def get_all_objs():
a, img_work = utils.take_workspace_img(client)
img_work = utils.standardize_img(img_work)
if not a:
a, img_work = debug_markers(img_work)
return img_work, None
mask = utils.objs_mask(img_work)
objs = utils.extract_objs(img_work, mask)
if len(objs) == 0:
return img_work, []
imgs = []
if model is None:
return img_work, objs
for x in range(len(objs)):
imgs.append(cv2.resize(objs[x].img, (64, 64)))
imgs = np.array(imgs)
predictions = model.predict(imgs)
for x in range(len(predictions)):
obj = objs[x]
obj.type = predictions[x].argmax()
# graphical debug
cv2.drawContours(img_work, [obj.box], 0, (0, 0, 255), 2)
pos = [obj.x + 20, obj.y]
# Text position
img_work = cv2.putText(img_work, objects_names[obj.type], tuple(pos), font, 0.5, (255, 0, 0), 1, cv2.LINE_AA)
pos[0] += img_work.shape[0]
img_work = cv2.putText(img_work, objects_names[obj.type], tuple(pos), font, 0.5, (0, 255, 0), 1, cv2.LINE_AA)
return img_work, objs
def prepare_pneumonia_data(self):
print('== Preparing kaggle data')
normal_fn = os.path.join(self.pneumonia_path,
'stage_2_detailed_class_info.csv')
pneu_fn = os.path.join(self.pneumonia_path, 'stage_2_train_labels.csv')
csv_normal = pd.read_csv(normal_fn, nrows=None)
csv_pneu = pd.read_csv(pneu_fn, nrows=None)
labels = {'normal': [], 'pneumonia': []}
train = []
test = []
for index, row in csv_normal.iterrows():
if row['class'] == 'Normal':
labels['normal'].append(row['patientId'])
for index, row in csv_pneu.iterrows():
if int(row['Target']) == 1:
labels['pneumonia'].append(row['patientId'])
for key in labels.keys():
list_fns = labels[key]
if len(list_fns) == 0:
continue
# fixed evaluation
test_patients = np.load(
os.path.join(self.pneumonia_path,
'rsna_test_patients_{}.npy'.format(key)))
# Counts would be:
# test
# count: {'normal': 885, 'pneumonia': 1058, 'COVID-19': 10}
# train
# count: {'normal': 7966, 'pneumonia': 8526, 'COVID-19': 105}
for patient in list_fns:
fn = os.path.join(self.pneumonia_path, 'stage_2_train_images',
patient + '.dcm')
ds = dicom.dcmread(fn)
img = cv2.cvtColor(ds.pixel_array, cv2.COLOR_GRAY2RGB)
img = utils.standardize_img(
cv2.resize(img,
(self.resize, self.resize)).astype('float32'))
meta = {
'dataset': 'neumonia_kaggle_Dataset',
'id': patient,
'filename': fn,
'image': img,
'label': key,
'train': 0 if patient in test_patients else 1
}
if patient in test_patients:
test.append(meta)
self.test_count[key] += 1
else:
train.append(meta)
self.train_count[key] += 1
print('label {} read and packed!'.format(key))
training_patients = [ex['id'] for ex in train]
test_patients = [ex['id'] for ex in test]
assert len(set(training_patients).intersection(
set(test_patients))) == 0
for lb in labels.keys():
train_label = [ex for ex in train if ex['label'] == lb]
test_label = [ex for ex in test if ex['label'] == lb]
print('saving records for label: {}'.format(lb))
if len(train_label) != 0:
self.save_data(train_label, lb, 'kaggle')
if len(test_label) != 0:
self.save_data(test_label, lb, 'kaggle', training=False)
print('test count: ', self.test_count)
print('train count: ', self.train_count)
def prepare_covid_xray(self):
print('== Preparing covid19 data')
csv = pd.read_csv(os.path.join(args.covid_path, 'metadata.csv'),
nrows=None)
idx_pa = csv["view"] == "PA"
csv = csv[idx_pa]
train = []
test = []
labels = {'normal': [], 'pneumonia': [], 'COVID-19': []}
for index, row in csv.iterrows():
f = row['finding']
if f in self.mapping:
self.data_counts[self.mapping[f]] += 1
entry = [
int(row['patientid']), row['filename'], self.mapping[f]
]
labels[self.mapping[f]].append(entry)
print('Data distribution from covid-chestxray-dataset:')
print(self.data_counts)
for key in labels.keys():
data_list = labels[key]
if len(data_list) == 0:
print('No data for {} in this dataset'.format(key))
continue
print('Key: ', key)
print('Test patients: ', self.test_dict_persons[key])
# go through all the patients
for patient in data_list:
fn = os.path.join(args.covid_path, 'images', patient[1])
img = cv2.resize(utils.imread(fn), (self.resize, self.resize))
# print(img.shape)
meta = {
'dataset':
'covid-chestxray-dataset',
'patient_id':
patient,
'filename':
fn,
'image':
utils.standardize_img(img.astype('float32')),
'label':
key,
'train':
0 if str(patient[0]) in self.test_dict_persons[key] else 1
}
# print(0 if patient[0] in self.test_dict_persons[key] else 1)
# print(patient[0], self.test_dict_persons[key])
if meta['train'] == 0:
test.append(meta)
self.test_count[patient[2]] += 1
else:
train.append(meta)
self.train_count[patient[2]] += 1
for lb in labels.keys():
train_label = [ex for ex in train if ex['label'] == lb]
test_label = [ex for ex in test if ex['label'] == lb]
print('saving records for label: {}'.format(lb))
if len(train_label) != 0:
self.save_data(train_label, lb, 'chestxray')
if len(test_label) != 0:
self.save_data(test_label, lb, 'chestxray', False)
print('test count: ', self.test_count)
print('train count: ', self.train_count)
def load_dataset(data_path):
#list all directories in data_path
objects_names = os.listdir(data_path)
print(objects_names)
objects_list = [] #list containing all objects images
labels_list = [] #list containing all objects Labels
files_names = [] #list containing all objects names
obj_id = 0
#make the data_mask directorie
try:
os.mkdir("./data_mask1")
except:
pass
for name in objects_names:
# list all image for each objects
list_dir = os.listdir(data_path + name)
print(name + " " + str(len(list_dir)))
# make a "data_mask/<obj_name>" subdirectory for each object
try:
os.mkdir("./data_mask1/" + name)
except:
pass
#for each file in in "data/<obj_name>""
for file_name in list_dir:
# read the file
img = cv2.imread(data_path + name + "/" + file_name)
# img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = utils.standardize_img(img)
# extract all objects from the image
mask = utils.objs_mask(img)
objs = utils.extract_objs(img, mask)
#for each object in the image
for x in range(0, len(objs)):
img = objs[x].img
#write the image in data_mask/<file_name>"
cv2.imwrite(
"data_mask1/" + name + "/" + str(x) + "_" + file_name, img)
#resize the image to match our model input
img = cv2.resize(img, (64, 64))
#create a numpy array of float with a size of (64, 64, 3)
img_float = np.zeros((64, 64, 3), np.float32)
#scale the image color between 0.0 and 1.0 and copy it in the numpy array
img_float[:][:][:] = img[:][:][:] / 255.0
#create a numpy array full of 0 with a shape of (len(objects_names))
label = np.zeros((len(objects_names)), np.float32)
#set is corresponding id to 1
label[obj_id] = 1
#insert all our numpy array in the data set
objects_list.append(img_float)
labels_list.append(label)
files_names.append(str(x) + "_" + file_name)
print(len(objs), end='')
print("|", end="", flush=True)
print("")
obj_id += 1
return [objects_list, labels_list, files_names, objects_names]
except:
print("calibration failed")
client.change_tool(RobotTool.GRIPPER_2)
while 1:
#move to observation_pose
client.move_pose(*observation_pose.to_list())
a = False
#take a picture of the workspace (infinite loop if the robot can't see the workspace)
while not a:
a, img_work = utils.take_workspace_img(client)
img_work = utils.standardize_img(img_work)
#calculate the mask for img_work (black and white image)
mask = utils.objs_mask(img_work)
#aply the mask to the image
img_masked = cv2.bitwise_and(img_work, img_work, mask=mask)
img_db = concat_imgs([img_work, mask, img_masked], 1)
#get all opbject from the image
objs = utils.extract_objs(img_work, mask)
if len(objs) == 0:
continue
imgs = []
#resize all objects img to 64*64 pixels