diff --git a/src/odroid/catkin_ws/src/car/scripts/saliency.py b/src/odroid/catkin_ws/src/car/scripts/saliency.py
index 4d9b7c7a05613f230ea80d0ae13b35674e699e6e..6765a65de31cb20f59412ad90579709cf5a353f5 100644
--- a/src/odroid/catkin_ws/src/car/scripts/saliency.py
+++ b/src/odroid/catkin_ws/src/car/scripts/saliency.py
@@ -1,7 +1,7 @@
import sys
sys.path.insert(0, "./deepviz")
-from vis.visualization import visualize_saliency, visualize_cam
+from vis.visualization import visualize_saliency, visualize_cam, overlay
from keras import activations
from utils import *
from vis.utils import utils
@@ -54,32 +54,35 @@ from keras.models import Model
# ax2.imshow(mask)
images = [
- load_image("I:/recordings/data-2018-08-31-10-42-27", "img-819.jpg"), # left, train
- load_image("I:/recordings/data-2018-08-31-10-42-27", "img-625.jpg"), # straight, train
- load_image("I:/recordings/data-2018-08-31-10-42-27", "img-9640.jpg"), # right, train
- load_image("I:/recordings/data-2018-09-07-15-59-32", "img-700.jpg"), # left, valid
+ load_image("I:/recordings/data-2018-08-31-12-12-14", "img-300.jpg"), # left, ideal line
+ load_image("I:/recordings/data-2018-08-31-12-12-14", "img-500.jpg"), # straight, ideal line
+ load_image("I:/recordings/data-2018-08-31-12-12-14", "img-6175.jpg"), # right, ideal line
+
+ load_image("I:/recordings/data-2018-08-31-10-42-27", "img-819.jpg"), # left, center line
+ load_image("I:/recordings/data-2018-08-31-10-42-27", "img-625.jpg"), # straight, center line
+ load_image("I:/recordings/data-2018-08-31-10-42-27", "img-9640.jpg"), # right, center line
]
models = [
- load_model("G:/car/models/model-001.h5"),
- load_model("G:/car/models/model-005.h5"),
load_model("G:/car/models/model-010.h5"),
+ load_model("G:/car/models/model-050.h5"),
+ load_model("G:/car/models/model-100.h5"),
]
-fig, axes = plt.subplots(len(images) * 2, len(models))
+fig, axes = plt.subplots(len(images), len(models))
for j, model in enumerate(models):
- layer_idx = -1
+ layer_idx = 6
model.layers[layer_idx].activation = activations.linear
model = utils.apply_modifications(model)
for i, image in enumerate(images):
image = preprocess_image(image)
- grads = visualize_cam(model, layer_idx, filter_indices=None, seed_input=image, grad_modifier=None)
- axes[(2 * i) + 1, j].imshow(grads)
+ grads = visualize_saliency(model, layer_idx, filter_indices=None, seed_input=image, backprop_modifier="guided")
image = (image + 1.0) * 127.5
image = image.astype(np.uint8)
image = cv2.cvtColor(image, cv2.COLOR_YUV2RGB)
- axes[2 * i, j].imshow(image)
+
+ axes[i, j].imshow(overlay(image, grads, alpha=0.3))
plt.show()
diff --git a/src/odroid/catkin_ws/src/car/scripts/train.py b/src/odroid/catkin_ws/src/car/scripts/train.py
index 20b3ba9319fa9ce46160578b5622e75bde4ea178..8592f67f74481e5336733dc97eedec228f1f2925 100644
--- a/src/odroid/catkin_ws/src/car/scripts/train.py
+++ b/src/odroid/catkin_ws/src/car/scripts/train.py
@@ -63,12 +63,10 @@ def build_model_nvidia(args):
x = Dense(10, activation="elu")(x)
angle_out = Dense(1, name="angle_out", activation="linear")(x)
- throttle_out = Dense(1, name="throttle_out", activation="linear")(x)
- model = Model(inputs=[img_in], outputs=[angle_out, throttle_out])
+ model = Model(inputs=[img_in], outputs=[angle_out])
model.compile(optimizer="adam",
- loss={"angle_out": "mean_squared_error",
- "throttle_out": "mean_squared_error"})
+ loss={"angle_out": "mean_squared_error"})
return model
@@ -113,12 +111,10 @@ def build_model_custom(args):
x = Dense(50, activation="relu")(x)
angle_out = Dense(1, name="angle_out", activation="linear")(x)
- throttle_out = Dense(1, name="throttle_out", activation="linear")(x)
- model = Model(inputs=[img_in], outputs=[angle_out, throttle_out])
+ model = Model(inputs=[img_in], outputs=[angle_out])
model.compile(optimizer="adam",
- loss={"angle_out": "mean_squared_error",
- "throttle_out": "mean_squared_error"})
+ loss={"angle_out": "mean_squared_error"})
return model
@@ -194,7 +190,7 @@ def main():
parser.add_argument('-d', help='data directory', dest='data_dir', type=str, default=default_training_dir)
parser.add_argument('-t', help='test size fraction', dest='test_size', type=float, default=0.2)
parser.add_argument('-k', help='drop out probability', dest='keep_prob', type=float, default=0.5)
- parser.add_argument('-n', help='number of epochs', dest='nb_epoch', type=int, default=100)
+ parser.add_argument('-n', help='number of epochs', dest='nb_epoch', type=int, default=10)
parser.add_argument('-b', help='batch size', dest='batch_size', type=int, default=40)
parser.add_argument('-o', help='save best models only', dest='save_best_only', type=s2b, default='false')
parser.add_argument('-l', help='learning rate', dest='learning_rate', type=float, default=3.0e-4)
@@ -215,7 +211,7 @@ def main():
#load data
data = load_data(args)
#build model
- model = build_model_custom(args)
+ model = build_model_nvidia(args)
#train model on data, it saves as model.h5
train_model(model, args, *data)
diff --git a/src/odroid/catkin_ws/src/car/scripts/utils.py b/src/odroid/catkin_ws/src/car/scripts/utils.py
index 08510829897b7aa426f3f377a93baead9ed9274d..b8bee47a8fc19d2708d959f27bbbc73422551cc8 100644
--- a/src/odroid/catkin_ws/src/car/scripts/utils.py
+++ b/src/odroid/catkin_ws/src/car/scripts/utils.py
@@ -53,7 +53,7 @@ def resize(image):
def normalize(image):
- return ne.evaluate("image / 127.5 - 1.0")
+ return ne.evaluate("(image / 127.5) - 1.0")
def convert(image):
@@ -185,12 +185,12 @@ class BatchSequence(keras.utils.Sequence):
def __getitem__(self, idx):
inputs = np.empty([self.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, IMAGE_CHANNELS])
- outputs = [np.empty(self.batch_size), np.empty(self.batch_size)]
+ outputs = [np.empty(self.batch_size)]
for batch_idx in range(self.batch_size):
seq_idx = (idx * self.batch_size + batch_idx) % len(self.sequence)
sample_idx = self.sequence[seq_idx]
- angle, throttle = self.y[sample_idx]
+ angle, _ = self.y[sample_idx]
image_idx = self.X[sample_idx]
image = self.image_cache.get_image(image_idx)
# augumentation
@@ -198,7 +198,6 @@ class BatchSequence(keras.utils.Sequence):
image, angle = augument(image, angle)
inputs[batch_idx] = normalize(convert(image))
outputs[0][batch_idx] = preprocess_angle(angle)
- outputs[1][batch_idx] = preprocess_throttle(throttle)
return inputs, outputs