def __init__(self, project):
self.project = project
self._dumps = None
self.deploy_path = join(project.path, 'deploy')
self.lock_path = join(self.deploy_path, 'locked.txt')
self.dumps_path = join(self.deploy_path, 'dumps')
utils.maybe_makedirs(self.dumps_path)
self.remote_dump_local_path = join('deploy', 'dump.sql.gz')
def __init__(self,
config_dir=os.path.expanduser('~/.synthese_kiosk'),
verbose=False):
self._verbose = verbose
self._config = None
self._config_dir = config_dir
self._config_dir = os.path.normpath(os.path.abspath(self._config_dir))
utils.maybe_makedirs(self._config_dir)
self._online = False
self._force_offline = False
self._force_offline_url_to_check = None
self._sched = sched.scheduler(time.time, time.sleep)
self._config_path = os.path.join(self._config_dir, 'config.json')
self._init_logging()
log.debug('Config path: %s', self._config_path)
if self.config.has_key('webapp_port'):
self.WEBAPP_PORT = self.config['webapp_port']
if not self.config['offline_cache_dir']:
print "ERROR: Missing parameter 'offline_cache_dir' in your json config"
sys.exit(1)
self._cache_manager = CacheManager(self.config['offline_cache_dir'],
self.config['synthese_url'],
self.config['kiosk_config'])
self._cache_manager.refresh_kiosk_config()
self._kiosk_config = KioskConfig(self.config['offline_cache_dir'],
self.config['kiosk_config'])
# If we don't have the kiosk_config.json file localy, we
# download it now.
while not self._kiosk_config.load():
log.info("We have no config yet, will retry later in 10s")
time.sleep(10)
self._cache_manager.refresh_kiosk_config()
# Init refresh timers
self._next_config_refresh_date = datetime.now() + \
timedelta(seconds=self._kiosk_config.getConfigRefreshTimeout())
self._next_fallback_refresh_date = datetime.now() + \
timedelta(seconds=self._kiosk_config.getFallBackRefreshTimeout())
if self.config['auto_refresh_timeout']:
self._next_browser_refresh_date = datetime.now() + \
timedelta(seconds=self.config['auto_refresh_timeout'])
else:
self._next_browser_refresh_date = None
self._init_displays()
def __init__(self, config_dir=os.path.expanduser('~/.synthese_kiosk'), verbose=False):
self._verbose = verbose
self._config = None
self._config_dir = config_dir
self._config_dir = os.path.normpath(os.path.abspath(self._config_dir))
utils.maybe_makedirs(self._config_dir)
self._online = False
self._force_offline = False
self._force_offline_url_to_check = None
self._sched = sched.scheduler(time.time, time.sleep)
self._config_path = os.path.join(self._config_dir, 'config.json')
self._init_logging()
log.debug('Config path: %s', self._config_path)
if not self.config['offline_cache_dir']:
print "ERROR: Missing parameter 'offline_cache_dir' in your json config"
sys.exit(1)
self._cache_manager = CacheManager(self.config['offline_cache_dir'],
self.config['synthese_url'])
self._cache_manager.refresh_kiosk_config()
self._kiosk_config = KioskConfig(self.config['offline_cache_dir'])
# If we don't have the kiosk_config.json file localy, we
# download it now.
while not self._kiosk_config.load():
log.info("We have no config yet, will retry later in 10s")
time.sleep(10)
self._cache_manager.refresh_kiosk_config()
# Init refresh timers
self._next_config_refresh_date = datetime.now() + \
timedelta(seconds=self._kiosk_config.getConfigRefreshTimeout())
self._next_fallback_refresh_date = datetime.now() + \
timedelta(seconds=self._kiosk_config.getFallBackRefreshTimeout())
if self.config['auto_refresh_timeout']:
self._next_browser_refresh_date = datetime.now() + \
timedelta(seconds=self.config['auto_refresh_timeout'])
else:
self._next_browser_refresh_date = None
self._init_displays()
def _wget(self, directory, url):
log.debug('running wget in ' + directory + " for url " + url)
utils.maybe_makedirs(directory)
cmd_line = ["wget"]
cmd_line.append('--recursive')
cmd_line.append('--quiet')
cmd_line.append('--no-host-directories')
cmd_line.append('--no-parent')
cmd_line.append('--page-requisites')
cmd_line.append('--convert-links')
cmd_line.append('--no-verbose')
cmd_line.append('--directory-prefix')
cmd_line.append(directory)
cmd_line.append(url)
log.debug('Wget command line: %s', cmd_line)
try:
retval = subprocess.call(cmd_line)
# Return True if the wget succeeded
return retval == 0
except Exception, e:
log.error("Failed to launch wget: %s", e)
def _wget(self, directory, url):
log.debug('running wget in ' + directory + " for url " + url)
utils.maybe_makedirs(directory)
cmd_line = ["wget"]
cmd_line.append('--recursive')
cmd_line.append('--quiet')
cmd_line.append('--no-host-directories')
cmd_line.append('--no-parent')
cmd_line.append('--page-requisites')
cmd_line.append('--convert-links')
cmd_line.append('--no-verbose')
cmd_line.append('--directory-prefix')
cmd_line.append(directory)
cmd_line.append(url)
log.debug('Wget command line: %s', cmd_line)
try:
retval = subprocess.call(cmd_line)
# Return True if the wget succeeded
return retval == 0
except Exception, e:
log.error("Failed to launch wget: %s", e)
def deploy_remote_prepare(self):
utils.maybe_makedirs(self.deploy_path)
'position': {
'x': {'mean': 0, 'std': 5},
'y': {'mean': 0, 'std': 5}
},
'velocity': {
'x': {'mean': 0, 'std': 2},
'y': {'mean': 0, 'std': 2}
},
'acceleration': {
'x': {'mean': 0, 'std': 1},
'y': {'mean': 0, 'std': 1}
}
}
}
maybe_makedirs('../processed')
data_columns = pd.MultiIndex.from_product([['position', 'velocity', 'acceleration'], ['x', 'y']])
for desired_source in ['particles']:
dataset = dict()
for data_class in ['train', 'val', 'test']:
with open(data_class + '_data_2_robot.pkl', 'rb') as f:
dataset[data_class] = dill.load(f, encoding='latin1')
env = Environment(node_type_list=['PARTICLE'], standardization=standardization)
attention_radius = dict()
attention_radius[(env.NodeType.PARTICLE, env.NodeType.PARTICLE)] = 10.0
env.attention_radius = attention_radius
env.robot_type = env.NodeType.PARTICLE
scenes = []
data_dict_path = os.path.join('../processed', '_'.join([desired_source, data_class]) + '_2_robot.pkl')
# open dataset
def retrain(image_dir):
# Create the base model from the pre-trained model InceptionV3
base_model = tf.keras.applications.InceptionV3(input_shape=IMG_SHAPE,
include_top=False,
pooling='avg',
weights='imagenet')
# base_model.summary()
base_model.compile(loss='mse')
input_shape = base_model.output_shape[1:]
print("Generating Bottleneck Dataset... this may take some minutes.")
bottleneck_train_ds, num_train = get_bottleneck_dataset(base_model,
img_dir=image_dir,
img_size=IMG_SIZE)
train_batches = bottleneck_train_ds.shuffle(10000).batch(
BATCH_SIZE).repeat()
print("Done generating Bottleneck Dataset")
######### Your code starts here #########
# We want to create a linear classifier which takes the bottleneck data as input
# 1. Get the size of the bottleneck tensor. Hint: You can get the shape of a tensor via tensor.get_shape().as_list()
# 2. Define a new tf.keras Model which is a linear classifier
# 2.1 Define a keras Input (retrain_input)
# 2.2 Define the trainable layer (retrain_layer)
# 2.3 Define the activation function (retrain activation)
# 2.4 Create a new model
# 3. Define a loss and a evaluation metric
# Fill in the parts indicated by #FILL#. No additional lines are required.
input_shape = base_model.output_shape[1:]
retrain_model = tf.keras.models.Sequential(
[tf.keras.layers.Dense(len(LABELS), input_shape=input_shape)])
loss = tf.keras.losses.CategoricalCrossentropy(from_logits=True)
metric = 'accuracy'
######### Your code ends here #########
retrain_model.compile(optimizer=tf.keras.optimizers.SGD(lr=lr),
loss=loss,
metrics=[metric])
retrain_model.summary()
EPOCHS = 1
steps_per_epoch = 5000
tensorboard_callback = tf.keras.callbacks.TensorBoard(
log_dir='./retrain_logs', update_freq='batch')
retrain_model.fit(train_batches,
epochs=EPOCHS,
steps_per_epoch=steps_per_epoch)
# callbacks=[tensorboard_callback])
######### Your code starts here #########
# We now want to create the full model using the newly trained classifier
# Use tensorflow keras Sequential to stack the base_model and the new layers
# Fill in the parts indicated by #FILL#. No additional lines are required.
model = tf.keras.models.Sequential([base_model, retrain_model])
######### Your code ends here #########
model.compile(loss=loss, metrics=[metric])
maybe_makedirs('./trained_models')
model.save('./trained_models/trained.h5')
def compute_and_plot_saliency(model, image_path):
"""
This function computes and plots the saliency plot.
You need to compute the matrix M detailed in section 3.1 in
K. Simonyan, A. Vedaldi, and A. Zisserman,
"Deep inside convolutional networks: Visualising imageclassification models and saliency maps,"
2013, Available at https://arxiv.org/abs/1312.6034.
:param model: Model which is used
:param image_path: Path to the image to be analysed
:return: None
"""
raw_image = tf.dtypes.cast(decode_jpeg(image_path), tf.float32)
logits_tensor = model.get_layer("classifier")
logits_model = tf.keras.Model(model.input, logits_tensor.output)
with tf.GradientTape() as t:
######### Your code starts here #########
# Fill in the parts indicated by #FILL#. No additional lines are
# required.
######### Your code ends here #########
plt.subplot(2, 1, 1)
plt.imshow(M)
plt.title("Saliency with respect to predicted class %s" % LABELS[top_class])
plt.subplot(2, 1, 2)
plt.imshow(decode_jpeg(image_path).numpy())
plt.savefig("../plots/saliency.png")
plt.show()
def plot_classification(image_path, classification_array):
nH, nW, _ = classification_array.shape
image_data = decode_jpeg(image_path).numpy()
aspect_ratio = float(image_data.shape[0]) / image_data.shape[1]
plt.figure(figsize=(8, 8 * aspect_ratio))
p1 = plt.subplot(2, 2, 1)
plt.imshow(classification_array[:, :, 0], interpolation="none", cmap="jet")
plt.title("%s probability" % LABELS[0])
p1.set_aspect(aspect_ratio * nW / nH)
plt.colorbar()
p2 = plt.subplot(2, 2, 2)
plt.imshow(classification_array[:, :, 1], interpolation="none", cmap="jet")
plt.title("%s probability" % LABELS[1])
p2.set_aspect(aspect_ratio * nW / nH)
plt.colorbar()
p2 = plt.subplot(2, 2, 3)
plt.imshow(classification_array[:, :, 2], interpolation="none", cmap="jet")
plt.title("%s probability" % LABELS[2])
p2.set_aspect(aspect_ratio * nW / nH)
plt.colorbar()
plt.subplot(2, 2, 4)
plt.imshow(image_data)
plt.savefig("../plots/detect.png")
plt.show()
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--image", type=str)
parser.add_argument("--scheme", type=str)
FLAGS, _ = parser.parse_known_args()
maybe_makedirs("../plots")
model = tf.keras.models.load_model("./trained_models/trained.h5")
model.__call__ = tf.function(model.__call__)
writer = tf.summary.create_file_writer("retrain_logs")
tf.summary.trace_on()
if FLAGS.scheme == "brute":
plot_classification(
FLAGS.image,
compute_brute_force_classification(model, FLAGS.image, 8, 8),
)
elif FLAGS.scheme == "conv":
plot_classification(
FLAGS.image,
compute_convolutional_KxK_classification(model, FLAGS.image),
)
elif FLAGS.scheme == "saliency":
compute_and_plot_saliency(model, FLAGS.image)
else:
print("Unrecognized scheme:", FLAGS.scheme)
with writer.as_default():
tf.summary.trace_export("detect_%s" % FLAGS.scheme, step=0)
NUM_TEST = 1000
Z_HIDDENS = [20, 20]
T_HIDDENS = []
R_HIDDENS = []
A_HIDDENS = None
LEARNING_RATE = 0.01
Z_SIZE = 4
BATCH_SIZE = 32
TRAIN_STEPS = 250
NORMS = [0.5, 0.6, 0.7, 0.8, 0.9, 1.0]
LAMBDAS = [0.5, 0.1, 0.05, 0.01, 0.005, 0.001]
RUNS = 20
SAVE_DIR = "results/env_2d"
utils.maybe_makedirs(SAVE_DIR)
SAVE_FILE = "1a_medium_norm_lambda.pickle"
SAVE_PATH = os.path.join(SAVE_DIR, SAVE_FILE)
results = utils.maybe_read_pickle(SAVE_PATH)
for norm in NORMS:
for lambda_1 in LAMBDAS:
for run in range(RUNS):
key = (norm, lambda_1, run)
if key not in results:
def compute_and_plot_saliency(model, image_path):
"""
This function computes and plots the saliency plot.
You need to compute the matrix M detailed in section 3.1 in
K. Simonyan, A. Vedaldi, and A. Zisserman,
"Deep inside convolutional networks: Visualising imageclassification models and saliency maps,"
2013, Available at https://arxiv.org/abs/1312.6034.
:param model: Model which is used
:param image_path: Path to the image to be analysed
:return: None
"""
raw_image = tf.dtypes.cast(decode_jpeg(image_path), tf.float32)
logits_tensor = model.get_layer('classifier')
logits_model = tf.keras.Model(model.input, logits_tensor.output)
with tf.GradientTape() as t:
######### Your code starts here #########
######### Your code ends here #########
plt.subplot(2, 1, 1)
plt.imshow(M)
plt.title('Saliency with respect to predicted class %s' % LABELS[top_class])
plt.subplot(2, 1, 2)
plt.imshow(decode_jpeg(image_path).numpy())
plt.savefig("../plots/saliency.png")
plt.show()
def plot_classification(image_path, classification_array):
nH, nW, _ = classification_array.shape
image_data = decode_jpeg(image_path).numpy()
aspect_ratio = float(image_data.shape[0]) / image_data.shape[1]
plt.figure(figsize=(8, 8*aspect_ratio))
p1 = plt.subplot(2,2,1)
plt.imshow(classification_array[:,:,0], interpolation='none', cmap='jet')
plt.title('%s probability' % LABELS[0])
p1.set_aspect(aspect_ratio*nW/nH)
plt.colorbar()
p2 = plt.subplot(2,2,2)
plt.imshow(classification_array[:,:,1], interpolation='none', cmap='jet')
plt.title('%s probability' % LABELS[1])
p2.set_aspect(aspect_ratio*nW/nH)
plt.colorbar()
p2 = plt.subplot(2,2,3)
plt.imshow(classification_array[:,:,2], interpolation='none', cmap='jet')
plt.title('%s probability' % LABELS[2])
p2.set_aspect(aspect_ratio*nW/nH)
plt.colorbar()
plt.subplot(2,2,4)
plt.imshow(image_data)
plt.savefig("../plots/detect.png")
plt.show()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--image', type=str)
parser.add_argument('--scheme', type=str)
FLAGS, _ = parser.parse_known_args()
maybe_makedirs("../plots")
model = tf.keras.models.load_model('./trained_models/trained.h5')
if FLAGS.scheme == 'brute':
plot_classification(FLAGS.image, compute_brute_force_classification(model, FLAGS.image, 8, 8))
elif FLAGS.scheme == 'conv':
plot_classification(FLAGS.image, compute_convolutional_KxK_classification(model, FLAGS.image))
elif FLAGS.scheme == 'saliency':
compute_and_plot_saliency(model, FLAGS.image)
else:
print('Unrecognized scheme:', FLAGS.scheme)
def retrain(image_dir):
# Create the base model from the pre-trained model InceptionV3
base_model = tf.keras.applications.InceptionV3(
input_shape=IMG_SHAPE,
include_top=False,
pooling="avg",
weights="imagenet",
)
base_model.summary()
base_model.compile(loss="mse")
print("Generating Bottleneck Dataset... this may take some minutes.")
bottleneck_train_ds, num_train = get_bottleneck_dataset(
base_model, img_dir=image_dir, img_size=IMG_SIZE
)
train_batches = (
bottleneck_train_ds.shuffle(10000).batch(BATCH_SIZE).repeat()
)
print("Done generating Bottleneck Dataset")
######### Your code starts here #########
# We want to create a linear classifier which takes the bottleneck data as input
# 1. Get the size of the bottleneck tensor. Hint: You can get the shape of a tensor via tensor.get_shape().as_list()
# 2. Define a new tf.keras Model which is a linear classifier
# 2.1 Define a keras Input (retrain_input)
# 2.2 Define the trainable layer (retrain_layer)
# 2.3 Define the activation function (retrain activation)
# 2.4 Create a new model
# 3. Define a loss and a evaluation metric
# Fill in the parts indicated by #FILL#. No additional lines are required.
######### Your code ends here #########
retrain_model.compile(
optimizer=tf.keras.optimizers.SGD(lr=lr), loss=loss, metrics=[metric]
)
retrain_model.summary()
EPOCHS = 1
steps_per_epoch = 5000
tensorboard_callback = tf.keras.callbacks.TensorBoard(
log_dir="retrain_logs", update_freq="batch"
)
retrain_model.fit(
train_batches,
epochs=EPOCHS,
steps_per_epoch=steps_per_epoch,
callbacks=[tensorboard_callback],
)
######### Your code starts here #########
# We now want to create the full model using the newly trained classifier
# Use tensorflow keras Sequential to stack the base_model and the new layers
# Fill in the parts indicated by #FILL#. No additional lines are required.
######### Your code ends here #########
model.compile(loss=loss, metrics=[metric])
maybe_makedirs("trained_models")
model.save("trained_models/trained.h5")
