def convert_mat_to_png(mat_folder, b_runmatlab=True):
# need to run matlab script
# original_path = str(os.getcwd())
# path = os.path.join(codes_path,'DeepLab-Context','matlab','my_script')
# os.chdir(path)
# # convert the mat files to png or tif
# if b_runmatlab :
# subprocess.call("matlab -r 'converttoPng; exit;'", shell=True)
# os.chdir(original_path)
# using python script instead of matlab script
result = mat_To_png.convert_mat_to_png(mat_folder)
if result is False:
return False
# read the png or tif files list
# result = io_function.get_file_list_by_ext('.tif',mat_file_foler,bsub_folder=False)
# result = io_function.get_file_list_by_ext('.mat', mat_file_foler, bsub_folder=False)
# if len(result) < 1:
# result = io_function.get_file_list_by_ext('.png', mat_file_foler, bsub_folder=False)
# if len(result) < 1:
# basic.outputlogMessage('error, Not result (.tif or .png) in Mat folder:%s'%mat_file_foler)
# return False
if len(result) != len(test_data):
basic.outputlogMessage(
'error, the count of results is not the same as input test file')
return False
return result
def model_finder(path):
mtime = lambda f: os.stat(os.path.join(path, f)).st_mtime
files = list(sorted(glob.glob(path + '/*.pth'), key=mtime))
if len(files) >= 1:
file_ = files[-1]
else:
basic.outputlogMessage('error, can not find pth model file')
return False
return file_
def run_train():
if len(train_data)<1:
basic.outputlogMessage('error, No input image for training')
return False
init_model_path ='data/MS_DeepLab_resnet_pretrained_COCO_init.pth'
train.run_training(gpuid,init_model_path,train_data,batch_size=1)
pass
def merge_edge_to_detected_result(edgemap_list, detected_png_list):
if io_function.is_file_exist(edgemap_list) is False:
return False
edge_data = []
f_obj = open(edgemap_list)
f_lines = f_obj.readlines()
f_obj.close()
for i in range(0, len(f_lines)):
temp = f_lines[i].split()
sample = SampleClass()
sample.image = temp[0]
sample.edge_map = temp[1]
edge_data.append(sample)
if len(edge_data) != len(detected_png_list):
basic.outputlogMessage(
'error, the count of edge map is different from the one of detected result'
)
return False
# make sure edge_data have the same order of png list
for i in range(0, len(edge_data)):
if edge_data[i].image != test_data[i].image:
basic.outputlogMessage(
'error, the source data for edge and building detection are different'
)
return False
#merge two png (edge and detected)
edge_file = edge_data[i].edge_map
im_edge = Image.open(edge_file)
in_edge = numpy.array(im_edge, dtype=numpy.uint8)
det_file = os.path.join(
os.path.split(detected_png_list[i])[0],
test_data[i].id + '_blob_0.png')
if det_file not in detected_png_list:
basic.outputlogMessage('result_file file not in the list %s' %
det_file)
return False
im_det = Image.open(det_file)
in_det = numpy.array(im_det, dtype=numpy.uint8)
in_det[in_edge == 0] = 0 # 0 is the value of edge pixel
#backup for test
backcup_name = io_function.get_name_by_adding_tail(det_file, 'bak')
io_function.copy_file_to_dst(det_file, backcup_name)
basic.outputlogMessage('merge edge to detection result: %d /%d: %s to %s' % (i + 1, len(edge_data),\
os.path.basename(edge_file),os.path.basename(det_file)))
#over write original image
# cv2.imwrite(det_file, in_det)
misc.imsave(det_file, in_det)
return True
def run_test():
"""
run test with pytorch_deeplab_resnet
:return: result file list
"""
if len(test_data) < 1:
basic.outputlogMessage('error, no input data in the list ')
return False
find_model = model_finder(os.path.join(expr_folder, 'data', 'snapshots'))
if find_model is False:
return False
result_list = evalpyt.run_evalpyt(gpuid, find_model, test_data,
save_file_folder)
return result_list
def test_plot_truth_coords():
import SpaceNetData.plot_truth_coords as ptcoords
import SpaceNetData.geojson_to_pixel_arr as geojson_to_pixel_arr
img_path = "/Users/huanglingcao/Data/aws_SpaceNet/AOI_3_Paris_Train/RGB-PanSharpen/RGB-PanSharpen_AOI_3_Paris_img183.tif"
# img_path = "/Users/huanglingcao/Data/aws_SpaceNet/AOI_3_Paris_Train/PAN/PAN_AOI_3_Paris_img183.tif"
geojson = '/Users/huanglingcao/Data/aws_SpaceNet/AOI_3_Paris_Train/geojson/buildings/buildings_AOI_3_Paris_img183.geojson'
pixel_coords, latlon_coords = \
geojson_to_pixel_arr.geojson_to_pixel_arr(img_path,geojson,pixel_ints=True,verbose=True)
# img_obj = RSImageclass()
# if img_obj.open(img_path) is False:
# basic.outputlogMessage('Open image %s failed'%img_path)
# return False
# Datatype = img_obj.GetGDALDataType()
# width = img_obj.GetWidth()
# height = img_obj.GetHeight()
img_pro_obj = RSImgProclass()
input_image_band1 =img_pro_obj.Read_Image_band_data_to_numpy_array_all_pixel(1,img_path)
print numpy.min(input_image_band1), numpy.max(input_image_band1)
input_image_band1 = (input_image_band1- numpy.min(input_image_band1))*1.0/(numpy.max(input_image_band1) - numpy.min(input_image_band1))
print numpy.min(input_image_band1),numpy.max(input_image_band1)
print input_image_band1.shape
input_image_band2 = img_pro_obj.Read_Image_band_data_to_numpy_array_all_pixel(2, img_path)
input_image_band2 = (input_image_band2 - numpy.min(input_image_band2)) * 1.0 / (numpy.max(input_image_band2) - numpy.min(input_image_band2))
print numpy.min(input_image_band2), numpy.max(input_image_band2)
print input_image_band2.shape
input_image_band3 = img_pro_obj.Read_Image_band_data_to_numpy_array_all_pixel(3, img_path)
input_image_band3 = (input_image_band3 - numpy.min(input_image_band3)) * 1.0 / (numpy.max(input_image_band3) - numpy.min(input_image_band3))
print numpy.min(input_image_band3), numpy.max(input_image_band3)
print input_image_band3.shape
if input_image_band1 is False:
basic.outputlogMessage('Open image %s failed' % img_path)
return False
input_image = numpy.stack((input_image_band1,input_image_band2,input_image_band3),axis=2)
# input_image = input_image_band1
# input_image.reshape(650,650,3)
input_image = input_image.astype(numpy.float32) # the value of each pixel is between 0-1
print input_image.shape
ptcoords.plot_truth_coords(input_image,pixel_coords)
pass
def read_test_data(test_file, file_id):
if os.path.isfile(test_file) is False:
basic.outputlogMessage('error: file not exist %s' % test_file)
return False
f_obj = open(test_file)
fid_obj = open(file_id)
f_lines = f_obj.readlines()
fid_lines = fid_obj.readlines()
fid_obj.close()
f_obj.close()
if len(f_lines) != len(fid_lines):
basic.outputlogMessage(
'the number of lines in test_file and test_file_id is not the same'
)
return False
for i in range(0, len(f_lines)):
temp = f_lines[i].split()
# temp[1].
temp.append(fid_lines[i].strip())
test_data.append(temp)
return True
def combine_IMGpath_and_Groundpath(img_only_list_file,
img_groud_true_list_file, result_path):
if io_function.is_file_exist(img_only_list_file) is False \
or io_function.is_file_exist(img_groud_true_list_file) is False:
return False
f_img_only = open(img_only_list_file, 'r')
f_img_ground = open(img_groud_true_list_file, 'r')
fw_obj = open(result_path, 'w')
mg_ground_lines = f_img_ground.readlines()
for line in f_img_only.readlines():
line = line.strip()
# get imgid
imgid = os.path.splitext(os.path.basename(line))[0]
imgid = imgid[imgid.find('img'):len(imgid)]
write_line = ''
for save_line in mg_ground_lines:
bfound = False
if save_line.find(imgid + 'segcls.png') >= 0:
temp = save_line.split()
write_line = line + ' ' + temp[1] + '\n'
bfound = True
break
if bfound:
fw_obj.writelines(write_line)
else:
basic.outputlogMessage('Image %s can not found ground True' % line)
f_img_only.close()
f_img_ground.close()
fw_obj.close()
return True
def read_test_data(test_file, file_id=None):
"""
read the file list
:param test_file: file tontains the test file list
:param file_id: id is need in caffe for output result, not need in pytorch_deeplab_resnet
:return: True if succeful, False otherwise
"""
if os.path.isfile(test_file) is False:
basic.outputlogMessage('error: file not exist %s' % test_file)
return False
f_obj = open(test_file)
f_lines = f_obj.readlines()
f_obj.close()
if file_id is not None:
fid_obj = open(file_id)
fid_lines = fid_obj.readlines()
fid_obj.close()
if len(f_lines) != len(fid_lines):
basic.outputlogMessage(
'the number of lines in test_file and test_file_id is not the same'
)
return False
for i in range(0, len(f_lines)):
temp = f_lines[i].split()
sample = SampleClass()
sample.image = temp[0]
if len(temp) > 1:
sample.groudT = temp[1]
sample.id = fid_lines[i].strip()
test_data.append(sample)
else:
for i in range(0, len(f_lines)):
temp = f_lines[i].split()
sample = SampleClass()
sample.image = temp[0]
if len(temp) > 1:
sample.groudT = temp[1]
test_data.append(sample)
# prepare file for pytorch_deeplab_resnet
if len(test_data) < 1:
basic.outputlogMessage('error, not input test data ')
return False
# check all image file and ground true file
for sample in test_data:
# check image path
image_basename = os.path.basename(sample.image)
if os.path.isfile(sample.image) is False:
# sample.image = os.path.basename(sample.image)
sample.image = os.path.join(IMpath, image_basename)
if os.path.isfile(sample.image) is False:
basic.outputlogMessage('error, file not exist: %s' % sample.image)
return False
# check ground path
if len(sample.groudT) > 0 and os.path.isfile(sample.groudT) is False:
sample.groudT = os.path.basename(sample.groudT)
sample.groudT = os.path.join(GTpath, sample.groudT)
if len(sample.id) < 1:
sample.id = os.path.splitext(image_basename)[0]
basic.outputlogMessage('read test data completed, sample count %d' %
len(test_data))
return True
def convert_png_result_to_geojson(result_list):
rsimg_obj = RSImageclass()
rsimgPro_obj = RSImgProclass()
if len(result_list) != len(test_data):
basic.outputlogMessage(
'error, the count of results is not the same as input test file')
return False
geojson_without_fix_folder = os.path.join(save_file_folder,
'geojson_without_fix')
io_function.mkdir(geojson_without_fix_folder)
geojson_folder = os.path.join(save_file_folder, 'geojson')
io_function.mkdir(geojson_folder)
basic.outputlogMessage('geojson_without_fix_folder: %s' %
geojson_without_fix_folder)
basic.outputlogMessage('geojson_folder: %s' % geojson_folder)
geojson_list = []
for i in range(0, len(result_list)):
# read geo information
if rsimg_obj.open(test_data[i].image) is False:
return False
result_file = os.path.join(save_file_folder, test_data[i].id + '.png')
if result_file not in result_list:
basic.outputlogMessage('result_file file not in the list %s' %
result_file)
return False
basic.outputlogMessage(
'png to geojson: %d / %d :Convert file: %s' %
(i + 1, len(result_list), os.path.basename(result_file)))
prj = rsimg_obj.GetProjection()
geom = rsimg_obj.GetGeoTransform()
# read pixel value
im_data = rsimgPro_obj.Read_Image_band_data_to_numpy_array_all_pixel(
1, result_file)
if im_data is False:
basic.outputlogMessage('Read image data failed: %s' % result_file)
return False
# im = Image.open(result_file) # Can be many different formats.
# pix = im.load()
# im_data = numpy.array(pix)
imgID = test_data[i].id #test_data[i][2]
geojson = geoJSONfromCluster.CreateGeoJSON(geojson_without_fix_folder,
imgID, im_data, geom, prj)
fix_geojson = FixGeoJSON.FixGeoJSON(geojson, geojson_folder)
geojson_list.append(fix_geojson)
return geojson_list
def read_test_data(test_file, file_id):
if os.path.isfile(test_file) is False:
basic.outputlogMessage('error: file not exist %s' % test_file)
return False
f_obj = open(test_file)
f_lines = f_obj.readlines()
f_obj.close()
if file_id is not None:
fid_obj = open(file_id)
fid_lines = fid_obj.readlines()
fid_obj.close()
if len(f_lines) != len(fid_lines):
basic.outputlogMessage(
'the number of lines in test_file and test_file_id is not the same'
)
return False
for i in range(0, len(f_lines)):
temp = f_lines[i].split()
sample = SampleClass()
sample.image = temp[0]
if len(temp) > 1:
sample.groudT = temp[1]
sample.id = fid_lines[i].strip()
test_data.append(sample)
else:
for i in range(0, len(f_lines)):
temp = f_lines[i].split()
sample = SampleClass()
sample.image = temp[0]
if len(temp) > 1:
sample.groudT = temp[1]
test_data.append(sample)
# prepare file for pytorch_deeplab_resnet
if len(test_data) < 1:
basic.outputlogMessage('error, not input test data ')
return False
# check all image file and ground true file
for sample in test_data:
# check image path
image_basename = os.path.basename(sample.image)
if os.path.isfile(sample.image) is False:
basic.outputlogMessage('error, file not exist: %s' % sample.image)
return False
# check ground path
if len(sample.groudT) > 0 and os.path.isfile(sample.groudT) is False:
basic.outputlogMessage('error, file not exist: %s' % sample.groudT)
return False
# if len(sample.id)< 1:
# sample.id = os.path.splitext(image_basename)[0]
basic.outputlogMessage('read test data completed, sample count %d' %
len(test_data))
return True
import basic.basic as basic
import basic.io_function as io_function
from basic.RSImage import RSImageclass
from basic.RSImageProcess import RSImgProclass
import basic.mat_To_png as mat_To_png
# sys.path.insert(0, codes_path + '/SpaceNetData')
import SpaceNetData.geoJSONfromCluster as geoJSONfromCluster
import SpaceNetData.FixGeoJSON as FixGeoJSON
#sys.path.insert(0, os.getcwd() + '../SpaceNetChallenge/')
import SpaceNetChallenge.utilities.python.createCSVFromGEOJSON as createCSVFromGEOJSON
if len(sys.argv) == 2:
expr = sys.argv[1]
basic.outputlogMessage('set expr as : %s' % expr)
if len(sys.argv) == 3:
expr = sys.argv[1]
gpuid = int(sys.argv[2])
basic.outputlogMessage('set expr as : %s' % expr)
basic.outputlogMessage('set gpuid as : %d' % gpuid)
if os.path.isdir(expr) is False:
print 'error, %s not exist ' % expr
exit(1)
run_deeplab.EXP = expr
run_deeplab.DEV_ID = gpuid
run_deeplab.NET_ID = NET_ID
test_file = expr + '/list/val.txt'
def run_evalpyt(gpu0, model_path, test_data, save_dir):
model = getattr(deeplab_resnet, 'Res_Deeplab')()
model.eval()
model.cuda(gpu0)
img_rows = 650
img_cols = 650
band = 3
# model_path = os.path.join('data/snapshots/', snapPrefix + str(iter) + '000.pth')
basic.outputlogMessage('the trained model file used in this test is: %s' %
model_path)
saved_state_dict = torch.load(model_path)
model.load_state_dict(saved_state_dict)
save_png_dir = save_dir
if os.path.isdir(save_png_dir) is False:
io_function.mkdir(save_png_dir)
pytorch_list = []
result_list = []
count = 0
for test_sample in test_data:
img = np.zeros((663, 663, band))
count = count + 1
img_path = test_sample.image
# img_temp = cv2.imread(os.path.join(im_path,i[:-1]+'.jpg')).astype(float)
img_temp = cv2.imread(img_path).astype(float)
img_original = img_temp
img_temp[:, :, 0] = img_temp[:, :, 0] - 104.008
img_temp[:, :, 1] = img_temp[:, :, 1] - 116.669
img_temp[:, :, 2] = img_temp[:, :, 2] - 122.675
img[:img_temp.shape[0], :img_temp.shape[1], :] = img_temp
# gt = cv2.imread(os.path.join(gt_path,i[:-1]+'.png'),0)
output = model(
Variable(torch.from_numpy(img[np.newaxis, :].transpose(0, 3, 1,
2)).float(),
volatile=True).cuda(gpu0))
interp = nn.UpsamplingBilinear2d(size=(663, 663))
output = interp(output[3]).cpu().data[0].numpy()
output = output[:, :img_temp.shape[0], :img_temp.shape[1]]
output = output.transpose(1, 2, 0)
output = np.argmax(output, axis=2)
# save to png
result = np.uint8(output)
result = result.reshape(img_rows, img_cols)
result[np.where(result == 1)] = 255
save_png = os.path.join(save_png_dir, test_sample.id + '.png')
misc.imsave(save_png, result)
result_list.append(save_png)
basic.outputlogMessage('%d/%d : saved %s' %
(count, len(test_data), save_png))
# if args['--visualize']:
# plt.subplot(3, 1, 1)
# plt.imshow(img_original)
# plt.subplot(3, 1, 2)
# plt.imshow(gt)
# plt.subplot(3, 1, 3)
# plt.imshow(output)
# plt.show()
png_path = test_sample.groudT
if len(png_path) > 0:
gt = cv2.imread(png_path, 0)
gt[gt == 255] = 0
iou_pytorch = get_iou(output, gt)
pytorch_list.append(iou_pytorch)
if len(pytorch_list) > 0:
print('pytorch', iter,
np.sum(np.asarray(pytorch_list)) / len(pytorch_list))
basic.outputlogMessage('pytorch {} , {} \n'.format(
iter,
np.sum(np.asarray(pytorch_list)) / len(pytorch_list)))
return result_list
