class MakeMovie(BaseCommand):
def parse_args(self, args):
parser = argparse.ArgumentParser(prog='makemovie')
parser.add_argument('--tub', help='The tub to make movie from')
parser.add_argument(
'--out',
default='tub_movie.mp4',
help='The movie filename to create. default: tub_movie.mp4')
parser.add_argument(
'--config',
default='./config.py',
help='location of config file to use. default: ./config.py')
parser.add_argument('--model',
default='None',
help='the model to use to show control outputs')
parser.add_argument('--model_type',
default='categorical',
help='the model type to load')
parser.add_argument(
'--salient',
action="store_true",
help='should we overlay salient map showing avtivations')
parser.add_argument('--start',
type=int,
default=1,
help='first frame to process')
parser.add_argument('--end',
type=int,
default=-1,
help='last frame to process')
parser.add_argument('--scale',
type=int,
default=2,
help='make image frame output larger by X mult')
parsed_args = parser.parse_args(args)
return parsed_args, parser
def run(self, args):
'''
Load the images from a tub and create a movie from them.
Movie
'''
import moviepy.editor as mpy
args, parser = self.parse_args(args)
if args.tub is None:
parser.print_help()
return
if args.salient:
#imported like this, we make TF conditional on use of --salient
#and we keep the context maintained throughout our callbacks to
#compute the salient mask
from keras import backend as K
import tensorflow as tf
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
conf = os.path.expanduser(args.config)
if not os.path.exists(conf):
print("No config file at location: %s. Add --config to specify\
location or run from dir containing config.py." % conf)
return
try:
cfg = dk.load_config(conf)
except:
print("Exception while loading config from", conf)
return
self.tub = Tub(args.tub)
self.num_rec = self.tub.get_num_records()
if args.start == 1:
self.start = self.tub.get_index(shuffled=False)[0]
else:
self.start = args.start
if args.end != -1:
self.end = args.end
else:
self.end = self.num_rec - self.start
self.num_rec = self.end - self.start
self.iRec = args.start
self.scale = args.scale
self.keras_part = None
self.convolution_part = None
if not args.model == "None":
self.keras_part = get_model_by_type(args.model_type, cfg=cfg)
self.keras_part.load(args.model)
self.keras_part.compile()
if args.salient:
self.init_salient(self.keras_part.model)
#This method nested in this way to take the conditional import of TF
#in a manner that extends to this callback. Done this way, we avoid
#importing in the below method, which triggers a new cuda device allocation
#each call.
def compute_visualisation_mask(img):
#from https://github.com/ermolenkodev/keras-salient-object-visualisation
activations = self.functor([np.array([img])])
activations = [
np.reshape(
img, (1, img.shape[0], img.shape[1], img.shape[2]))
] + activations
upscaled_activation = np.ones((3, 6))
for layer in [5, 4, 3, 2, 1]:
averaged_activation = np.mean(
activations[layer],
axis=3).squeeze(axis=0) * upscaled_activation
output_shape = (activations[layer - 1].shape[1],
activations[layer - 1].shape[2])
x = tf.constant(
np.reshape(averaged_activation,
(1, averaged_activation.shape[0],
averaged_activation.shape[1], 1)),
tf.float32)
conv = tf.nn.conv2d_transpose(
x,
self.layers_kernels[layer],
output_shape=(1, output_shape[0], output_shape[1],
1),
strides=self.layers_strides[layer],
padding='VALID')
with tf.Session() as session:
result = session.run(conv)
upscaled_activation = np.reshape(result, output_shape)
final_visualisation_mask = upscaled_activation
return (final_visualisation_mask -
np.min(final_visualisation_mask)) / (
np.max(final_visualisation_mask) -
np.min(final_visualisation_mask))
self.compute_visualisation_mask = compute_visualisation_mask
print('making movie', args.out, 'from', self.num_rec, 'images')
clip = mpy.VideoClip(self.make_frame,
duration=(self.num_rec // cfg.DRIVE_LOOP_HZ) - 1)
clip.write_videofile(args.out, fps=cfg.DRIVE_LOOP_HZ)
print('done')
def draw_model_prediction(self, record, img):
'''
query the model for it's prediction, draw the user input and the predictions
as green and blue lines on the image
'''
if self.keras_part is None:
return
import cv2
user_angle = float(record["user/angle"])
user_throttle = float(record["user/throttle"])
pilot_angle, pilot_throttle = self.keras_part.run(img)
a1 = user_angle * 45.0
l1 = user_throttle * 3.0 * 80.0
a2 = pilot_angle * 45.0
l2 = pilot_throttle * 3.0 * 80.0
p1 = tuple((74, 119))
p2 = tuple((84, 119))
p11 = tuple(
(int(p1[0] + l1 * math.cos((a1 + 270.0) * math.pi / 180.0)),
int(p1[1] + l1 * math.sin((a1 + 270.0) * math.pi / 180.0))))
p22 = tuple(
(int(p2[0] + l2 * math.cos((a2 + 270.0) * math.pi / 180.0)),
int(p2[1] + l2 * math.sin((a2 + 270.0) * math.pi / 180.0))))
cv2.line(img, p1, p11, (0, 255, 0), 2)
cv2.line(img, p2, p22, (0, 0, 255), 2)
def init_salient(self, model):
#from https://github.com/ermolenkodev/keras-salient-object-visualisation
from keras.layers import Input, Dense, merge
from keras.models import Model
from keras.layers import Convolution2D, MaxPooling2D, Reshape, BatchNormalization
from keras.layers import Activation, Dropout, Flatten, Dense
img_in = Input(shape=(120, 160, 3), name='img_in')
x = img_in
x = Convolution2D(24, (5, 5),
strides=(2, 2),
activation='relu',
name='conv1')(x)
x = Convolution2D(32, (5, 5),
strides=(2, 2),
activation='relu',
name='conv2')(x)
x = Convolution2D(64, (5, 5),
strides=(2, 2),
activation='relu',
name='conv3')(x)
x = Convolution2D(64, (3, 3),
strides=(2, 2),
activation='relu',
name='conv4')(x)
conv_5 = Convolution2D(64, (3, 3),
strides=(1, 1),
activation='relu',
name='conv5')(x)
self.convolution_part = Model(inputs=[img_in], outputs=[conv_5])
for layer_num in ('1', '2', '3', '4', '5'):
self.convolution_part.get_layer('conv' + layer_num).set_weights(
model.get_layer('conv2d_' + layer_num).get_weights())
from keras import backend as K
import tensorflow as tf
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
self.inp = self.convolution_part.input # input placeholder
self.outputs = [
layer.output for layer in self.convolution_part.layers[1:]
] # all layer outputs
self.functor = K.function([self.inp], self.outputs)
kernel_3x3 = tf.constant(
np.array([[[[1]], [[1]], [[1]]], [[[1]], [[1]], [[1]]],
[[[1]], [[1]], [[1]]]]), tf.float32)
kernel_5x5 = tf.constant(
np.array([[[[1]], [[1]], [[1]], [[1]], [[1]]],
[[[1]], [[1]], [[1]], [[1]], [[1]]],
[[[1]], [[1]], [[1]], [[1]], [[1]]],
[[[1]], [[1]], [[1]], [[1]], [[1]]],
[[[1]], [[1]], [[1]], [[1]], [[1]]]]), tf.float32)
self.layers_kernels = {
5: kernel_3x3,
4: kernel_3x3,
3: kernel_5x5,
2: kernel_5x5,
1: kernel_5x5
}
self.layers_strides = {
5: [1, 1, 1, 1],
4: [1, 2, 2, 1],
3: [1, 2, 2, 1],
2: [1, 2, 2, 1],
1: [1, 2, 2, 1]
}
def draw_salient(self, img):
#from https://github.com/ermolenkodev/keras-salient-object-visualisation
import cv2
alpha = 0.004
beta = 1.0 - alpha
salient_mask = self.compute_visualisation_mask(img)
salient_mask_stacked = np.dstack((salient_mask, salient_mask))
salient_mask_stacked = np.dstack((salient_mask_stacked, salient_mask))
blend = cv2.addWeighted(img.astype('float32'), alpha,
salient_mask_stacked, beta, 0.0)
return blend
def make_frame(self, t):
'''
Callback to return an image from from our tub records.
This is called from the VideoClip as it references a time.
We don't use t to reference the frame, but instead increment
a frame counter. This assumes sequential access.
'''
if self.iRec >= self.end:
return None
rec = None
while rec is None and self.iRec < self.end:
try:
rec = self.tub.get_record(self.iRec)
except Exception as e:
print(e)
print("Failed to get image for frame", self.iRec)
self.iRec = self.iRec + 1
rec = None
image = rec['cam/image_array']
if self.convolution_part:
image = self.draw_salient(image)
image = image * 255
image = image.astype('uint8')
self.draw_model_prediction(rec, image)
if self.scale != 1:
import cv2
h, w, d = image.shape
dsize = (w * self.scale, h * self.scale)
image = cv2.resize(image,
dsize=dsize,
interpolation=cv2.INTER_CUBIC)
self.iRec = self.iRec + 1
return image # returns a 8-bit RGB array
class MakeMovie(object):
def __init__(self):
self.deg_to_rad = math.pi / 180.0
def run(self, args, parser):
'''
Load the images from a tub and create a movie from them.
Movie
'''
if args.tub is None:
print("ERR>> --tub argument missing.")
parser.print_help()
return
if args.type is None and args.model is not None:
print("ERR>> --type argument missing. Required when providing a model.")
parser.print_help()
return
if args.salient:
if args.model is None:
print("ERR>> salient visualization requires a model. Pass with the --model arg.")
parser.print_help()
conf = os.path.expanduser(args.config)
if not os.path.exists(conf):
print("No config file at location: %s. Add --config to specify\
location or run from dir containing config.py." % conf)
return
self.cfg = dk.load_config(conf)
self.tub = Tub(args.tub)
self.index = self.tub.get_index(shuffled=False)
start = args.start
self.end = args.end if args.end != -1 else len(self.index)
if self.end >= len(self.index):
self.end = len(self.index) - 1
num_frames = self.end - start
self.iRec = start
self.scale = args.scale
self.keras_part = None
self.do_salient = False
if args.model is not None:
self.keras_part = get_model_by_type(args.type, cfg=self.cfg)
self.keras_part.load(args.model)
self.keras_part.compile()
if args.salient:
self.do_salient = self.init_salient(self.keras_part.model)
print('making movie', args.out, 'from', num_frames, 'images')
clip = mpy.VideoClip(self.make_frame,
duration=((num_frames - 1) / self.cfg.DRIVE_LOOP_HZ))
clip.write_videofile(args.out, fps=self.cfg.DRIVE_LOOP_HZ)
def draw_user_input(self, record, img):
'''
Draw the user input as a green line on the image
'''
import cv2
user_angle = float(record["user/angle"])
user_throttle = float(record["user/throttle"])
height, width, _ = img.shape
length = height
a1 = user_angle * 45.0
l1 = user_throttle * length
mid = width // 2 - 1
p1 = tuple((mid - 2, height - 1))
p11 = tuple((int(p1[0] + l1 * math.cos((a1 + 270.0) * self.deg_to_rad)),
int(p1[1] + l1 * math.sin((a1 + 270.0) * self.deg_to_rad))))
# user is green, pilot is blue
cv2.line(img, p1, p11, (0, 255, 0), 2)
def draw_model_prediction(self, record, img):
'''
query the model for it's prediction, draw the predictions
as a blue line on the image
'''
if self.keras_part is None:
return
import cv2
expected = self.keras_part.model.inputs[0].shape[1:]
actual = img.shape
# normalize image before prediction
pred_img = img.astype(np.float32) / 255.0
# check input depth
if expected[2] == 1 and actual[2] == 3:
pred_img = rgb2gray(pred_img)
pred_img = pred_img.reshape(pred_img.shape + (1,))
actual = pred_img.shape
if expected != actual:
print("expected input dim", expected, "didn't match actual dim", actual)
return
pilot_angle, pilot_throttle = self.keras_part.run(pred_img)
height, width, _ = pred_img.shape
length = height
a2 = pilot_angle * 45.0
l2 = pilot_throttle * length
mid = width // 2 - 1
p2 = tuple((mid + 2, height - 1))
p22 = tuple((int(p2[0] + l2 * math.cos((a2 + 270.0) * self.deg_to_rad)),
int(p2[1] + l2 * math.sin((a2 + 270.0) * self.deg_to_rad))))
# user is green, pilot is blue
cv2.line(img, p2, p22, (0, 0, 255), 2)
def draw_steering_distribution(self, record, img):
'''
query the model for it's prediction, draw the distribution of steering choices
'''
from donkeycar.parts.keras import KerasCategorical
if self.keras_part is None or type(self.keras_part) is not KerasCategorical:
return
import cv2
pred_img = img.reshape((1,) + img.shape)
angle_binned, _ = self.keras_part.model.predict(pred_img)
x = 4
dx = 4
y = 120 - 4
iArgMax = np.argmax(angle_binned)
for i in range(15):
p1 = (x, y)
p2 = (x, y - int(angle_binned[0][i] * 100.0))
if i == iArgMax:
cv2.line(img, p1, p2, (255, 0, 0), 2)
else:
cv2.line(img, p1, p2, (200, 200, 200), 2)
x += dx
def init_salient(self, model):
# Utility to search for layer index by name.
# Alternatively we can specify this as -1 since it corresponds to the last layer.
first_output_name = None
for i, layer in enumerate(model.layers):
if first_output_name is None and "dropout" not in layer.name.lower() and "out" in layer.name.lower():
first_output_name = layer.name
layer_idx = i
if first_output_name is None:
print("Failed to find the model layer named with 'out'. Skipping salient.")
return False
print("####################")
print("Visualizing activations on layer:", first_output_name)
print("####################")
# ensure we have linear activation
model.layers[layer_idx].activation = activations.linear
# build salient model and optimizer
sal_model = utils.apply_modifications(model)
modifier_fn = get('guided')
sal_model_mod = modifier_fn(sal_model)
losses = [
(ActivationMaximization(sal_model_mod.layers[layer_idx], None), -1)
]
self.opt = Optimizer(sal_model_mod.input, losses, norm_grads=False)
return True
def compute_visualisation_mask(self, img):
grad_modifier = 'absolute'
grads = self.opt.minimize(seed_input=img, max_iter=1, grad_modifier=grad_modifier, verbose=False)[1]
channel_idx = 1 if K.image_data_format() == 'channels_first' else -1
grads = np.max(grads, axis=channel_idx)
res = utils.normalize(grads)[0]
return res
def draw_salient(self, img):
import cv2
alpha = 0.004
beta = 1.0 - alpha
expected = self.keras_part.model.inputs[0].shape[1:]
actual = img.shape
pred_img = img.astype(np.float32) / 255.0
# check input depth
if expected[2] == 1 and actual[2] == 3:
pred_img = rgb2gray(pred_img)
pred_img = pred_img.reshape(pred_img.shape + (1,))
salient_mask = self.compute_visualisation_mask(pred_img)
z = np.zeros_like(salient_mask)
salient_mask_stacked = np.dstack((z, z))
salient_mask_stacked = np.dstack((salient_mask_stacked, salient_mask))
blend = cv2.addWeighted(img.astype('float32'), alpha, salient_mask_stacked, beta, 0.0)
return blend
def make_frame(self, t):
'''
Callback to return an image from from our tub records.
This is called from the VideoClip as it references a time.
We don't use t to reference the frame, but instead increment
a frame counter. This assumes sequential access.
'''
if self.iRec >= self.end or self.iRec >= len(self.index):
return None
rec_ix = self.index[self.iRec]
rec = self.tub.get_record(rec_ix)
image = rec['cam/image_array']
if self.cfg.ROI_CROP_TOP != 0 or self.cfg.ROI_CROP_BOTTOM != 0:
image = img_crop(image, self.cfg.ROI_CROP_TOP, self.cfg.ROI_CROP_BOTTOM)
if self.do_salient:
image = self.draw_salient(image)
image = image * 255
image = image.astype('uint8')
self.draw_user_input(rec, image)
if self.keras_part is not None:
self.draw_model_prediction(rec, image)
self.draw_steering_distribution(rec, image)
if self.scale != 1:
import cv2
h, w, d = image.shape
dsize = (w * self.scale, h * self.scale)
image = cv2.resize(image, dsize=dsize, interpolation=cv2.INTER_CUBIC)
self.iRec += 1
# returns a 8-bit RGB array
return image
class MakeMovie(BaseCommand):
def __init__(self):
self.deg_to_rad = math.pi / 180.0
def parse_args(self, args):
parser = argparse.ArgumentParser(prog='makemovie')
parser.add_argument('--tub', help='The tub to make movie from')
parser.add_argument(
'--out',
default='tub_movie.mp4',
help='The movie filename to create. default: tub_movie.mp4')
parser.add_argument(
'--config',
default='./config.py',
help='location of config file to use. default: ./config.py')
parser.add_argument('--model',
help='the model to use to show control outputs')
parser.add_argument('--type', help='the model type to load')
parser.add_argument(
'--salient',
action="store_true",
help='should we overlay salient map showing avtivations')
parser.add_argument('--start',
type=int,
default=1,
help='first frame to process')
parser.add_argument('--end',
type=int,
default=-1,
help='last frame to process')
parser.add_argument('--scale',
type=int,
default=2,
help='make image frame output larger by X mult')
parsed_args = parser.parse_args(args)
return parsed_args, parser
def run(self, args):
'''
Load the images from a tub and create a movie from them.
Movie
'''
import moviepy.editor as mpy
args, parser = self.parse_args(args)
if args.tub is None:
print("ERR>> --tub argument missing.")
parser.print_help()
return
if args.model is not None and args.type is None:
print("ERR>> --type argument missing.")
parser.print_help()
return
if args.salient:
if args.model is None in args.model:
print(
"ERR>> salient visualization requires a model. Pass with the --model arg."
)
parser.print_help()
return
# imported like this, we make TF conditional on use of --salient
# and we keep the context maintained throughout our callbacks to
# compute the salient mask
from tensorflow.python.keras import backend as K
import tensorflow as tf
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
conf = os.path.expanduser(args.config)
if not os.path.exists(conf):
print("No config file at location: %s. Add --config to specify\
location or run from dir containing config.py." % conf)
return
self.cfg = dk.load_config(conf)
self.tub = Tub(args.tub)
self.index = self.tub.get_index(shuffled=False)
start = args.start
self.end = args.end if args.end != -1 else len(self.index)
if self.end >= len(self.index):
self.end = len(self.index) - 1
num_frames = self.end - start
self.iRec = start
self.scale = args.scale
self.keras_part = None
self.convolution_part = None
if not args.model is None:
self.keras_part = get_model_by_type(args.type, cfg=self.cfg)
self.keras_part.load(args.model)
self.keras_part.compile()
if args.salient:
self.init_salient(self.keras_part.model)
#This method nested in this way to take the conditional import of TF
#in a manner that extends to this callback. Done this way, we avoid
#importing in the below method, which triggers a new cuda device allocation
#each call.
def compute_visualisation_mask(img):
#from https://github.com/ermolenkodev/keras-salient-object-visualisation
activations = self.functor([np.array([img])])
activations = [
np.reshape(
img, (1, img.shape[0], img.shape[1], img.shape[2]))
] + activations
upscaled_activation = np.ones((3, 6))
for layer in [5, 4, 3, 2, 1]:
averaged_activation = np.mean(
activations[layer],
axis=3).squeeze(axis=0) * upscaled_activation
output_shape = (activations[layer - 1].shape[1],
activations[layer - 1].shape[2])
x = tf.constant(
np.reshape(averaged_activation,
(1, averaged_activation.shape[0],
averaged_activation.shape[1], 1)),
tf.float32)
conv = tf.nn.conv2d_transpose(
x,
self.layers_kernels[layer],
output_shape=(1, output_shape[0], output_shape[1],
1),
strides=self.layers_strides[layer],
padding='VALID')
with tf.Session() as session:
result = session.run(conv)
upscaled_activation = np.reshape(result, output_shape)
final_visualisation_mask = upscaled_activation
return (final_visualisation_mask -
np.min(final_visualisation_mask)) / (
np.max(final_visualisation_mask) -
np.min(final_visualisation_mask))
self.compute_visualisation_mask = compute_visualisation_mask
print('making movie', args.out, 'from', num_frames, 'images')
clip = mpy.VideoClip(self.make_frame,
duration=((num_frames - 1) /
self.cfg.DRIVE_LOOP_HZ))
clip.write_videofile(args.out, fps=self.cfg.DRIVE_LOOP_HZ)
def draw_model_prediction(self, record, img):
'''
query the model for it's prediction, draw the user input and the predictions
as green and blue lines on the image
'''
if self.keras_part is None:
return
import cv2
user_angle = float(record["user/angle"])
user_throttle = float(record["user/throttle"])
expected = self.keras_part.model.inputs[0].shape[1:]
actual = img.shape
pred_img = img
# check input depth
if expected[2] == 1 and actual[2] == 3:
pred_img = rgb2gray(pred_img)
pred_img = pred_img.reshape(pred_img.shape + (1, ))
actual = pred_img.shape
if expected != actual:
print("expected input dim", expected, "didn't match actual dim",
actual)
return
pilot_angle, pilot_throttle = self.keras_part.run(pred_img)
length = self.cfg.IMAGE_H
a1 = user_angle * 45.0
l1 = user_throttle * length
a2 = pilot_angle * 45.0
l2 = pilot_throttle * length
mid = self.cfg.IMAGE_W // 2 - 1
p1 = tuple((mid - 2, self.cfg.IMAGE_H - 1))
p2 = tuple((mid + 2, self.cfg.IMAGE_H - 1))
p11 = tuple(
(int(p1[0] + l1 * math.cos((a1 + 270.0) * self.deg_to_rad)),
int(p1[1] + l1 * math.sin((a1 + 270.0) * self.deg_to_rad))))
p22 = tuple(
(int(p2[0] + l2 * math.cos((a2 + 270.0) * self.deg_to_rad)),
int(p2[1] + l2 * math.sin((a2 + 270.0) * self.deg_to_rad))))
# user is green, pilot is blue
cv2.line(img, p1, p11, (0, 255, 0), 2)
cv2.line(img, p2, p22, (0, 0, 255), 2)
def draw_steering_distribution(self, record, img):
'''
query the model for it's prediction, draw the distribution of steering choices
'''
from donkeycar.parts.keras import KerasCategorical
if self.keras_part is None or type(
self.keras_part) is not KerasCategorical:
return
import cv2
pred_img = img.reshape((1, ) + img.shape)
angle_binned, throttle = self.keras_part.model.predict(pred_img)
x = 4
dx = 4
y = 120 - 4
iArgMax = np.argmax(angle_binned)
for i in range(15):
p1 = (x, y)
p2 = (x, y - int(angle_binned[0][i] * 100.0))
if i == iArgMax:
cv2.line(img, p1, p2, (255, 0, 0), 2)
else:
cv2.line(img, p1, p2, (200, 200, 200), 2)
x += dx
def init_salient(self, model):
#from https://github.com/ermolenkodev/keras-salient-object-visualisation
from tensorflow.python.keras.layers import Input, Dense, merge
from tensorflow.python.keras.models import Model
from tensorflow.python.keras.layers import Convolution2D, MaxPooling2D, Reshape, BatchNormalization
from tensorflow.python.keras.layers import Activation, Dropout, Flatten, Dense
input_shape = model.inputs[0].shape
img_in = Input(shape=(input_shape[1], input_shape[2], input_shape[3]),
name='img_in')
x = img_in
x = Convolution2D(24, (5, 5),
strides=(2, 2),
activation='relu',
name='conv1')(x)
x = Convolution2D(32, (5, 5),
strides=(2, 2),
activation='relu',
name='conv2')(x)
x = Convolution2D(64, (5, 5),
strides=(2, 2),
activation='relu',
name='conv3')(x)
x = Convolution2D(64, (3, 3),
strides=(2, 2),
activation='relu',
name='conv4')(x)
conv_5 = Convolution2D(64, (3, 3),
strides=(1, 1),
activation='relu',
name='conv5')(x)
self.convolution_part = Model(inputs=[img_in], outputs=[conv_5])
for layer_num in ('1', '2', '3', '4', '5'):
try:
self.convolution_part.get_layer(
'conv' + layer_num).set_weights(
model.get_layer('conv2d_' + layer_num).get_weights())
except Exception as e:
print(e)
print("Failed to load layer weights for layer", layer_num)
raise Exception("Failed to load weights")
from tensorflow.python.keras import backend as K
import tensorflow as tf
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
self.inp = self.convolution_part.input # input placeholder
self.outputs = [
layer.output for layer in self.convolution_part.layers[1:]
] # all layer outputs
self.functor = K.function([self.inp], self.outputs)
kernel_3x3 = tf.constant(
np.array([[[[1]], [[1]], [[1]]], [[[1]], [[1]], [[1]]],
[[[1]], [[1]], [[1]]]]), tf.float32)
kernel_5x5 = tf.constant(
np.array([[[[1]], [[1]], [[1]], [[1]], [[1]]],
[[[1]], [[1]], [[1]], [[1]], [[1]]],
[[[1]], [[1]], [[1]], [[1]], [[1]]],
[[[1]], [[1]], [[1]], [[1]], [[1]]],
[[[1]], [[1]], [[1]], [[1]], [[1]]]]), tf.float32)
self.layers_kernels = {
5: kernel_3x3,
4: kernel_3x3,
3: kernel_5x5,
2: kernel_5x5,
1: kernel_5x5
}
self.layers_strides = {
5: [1, 1, 1, 1],
4: [1, 2, 2, 1],
3: [1, 2, 2, 1],
2: [1, 2, 2, 1],
1: [1, 2, 2, 1]
}
def draw_salient(self, img):
# from https://github.com/ermolenkodev/keras-salient-object-visualisation
import cv2
alpha = 0.004
beta = 1.0 - alpha
expected = self.keras_part.model.inputs[0].shape[1:]
actual = img.shape
pred_img = img
# check input depth
if expected[2] == 1 and actual[2] == 3:
pred_img = rgb2gray(pred_img)
pred_img = pred_img.reshape(pred_img.shape + (1, ))
actual = pred_img.shape
salient_mask = self.compute_visualisation_mask(pred_img)
salient_mask_stacked = np.dstack((salient_mask, salient_mask))
salient_mask_stacked = np.dstack((salient_mask_stacked, salient_mask))
blend = cv2.addWeighted(img.astype('float32'), alpha,
salient_mask_stacked, beta, 0.0)
return blend
def make_frame(self, t):
'''
Callback to return an image from from our tub records.
This is called from the VideoClip as it references a time.
We don't use t to reference the frame, but instead increment
a frame counter. This assumes sequential access.
'''
if self.iRec >= self.end or self.iRec >= len(self.index):
return None
rec_ix = self.index[self.iRec]
rec = self.tub.get_record(rec_ix)
image = rec['cam/image_array']
if self.convolution_part:
image = self.draw_salient(image)
image = image * 255
image = image.astype('uint8')
self.draw_model_prediction(rec, image)
self.draw_steering_distribution(rec, image)
if self.scale != 1:
import cv2
h, w, d = image.shape
dsize = (w * self.scale, h * self.scale)
image = cv2.resize(image,
dsize=dsize,
interpolation=cv2.INTER_CUBIC)
self.iRec += 1
# returns a 8-bit RGB array
return image
class TubFormat(DriveFormat):
""" A class to represent a DonkeyCar Tub drive on disc.
Current assumptions:
Tub records are 1 indexed and sequential with no gaps.
We only care about editing steering and throttle.
Steering and throttle should be clipped to -1/1.
TODO: Change actions to be a dictionary of dictionaries, with the outer key being a record's real index.
Images would need to be included in that (which is how the Tub class does it).
"""
def __init__( self, path ):
DriveFormat.__init__(self)
if not os.path.exists(path):
raise IOError( "TubFormat directory does not exist: {}".format( path ) )
if not os.path.isdir(path):
raise IOError( "TubFormat path is not a directory: {}".format( path ) )
self.path = path
self.tub = Tub(path)
self.meta = self.tub.meta
self.edit_list = set()
self.shape = None
self.auxMeta = {}
self.aux_clean = True
#(self.images, self.actions) = self._load(path)
def _load( self, path, image_norm=True, progress=None ):
records = {}
indexes = self.tub.get_index(shuffled=False)
for idx in indexes:
rec = self.tub.get_record(idx)
if self.shape is None:
self.shape = rec['cam/image_array'].shape
records[idx] = rec
self.records = records
self.indexes = indexes
if 'auxiliary' in self.tub.meta:
self.auxMeta = self.tub.meta['auxiliary']
# "sim/info": {"done": true, "pos": [32.82384, 5.567082, -9.720116], "reward": -1.0, "hit": "none", "cte": 2.948259, "speed": 9.52644}
if 'sim/info' in self.tub.meta['inputs']:
self.auxMeta['done'] = { "name": "done", "type": "categorical", "categories": ["True", "False"]}
self.auxMeta['reward'] = { "name": "reward", "type": "continuous"}
def load( self, progress=None ):
self._load(self.path, progress=progress)
self.setClean()
def save( self ):
if not self.aux_clean:
# update meta with new aux meta and write it out
for name, aux in self.auxMeta.items():
if name not in self.tub.meta['inputs']:
self.tub.meta['inputs'].append(name)
if 'continuous' == aux['type']:
aux_type = 'float'
elif 'categorical' == aux['type']:
aux_type = 'int'
else:
raise ValueError( "Unknown auxiliary data type: {}".format( aux['type'] ) )
self.tub.meta['types'].append(aux_type)
self.tub.meta['auxiliary'] = self.auxMeta
with open(self.tub.meta_path, 'w') as f:
json.dump(self.tub.meta, f)
self.aux_clean = True
if self.isClean():
return
self.tub.write_exclude()
for ix in self.edit_list:
rec = self.records[ix]
path = self.tub.get_json_record_path(ix)
try:
with open(path, 'r') as fp:
old_rec = json.load(fp)
except TypeError:
print('troubles with record:', path)
except FileNotFoundError:
raise
except:
print("Unexpected error:", sys.exc_info()[0])
raise
# Copy over only the keys we might have modified
chg_keys = ['user/angle', 'user/throttle', 'orig/angle', 'orig/throttle']
for key in chg_keys:
if key in rec:
old_rec[key] = rec[key]
# Now do any auxiliary data
for key in self.auxMeta.keys():
if key in rec:
if rec[key] is None:
old_rec.pop(key,None)
else:
#if self.auxMeta[key]['type'] == "categorical":
# val = self.auxMeta[key]['categories'].index(rec[key])
#else:
# val = rec[key]
old_rec[key] = rec[key]
try:
with open(path, 'w') as fp:
json.dump(old_rec, fp)
except TypeError:
print('troubles with record:', path)
except FileNotFoundError:
raise
except:
print("Unexpected error:", sys.exc_info()[0])
raise
self.edit_list.clear()
self.setClean()
def count( self ):
return len(self.records)
def imageForIndex( self, index ):
idx = self.indexes[index]
img = self.records[idx]['cam/image_array']
if self.tub.excluded(index + 1):
# This grayed out image ends up looking ugly, can't figure out why
tmp = img.mean(axis=-1,dtype=img.dtype,keepdims=False)
tmp = np.repeat( tmp[:,:,np.newaxis], 3, axis=2 )
return tmp
return img
def actionForIndex( self, index ):
idx = self.indexes[index]
rec = self.records[idx]
angle, throttle = Tub.get_angle_throttle(rec)
return [angle, throttle]
def setActionForIndex( self, new_action, index ):
idx = self.indexes[index]
rec = self.records[idx]
angle, throttle = Tub.get_angle_throttle(rec)
old_action = [angle, throttle]
if not np.array_equal( old_action, new_action ):
if (rec["user/angle"] != new_action[0]) or (rec["user/throttle"] != new_action[1]):
# Save the original values if not already done
if "orig/angle" not in rec:
rec["orig/angle"] = rec["user/angle"]
if "orig/throttle" not in rec:
rec["orig/throttle"] = rec["user/throttle"]
rec["user/angle"] = new_action[0]
rec["user/throttle"] = new_action[1]
self.edit_list.add(idx)
self.setDirty()
def actionForKey(self,keybind,oldAction=None):
oldAction = copy.copy(oldAction)
if keybind == 'w':
oldAction[1] += 0.1
elif keybind == 'x':
oldAction[1] -= 0.1
elif keybind == 'a':
oldAction[0] -= 0.1
elif keybind == 'd':
oldAction[0] += 0.1
elif keybind == 's':
oldAction[0] = 0.0
oldAction[1] = 0.0
else:
return None
return np.clip(oldAction, -1.0, 1.0)
def deleteIndex( self, index ):
if index >= 0 and index < self.count():
index += 1
if self.tub.excluded(index):
self.tub.include_index(index)
else:
self.tub.exclude_index(index)
self.setDirty()
def isIndexDeleted(self, index):
if index >= 0 and index < self.count():
index += 1
return self.tub.excluded(index)
return False
def metaString(self):
#{"inputs": ["cam/image_array", "user/angle", "user/throttle", "user/mode"], "start": 1550950724.8622544, "types": ["image_array", "float", "float", "str"]}
ret = ""
for k, v in self.tub.meta.items():
ret += "{}: {}\n".format( k, v )
return ret
def actionStats(self):
stats = defaultdict(int)
if self.count() > 0:
actions = []
for i in range(self.count()):
act = self.actionForIndex( i )
actions.append(act)
stats["Min"] = np.min(actions)
stats["Max"] = np.max(actions)
stats["Mean"] = np.mean(actions)
stats["StdDev"] = np.std(actions)
return stats
def supportsAuxData(self):
return True
def getAuxMeta(self):
return self.auxMeta
def addAuxData(self, meta):
# TODO Check to make sure the meta data is all the same
if meta["name"] not in self.auxMeta:
self.auxMeta[meta["name"]] = meta
self.aux_clean = False
def auxDataAtIndex(self, auxName, index):
if not auxName in self.auxMeta:
return None
idx = self.indexes[index]
rec = self.records[idx]
if auxName in rec:
if rec[auxName] is not None and self.auxMeta[auxName]['type'] == "categorical":
return self.auxMeta[auxName]['categories'][rec[auxName]]
return rec[auxName]
elif 'sim/info' in rec and auxName in rec['sim/info']:
rec = rec['sim/info']
if rec[auxName] is not None and self.auxMeta[auxName]['type'] == "categorical":
return str(rec[auxName])
return rec[auxName]
return None
def setAuxDataAtIndex(self, auxName, auxData, index):
if not auxName in self.auxMeta:
return False
idx = self.indexes[index]
rec = self.records[idx]
if auxName not in rec or rec[auxName] != auxData:
if auxData is not None and self.auxMeta[auxName]['type'] == "categorical":
auxData = self.auxMeta[auxName]['categories'].index(auxData)
rec[auxName] = auxData
self.edit_list.add(idx)
self.setDirty()
return True
@classmethod
def canOpenFile( cls, path ):
if not os.path.exists(path):
return False
if not os.path.isdir(path):
return False
meta_file = os.path.join( path, "meta.json" )
if not os.path.exists(meta_file):
return False
#if os.path.splitext(path)[1] == ".tub":
# return True
return True
@staticmethod
def defaultInputTypes():
return [{"name":"Images", "type":"numpy image", "shape":(120,160,3)}]
def inputTypes(self):
res = TubFormat.defaultInputTypes()
if self.shape is not None:
res[0]["shape"] = self.shape
return res
@staticmethod
def defaultOutputTypes():
return [{"name":"Actions", "type":"continuous", "range":(-1.0,1.0)}]
def outputTypes(self):
res = []
for act in ["user/angle", "user/throttle"]:
display_name = act.split("/")[1]
res.append( {"name":display_name, "type":"continuous", "range":(-1.0,1.0)} )
return res