def experiment():
params = Params.Params()
df_error = pickle.load(open('df_error.p', 'rb'))
df_state = pickle.load(open('df_state.p', 'rb'))
model = False
trainFunctions = [
x for x in params.functions if x.name not in ['circle_4']
]
#trainFunctions = [x for x in params.functions if x.name in ['elipse']]
df_state, scaler_state = scale_data(df_state, trainFunctions)
df_error, scaler_error = scale_data(df_error, trainFunctions)
pickle.dump(scaler_state, open("scaler_state.p", "wb"))
pickle.dump(scaler_error, open("scaler_error.p", "wb"))
for function in trainFunctions:
name = function.name
value = df_state[name]
print("Training State for {} shape".format(name))
model = train_lstm(value.to_numpy(), 'state', model)
model = False
for function in trainFunctions:
name = function.name
value = df_error[name]
print("Training Error for {} shape".format(name))
model = train_lstm(value.to_numpy(), 'error', model)
def main():
import params
import sys
pa = params.Params()
# pa.simu_len = 1000 # 1000
# pa.num_ex = 100 # 100
# pa.num_nw = 10
# pa.num_seq_per_batch = 20
# pa.output_freq = 50
# # pa.max_nw_size = 5
# # pa.job_len = 5
# pa.new_job_rate = 0.3
# pa.episode_max_length = 10000 # 2000
# pa.compute_dependent_parameters()
pg_resume = None
# pg_resume = 'data/tmp_450.pkl'
render = False
launch(pa, pg_resume, render, repre='image', end='all_done')
def collect(self):
p = params.Params()
allShapes = [Shape(x) for x in p.functions]
shapesData = pd.Series(dtype='object')
for shape in allShapes:
data, _ = shape.trace()
shapesData = shapesData.append(pd.Series({shape.name: data}))
def train(self):
p = params.Params()
data = pkl.load(open('hessian_data.p', 'rb'))
nfeatures = 21
rootpath = "/tmp/"
scaler = MinMaxScaler(feature_range=(-1, 1))
model = Sequential()
model.add(LSTM(nfeatures+1, batch_input_shape=(p.batch_size, 1, nfeatures), stateful=True))
model.add(Dense(4))
model.compile(loss='mean_squared_error', optimizer='adam')
for function in p.functions:
scaler = MinMaxScaler(feature_range=(-1, 1))
train = scaler.fit_transform(data[function.name])
pkl.dump(scaler, open("hessian_scaler.p", "wb"))
#Skip last element, and upshift hessian for training next step
trainX, trainY = train[:-1,:], train[1:, -4:]
#Align to batch_size of 25
trainX, trainY = trainX[24:,:], trainY[24:,:]
trainX = np.reshape(trainX, (trainX.shape[0], 1, trainX.shape[1]))
es = EarlyStopping(monitor='loss', mode='min', verbose=1, patience=20)
rs = LambdaCallback(on_epoch_end=lambda epoch, logs: model.reset_states())
mc = ModelCheckpoint(os.path.join(rootpath, "hessian.h5"),
monitor='loss', mode='min', verbose=1, save_best_only=True)
#Reuse model in loops - incremental training.
history = model.fit(trainX, trainY, epochs=200, batch_size=p.batch_size, verbose=2, callbacks=[es, mc, rs],
shuffle=False)
trainPredict = model.predict(trainX, batch_size=p.batch_size)
# calculate root mean squared error
trainScore = math.sqrt(mean_squared_error(trainY, trainPredict))
print("Train Score: {} RMSE".format(trainScore))
shutil.copyfile(os.path.join(rootpath, "hessian.h5"), "/home/012532065/final/MotionControl/hessian.h5")
def formationCenter(r_c, z_c, dz_c, hessian, x_2, y_2, mu_f, z_desired):
# hessian = np.array([[2, 0],[2, 0]]) * np.random.rand(1)
# z_c, dz_c, p = kalmanFilter(z_c, dz_c, r, z_r, r_c, r_c_old, p, hessian, numSensors)
p = params.Params()
rotateLeft = p.rotateLeft
rotateRight = p.rotateRight
K4 = p.K4
dt = p.dt
y_1 = dz_c / norm(dz_c)
x_1 = rotateRight @ y_1
theta = atan2(x_2[1], x_2[0]) - atan2(x_1[1], x_1[0])
kappa_1 = (x_1.T @ hessian @ x_1) / norm(dz_c)
kappa_2 = (x_1.T @ hessian @ y_1) / norm(dz_c)
f_z = mu_f * (1 - (z_desired / z_c) ** 2)
u_c = kappa_1 * cos(theta) + \
kappa_2 * sin(theta) - \
(2 * f_z * norm(dz_c) * (cos(theta / 2) ** 2)) + \
K4 * sin(theta / 2)
# print([x_2, y_2])
x_2 = x_2 + dt * u_c * y_2
x_2 = x_2 / norm(x_2)
y_2 = rotateLeft @ x_2
pltVar.push(u_c)
r_c = r_c + dt * x_2
return r_c, x_2, y_2
def __init__(self, function):
p = params.Params()
self.params = p
self.function = function
self.name = self.function.name
self.r_c, self.r_c_old, self.r, = p.r_c, p.r_c, p.r
r_c = self.r_c
self.z_c = function.f(r_c[0], r_c[1])
self.dz_c = function.dz_f(r_c[0], r_c[1])
self.y_2 = self.dz_c / norm(self.dz_c)
self.x_2 = p.rotateRight @ self.y_2
self.q, self.dq, self.u_r, self.vel_q = p.q, p.dq, p.u_r, p.vel_q
self.p = np.zeros((3, 3))
z_r = np.array([self.function.f(*pt) for pt in self.r])
self.init_state = [
self.r_c, self.z_c, self.dz_c, self.r_c_old, self.p, self.r,
self.q, self.dq, self.u_r, self.vel_q, self.x_2, self.y_2, z_r
]
self.state = self.init_state
self.old_state = self.init_state
self.pltVarShape = False
def make_tf2_export(weights_path, export_dir):
if os.path.exists(export_dir):
log('TF2 export already exists in {}, skipping TF2 export'.format(
export_dir))
return
# Create a TF2 Module wrapper around YAMNet.
log('Building and checking TF2 Module ...')
params = yamnet_params.Params()
yamnet = YAMNet(weights_path, params)
check_model(yamnet, yamnet.class_map_path(), params)
log('Done')
# Make TF2 SavedModel export.
log('Making TF2 SavedModel export ...')
tf.saved_model.save(yamnet, export_dir)
log('Done')
# Check export with TF-Hub in TF2.
log('Checking TF2 SavedModel export in TF2 ...')
model = tfhub.load(export_dir)
check_model(model, model.class_map_path(), params)
log('Done')
# Check export with TF-Hub in TF1.
log('Checking TF2 SavedModel export in TF1 ...')
with tf.compat.v1.Graph().as_default(), tf.compat.v1.Session() as sess:
model = tfhub.load(export_dir)
sess.run(tf.compat.v1.global_variables_initializer())
def run_model(waveform):
return sess.run(model(waveform))
check_model(run_model, model.class_map_path().eval(), params)
log('Done')
def embedding(self, input_paths, output_paths, embed_paths=""):
"""Extract YAMnet features with opensmile using a single process."""
if embed_paths == "":
embed_paths = [""] * len(input_paths)
save_embedding = False
else:
save_embedding = True
paths = list(zip(input_paths, embed_paths, output_paths))
params = yamnet_params.Params(sample_rate=self.sample_rate,
patch_hop_seconds=0.48)
class_names = yamnet_model.class_names(self.class_names)
yamnet = yamnet_model.yamnet_frames_model(params)
yamnet.load_weights(self.model_checkpoint)
func = partial(
self._embed,
yamnet=yamnet,
params=params,
class_names=class_names,
save_embedding=save_embedding,
)
self.single_process(func, paths)
def main(argv):
assert argv, 'Usage: inference.py <wav file> <wav file> ...'
params = yamnet_params.Params()
yamnet = yamnet_model.yamnet_frames_model(params)
yamnet.load_weights('yamnet.h5')
yamnet_classes = yamnet_model.class_names('yamnet_class_map.csv')
for file_name in argv:
# Decode the WAV file.
wav_data, sr = sf.read(file_name, dtype=np.int16)
assert wav_data.dtype == np.int16, 'Bad sample type: %r' % wav_data.dtype
waveform = wav_data / 32768.0 # Convert to [-1.0, +1.0]
waveform = waveform.astype('float32')
# Convert to mono and the sample rate expected by YAMNet.
if len(waveform.shape) > 1:
waveform = np.mean(waveform, axis=1)
if sr != params.sample_rate:
waveform = resampy.resample(waveform, sr, params.sample_rate)
# Predict YAMNet classes.
scores, embeddings, spectrogram = yamnet(waveform)
# Scores is a matrix of (time_frames, num_classes) classifier scores.
# Average them along time to get an overall classifier output for the clip.
prediction = np.mean(scores, axis=0)
# Report the highest-scoring classes and their scores.
top5_i = np.argsort(prediction)[::-1][:5]
print(
file_name, ':\n' + '\n'.join(
' {:12s}: {:.3f}'.format(yamnet_classes[i], prediction[i])
for i in top5_i))
def formationControl(r_c, r, dz_c, q, dq, u_r, vel_q, t):
p = params.Params()
a = p.a
b = p.b
dt = p.dt
K2 = p.K2
K3 = p.K3
phi_inv = p.phi_inv
# Moving opposite to gradient direction : Source seeking
# velocity_center = - dz_c / norm(dz_c)
# Moving perpendiular to gradient direction : Boundary tracking
velocity_center = p.rotateLeft @ (dz_c / norm(dz_c))
y1 = velocity_center
x1 = p.rotateRight @ y1
# e_1 = r[1] - r[0]
# e_1 = e_1 / np.linalg.norm(e_1)
# e_2 = r[2] - r[3]
# e_2 = e_2 / np.linalg.norm(e_2)
# q_0 = np.array([
# [0, 0],
# (a / sqrt(2)) * e_1,
# (b / sqrt(2)) * e_2,
# [0, 0]
# ])
if p.numSensors == 4:
q_0 = np.array([
[0, 0],
(a / sqrt(2)) * y1,
(b / sqrt(2)) * x1,
[0, 0]
])
if p.numSensors == 2:
sigma = 10.0
q_0 = np.array([
[0, 0],
(a / sqrt(2)) * x1,
])
if floor( t / 4 ) % 2 == 0:
q_0[1] = array([[sin(sigma/pi), cos(sigma/pi)],
[-cos(sigma/pi), sin(sigma/pi)]]) @ q_0[1]
else:
q_0[1] = array([[sin(sigma), -cos(sigma)],
[cos(sigma), sin(sigma)]]) @ q_0[1]
# dq[1:] = dq[1:] + dt * u_r[1:]
# u_r[1:] = -K2 * (q[1:] - q_0[1:]) - K3 * dq[1:]
# vel_q[1:] = vel_q[1:] + dt * u_r[1:]
vel_q[1:] = -10 * (q[1:] - q_0[1:])
q[1:] = q[1:] + dt * vel_q[1:]
q_N = np.append([r_c], q[1:], axis=0)
r = phi_inv @ q_N
return r, q, dq, u_r, vel_q
def make_tflite_export(weights_path, model_path, export_dir):
if os.path.exists(export_dir):
log('TF-Lite export already exists in {}, skipping TF-Lite export'.
format(export_dir))
return
# Create a TF-Lite compatible Module wrapper around YAMNet.
log('Building and checking TF-Lite Module ...')
params = yamnet_params.Params(tflite_compatible=True)
yamnet = YAMNet(weights_path, params, model_path)
check_model(yamnet, yamnet.class_map_path(), params)
log('Done')
# Make TF-Lite SavedModel export.
log('Making TF-Lite SavedModel export ...')
saved_model_dir = os.path.join(export_dir, 'saved_model')
os.makedirs(saved_model_dir)
tf.saved_model.save(yamnet, saved_model_dir)
log('Done')
# Check that the export can be loaded and works.
log('Checking TF-Lite SavedModel export in TF2 ...')
model = tf.saved_model.load(saved_model_dir)
check_model(model, model.class_map_path(), params)
log('Done')
# Make a TF-Lite model from the SavedModel.
log('Making TF-Lite model ...')
tflite_converter = tf.lite.TFLiteConverter.from_saved_model(
saved_model_dir)
tflite_model = tflite_converter.convert()
tflite_model_path = os.path.join(export_dir, 'yamnet.tflite')
with open(tflite_model_path, 'wb') as f:
f.write(tflite_model)
log('Done')
# Check the TF-Lite export.
log('Checking TF-Lite model ...')
interpreter = tf.lite.Interpreter(tflite_model_path)
audio_input_index = interpreter.get_input_details()[0]['index']
scores_output_index = interpreter.get_output_details()[0]['index']
embeddings_output_index = interpreter.get_output_details()[1]['index']
#spectrogram_output_index = interpreter.get_output_details()[2]['index']
def run_model(waveform):
interpreter.resize_tensor_input(audio_input_index, [len(waveform)],
strict=True)
interpreter.allocate_tensors()
interpreter.set_tensor(audio_input_index, waveform)
interpreter.invoke()
return (interpreter.get_tensor(scores_output_index),
interpreter.get_tensor(embeddings_output_index)) #,
# interpreter.get_tensor(spectrogram_output_index))
check_model(run_model, 'yamnet_class_map.csv', params)
log('Done')
return saved_model_dir
def collect(self):
p = params.Params()
dataAll = pd.Series(dtype='object')
for function in p.functions:
data = self.stepFunction(function, collect=True)
dataAll = dataAll.append(pd.Series({function.name: data}))
pkl.dump(dataAll, open("hessian_data.p", "wb"))
return dataAll
def create_trials(num_hhld, num_days, num_hours, num_min, trial_code, chad_activity_params, \
demographic, num_people, do_minute_by_minute, do_print=False):
"""
This function creates the input data for each household in the simulation.
:param int num_hhld: the number of households simulated
:param int num_days: the number of days in the simulation
:param int num_hours: the number of additional hours
:param int num_min: the number of additional minutes
:param int trial_code: the trial identifier
:param chad_params.CHAD_params chad_activity_params: the activity parameters \
used to sample "good" CHAD data
:param int demographic: the demographic identifier
:param int num_people: the number of people per household
:param bool do_minute_by_minute: a flag for how the time steps progress in the scheduler
:param bool do_print: flag whether to print messages to the console
:returns: input data where each entry corresponds to the input \
for the respective household in the simulation
:rtype: list of :class:`trial.Trial`
"""
# load the parameters necessary for the runs comparing the ABMHAP to CHAD
# run the simulation using default parameters
param_list = [params.Params(num_days=num_days, num_hours=num_hours, num_min=num_min, \
num_people=num_people, do_minute_by_minute=do_minute_by_minute) \
for _ in range(num_hhld)]
# print message
if do_print:
print('initializing trials...')
#
# create the conditions for each trial
#
# start timing
start = time.time()
# initialize the simulation inputs
trials = initialize_trials(param_list, trial_code, chad_activity_params,
demographic)
# end timing
end = time.time()
# calculate the elapsed time
dt_elapsed = end - start
# print message
if do_print:
print('elapsed time to initialize %d trials:\t%.3f [s]\n' %
(num_hhld, dt_elapsed))
return trials
def __enter__(self):
self.params = params.Params()
self.fridge = fridge.Fridge(self, self._nowait)
self.uploader = uploader.Uploader(self)
self.configwaiter = configwaiter.ConfigWaiter(self)
self.params.load()
self.set_sensors(sensor.make_sensor(self))
asyncio.get_event_loop().add_signal_handler(signal.SIGHUP,
self.reload_signal)
return self
def main(): import params import sys pa=params.Params() #pg_resume='/home/dell/testing_part2/880_10ex.pkl_' pg_resume = None test_only = True if len(sys.argv) == 2: pg_resume=sys.argv[1] #give the path of weights file test_only=True render=False launch(pa,pg_resume,render=render,repre='image',end='all_done', test_only=test_only)
def main(): import params import sys pa=params.Params() pg_resume='/home/dell/rl_sudoku/4x4sudoku_4_5_6_7_8_saved_weights/380.pkl_' #pg_resume = None test_only = False if len(sys.argv) == 2: pg_resume=sys.argv[1] #give the path of weights file test_only=True render=False launch(pa,pg_resume,render=render,repre='image',end='all_done', test_only=test_only)
def _setup_job(self, job):
if not os.path.isdir(job['work_dir']):
os.makedirs(job['work_dir'])
job_params = params.Params((p.replace('job_params.', ''), v)
for p, v in job.items()
if p.startswith('job_params'))
job_params.name = job['job_name']
job_scripts.write_job_script(job['job_file'], self.job_template,
job_params)
print(job['job_file'])
def buildpars(self, ):
"""
Build the model parameter object.
Refer to params.py for the data that self.p contains. In general,
this contains the electrochemical parameters necessary for the
simulation of the full Pseudo-2D Newman-style model.
"""
self.p = params.Params()
self.p.buildpars(self.V_init, self.Pdat)
self.p.Ac = self.p.Area
self.pars = self.p
def __init__(self, function):
p = params.Params()
self.params = p
self.function = function
self.name = self.function.name
self.numSensors = p.numSensors
self.r_c, self.r_c_old, self.r, = p.r_c, p.r_c, p.r
r_c = self.r_c
self.z_c = function.f(r_c[0], r_c[1])
self.dz_c = function.dz_f(r_c[0], r_c[1])
self.y_2 = self.dz_c / norm(self.dz_c)
self.x_2 = p.rotateRight @ self.y_2
self.q, self.dq, self.u_r, self.vel_q = p.q, p.dq, p.u_r, p.vel_q
self.p = np.zeros((3, 3))
def main():
import params
import sys
pa = params.Params()
pg_resume = None
test_only = False
if len(sys.argv) == 2:
pg_resume = sys.argv[1] #give the path of weights file
test_only = True
render = False
launch(pa,
pg_resume,
render=render,
repre='image',
end='all_done',
test_only=test_only)
def test_params_basic(self):
defparams = params.Params()
self.assertEqual(defparams.overshoot_factor,
params._FIELD_DEFAULTS['overshoot_factor'])
# fetching a bad parameter fails
with self.assertRaises(KeyError):
x = self.params.param_that_doesnt_exist
# setting a parameter
defparams.overshoot_factor = 8877
self.assertEqual(defparams.overshoot_factor, 8877)
# setting a bad parameter fails
with self.assertRaises(KeyError):
self.params.somewrongthing = 5
def __init__(self, config_path="./config.yaml"):
"""Init method for the Searcher."""
super().__init__()
# Load the configuration
conf = OmegaConf.load(config_path)
self.dataset_path = conf.dataset_path
self.audio_path = os.path.join(conf.dataset_path, "podcasts-audio")
self.es_url = conf.search_es_url # URL of Elasticsearch to query
self.es_num = (conf.search_es_num
) # Number of segments to request from Elasticsearch
self.sample_rate = 44100 # Hardcoded sample rate of all podcast audio
# Load the podcast metadata
self.metadata = load_metadata(self.dataset_path)
# Set up the reranking model
self.rerank_tokenizer = AutoTokenizer.from_pretrained(
conf.search_rerank_model,
use_fast=True,
cache_dir=conf.search_cache_dir)
self.rerank_model = AutoModelForSequenceClassification.from_pretrained(
conf.search_rerank_model, cache_dir=conf.search_cache_dir)
self.rerank_model.to("cpu", non_blocking=True)
self.rerank_max_seq_len = 512
# Set up the openSMILE extractor
self.smile = opensmile.Smile(
feature_set=opensmile.FeatureSet.eGeMAPSv02,
feature_level=opensmile.FeatureLevel.Functionals,
options={
"frameModeFunctionalsConf":
os.path.join(
os.getenv("PODCAST_PATH"),
"data/custom_FrameModeFunctionals.conf.inc",
)
},
)
# Set up the YAMNet model
params = yamnet_params.Params(sample_rate=self.sample_rate,
patch_hop_seconds=0.48)
self.yamnet_classes = yamnet_model.class_names(
os.path.join(os.getenv("YAMNET_PATH"), "yamnet_class_map.csv"))
self.yamnet_model = yamnet_model.yamnet_frames_model(params)
self.yamnet_model.load_weights(
os.path.join(os.getenv("PODCAST_PATH"), "data/yamnet.h5"))
def processData(self, df, learn=True):
if isinstance(df, pd.DataFrame):
raw_data = df.to_numpy()
else:
raw_data = df
# ['s', 'hessian', 'r_c_delta', 'r_delta', 'z_r']
px = params.Params()
numSensors = px.numSensors
unscaled_data = np.array([[
x[0][0], x[0][1], x[0][2], x[1][0][0], x[1][0][1], x[1][1][0],
x[1][1][1], x[2][0], x[2][1], x[3][0][0], x[3][0][1], x[4][0],
x[3][1][0], x[3][1][1], x[4][1], x[3][2][0], x[3][2][1], x[4][2],
x[3][3][0], x[3][3][1], x[4][3]
] for x in raw_data])
#Retain only every nth data element
# unscaled_data = unscaled_data[::10,:]
lagWindowSize = self.lagWindowSize
numFeaturesState = self.numFeaturesState
numFeaturesStateAdded = self.numFeaturesStateAdded
datalength = unscaled_data.shape[0]
if learn:
scaler = MinMaxScaler(feature_range=(-1, 1))
scaled_data = scaler.fit_transform(unscaled_data)
rootpath = "/tmp" if os.name == 'posix' else ""
pkl.dump(scaler, open(os.path.join(rootpath, "scaler.p"), "wb"))
else:
self.history.append(unscaled_data[0])
self.history = self.history[1:]
scaled_data = self.scaler.transform(self.history)
# LSTM needs input in [samples, timestep, features] format.
if learn:
X = np.array([
scaled_data[i - lagWindowSize:i, :]
for i in range(lagWindowSize, datalength)
])
Y = scaled_data[lagWindowSize:, :]
# We dont need 4th parameter for prediction so -1
Y = Y[:, 0:numFeaturesState]
# Y = [Y[:, 0:numFeaturesState], Y[:, numFeaturesStateAdded:]]
else:
X = np.array([scaled_data])
# X = [X[:, :, 0:numFeaturesStateAdded], X[:, :, numFeaturesStateAdded:]]
Y = False
return X, Y
def main(argv):
assert argv, 'Usage: inference.py <wav file> <wav file> ...'
model_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'yamnet.h5')
classes_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'yamnet_class_map.csv')
event_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'event.json')
params = yamnet_params.Params()
yamnet = yamnet_model.yamnet_frames_model(params)
yamnet.load_weights(model_path)
yamnet_classes = yamnet_model.class_names(classes_path)
for file_name in argv:
# Decode the WAV file.
wav_data, sr = sf.read(file_name, dtype=np.int16)
assert wav_data.dtype == np.int16, 'Bad sample type: %r' % wav_data.dtype
waveform = wav_data / 32768.0 # Convert to [-1.0, +1.0]
waveform = waveform.astype('float32')
# Convert to mono and the sample rate expected by YAMNet.
if len(waveform.shape) > 1:
waveform = np.mean(waveform, axis=1)
if sr != params.sample_rate:
waveform = resampy.resample(waveform, sr, params.sample_rate)
# Predict YAMNet classes.
scores, embeddings, spectrogram = yamnet(waveform)
# Scores is a matrix of (time_frames, num_classes) classifier scores.
# Average them along time to get an overall classifier output for the clip.
prediction = np.mean(scores, axis=0)
# Report the highest-scoring classes and their scores.
top5_i = np.argsort(prediction)[::-1][:5]
print(file_name, ':\n' +
'\n'.join(' {:12s}: {:.3f}'.format(yamnet_classes[i], prediction[i])
for i in top5_i))
# print all classes
b = prediction.tolist() # nested lists with same data, indices
pred = []
for (i,cls) in enumerate(yamnet_classes):
item={}
item['label']=cls
item['value']=round(b[i], 6)
pred.append(item)
pred = sorted(pred, key=lambda x: x['value'], reverse=True)
json.dump(pred, codecs.open(event_path, 'w', encoding='utf-8'), separators=(',', ':'), sort_keys=True, indent=4) ### this saves the array in .json format
def __init__(self,
setType,
transform=None,
normalize=False,
crop_size=(256, 512),
root_dir=None): # setType: "train" or "test"
p = params.Params()
if root_dir is None:
root_dir = p.DATA_PATH
else:
root_dir = root_dir
self.setType = setType
self.normalize = normalize
self.crop_size = crop_size
if setType == "train":
path_paths_img_left = root_dir + p.sceneflow_paths_train_img_left
path_paths_img_right = root_dir + p.sceneflow_paths_train_img_right
path_paths_disp_left = root_dir + p.sceneflow_paths_train_disp_left
path_paths_disp_right = root_dir + p.sceneflow_paths_train_disp_right
if setType == "test":
path_paths_img_left = root_dir + p.sceneflow_paths_test_img_left
path_paths_img_right = root_dir + p.sceneflow_paths_test_img_right
path_paths_disp_left = root_dir + p.sceneflow_paths_test_disp_left
path_paths_disp_right = root_dir + p.sceneflow_paths_test_disp_right
finl = open(path_paths_img_left, 'rb')
finr = open(path_paths_img_right, 'rb')
self.paths_img_left = pickle.load(finl)
self.paths_img_right = pickle.load(finr)
finl.close()
finr.close()
finl = open(path_paths_disp_left, 'rb')
finr = open(path_paths_disp_right, 'rb')
self.paths_disp_left = pickle.load(finl)
self.paths_disp_right = pickle.load(finr)
finl.close()
finr.close()
assert len(self.paths_img_left) == len(self.paths_img_right) == len(
self.paths_disp_left) == len(self.paths_disp_right)
self.transform = transform
def scaffold_model(model_file, force=True):
scaff_out_file = model_file + '.scaffold_output'
try:
assert not force
scaff_params = params.read_params(scaff_out_file)
print(scaff_out_file)
except (IOError, AssertionError):
net = caffe_util.Net(model_file, phase=caffe.TRAIN)
scaff_params = params.Params()
scaff_params['n_params'] = net.get_n_params()
scaff_params['n_activs'] = net.get_n_activs()
scaff_params['size'] = net.get_approx_size()
scaff_params['min_width'] = net.get_min_width()
params.write_params(scaff_out_file, scaff_params)
return scaff_params
def make_tflite_export(weights_path, export_dir):
if os.path.exists(export_dir):
log('TF-Lite export already exists in {}, skipping TF-Lite export'.format(
export_dir))
return
# Create a TF-Lite compatible Module wrapper around YAMNet.
log('Building and checking TF-Lite Module ...')
params = yamnet_params.Params(tflite_compatible=True)
yamnet = YAMNet(weights_path, params)
check_model(yamnet, yamnet.class_map_path(), params)
log('Done')
# Make TF-Lite SavedModel export.
log('Making TF-Lite SavedModel export ...')
saved_model_dir = os.path.join(export_dir, 'saved_model')
os.makedirs(saved_model_dir)
tf.saved_model.save(
yamnet, saved_model_dir,
signatures={'serving_default': yamnet.__call__.get_concrete_function()})
log('Done')
# Check that the export can be loaded and works.
log('Checking TF-Lite SavedModel export in TF2 ...')
model = tf.saved_model.load(saved_model_dir)
check_model(model, model.class_map_path(), params)
log('Done')
# Make a TF-Lite model from the SavedModel.
log('Making TF-Lite model ...')
tflite_converter = tf.lite.TFLiteConverter.from_saved_model(
saved_model_dir, signature_keys=['serving_default'])
tflite_model = tflite_converter.convert()
tflite_model_path = os.path.join(export_dir, 'yamnet.tflite')
with open(tflite_model_path, 'wb') as f:
f.write(tflite_model)
log('Done')
# Check the TF-Lite export.
log('Checking TF-Lite model ...')
interpreter = tf.lite.Interpreter(tflite_model_path)
runner = interpreter.get_signature_runner('serving_default')
check_model(runner, 'yamnet_class_map.csv', params)
log('Done')
return saved_model_dir
def main(argv):
global analysisdata, frame_counter
log = open('/tmp/sound.log', 'w')
# Set up yamnet
params = yamnet_params.Params(sample_rate=ANALYSIS_SAMPLE_RATE,
patch_hop_seconds=0.1)
yamnet = yamnet_model.yamnet_frames_model(params)
yamnet.load_weights('/home/pi/models/research/audioset/yamnet/yamnet.h5')
yamnet_classes = yamnet_model.class_names(
'/home/pi/models/research/audioset/yamnet/yamnet_class_map.csv')
# Set up a live callback stream from the microphone
stream = sd.InputStream(device=1,
channels=1,
samplerate=RECORD_SAMPLE_RATE,
callback=audio_callback,
blocksize=BUFFER_SIZE_F)
with stream:
while True:
update_analysis_window()
if (frame_counter >= int(
ANALYSIS_LENGTH_S * ANALYSIS_SAMPLE_RATE)):
frame_counter = 0
scores = yamnet.predict(analysisdata, steps=1)[0]
if (len(scores)):
prediction = np.mean(scores, axis=0)
top5_i = np.argsort(prediction)[::-1][:1]
for x in top5_i:
if (prediction[x] > THRESHOLD):
top_class_str = yamnet_classes[x]
# Write any detected class (outside these noisy ones) to the log
if (not top_class_str in [
"Fireworks", "Silence",
"Inside, small room"
]):
log.write("[%s] %s %0.4f\n" %
(datetime.now().strftime(
"%m/%d/%Y %H:%M:%S"),
top_class_str, prediction[x]))
log.flush()
# And if it's one of the doorbell ones, ping the homebridge server
if (top_class_str in [
"Beep, bleep", "Doorbell", "Glass",
"Ding"
]):
trigger_homekit_motion()
def scale_data(dataset, trainFunctions):
params = Params.Params()
scaler = MinMaxScaler(feature_range=(-1, 1))
training_functions = [function.name for function in trainFunctions]
#fit_data = pd.concat([dataset[name] for name in training_functions])
#scaler = scaler.fit(fit_data)
#for name in training_functions:
# dataset[name] = pd.DataFrame(scaler.transform(dataset[name]))
scaler_return = False
for name in training_functions:
scaler = MinMaxScaler(feature_range=(-1, 1))
# print(dataset[name].describe(include='all'))
dataset[name] = pd.DataFrame(scaler.fit_transform(dataset[name]))
# print(dataset[name].describe(include='all'))
# Returning ellipse's scaler as it works better than rhombus in general.
if name == "elipse":
scaler_return = scaler
return dataset, scaler_return
def estimate_rot(data_num=1):
# Read the data
imu_data = read_data_imu(data_num)
q = quat.Quaternion(np.array([1, 0, 0, 0]))
omega = np.array([0, 0, 0])
curr_state = State(q, omega)
param = params.Params()
ts = np.squeeze(imu_data['ts'])
real_measurement = np.squeeze(imu_data['vals']).transpose()
size_ts = ts.shape[0]
sigma = Sigma_pts()
rpy = []
for i in range(1, size_ts):
dt = ts[i] - ts[i - 1]
sigma.find_points(curr_state, dt)
sigma.find_measurements()
corrected_measurements = convert_measurements(real_measurement[i, :3],
real_measurement[i, 3:])
curr_state.kalman_update(sigma, corrected_measurements)
rpy.append(curr_state.q.quat2rpy())
# plot vicon data
vicon_data = read_data_vicon(data_num)
v_ts = np.squeeze(vicon_data['ts'])
x, y, z = rotationMatrixToEulerAngles(vicon_data['rots'])
# plt.plot(v_ts, x, 'r')
plt.figure()
vicon, = plt.plot(v_ts, x, label="Vicon")
ukf, = plt.plot(ts[1:], np.asarray(rpy)[:, 0], label="UKF")
plt.legend(handles=[vicon, ukf])
plt.title("Roll")
plt.figure()
vicon, = plt.plot(v_ts, y, label="Vicon")
ukf, = plt.plot(ts[1:], np.asarray(rpy)[:, 1], label="UKF")
plt.legend(handles=[vicon, ukf])
plt.title("Pitch")
plt.figure()
vicon, = plt.plot(v_ts, z, label="Vicon")
ukf, = plt.plot(ts[1:], np.asarray(rpy)[:, 2], label="UKF")
plt.legend(handles=[vicon, ukf])
plt.title("Yaw")
plt.show()