 객체의 최소점, 최대점을 계산하여 하나로 합침
cv2.rectangle(img, (min_x1, min_y1), (max_x2, max_y2), (255, 0, 255),
2)
print('Elapsed time = {}'.format(time.time() - st))
print(all_dets)
cv2.imshow('img', img)
cv2.waitKey(0)
def run_prediction(config_filename,
model_path,
test_path,
out_path,
network='resnet50',
num_rois=32):
with open(config_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()}
#print(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_dim_ordering() == 'th':
input_shape_img = (3, None, None)
input_shape_features = (num_features, None, None)
else:
input_shape_img = (None, None, 3)
input_shape_features = (None, None, num_features)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(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)
C.model_path = model_path
#print('Loading weights from {}'.format(C.model_path))
model_rpn.load_weights(C.model_path, by_name=True)
model_classifier.load_weights(C.model_path, by_name=True)
model_rpn.compile(optimizer='sgd', loss='mse')
model_classifier.compile(optimizer='sgd', loss='mse')
all_imgs = []
classes = {}
bbox_threshold = 0.9 # default is 0.8
visualise = True
preresults = [] # add this
for idx, img_name in enumerate(sorted(os.listdir(img_path))):
if not img_name.lower().endswith(
('.bmp', '.jpeg', '.jpg', '.png', '.tif', '.tiff')):
continue
#print(img_name)
#st = time.time()
filepath = os.path.join(img_path, img_name)
img = cv2.imread(filepath)
X, ratio = 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) # default is 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.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, :])]
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 = []
bbox_results = []
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=0.4) # default is 0.5
for jk in range(new_boxes.shape[0]):
img_use = img.copy()
(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)
cv2.rectangle(img_use, (real_x1, real_y1), (real_x2, real_y2),
(0, 0, 255), 4)
#textLabel = '{}: {}'.format(key,int(100*new_probs[jk]))
all_dets.append((key, 100 * new_probs[jk]))
bbox_results.append(
(key, (real_x1, real_y1, real_x2, real_y2)))
#(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)
if not os.path.exists(out_path):
os.makedirs(out_path)
page_id = int(img_name.split('.')[0].split('_')[-1]) + 1
page_name = 'page_' + str(page_id)
cv2.imwrite(
os.path.join(
out_path,
'{}.jpg'.format(page_name + '_id_' + str(jk + 1))),
img_use) #format(idx)
#print('Elapsed time = {}'.format(time.time() - st))
#print(all_dets)
#print(bbox_results)
preresults.append([img_name, bbox_results]) ## add this
# cv2.imshow('img', img)
# cv2.waitKey(0)
#outpath = '/data/projects/table_parser/results/result_scanned/'
#if not os.path.exists(out_path):
#os.makedirs(out_path)
#cv2.imwrite(out_path + '{}.jpg'.format(img_name.split('.')[0]),img) #format(idx)
preresults = pd.DataFrame(preresults)
preresults.to_csv(os.path.join(out_path, 'preresults.txt'),
header=None,
index=None)
def detect_known_objects(self, img):
print("HELLLOOOOOO")
#img = self.image_resize(img, height=int(img.shape[0]/3.0))
#img_yuv = cv2.cvtColor(img, cv2.COLOR_BGR2YUV)
#img_yuv[:,:,0] = cv2.equalizeHist(img_yuv[:,:,0])
#img=cv2.cvtColor(img_yuv,cv2.COLOR_YUV2BGR)
X, ratio = self.format_img(img, self.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] = self.model_rpn.predict(X)
#print Y1, Y2, F
a = datetime.datetime.now()
R = roi_helpers.rpn_to_roi(Y1,
Y2,
self.C,
K.image_dim_ordering(),
overlap_thresh=0.7)
b = datetime.datetime.now()
delta = b - a
#print("roi_helpers.rpn_to_roi took:", int(delta.total_seconds() * 1000)) # milliseconds
#print R
#for i in R:
# cv2.rectangle(img,(i[0],i[1]),(i[2],i[3]),(0,255,0),3)
# 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 idx, jk in enumerate(range(R.shape[0] // self.C.num_rois + 1)):
ROIs = np.expand_dims(R[self.C.num_rois * jk:self.C.num_rois *
(jk + 1), :],
axis=0)
if ROIs.shape[1] == 0:
break
if jk == R.shape[0] // self.C.num_rois:
#pad R
curr_shape = ROIs.shape
target_shape = (curr_shape[0], self.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
#print("ROIs shape", np.array(ROIs).shape)
#print("F", np.array(F).shape)
a = datetime.datetime.now()
[P_cls, P_regr,
P_clust] = self.model_classifier_only.predict([F, ROIs])
b = datetime.datetime.now()
delta = b - a
#print("prediction of roi took: :", int(delta.total_seconds() * 1000)) # milliseconds
#print P_cls, P_regr
#print P_cls.shape, P_regr.shape
for ii in range(P_cls.shape[1]):
#print P_cls[0,ii,:]
if np.max(P_cls[0, ii, :]) < self.bbox_threshold or np.argmax(
P_cls[0, ii, :]) == (P_cls.shape[2] - 1):
continue
cls_name = self.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, :]
#print x, y, w, h
cls_num = np.argmax(P_cls[0, ii, :])
#print "something", cls_num
try:
(tx, ty, tw, th) = P_regr[0, ii,
4 * cls_num:4 * (cls_num + 1)]
tx /= self.C.classifier_regr_std[0]
ty /= self.C.classifier_regr_std[1]
tw /= self.C.classifier_regr_std[2]
th /= self.C.classifier_regr_std[3]
x, y, w, h = roi_helpers.apply_regr(
x, y, w, h, tx, ty, tw, th)
except:
print("exception")
pass
bboxes[cls_name].append([
self.C.rpn_stride * x, self.C.rpn_stride * y,
self.C.rpn_stride * (x + w), self.C.rpn_stride * (y + h)
])
probs[cls_name].append(np.max(P_cls[0, ii, :]))
detected_objects = []
for key in bboxes:
bbox = np.array(bboxes[key])
new_boxes, new_probs = roi_helpers.non_max_suppression_fast(
bbox, np.array(probs[key]), overlap_thresh=0.5)
for jk in range(new_boxes.shape[0]):
(x1, y1, x2, y2) = new_boxes[jk, :]
(real_x1, real_y1, real_x2,
real_y2) = self.get_real_coordinates(ratio, x1, y1, x2, y2)
#print "drawing detected rect at:", (real_x1, real_y1), (real_x2, real_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])),5)
#textLabel = '{}: {}'.format(key,int(100*new_probs[jk]))
#all_dets.append((key,100*new_probs[jk]))
#(retval,baseLine) = cv2.getTextSize(textLabel,cv2.FONT_HERSHEY_COMPLEX,1,1)
#textOrg = (real_x1-20, real_y1-20)
#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, 0.3, (0, 0, 0), 1)
height, width, channels = img.shape
#FOV horizontal = 62 degrees (from 90 on right to 33 on left)
angle_between_robot_centre_and_detected_object = self.angle_between(
(real_x1 + self.distance([real_x1, real_y1],
[real_x2, real_y1]) / 2.0,
(real_y1 + self.distance([real_x1, real_y1],
[real_x1, real_y2]) / 2.0)),
(width / 2.0, 0)) - 52.0
angle_between_robot_centre_and_detected_object = -angle_between_robot_centre_and_detected_object
angle_between_robot_centre_and_detected_object = calc_angle([
int((real_x1 + real_x2) / 2.0),
int((real_y1 + real_y2) / 2.0)
])
angle, distance = calculate_angle_and_distance(img,
real_x1,
real_x2,
real_y1,
real_y2,
obj_width=16)
angle_between_robot_centre_and_detected_object = angle
focal_length_mm = 1.0
average_real_object_height_mm = 1.0
image_height_px = height
object_height_px = self.distance([real_x1, real_y1],
[real_x1, real_y2])
sensor_height_mm = 314.2
distance_between_robot_centre_and_detected_object = (15.0 / (
(min(self.distance([real_x1, real_y1], [real_x2, real_y1]),
self.distance([real_x1, real_y1],
[real_x1, real_y2])) /
max(self.distance([real_x1, real_y1], [real_x2, real_y1]),
self.distance([real_x1, real_y1], [real_x1, real_y2]))
)) * 123) / self.distance([real_x1, real_y1],
[real_x2, real_y1])
distance_between_robot_centre_and_detected_object = distance_between_robot_centre_and_detected_object * 5.0
distance_between_robot_centre_and_detected_object = distance
detected_objects.append(
(key, "", real_x1, real_y1, real_x2, real_y2,
distance_between_robot_centre_and_detected_object,
angle_between_robot_centre_and_detected_object))
print("detected objects", len(detected_objects))
temporary_memory = []
for image_item in detected_objects:
seen_item_centroid = (image_item[2] + self.distance(
(image_item[2], image_item[3]),
(image_item[4], image_item[3])) / 2.0,
image_item[4] + self.distance(
(image_item[2], image_item[3]),
(image_item[2], image_item[5])) / 2.0)
tracking_uuid = None
#print ("items in memory", len(self.SHORT_TERM_MEMORY))
for memory_item in self.SHORT_TERM_MEMORY:
memory_centroid = (memory_item[2] + self.distance(
(memory_item[2], memory_item[3]),
(memory_item[4], memory_item[3])) / 2.0,
memory_item[4] + self.distance(
(memory_item[2], memory_item[3]),
(memory_item[2], memory_item[5])) / 2.0)
#print ("distance", self.distance(seen_item_centroid, memory_centroid))
if self.distance(seen_item_centroid,
memory_centroid) < self.distance(
[image_item[2], image_item[3]],
[image_item[4], image_item[5]
]) and image_item[0] == memory_item[0]:
tracking_uuid = memory_item[1]
continue
if tracking_uuid != None:
temporary_memory.append(
(self.KNOWN_OBJECTS[int(image_item[0])], tracking_uuid,
image_item[2], image_item[3], image_item[4],
image_item[5], image_item[6], image_item[7]))
else:
temporary_memory.append(
(self.KNOWN_OBJECTS[int(image_item[0])], uuid.uuid1(),
image_item[2], image_item[3], image_item[4],
image_item[5], image_item[6], image_item[7]))
#print ("temp memory items", len(temporary_memory))
if self.show_image:
for item in temporary_memory:
cv2.rectangle(img, (item[2], item[3]), (item[4], item[5]),
(0, 0, 0), 5)
textLabel = '{}: {}'.format(item[0], item[1])
(retval,
baseLine) = cv2.getTextSize(textLabel,
cv2.FONT_HERSHEY_COMPLEX, 1,
1)
textOrg = (image_item[2] - 20, image_item[3] - 20)
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, 0.3, (0, 0, 0), 1)
self.SHORT_TERM_MEMORY = temporary_memory
if self.show_image:
cv2.imshow('image', img)
cv2.waitKey(3000)
#time.sleep(1)
#cv2.destroyAllWindows()
return self.SHORT_TERM_MEMORY
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.')
def test_view_func(C, model_rpn, model_classifier):
base_dir = os.getcwd()
test_cls_all = ['aeroplane']
class_mapping = C.class_mapping
inv_class_mapping = {v: k for k, v in class_mapping.iteritems()}
backend = K.image_dim_ordering()
filename = '/home/gilad/bar/real7.p'
video_filename = "/home/gilad/ssd/keras-frcnn-master/a.mp4"
write_flag = False
class_to_color = {
class_mapping[v]: np.random.randint(0, 255, 3)
for v in class_mapping
}
# turn off any data augmentation at test time
save_flag = False
visualise = False
count = 0
good_img = 0
not_good = 0
mAP = 0
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)
img = cv2.resize(img, (new_width, new_height),
interpolation=cv2.INTER_CUBIC)
return img, ratio
def format_img_channels(img, C):
""" formats the image channels based on config """
img = img[:, :, (2, 1, 0)]
img = img.astype(np.float32)
img[:, :, 0] -= C.img_channel_mean[0]
img[:, :, 1] -= C.img_channel_mean[1]
img[:, :, 2] -= C.img_channel_mean[2]
img /= C.img_scaling_factor
img = np.transpose(img, (2, 0, 1))
img = np.expand_dims(img, axis=0)
return img
def draw_bbox(img, bbox, prob, azimuth, ratio):
# new_boxes, new_probs, new_az = roi_helpers.non_max_suppression_fast(bbox, prob, azimuth, overlap_thresh=0.3,use_az=True)
new_boxes = bbox
new_az = azimuth
new_probs = prob
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)
# cv2.rectangle(img,(bbox_gt['x1'], bbox_gt['y1']), (bbox_gt['x2'], bbox_gt['y2']), (int(class_to_color[key][0]), int(class_to_color[key][1]), int(class_to_color[key][2])),2)
# textLabel = '{}: {},azimuth : {}'.format(key,int(100*new_probs[jk]),new_az[jk])
textLabel = 'azimuth : {}'.format(new_az[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 + 15)
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)
return img
def format_img(img, C):
""" formats an image for model prediction based on config """
img, ratio = format_img_size(img, C)
img = format_img_channels(img, C)
return img, ratio
def display_image(img):
img1 = img[:, :, (2, 1, 0)]
# img1=img
im = Image.fromarray(img1.astype('uint8'), 'RGB')
im.show()
# Method to transform the coordinates of the bounding box to its original size
def get_real_coordinates(ratio, x1, y1, x2, y2):
## read the training data from pickle file or from annotations
real_x1 = int(round(x1 // ratio))
real_y1 = int(round(y1 // ratio))
real_x2 = int(round(x2 // ratio))
real_y2 = int(round(y2 // ratio))
return (real_x1, real_y1, real_x2, real_y2)
vnum_test = 24
azimuth_vec = np.concatenate(
([0],
np.linspace((360. / (vnum_test * 2)), 360. -
(360. / (vnum_test * 2)), vnum_test)),
axis=0)
def find_interval(azimuth, azimuth_vec):
for i in range(len(azimuth_vec)):
if azimuth < azimuth_vec[i]:
break
ind = i
if azimuth > azimuth_vec[-1]:
ind = 1
return ind
# print(rep)
obj_num = 0
bbox_threshold_orig = 0.6
th_bbox = 0.3
#### open images from folder
# for idx, img_name in enumerate(sorted(os.listdir(img_path))):
# if not img_name.lower().endswith(('.bmp', '.jpeg', '.jpg', '.png', '.tif', '.tiff')):
# continue
# print(img_name)
# filepath = os.path.join(img_path,img_name)
# img = cv2.imread(filepath)caricycle
#### open images from file
## read the training data from pickle file or from annotations
# class_mapping = C.class_mapping
for test_cls in test_cls_all:
good_img = 0
not_good = 0
count = 0
obj_num = 0
gt_cls_num = class_mapping[test_cls]
print('work on class {}'.format(test_cls))
test_pickle = os.path.join(
base_dir, 'pickle_data/test_data_{}.pickle'.format(test_cls))
if os.path.exists(test_pickle):
with open(test_pickle) as f:
all_imgs, classes_count, _ = pickle.load(f)
for im_file in all_imgs:
filepath = im_file['filepath']
img = cv2.imread(filepath)
img_gt = np.copy(img)
if img is None:
not_good += 1
continue
else:
good_img += 1
# print ('im num {}'.format(good_img))
if good_img % 50 == 0:
print("worked on {} images".format(good_img))
X, ratio = format_img(img, C)
if backend == 'tf':
X = np.transpose(X, (0, 2, 3, 1))
# get the feature maps and output from the RPN
[Y1, Y2] = 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]
width, height = int(im_file["width"]), int(im_file["height"])
resized_width, resized_height = data_generators.get_new_img_size(
width, height, C.im_size)
# [_,_, F] = model_rpn.predict(X)
ROIs = []
## pass on all the labels in the image, some of them are not equal to test_cls
for bbox_gt in im_file['bboxes']:
if not bbox_gt['class'] == test_cls:
continue
no_bbox_flag = 1
bbox_threshold = bbox_threshold_orig
while no_bbox_flag and bbox_threshold > th_bbox:
cls_gt = bbox_gt['class']
az_gt = bbox_gt['azimuth']
el_gt = bbox_gt['elevation']
t_gt = bbox_gt['tilt']
if bbox_gt[
'class'] == test_cls and bbox_threshold == bbox_threshold_orig:
obj_num += 1
if len(ROIs) == 0:
# apply the spatial pyramid pooling to the proposed regions
bboxes = {}
probs = {}
azimuths = {}
# print ('obj num {}'.format(obj_num))
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,
P_view] = model_classifier.predict([X, ROIs])
for ii in range(P_cls.shape[1]):
if np.max(P_cls[0, ii, :]
) < bbox_threshold or np.argmax(
P_cls[0, ii, :]) == (
P_cls.shape[2] - 1):
continue
## get class from the net
# cls_num = np.argmax(P_cls[0, ii, :])
## use gt class
cls_num = gt_cls_num
cls_name = inv_class_mapping[cls_num]
cls_view = P_view[0, ii, 4 * cls_num:4 *
(cls_num + 1)]
# cls_name_gt = cls_nimg = draw_bbox(img,bbox, prob, azimuth, ratio)ame
# if cls_name == cls_name_gt:
# print(np.argmax(cls_view,axis=0))
if cls_name not in bboxes:
bboxes[cls_name] = []
probs[cls_name] = []
azimuths[cls_name] = []
(x, y, w, h) = ROIs[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, :]))
azimuths[cls_name].append(Quat(cls_view).ra)
all_dets = []
if len(bboxes) == 0:
bbox_threshold -= 0.1
for key in bboxes:
# if 1:
if key == test_cls and bbox_gt['class'] == test_cls:
bbox = np.array(bboxes[key])
prob = np.array(probs[key])
azimuth = np.array(azimuths[key])
## get the azimuth from bbox that have more than 'overlap_thresh' overlap with gt_bbox
az = []
overlap_thresh = 0.5
try:
while np.size(
az) == 0 and overlap_thresh > 0.3:
_, prob_bbox, az = roi_helpers.overlap_with_gt(
bbox,
prob,
azimuth,
bbox_gt,
ratio=ratio,
overlap_thresh=overlap_thresh,
max_boxes=300,
use_az=True)
if np.size(
az) != 0 and overlap_thresh == 0.5:
mAP += 1
overlap_thresh -= 0.1
if overlap_thresh == 0:
print("No good Bbox was found")
counts = np.bincount(az)
except:
az = []
counts = []
try:
az_fin = np.argmax(counts)
true_bin = find_interval(az_gt, azimuth_vec)
prob_bin = find_interval(az_fin, azimuth_vec)
no_bbox_flag = 0
if true_bin == prob_bin:
count += 1
break
except:
# print('here')
no_bbox_flag = 1
bbox_threshold -= 0.1
## azimuth calculations
## display
# if visualise:
# display_image(img)
# # cv2.imshow('img', img)
# # cv2.waitKey(0)
# if save_flag:
# cv2.imwrite('./results_imgs/{}'.format(img_name),img)
# # img = img[:, :, (2, 1, 0)]
# # cv2.imwrite('./results_imgs/video/{}.png'.format(num),img)
# # print('save')
bbox_threshold -= 0.1
# if visualise:
# display_image(img)
succ = float(count) / float(obj_num) * 100.
string = 'for class {} -true count is {} out of {} from {} images . {} success'.format(
test_cls, count, obj_num, good_img, succ)
print(string)
mAP = float(mAP) / float(obj_num) * 100.
print("MAP is {}".format(mAP))
# if write_flag:
# f = open('{}_results.txt'.format(weight_name),'a')
# f.write(string+'\n')
# f.close()
return succ, mAP
def upload_file():
print("request is ", request.files)
st = time.time()
content_length = request.content_length
print(f"Content_length : {content_length}")
print("data type is ", type(request))
print("data type of request files ", type(request.files))
data_dict = request.form.to_dict()
#print(type(data_dict))
#print(data_dict['file'])
#print('data from frontend',data_dict)
data = (data_dict['file'].split(',')[1])
l, b = (data_dict['imgDimensions'].split(','))
l = int(l)
b = int(b)
print('width of image', l)
print('type of l ', type(l))
print('height of image', b)
#print(data)
#print(len(data_dict))
#print(data)
imgdata = base64.b64decode(data)
print("imagedata type is", type(imgdata))
img2 = Image.open(io.BytesIO(imgdata))
print(type(img2))
#img2.show()
#img = cv2.imread(img2)
#print('opencv type' , type(img))
#print(type(img))
a = np.array(img2.getdata()).astype(np.float64)
#print('datatype of w ', w.dtype)
#b = np.ones(172800,3)
#a = np.concatenate((w,b), axis=None)
print('type of data to model ', type(a))
print('shape of data from frontend', a.shape)
#r, c = a.shape
#print('Value of r', r)
"""
if a.shape == (480000, 3):
data = a.reshape(600, 800, 3)
else: data = a.reshape(480, 640, 3)
"""
data = a.reshape(b, l, 3)
st = time.time()
parser = OptionParser()
parser.add_option(
"-n",
"--num_rois",
type="int",
dest="num_rois",
help="Number of ROIs per iteration. Higher means more memory use.",
default=64)
parser.add_option(
"--config_filename",
dest="config_filename",
help=
"Location to read the metadata related to the training (generated when training).",
default="config.pickle")
parser.add_option("--network",
dest="network",
help="Base network to use. Supports vgg or resnet50.",
default='resnet50')
(options, args) = parser.parse_args()
config_output_filename = options.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
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)
class_to_color = {
class_mapping[v]: np.random.randint(0, 255, 3)
for v in class_mapping
}
C.num_rois = int(options.num_rois)
if C.network == 'resnet50':
num_features = 1024
elif C.network == 'vgg':
num_features = 512
if K.image_dim_ordering() == 'th':
input_shape_img = (3, None, None)
input_shape_features = (num_features, None, None)
else:
input_shape_img = (None, None, 3)
input_shape_features = (None, None, num_features)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(C.num_rois, 4))
feature_map_input = Input(shape=input_shape_features)
# define the base network (resnet here, can be VGG, Inception, etc)
shared_layers = nn.nn_base(img_input, trainable=True)
# define the RPN, built on the base layers
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn_layers = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(feature_map_input,
roi_input,
C.num_rois,
nb_classes=len(class_mapping),
trainable=True)
model_rpn = Model(img_input, rpn_layers)
model_classifier_only = Model([feature_map_input, roi_input], classifier)
model_classifier = Model([feature_map_input, roi_input], classifier)
print('Loading weights from {}'.format(C.model_path))
model_rpn.load_weights(C.model_path, by_name=True)
model_classifier.load_weights(C.model_path, by_name=True)
model_rpn.compile(optimizer='sgd', loss='mse')
model_classifier.compile(optimizer='sgd', loss='mse')
all_imgs = []
classes = {}
bbox_threshold = 0.6
visualise = True
#if not img_name.lower().endswith(('.bmp', '.jpeg', '.jpg', '.png', '.tif', '.tiff')):
# continue
#print(img_name)
#filepath = os.path.join(img_path,img_name)
img = data
#cv2.imshow('img', img)
#cv2.waitKey(0)
X, ratio = 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=0.6)
# 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.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, :])]
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=0.6)
for jk in range(new_boxes.shape[0]):
(x1, y1, x2, y2) = new_boxes[jk, :]
(real_x1, real_y1, real_x2,
real_y2) = get_real_coordinates(ratio, x1, y1, x2, y2)
cv2.rectangle(
img, (real_x1, real_y1), (real_x2, real_y2),
(int(class_to_color[key][0]), int(
class_to_color[key][1]), int(class_to_color[key][2])), 2)
textLabel = '{}: {}'.format(key, int(100 * new_probs[jk]))
all_dets.append((key, 100 * new_probs[jk]))
(retval, baseLine) = cv2.getTextSize(textLabel,
cv2.FONT_HERSHEY_COMPLEX, 1,
1)
textOrg = (real_x1, real_y1 - 0)
cv2.rectangle(
img, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5),
(0, 0, 0), 2)
cv2.rectangle(
img, (textOrg[0] - 5, textOrg[1] + baseLine - 5),
(textOrg[0] + retval[0] + 5, textOrg[1] - retval[1] - 5),
(255, 255, 255), -1)
cv2.putText(img, textLabel, textOrg, cv2.FONT_HERSHEY_DUPLEX, 1,
(0, 0, 0), 1)
print('Elapsed time = {}'.format(time.time() - st))
print('number of windoiws detected', len(all_dets))
print(all_dets)
r = len(all_dets)
img3 = normalize(img)
#plt.imshow(img)
#cv2.imshow('img3', img3)
#cv2.waitKey(0)
K.clear_session()
#data = process(data)
#print('after reshape',data.shape)
im2 = Image.fromarray(img.astype("uint8"), "RGB")
print("im2 data type is ", type(im2))
#to_frontend = (" ".join(str(x) for x in data))
db = data.tobytes()
print('type of data to database :', type(db))
todb = insertBLOB('Image007', db)
print('final data shape fed to model : ', data.shape)
# ImageFile img = db.b64encode()
# with open("t.png", "rb") as imageFile:
# str = base64.b64encode(imageFile.read())
#cv2.imshow('image', cv2.cvtColor(data, cv2.COLOR_BGR2GRAY))
#cv2.waitKey()
#str = base64.b64encode(data)
# return jsonify(to_frontend, r)
#img = Image.open( 'C:\Window Counter_Project\Flickr\Window_101 (131).jpg' )
#img.load()
#data = np.asarray( img, dtype="int32" )
#im = Image.fromarray(data.astype("uint8"))
#im.show() # uncomment to look at the image
rawBytes = io.BytesIO()
print(rawBytes)
im2.save(rawBytes, "jpeg")
#im2.show()
print('type of im2 is ', type(im2))
rawBytes.seek(0) # return to the start of the file
response_obj = {
'count': r,
'image':
"data:image/jpeg;base64," + str(base64.b64encode(rawBytes.read()))
}
#print("response is", type(response_obj))
return jsonify(Data=response_obj)
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.')
img = cv2.imread(filepath)
#imggg = Image.fromarray(img)
#imggg.show()
#cv2.imshow('img', img)
#cv2.waitKey()
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:
def Train_frcnn(
train_path, # path to the text file containing the data
network_arch, # the full faster rcnn network architecture object
num_epochs, # num of epochs
output_weight_path, # path to save the model_all.weights as hdf5
preprocessing_function=None,
config_filename="config.pickle",
input_weights_path=None,
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"
)
nn = network_arch
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'] == 'valid']
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:
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.')
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/resnet50_weights_tf_dim_ordering_tf_kernels.h5'
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 = 500
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 training(train_path,
input_weight_path,
parser="pascal_voc",
num_rois=32,
network="resnet50",
horizontal_flips=False,
vertical_flips=False,
rot_90=False,
num_epochs=2000,
config_filename="config.pickle",
output_weight_path='./model_frcnn.hdf5'):
'''
train_path = Path to training data
parser = Parser to use. One of simple or pascal_voc
num_rois = Number of RoIs to process at once
network = Base network to use. Supports vgg or resnet50
horizontal_flips = Augment with horizontal flips in training. (Default=false)
vertical_flips = Augment with vertical flips in training. (Default=false)
rot_90 = Augment with 90 degree rotations in training. (Default=false)
num_epochs = Number of epochs
config_filename = Location to store all the metadata related to the training (to be used when testing)
output_weight_path = Output path for weights
input_weight_path = Input path for weights. If not specified, will try to load default weights provided by keras
'''
if not train_path: # if filename is not given
print(
'Error: path to training data must be specified. Pass --path to command line'
)
if parser == 'pascal_voc':
from keras_frcnn.pascal_voc_parser import get_data
elif parser == 'simple':
from keras_frcnn.simple_parser import get_data
else:
raise ValueError(
"Command line option parser must be one of 'pascal_voc' or 'simple'"
)
# pass the settings from the command line, and persist them in the config object
C = config.Config()
C.use_horizontal_flips = bool(horizontal_flips)
C.use_vertical_flips = bool(vertical_flips)
C.rot_90 = bool(rot_90)
C.model_path = output_weight_path
C.num_rois = int(num_rois)
if network == 'vgg':
C.network = 'vgg'
from keras_frcnn import vgg as nn
elif network == 'resnet50':
from keras_frcnn import resnet as nn
C.network = 'resnet50'
else:
print('Not a valid model')
raise ValueError
# check if weight path was passed via command line
if input_weight_path != "":
C.base_net_weights = input_weight_path
else:
# set the path to weights based on backend and model
C.base_net_weights = nn.get_weight_path()
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
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 = config_filename
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 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)))
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_gen_val = data_generators.get_anchor_gt(val_imgs,
classes_count,
C,
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(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(shared_layers,
roi_input,
C.num_rois,
nb_classes=len(classes_count),
trainable=True)
model_rpn = Model(img_input, rpn[:2])
model_classifier = Model([img_input, roi_input], classifier)
# this is a model that holds both the RPN and the classifier, used to load/save weights for the models
model_all = Model([img_input, roi_input], rpn[:2] + classifier)
try:
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)
model_rpn.compile(optimizer=optimizer,
loss=[
losses.rpn_loss_cls(num_anchors),
losses.rpn_loss_regr(num_anchors)
])
model_classifier.compile(
optimizer=optimizer_classifier,
loss=[
losses.class_loss_cls,
losses.class_loss_regr(len(classes_count) - 1)
],
metrics={'dense_class_{}'.format(len(classes_count)): 'accuracy'})
model_all.compile(optimizer='sgd', loss='mae')
epoch_length = 1000
num_epochs = int(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 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)
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)
# 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)
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]
progbar.update(iter_num + 1,
[('rpn_cls', losses_[iter_num, 0]),
('rpn_regr', losses_[iter_num, 1]),
('detector_cls', losses_[iter_num, 2]),
('detector_regr', losses_[iter_num, 3])])
iter_num += 1
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))
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
model_all.save_weights((C.model_path))
break
except Exception as e:
print('Exception: {}'.format(e))
continue
print('Training complete, exiting.')
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 = tpu_model.train_on_batch(X, Y)
# loss_rpn = model_rpn.train_on_batch(X, Y)
P_rpn = tpu_model.predict_on_batch(X)
# 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)
# 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)
def handler(event, context):
img_name = event['img_process']
client.download_file('adaproject', img_name, '/tmp/' + img_name)
X = np.load('/tmp/' + img_name)
with open('config.pickle', 'rb') as f_in:
C = pickle.load(f_in)
class_mapping = C.class_mapping
num_features = 1024
input_shape_img = (None, None, 3)
input_shape_features = (None, None, num_features)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(C.num_rois, 4))
feature_map_input = Input(shape=input_shape_features)
# define the base network (resnet here, can be VGG, Inception, etc)
shared_layers = nn.nn_base(img_input, trainable=True)
# define the RPN, built on the base layers
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn_layers = nn.rpn(shared_layers, num_anchors)
classifier = nn.classifier(feature_map_input,
roi_input,
C.num_rois,
nb_classes=len(class_mapping),
trainable=True)
model_rpn = Model(img_input, rpn_layers)
model_classifier = Model([feature_map_input, roi_input], classifier)
BUCKET_NAME = 'adaproject' # replace with your bucket name
KEY = 'model_frcnn.hdf5' # replace with your object key
s3 = boto3.resource('s3')
try:
s3.Bucket(BUCKET_NAME).download_file(KEY, '/tmp/model_frcnn.hdf5')
print 'File Found'
except botocore.exceptions.ClientError as e:
if e.response['Error']['Code'] == "404":
print("The object does not exist.")
else:
raise
model_rpn.load_weights('/tmp/model_frcnn.hdf5', by_name=True)
model_classifier.load_weights('/tmp/model_frcnn.hdf5', by_name=True)
model_rpn.compile(optimizer='sgd', loss='mse')
model_classifier.compile(optimizer='sgd', loss='mse')
# Starting RPN prediction
[Y1, Y2, F] = model_rpn.predict(X)
R = roi_helpers.rpn_to_roi(Y1,
Y2,
C,
K.image_dim_ordering(),
overlap_thresh=0.7)
R[:, 2] -= R[:, 0]
R[:, 3] -= R[:, 1]
bboxes = {}
probs = {}
bbox_threshold = 0.8
class_mapping = {v: k for k, v in class_mapping.items()}
# print(class_mapping)
class_to_color = {
class_mapping[v]: np.random.randint(0, 255, 3)
for v in class_mapping
}
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.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, :])]
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, :]))
final_data = []
output = {}
for key in bboxes:
data = {}
bbox = np.array(bboxes[key])
new_boxes, new_probs = roi_helpers.non_max_suppression_fast(
bbox, np.array(probs[key]), overlap_thresh=0.5)
data[key] = {}
for i in range(new_boxes.shape[0]):
data[key]['x'] = str(new_boxes[i][0])
data[key]['y'] = str(new_boxes[i][1])
data[key]['w'] = str(new_boxes[i][2])
data[key]['z'] = str(new_boxes[i][3])
data[key]['prob'] = str(new_probs[i])
final_data.append(data)
output['bboxes'] = bboxes
output['rpn'] = final_data
timestamp = int(time.time() * 1000)
table = dynamodb.Table(os.environ['DYNAMODB_TABLE'])
result = table.update_item(
Key={'requestId': event['requestId']},
ExpressionAttributeNames={
'#status': 'status',
'#result': 'result',
},
ExpressionAttributeValues={
':status': 'DONE',
':result': output,
':updatedAt': timestamp,
},
UpdateExpression='SET #status = :status, '
'#result = :result, '
'updatedAt = :updatedAt',
ReturnValues='ALL_NEW',
)
response = {
"statusCode": 200,
"body": json.dumps(result['Attributes'], cls=DecimalEncoder)
}
return response
data_to_run = ref
break
X, Y = data_to_run['x'], data_to_run['y']
print('\n\n')
print('BATCHED X SHAPE: {}'.format(X.shape))
for ii in range(noutputs):
print('BATCHED Y {} SHAPE: {}'.format(ii, Y[ii].shape))
print('TRAINING, WATCH THE GPU UTILIZATION SPIKE ON ALL GPUS')
loss_rpn = model_rpn.train_on_batch(X, Y)
P_rpn = model_rpn.predict_on_batch(X)
print('\n\n ==================== WILL CRASH SOON. YOU FIX IT '
'=======================\n\n')
R = roi_helpers.rpn_to_roi(P_rpn[0], P_rpn[1], C, K.image_dim_ordering(
), use_regr=True, overlap_thresh=0.7, max_boxes=300)
# note: calc_iou converts from (x1,y1,x2,y2) to (x,y,w,h) format
X2, Y1, Y2, 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:
def main():
cleanup()
sys.setrecursionlimit(40000)
config_output_filename = 'config.pickle'
with open(config_output_filename, 'r') as f_in:
C = pickle.load(f_in)
# 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
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 = num_rois
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(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 = {}
bbox_threshold = 0.8
visualise = True
print("Converting video to images..")
convert_to_images()
print("anotating...")
list_files = sorted(get_file_names(img_path), key=lambda var:[int(x) if x.isdigit() else x for x in re.findall(r'[^0-9]|[0-9]+', var)])
for img_name in list_files:
if not img_name.lower().endswith(('.bmp', '.jpeg', '.jpg', '.png', '.tif', '.tiff')):
continue
print(img_name)
st = time.time()
filepath = os.path.join(img_path,img_name)
img = cv2.imread(filepath)
X = format_img(img, C)
img_scaled = np.transpose(X.copy()[0, (2, 1, 0), :, :], (1, 2, 0)).copy()
img_scaled[:, :, 0] += 123.68
img_scaled[:, :, 1] += 116.779
img_scaled[:, :, 2] += 103.939
img_scaled = img_scaled.astype(np.uint8)
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.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, :])]
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 = []
all_objects = []
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,:]
cv2.rectangle(img_scaled,(x1, y1), (x2, y2), class_to_color[key],2)
textLabel = '{}: {}'.format(key,int(100*new_probs[jk]))
all_dets.append((key,100*new_probs[jk]))
all_objects.append((key, 1))
(retval,baseLine) = cv2.getTextSize(textLabel,cv2.FONT_HERSHEY_COMPLEX,1,1)
textOrg = (x1, y1-0)
cv2.rectangle(img_scaled, (textOrg[0] - 5, textOrg[1]+baseLine - 5), (textOrg[0]+retval[0] + 5, textOrg[1]-retval[1] - 5), (0, 0, 0), 2)
cv2.rectangle(img_scaled, (textOrg[0] - 5,textOrg[1]+baseLine - 5), (textOrg[0]+retval[0] + 5, textOrg[1]-retval[1] - 5), (255, 255, 255), -1)
cv2.putText(img_scaled, textLabel, textOrg, cv2.FONT_HERSHEY_DUPLEX, 1, (0, 0, 0), 1)
print('Elapsed time = {}'.format(time.time() - st))
height, width, channels = img_scaled.shape
cv2.rectangle(img_scaled, (0,0), (width, 30), (0, 0, 0), -1)
cv2.putText(img_scaled, "Obj count: " + str(list(accumulate(all_objects))), (5, 19), cv2.FONT_HERSHEY_TRIPLEX, 0.5, (255, 255, 255), 1)
cv2.imwrite(os.path.join(output_path, img_name), img_scaled)
print(all_dets)
print("saving to video..")
save_to_video()