def main():
parser = argparse.ArgumentParser(description="Visualize AdmiralNet")
parser.add_argument("--model_file", type=str, required=True)
parser.add_argument("--input_file", type=str, required=True)
parser.add_argument("--layer", type=str, required=True)
args = parser.parse_args()
model_dir, model_file = os.path.split(args.model_file)
config_path = os.path.join(model_dir,'config.pkl')
config_file = open(config_path,'rb')
config = pickle.load(config_file)
print(config)
gpu = int(config['gpu'])
use_float32 = bool(config['use_float32'])
label_scale = float(config['label_scale'])
size = (66,200)
context_length = int(config['context_length'])
sequence_length = int(config['sequence_length'])
hidden_dim = int(config['hidden_dim'])
optical_flow = bool(config.get('optical_flow',''))
if(optical_flow):
prvs = load_image(args.input_file).astype(np.float32) / 255.0
prvs_grayscale = cv2.cvtColor(prvs,cv2.COLOR_BGR2GRAY)
prvs_resize = cv2.resize(prvs_grayscale, (self.im_size[1], self.im_size[0]), interpolation = cv2.INTER_CUBIC)
flow = cv2.calcOpticalFlowFarneback(prvs_resize,next_resize, None, 0.5, 3, 20, 8, 5, 1.2, 0)
inputfile=flow.transpose(2, 0, 1)
#inputfile=inputfile.reshape(-1,2,1825,300)
else:
inputfile = load_image(args.input_file).astype(np.float32) / 255.0
#inputfile=inputfile.reshape(-1,3,1825,300)
network = models.AdmiralNet(cell='lstm',context_length = context_length, sequence_length=sequence_length, hidden_dim = hidden_dim, use_float32 = use_float32, gpu = gpu, optical_flow=optical_flow)
state_dict = torch.load(args.model_file)
network.load_state_dict(state_dict)
print(network)
get_fmaps(network,inputfile,int(args.layer))
def main():
parser = argparse.ArgumentParser(description="Test AdmiralNet")
parser.add_argument("--model_file", type=str, required=True)
parser.add_argument("--annotation_file", type=str, required=True)
parser.add_argument("--write_images", action="store_true")
parser.add_argument("--plot", action="store_true")
args = parser.parse_args()
annotation_dir, annotation_file = os.path.split(args.annotation_file)
model_dir, model_file = os.path.split(args.model_file)
config_path = os.path.join(model_dir, 'config.pkl')
config_file = open(config_path, 'rb')
config = pickle.load(config_file)
print(config)
model_prefix, _ = model_file.split(".")
# return
gpu = int(config['gpu'])
use_float32 = bool(config['use_float32'])
label_scale = float(config['label_scale'])
size = (66, 200)
prefix, _ = annotation_file.split(".")
prefix = prefix + config['file_prefix']
context_length = int(config['context_length'])
sequence_length = int(config['sequence_length'])
hidden_dim = int(config['hidden_dim'])
optical_flow = bool(config.get('optical_flow', ''))
rnn_cell_type = 'lstm'
network = models.AdmiralNet(cell=rnn_cell_type,
context_length=context_length,
sequence_length=sequence_length,
hidden_dim=hidden_dim,
use_float32=use_float32,
gpu=gpu,
optical_flow=optical_flow)
state_dict = torch.load(args.model_file)
network.load_state_dict(state_dict)
print(network)
#result_data=[]
if (label_scale == 1.0):
label_transformation = None
else:
label_transformation = transforms.Compose(
[transforms.Lambda(lambda inputs: inputs.mul(label_scale))])
if (use_float32):
network.float()
trainset = loaders.F1SequenceDataset(annotation_dir,annotation_file,(66,200),\
context_length=context_length, sequence_length=sequence_length, use_float32=True, label_transformation = label_transformation, optical_flow=optical_flow)
else:
network.double()
trainset = loaders.F1SequenceDataset(annotation_dir, annotation_file,(66,200),\
context_length=context_length, sequence_length=sequence_length, label_transformation = label_transformation, optical_flow=optical_flow)
if (gpu >= 0):
network = network.cuda(gpu)
if optical_flow:
if ((not os.path.isfile("./" + prefix + "_opticalflows.pkl"))
or (not os.path.isfile("./" + prefix +
"_opticalflowannotations.pkl"))):
trainset.read_files_flow()
trainset.write_pickles(prefix + "_opticalflows.pkl",
prefix + "_opticalflowannotations.pkl")
else:
trainset.read_pickles(prefix + "_opticalflows.pkl",
prefix + "_opticalflowannotations.pkl")
else:
if ((not os.path.isfile("./" + prefix + "_images.pkl"))
or (not os.path.isfile("./" + prefix + "_annotations.pkl"))):
trainset.read_files()
trainset.write_pickles(prefix + "_images.pkl",
prefix + "_annotations.pkl")
else:
trainset.read_pickles(prefix + "_images.pkl",
prefix + "_annotations.pkl")
trainset.img_transformation = config['image_transformation']
loader = torch.utils.data.DataLoader(trainset,
batch_size=1,
shuffle=False,
num_workers=0)
cum_diff = 0.0
t = tqdm(enumerate(loader))
network.eval()
predictions = []
if args.write_images:
imdir = "admiralnet_prediction_images_" + model_prefix
os.mkdir(imdir)
annotation_file = open(args.annotation_file, 'r')
annotations = annotation_file.readlines()
annotation_file.close()
im, _, _, _, _ = annotations[0].split(",")
background = cv2.imread(
os.path.join(annotation_dir, 'raw_images_full', im),
cv2.IMREAD_UNCHANGED)
out_size = background.shape
fourcc = cv2.VideoWriter_fourcc(*'XVID')
videoout = cv2.VideoWriter(os.path.join(imdir, "video.avi"), fourcc,
30.0, (out_size[1], out_size[0]), True)
wheel_pred = cv2.imread('predicted_fixed.png', cv2.IMREAD_UNCHANGED)
wheelrows_pred = 65
wheelcols_pred = 65
wheel_pred = cv2.resize(wheel_pred, (wheelcols_pred, wheelrows_pred),
interpolation=cv2.INTER_CUBIC)
for idx, (inputs, throttle, brake, _, labels) in t:
if (gpu >= 0):
inputs = inputs.cuda(gpu)
throttle = throttle.cuda(gpu)
brake = brake.cuda(gpu)
labels = labels.cuda(gpu)
pred = torch.div(network(inputs, throttle, brake), label_scale)
#result_data.append([labels,pred])
if pred.shape[1] == 1:
angle = pred.item()
else:
angle = pred.squeeze()[0].item()
predictions.append(angle)
t.set_postfix(angle=angle)
if args.write_images:
scaled_pred_angle = 180.0 * angle
M_pred = cv2.getRotationMatrix2D(
(wheelrows_pred / 2, wheelcols_pred / 2), scaled_pred_angle, 1)
wheel_pred_rotated = cv2.warpAffine(
wheel_pred, M_pred, (wheelrows_pred, wheelcols_pred))
numpy_im = np.transpose(trainset.images[idx],
(1, 2, 0)).astype(np.float32)
# print(numpy_im.shape)
im, _, _, _, _ = annotations[idx].split(",")
background = cv2.imread(
os.path.join(annotation_dir, 'raw_images_full', im),
cv2.IMREAD_UNCHANGED)
out_size = background.shape
font = cv2.FONT_HERSHEY_SIMPLEX
bottomLeftCornerOfText = (int(
(out_size[1] - wheelcols_pred) / 2) - 15,
int((out_size[0] - wheelcols_pred) / 2) -
55)
fontScale = 0.45
fontColor = (0, 0, 0)
lineType = 1
overlay = background.copy()
cv2.rectangle(overlay,
(int((out_size[1] - wheelcols_pred) / 2) - 20,
int((out_size[0] - wheelcols_pred) / 2) - 53),
(int((out_size[1] - wheelcols_pred) / 2) + 100,
int((out_size[0] - wheelcols_pred) / 2) - 67),
(255, 255, 255, 0.2), -1)
alpha = 0.5
cv2.addWeighted(overlay, alpha, background, 1 - alpha, 0,
background)
cv2.putText(background, 'Predicted:' + "{0:.2f}".format(angle),
bottomLeftCornerOfText, font, fontScale, fontColor,
lineType)
overlayed_pred = imutils.annotation_utils.overlay_image(
background, wheel_pred_rotated,
int((out_size[1] - wheelcols_pred) / 2) + 10,
int((out_size[0] - wheelcols_pred) / 2) - 150)
name = "ouput_image_" + str(idx) + ".png"
output_path = os.path.join(imdir, name)
cv2.imwrite(output_path, overlayed_pred)
videoout.write(overlayed_pred)
predictions_array = np.array(predictions)
log_name = "ouput_log.txt"
imdir = "admiralnet_prediction_images_" + model_prefix
if (os.path.exists(imdir) == False):
os.mkdir(imdir)
log_output_path = os.path.join(imdir, log_name)
log = predictions_array
with open(log_output_path, "a") as myfile:
for x in log:
myfile.write("{0}\n".format(x))
if args.plot:
fig = plt.figure()
ax = plt.subplot(111)
t = np.linspace(0, len(loader) - 1, len(loader))
ax.plot(t, predictions_array, 'r', label='Predicted')
ax.legend()
plt.savefig("admiralnet_prediction_images_" + model_prefix +
"\plot.jpeg")
plt.show()
def main():
parser = argparse.ArgumentParser(
description="Steering prediction with PilotNet")
parser.add_argument("--config_file",
type=str,
required=True,
help="Config file to use")
args = parser.parse_args()
config_fp = args.config_file
config = load_config(config_fp)
#mandatory parameters
learning_rate = float(config['learning_rate'])
root_dir, annotation_file = os.path.split(config['annotation_file'])
prefix, _ = annotation_file.split(".")
#optional parameters
file_prefix = config['file_prefix']
checkpoint_file = config['checkpoint_file']
load_files = bool(config['load_files'])
use_float32 = bool(config['use_float32'])
optical_flow = bool(config['optical_flow'])
apply_norm = bool(config['apply_normalization'])
label_scale = float(config['label_scale'])
momentum = float(config['momentum'])
batch_size = int(config['batch_size'])
gpu = int(config['gpu'])
epochs = int(config['epochs'])
workers = int(config['workers'])
context_length = int(config['context_length'])
sequence_length = int(config['sequence_length'])
hidden_dim = int(config['hidden_dim'])
_, config_file = os.path.split(config_fp)
config_file_name, _ = config_file.split(".")
output_dir = config_file_name.replace("\n", "")
prefix = prefix + file_prefix
rnn_cell_type = 'lstm'
output_dir = output_dir + "_" + rnn_cell_type
network = models.AdmiralNet(cell=rnn_cell_type,
context_length=context_length,
sequence_length=sequence_length,
hidden_dim=hidden_dim,
use_float32=use_float32,
gpu=gpu)
starting_epoch = 0
if (checkpoint_file != ''):
dir, file = os.path.split(checkpoint_file)
_, ep = file.split("epoch")
num, ext = ep.split(".")
starting_epoch = int(num)
print("Starting Epoch number:", starting_epoch)
state_dict = torch.load(checkpoint_file)
network.load_state_dict(state_dict)
if (label_scale == 1.0):
label_transformation = None
else:
label_transformation = transforms.Compose(
[transforms.Lambda(lambda inputs: inputs.mul(label_scale))])
if (use_float32):
network.float()
trainset = loaders.F1SequenceDataset(root_dir,annotation_file,(66,200),\
context_length=context_length, sequence_length=sequence_length, use_float32=True, label_transformation = label_transformation, optical_flow = optical_flow)
else:
network.double()
trainset = loaders.F1SequenceDataset(root_dir, annotation_file,(66,200),\
context_length=context_length, sequence_length=sequence_length, label_transformation = label_transformation, optical_flow = optical_flow)
if (gpu >= 0):
network = network.cuda(gpu)
# trainset.read_files()
if optical_flow:
if (load_files
or (not os.path.isfile("./" + prefix + "_opticalflows.pkl"))
or (not os.path.isfile("./" + prefix +
"_opticalflowannotations.pkl"))):
trainset.read_files_flow()
trainset.write_pickles(prefix + "_opticalflows.pkl",
prefix + "_opticalflowannotations.pkl")
else:
trainset.read_pickles(prefix + "_opticalflows.pkl",
prefix + "_opticalflowannotations.pkl")
else:
if (load_files or (not os.path.isfile("./" + prefix + "_images.pkl"))
or (not os.path.isfile("./" + prefix + "_annotations.pkl"))):
trainset.read_files()
trainset.write_pickles(prefix + "_images.pkl",
prefix + "_annotations.pkl")
else:
trainset.read_pickles(prefix + "_images.pkl",
prefix + "_annotations.pkl")
im = trainset.images[66]
im = im.transpose(1, 2, 0)
#cv2.namedWindow("im",cv2.WINDOW_AUTOSIZE)
#cv2.imshow("im",im/255.0)
# cv2.waitKey(0)
# cv2.destroyWindow("im")
print("Dataset has type: " + str(im.dtype))
# cv2.imwrite("img.jpg",im)
if apply_norm:
mean, stdev = trainset.statistics()
mean_ = torch.from_numpy(mean).float()
stdev_ = torch.from_numpy(stdev).float()
print("Mean")
print(mean_)
print("Stdev")
print(stdev_)
trainset.img_transformation = transforms.Normalize(mean_, stdev_)
else:
print("Skipping Normalize")
trainset.img_transformation = None
# trainset.img_transformation = transforms.Normalize([2.5374081e-06, -3.1837547e-07] , [3.0699273e-05, 5.9349504e-06])
config['image_transformation'] = trainset.img_transformation
config['label_transformation'] = trainset.label_transformation
print("Using configuration: ", config)
if (not os.path.isdir(output_dir)):
os.makedirs(output_dir)
config_dump = open(os.path.join(output_dir, "config.pkl"), 'wb')
pickle.dump(config, config_dump)
config_dump.close()
trainLoader = torch.utils.data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
num_workers=workers)
print(trainLoader)
#Definition of our loss.
criterion = nn.MSELoss()
# Definition of optimization strategy.
optimizer = optim.SGD(network.parameters(),
lr=learning_rate,
momentum=momentum)
losses = train_model(network,
criterion,
optimizer,
trainLoader,
rnn_cell_type,
prefix,
output_dir,
n_epochs=epochs,
gpu=gpu,
starting_epoch=starting_epoch)
if (optical_flow):
loss_path = os.path.join(output_dir,
"" + prefix + "_" + rnn_cell_type + "_OF.txt")
else:
loss_path = os.path.join(output_dir,
"" + prefix + "_" + rnn_cell_type + ".txt")
f = open(loss_path, "w")
f.write("\n".join(map(lambda x: str(x), losses)))
f.close()
def main():
parser = argparse.ArgumentParser(description="Test AdmiralNet")
parser.add_argument("--model_file", type=str, required=True)
args = parser.parse_args()
model_dir, model_file = os.path.split(args.model_file)
config_path = os.path.join(model_dir, 'config.pkl')
config_file = open(config_path, 'rb')
config = pickle.load(config_file)
model_prefix, _ = model_file.split(".")
gpu = int(config['gpu'])
use_float32 = bool(config['use_float32'])
label_scale = float(config['label_scale'])
context_length = int(config['context_length'])
sequence_length = int(config['sequence_length'])
hidden_dim = int(config['hidden_dim'])
optical_flow = bool(config.get('optical_flow', ''))
rnn_cell_type = 'lstm'
network = models.AdmiralNet(cell=rnn_cell_type,
context_length=context_length,
sequence_length=sequence_length,
hidden_dim=hidden_dim,
use_float32=use_float32,
gpu=gpu,
optical_flow=optical_flow)
state_dict = torch.load(args.model_file)
network.load_state_dict(state_dict)
network = network.float()
network = network.cuda(0)
print(network)
vjoy_max = 32000
throttle = torch.Tensor(1, 10)
brake = torch.Tensor(1, 10)
if (use_float32):
network.float()
else:
network.double()
if (gpu >= 0):
network = network.cuda(gpu)
network.eval()
vj = py_vjoy.vJoy()
vj.capture(1) #1 is the device ID
vj.reset()
js = py_vjoy.Joystick()
js.setAxisXRot(int(round(vjoy_max / 2)))
js.setAxisYRot(int(round(vjoy_max / 2)))
vj.update(js)
time.sleep(2)
inputs = []
'''
'''
wheel_pred = cv2.imread('predicted_fixed.png', cv2.IMREAD_UNCHANGED)
wheelrows_pred = 66
wheelcols_pred = 66
wheel_pred = cv2.resize(wheel_pred, (wheelcols_pred, wheelrows_pred),
interpolation=cv2.INTER_CUBIC)
buffer = numpy_ringbuffer.RingBuffer(capacity=context_length,
dtype=(np.float32, (2, 66, 200)))
dt = 12
context_length = 10
debug = True
app = "F1 2017"
dl = pyf1_datalogger.ScreenVideoCapture()
dl.open(app, 0, 200, 1700, 300)
interp = cv2.INTER_AREA
if debug:
cv2.namedWindow(app, cv2.WINDOW_AUTOSIZE)
pscreen = fill_buffer(buffer,
dl,
dt=dt,
context_length=context_length,
interp=interp)
buffer_torch = torch.rand(1, 10, 2, 66, 200).float()
buffer_torch = buffer_torch.cuda(0)
while (True):
cv2.waitKey(dt)
screen = grab_screen(dl)
screen_grey = cv2.cvtColor(screen, cv2.COLOR_BGR2GRAY)
screen_grey = cv2.resize(screen_grey, (200, 66), interpolation=interp)
flow = cv2.calcOpticalFlowFarneback(pscreen, screen_grey, None, 0.5, 3,
20, 8, 5, 1.2, 0)
im = flow.transpose(2, 0, 1).astype(np.float32)
buffer.append(im)
pscreen = screen_grey
buffer_torch[0] = torch.from_numpy(np.array(buffer))
#print("Input Size: " + str(buffer_torch.size()))
outputs = network(buffer_torch, throttle=None, brake=None)
angle = outputs[0][0].item()
print("Output: " + str(angle))
scaled_pred_angle = 180.0 * angle + 7
M_pred = cv2.getRotationMatrix2D(
(wheelrows_pred / 2, wheelcols_pred / 2), scaled_pred_angle, 1)
wheel_pred_rotated = cv2.warpAffine(wheel_pred, M_pred,
(wheelrows_pred, wheelcols_pred))
background = screen
out_size = background.shape
print(out_size)
print(wheel_pred_rotated.shape)
overlayed_pred = imutils.annotation_utils.overlay_image(
background, wheel_pred_rotated,
int((out_size[1] - wheelcols_pred) / 2),
int((out_size[0] - wheelcols_pred) / 2))
if debug:
cv2.imshow(app, overlayed_pred)
vjoy_angle = -angle * vjoy_max + vjoy_max / 2.0
js.setAxisXRot(int(round(vjoy_angle)))
js.setAxisYRot(int(round(vjoy_angle)))
vj.update(js)
'''
'''
print(buffer.shape)
def main():
parser = argparse.ArgumentParser(description="Test AdmiralNet")
parser.add_argument("--model_file", type=str, required=True)
parser.add_argument("--annotation_file", type=str, required=True)
parser.add_argument("--write_images", action="store_true")
parser.add_argument("--plot", action="store_true")
args = parser.parse_args()
annotation_dir, annotation_file = os.path.split(args.annotation_file)
model_dir, model_file = os.path.split(args.model_file)
config_path = os.path.join(model_dir, 'config.pkl')
config_file = open(config_path, 'rb')
config = pickle.load(config_file)
print(config)
model_prefix, _ = model_file.split(".")
# return
gpu = int(config['gpu'])
use_float32 = bool(config['use_float32'])
label_scale = float(config['label_scale'])
#size = (66,200)
prefix, _ = annotation_file.split(".")
prefix = prefix + config['file_prefix']
context_length = int(config['context_length'])
sequence_length = int(config['sequence_length'])
hidden_dim = int(config['hidden_dim'])
optical_flow = bool(config.get('optical_flow', ''))
rnn_cell_type = 'lstm'
network = models.AdmiralNet(cell=rnn_cell_type,
context_length=context_length,
sequence_length=sequence_length,
hidden_dim=hidden_dim,
use_float32=use_float32,
gpu=gpu)
state_dict = torch.load(args.model_file)
network.load_state_dict(state_dict)
print(network)
#result_data=[]
if (label_scale == 1.0):
label_transformation = None
else:
label_transformation = transforms.Compose(
[transforms.Lambda(lambda inputs: inputs.mul(label_scale))])
if (use_float32):
network.float()
trainset = loaders.F1SequenceDataset(annotation_dir,annotation_file,(66,200),\
context_length=context_length, sequence_length=sequence_length, use_float32=True, label_transformation = label_transformation, optical_flow=optical_flow)
else:
network.double()
trainset = loaders.F1SequenceDataset(annotation_dir, annotation_file,(66,200),\
context_length=context_length, sequence_length=sequence_length, label_transformation = label_transformation, optical_flow=optical_flow)
if (gpu >= 0):
network = network.cuda(gpu)
pickle_dir, _ = annotation_file.split('.')
pickle_dir += '_data'
if optical_flow:
load_files = glob.glob(pickle_dir + '\saved_image_opticalflow*.pkl')
else:
load_files = glob.glob(pickle_dir + '\saved_image*.pkl')
if (len(load_files) == 0):
if optical_flow:
trainset.read_files_flow()
load_files = glob.glob(pickle_dir +
'\saved_image_opticalflow*.pkl')
else:
trainset.read_files()
load_files = glob.glob(pickle_dir + '\saved_image*.pkl')
load_files.sort()
predictions = []
ground_truths = []
losses = []
criterion = nn.MSELoss()
cum_diff = 0.0
if (gpu >= 0):
criterion = criterion.cuda(gpu)
network.eval()
for file in load_files:
#Load partitioned Dataset
if optical_flow:
dir, file = file.split('\\')
prefix, data_type, op, suffix = file.split('_')
data_type = 'labels'
label_file = prefix + '_' + data_type + '_' + op + '_' + suffix
else:
dir, file = file.split('\\')
prefix, data_type, suffix = file.split('_')
data_type = 'labels'
label_file = prefix + '_' + data_type + '_' + suffix
trainset.read_pickles(os.path.join(dir, file),
os.path.join(dir, label_file))
trainset.img_transformation = config['image_transformation']
loader = torch.utils.data.DataLoader(trainset,
batch_size=1,
shuffle=False,
num_workers=0)
t = tqdm(enumerate(loader))
if args.write_images:
imdir = "admiralnet_prediction_images_" + model_prefix
if (not os.path.exists(imdir)):
os.mkdir(imdir)
annotation_file = open(args.annotation_file, 'r')
annotations = annotation_file.readlines()
annotation_file.close()
im, _, _, _, _ = annotations[0].split(",")
background = cv2.imread(
os.path.join(annotation_dir, 'raw_images', im),
cv2.IMREAD_UNCHANGED)
out_size = background.shape
fourcc = cv2.VideoWriter_fourcc(*'XVID')
fps = 60
videoout = cv2.VideoWriter(os.path.join(imdir,
"video.avi"), fourcc, fps,
(out_size[1], out_size[0]), True)
wheel_pred = cv2.imread('predicted_fixed.png',
cv2.IMREAD_UNCHANGED)
wheel_ground = cv2.imread('ground_truth_fixed.png',
cv2.IMREAD_UNCHANGED)
wheelrows_pred = 65
wheelcols_pred = 65
wheel_pred = cv2.resize(wheel_pred,
(wheelcols_pred, wheelrows_pred),
interpolation=cv2.INTER_CUBIC)
wheelrows_ground = 65
wheelcols_ground = 65
wheel_ground = cv2.resize(wheel_ground,
(wheelcols_ground, wheelrows_ground),
interpolation=cv2.INTER_CUBIC)
for idx, (inputs, throttle, brake, _, labels, flag) in t:
if (all(flag.numpy())):
if (gpu >= 0):
inputs = inputs.cuda(gpu)
throttle = throttle.cuda(gpu)
brake = brake.cuda(gpu)
labels = labels.cuda(gpu)
pred = torch.div(network(inputs, throttle, brake), label_scale)
#result_data.append([labels,pred])
if pred.shape[1] == 1:
angle = pred.item()
ground_truth = labels.item()
else:
angle = pred.squeeze()[0].item()
ground_truth = labels.squeeze()[0].item()
predictions.append(angle)
ground_truths.append(ground_truth)
loss = criterion(pred, labels)
losses.append(loss.item())
t.set_postfix(angle=angle, ground_truth=ground_truth)
#print("Ground Truth: %f. Prediction: %f.\n" %(scaled_ground_truth, scaled_angle))
if args.write_images:
scaled_pred_angle = 180.0 * angle
scaled_truth_angle = 180.0 * ground_truth
M_pred = cv2.getRotationMatrix2D(
(wheelrows_pred / 2, wheelcols_pred / 2),
scaled_pred_angle, 1)
wheel_pred_rotated = cv2.warpAffine(
wheel_pred, M_pred, (wheelrows_pred, wheelcols_pred))
M_ground = cv2.getRotationMatrix2D(
(wheelrows_ground / 2, wheelcols_ground / 2),
scaled_truth_angle, 1)
wheel_ground_rotated = cv2.warpAffine(
wheel_ground, M_ground,
(wheelrows_ground, wheelcols_ground))
numpy_im = np.transpose(trainset.images[idx],
(1, 2, 0)).astype(np.float32)
#print(numpy_im.shape)
im, _, _, _, _ = annotations[idx].split(",")
background = cv2.imread(
os.path.join(annotation_dir, 'raw_images', im),
cv2.IMREAD_UNCHANGED)
out_size = background.shape
font = cv2.FONT_HERSHEY_SIMPLEX
bottomLeftCornerOfText = (
int((out_size[1] - wheelcols_pred) / 2) - 90,
int((out_size[0] - wheelcols_pred) / 3) - 25)
bottomLeftCornerOfText2 = (
int((out_size[1] - wheelcols_pred) / 2) + 40,
int((out_size[0] - wheelcols_pred) / 3) - 25)
fontScale = 0.45
fontColor = (0, 0, 0)
lineType = 1
overlay = background.copy()
cv2.rectangle(
overlay,
(int((out_size[1] - wheelcols_pred) / 2) - 95,
int((out_size[0] - wheelcols_pred) / 3) - 23),
(int((out_size[1] - wheelcols_pred) / 2) + 25,
int((out_size[0] - wheelcols_pred) / 3) - 37),
(255, 255, 255, 0.2), -1)
cv2.rectangle(
overlay,
(int((out_size[1] - wheelcols_pred) / 2) + 35,
int((out_size[0] - wheelcols_pred) / 3) - 23),
(int((out_size[1] - wheelcols_pred) / 2) + 180,
int((out_size[0] - wheelcols_pred) / 3) - 37),
(255, 255, 255, 0.2), -1)
alpha = 0.5
cv2.addWeighted(overlay, alpha, background, 1 - alpha, 0,
background)
cv2.putText(background,
'Predicted:' + "{0:.2f}".format(angle),
bottomLeftCornerOfText, font, fontScale,
fontColor, lineType)
cv2.putText(
background,
'Ground Truth:' + "{0:.2f}".format(ground_truth),
bottomLeftCornerOfText2, font, fontScale, fontColor,
lineType)
#print(background.shape)
overlayed_pred = imutils.annotation_utils.overlay_image(
background, wheel_pred_rotated,
int((out_size[1] - wheelcols_pred) / 2) - 60,
int((out_size[0] - wheelcols_pred) / 3))
overlayed_ground = imutils.annotation_utils.overlay_image(
overlayed_pred, wheel_ground_rotated,
int((out_size[1] - wheelcols_ground) / 2) + 75,
int((out_size[0] - wheelcols_ground) / 3))
name = "ouput_image_" + str(idx) + ".png"
output_path = os.path.join(imdir, name)
cv2.imwrite(output_path, overlayed_ground)
videoout.write(overlayed_ground)
else:
break
predictions_array = np.array(predictions)
ground_truths_array = np.array(ground_truths)
log_name = "ouput_log.txt"
imdir = "admiralnet_prediction_images_" + model_prefix
if (os.path.exists(imdir) == False):
os.mkdir(imdir)
log_output_path = os.path.join(imdir, log_name)
log = list(zip(ground_truths_array, predictions_array))
with open(log_output_path, "a") as myfile:
for x in log:
log_item = [x[0], x[1]]
myfile.write("{0},{1}\n".format(log_item[0], log_item[1]))
diffs = np.subtract(predictions_array, ground_truths_array)
rms = np.sqrt(np.mean(np.array(losses)))
nrms = np.sqrt(
np.mean(
np.divide(
np.square(np.array(losses)),
np.multiply(np.mean(np.array(predictions)),
np.mean(np.array(ground_truths))))))
print("RMS Error: ", rms)
print("NRMS Error: ", nrms)
if args.plot:
fig = plt.figure()
ax = plt.subplot(111)
t = np.linspace(0, len(predictions_array) - 1, len(predictions_array))
ax.plot(t, predictions_array, 'r', label='Predicted')
ax.plot(t, ground_truths_array, 'b', label='Ground Truth')
ax.legend()
ax.set_xlabel("Frames")
ax.set_ylabel("Steering")
plt.savefig("admiralnet_prediction_images_" + model_prefix +
"\plot.jpeg")
plt.show()