def main():
""" Execution logic """
x_0, y_0 = 0.5, 2.2
segment_end = 5.5
step = 0.1
x_nodes_list, y_nodes_list = [], []
for accuracy in range(2, 5):
x_nodes, y_nodes = runge_kutta_interpolation(given_function, x_0, y_0,
segment_end, step,
accuracy)
x_nodes_list.append(x_nodes)
y_nodes_list.append(y_nodes)
display_plot_async(x_nodes_list, y_nodes_list)
print('Runge-Kutta interpolation results')
print(tp.header(['X', 'Y', 'Y2', 'Y3', 'Y4']))
for x, y, y2, y3, y4 in zip(
x_nodes_list[0], [explicit_solution(x) for x in x_nodes_list[0]],
y_nodes_list[0], y_nodes_list[1], y_nodes_list[2]):
print(tp.row([
x,
y,
y2,
y3,
y4,
]))
def minimize(self, f_df, x0, display=sys.stdout, maxiter=1e3):
self.display = display
self.theta = x0
# setup
xk = self.algorithm.send(destruct(x0).copy())
store = defaultdict(list)
runtimes = []
if len(self.operators) == 0:
self.operators = [proxops.identity()]
# setup
obj, grad = wrap(f_df, x0)
transform = compose(destruct, *reversed(self.operators), self.restruct)
self.optional_print(tp.header(['Iteration', 'Objective', '||Grad||', 'Runtime']))
try:
for k in count():
# setup
tstart = perf_counter()
f = obj(xk)
df = grad(xk)
xk = transform(self.algorithm.send(df))
runtimes.append(perf_counter() - tstart)
store['f'].append(f)
# Update display
self.optional_print(tp.row([k,
f,
np.linalg.norm(destruct(df)),
tp.humantime(runtimes[-1])]))
if k >= maxiter:
break
except KeyboardInterrupt:
pass
self.optional_print(tp.bottom(4))
# cleanup
self.optional_print(u'\u279b Final objective: {}'.format(store['f'][-1]))
self.optional_print(u'\u279b Total runtime: {}'.format(tp.humantime(sum(runtimes))))
self.optional_print(u'\u279b Per iteration runtime: {} +/- {}'.format(
tp.humantime(np.mean(runtimes)),
tp.humantime(np.std(runtimes)),
))
# result
return OptimizeResult({
'x': self.restruct(xk),
'f': f,
'df': self.restruct(df),
'k': k,
'obj': np.array(store['f']),
})
def update_display(self, iteration, disp_level, col_width=12):
"""
Prints information about the optimization procedure to standard output
Parameters
----------
iteration : int
The current iteration. Must either a positive integer or -1, which indicates the end of the algorithm
disp_level : int
An integer which controls how much information to display, ranging from 0 (nothing) to 3 (lots of stuff)
col_width : int
The width of each column in the data table, used if disp_level > 1
"""
# exit and print nothing if disp_level is zero
if disp_level == 0:
return
else:
# simple update, no table
if disp_level == 1 and iteration >= 0:
print('[Iteration %i]' % iteration)
# fancy table updates
if disp_level > 1:
# get the metadata from this iteration
data = self.metadata.tail(1).irow(0).to_dict()
# choose what keys to use
keys = ['Time (s)', 'Primal resid', 'Dual resid']
if disp_level > 2:
keys += ['Momentum (rho)', 'Primal runtimes']
# initial update. print out table headers
if iteration == 1:
hr = tableprint.hr(len(keys), column_width=col_width)
print('\n' + hr)
print(tableprint.header(keys, column_width=col_width))
print(hr)
# print data
tabledata = map(lambda d: float(d) if d.size == 1 else ', '.join(map(lambda i: '{:.2f}'.format(i), d)),
[np.array(data[key]) for key in keys])
print(tableprint.row(tabledata, column_width=col_width, precision='4g'))
if iteration == -1:
print(tableprint.hr(len(keys), column_width=col_width) + '\n')
# print convergence statement
if iteration == -1 and self.converged:
print('Converged after %i iterations!' % len(self.metadata))
def sample(data_path: str,
encoder_path: str,
vocab_path: str,
sample_length: int = 30,
output: str = None):
dump = torch.load(encoder_path, map_location=lambda storage, loc: storage)
encodermodel = dump['encodermodel']
decodermodel = dump['decodermodel']
# Some scaler (sklearn standardscaler)
scaler = dump['scaler']
# Also load previous training config
config_parameters = dump['config']
vocab = torch.load(vocab_path)
print(encodermodel)
print(decodermodel)
# load images from previous
encodermodel = encodermodel.to(DEVICE).eval()
decodermodel = decodermodel.to(DEVICE).eval()
kaldi_string = parsecopyfeats(data_path,
**config_parameters['feature_args'])
width_length = sample_length * 4
with stdout_or_file(output) as writer:
writer.write(
tp.header(["InputUtterance", "Output Sentence"],
style='grid',
width=width_length))
writer.write('\n')
for k, features in kaldi_io.read_mat_ark(kaldi_string):
features = scaler.transform(features)
# Add single batch dimension
features = torch.from_numpy(features).to(DEVICE).unsqueeze(0)
# Generate an caption embedding
encoded_feature, hiddens = encodermodel(features)
sampled_ids = decodermodel.sample(encoded_feature,
states=hiddens,
maxlength=sample_length)
# (1, max_seq_length) -> (max_seq_length)
sampled_ids = sampled_ids[0].cpu().numpy()
# Convert word_ids to words
sampled_caption = []
for word_id in sampled_ids:
word = vocab.idx2word[word_id]
sampled_caption.append(word)
if word == '<end>':
break
sentence = ''.join(sampled_caption)
# Print out the image and the generated caption
writer.write(
tp.row([k, sentence], style='grid', width=width_length))
writer.write('\n')
writer.flush()
writer.write(tp.bottom(2, style='grid', width=width_length))
def update_display(self,
iteration,
disp_level,
col_width=12): # pragma: no cover
"""
Prints information about the optimization procedure to standard output
Parameters
----------
iteration : int
The current iteration. Must either a positive integer or -1, which indicates the end of the algorithm
disp_level : int
An integer which controls how much information to display, ranging from 0 (nothing) to 3 (lots of stuff)
col_width : int
The width of each column in the data table, used if disp_level > 1
"""
# exit and print nothing if disp_level is zero
if disp_level == 0:
return
else:
# simple update, no table
if disp_level == 1 and iteration >= 0:
print('[Iteration %i]' % iteration)
# fancy table updates
if disp_level > 1:
# get the metadata from this iteration
data = valmap(last, self.metadata)
# choose what keys to use
keys = ['Time (s)', 'Primal resid', 'Dual resid', 'rho']
# initial update. print out table headers
if iteration == 1:
print(tableprint.header(keys, width=col_width))
# print data
print(
tableprint.row([data[k] for k in keys],
width=col_width,
format_spec='4g'))
if iteration == -1:
print(tableprint.bottom(len(keys), width=col_width) + '\n')
# print convergence statement
if iteration == -1 and self.converged:
print('Converged after %i iterations!' %
len(self.metadata['Primal resid']))
def buildTable(a, b, c, d):
n = len(a)
print(tp.header(['i', 'ai', 'bi', 'ci', 'di'], 10))
num = 1
for i in range(1, n - 1):
print(tp.row([num, "{:.3f}".format(a[i]), "{:.3f}".format(b[i]), \
"{:.3f}".format(c[i]), "{:.3f}".format(d[i-1])], 10))
num += 1
print(tp.row([num, "{:.3f}".format(a[n-1]), "{:.3f}".format(b[0]), \
"{:.3f}".format(c[n-1]), "{:.3f}".format(d[n-1])], 10))
print(tp.bottom(5, 10))
def _display_result(sample_mean, sample_variance, standard_deviation):
print('\n')
print(tp.header([
'sample mean',
'sample variance',
'standard deviation',
]))
print(tp.row([
sample_mean,
sample_variance,
standard_deviation,
]))
def _printfields(fields, print_header=True):
names, vals = [], []
for name, val, _ in fields:
names.append(name)
if val is None:
# Display None as empty
vals.append('')
else:
vals.append(val)
# Print header
if print_header:
print(tableprint.header(names, width=11))
# Print value row
print(tableprint.row(vals, width=11))
def geocode_crime_data(data_entries):
""" Geocodes a list of crime data entries """
print(tableprint.header(['City Name', 'Lat', 'Long'], width=30))
for entry in data_entries:
coords = geocode_city_name(entry['city_name'])
entry['coords'] = coords
# Add a delay for API limiting
time.sleep(0.1)
print(tableprint.bottom(3, width=30))
return data_entries
def run_explicit_euler_method():
""" Run implicit euler interpolation calculating """
step = 0.05
x_0, y_0 = 0, 1
segment_end = 1
x_nodes, y_nodes = explicit_euler_interpolation(given_function,
x_0,
y_0,
segment_start=x_0,
segment_end=segment_end,
step=step)
display_plot_async(x_nodes, y_nodes, "Explicit Eulerian interpolation")
print('\nExplicit Euler method results')
print(tp.header([
'X',
'Y',
'Interpolated',
]))
for x, y in zip(x_nodes, y_nodes):
print(tp.row([x, onhand_solution(x), y]))
def _display_result(observable_chi, critical_chi):
print('\nhypothesis of a normal distribution')
print(tp.header(['Observable chi', 'Critical chi', ]))
print(tp.row([observable_chi, critical_chi, ]))
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"serial_number",
type=int,
help=
"the 10-digit serial number found under the magnetic backplate of your Wave Plus",
)
parser.add_argument(
"--sample-period",
type=int,
default=300,
help=
"the number of seconds between reading the current values. Default: %(default)s",
)
parser.add_argument("--pipe",
action="store_true",
help="pipe the results to a file")
parser.add_argument(
"--statusbar",
action="store_true",
help="print air quality status suitable for statusbar",
)
parser.add_argument(
"--mac-addr",
help="the MAC address of the Wave Plus device",
)
args = parser.parse_args()
if len(str(args.serial_number)) != 10:
print("ERROR: Invalid SN format.")
parser.print_usage()
sys.exit(1)
if args.sample_period <= 0:
print("ERROR: Invalid SAMPLE-PERIOD. Must be larger than zero.")
parser.print_usage()
sys.exit(1)
try:
waveplus = WavePlus(args.serial_number, args.mac_addr)
header = [
"Humidity",
"Radon ST avg",
"Radon LT avg",
"Temperature",
"Pressure",
"CO2 level",
"VOC level",
]
if args.pipe:
print(*header, sep=",")
elif not args.statusbar:
print(tableprint.header(header, width=TABLEPRINT_WIDTH))
while True:
waveplus.connect()
sensors = waveplus.read()
data = {var: sensors.get_variable(var) for var in VARIABLES}
if args.statusbar:
statusbar_print(data)
sys.exit(0)
if args.pipe:
print(*data.values(), sep=",")
else:
print(
tableprint.row(list(map(str, data.values())),
width=TABLEPRINT_WIDTH))
waveplus.disconnect()
time.sleep(args.sample_period)
finally:
waveplus.disconnect()
def sample(self,
experiment_path: str,
feature_file: str,
feature_scp: str,
output: str = "output_word.txt",
**kwargs):
"""Generate captions given experiment model"""
"""kwargs: {'max_length': int, 'method': str, 'beam_size': int}"""
import tableprint as tp
dump = torch.load(os.path.join(experiment_path, "saved.pth"),
map_location="cpu")
# Load previous training config
config = dump["config"]
vocab_size = len(torch.load(config["vocab_file"]))
model = self._get_model(config, vocab_size)
model.load_state_dict(dump["model"])
# Some scaler (sklearn standardscaler)
scaler = dump["scaler"]
vocabulary = torch.load(config["vocab_file"])
zh = config["zh"]
model = model.to(self.device)
dataset = SJTUDatasetEval(feature=feature_file,
eval_scp=feature_scp,
transform=scaler.transform)
dataloader = torch.utils.data.DataLoader(dataset,
shuffle=False,
collate_fn=collate_fn((1, )),
batch_size=16,
num_workers=0)
width_length = 80
pbar = ProgressBar(persist=False, ascii=True)
writer = open(os.path.join(experiment_path, output), "w")
writer.write(
tp.header(["InputUtterance", "Output Sentence"],
width=[len("InputUtterance"), width_length]))
writer.write('\n')
sentences = []
def _sample(engine, batch):
# batch: [keys, feats, feat_lens]
with torch.no_grad():
model.eval()
keys = batch[0]
output = self._forward(model, batch, mode="sample", **kwargs)
seqs = output["seqs"].cpu().numpy()
for idx, seq in enumerate(seqs):
caption = self._convert_idx2sentence(seq,
vocabulary,
zh=zh)
if zh:
sentence = " ".join(caption)
else:
sentence = caption
writer.write(
tp.row([keys[idx], sentence],
width=[len("InputUtterance"), width_length]) +
"\n")
sentences.append(sentence)
sample_engine = Engine(_sample)
pbar.attach(sample_engine)
sample_engine.run(dataloader)
writer.write(
tp.bottom(2, width=[len("InputUtterance"), width_length]) + "\n")
writer.write("Unique sentence number: {}\n".format(len(
set(sentences))))
writer.close()
for serial in device_serials:
waveplus_devices.append(WavePlus(int(serial)))
if (Mode == 'terminal'):
print("\nPress ctrl+C to exit program\n")
# print "Device serial number: %s" %(SerialNumber)
header = [
'Timestamp', 'Sensor Serial Number', 'Humidity - %rH',
'Radon ST avg - Bq/m3', 'Radon LT avg - Bq/m3', 'T - degC', 'P - hPa',
'CO2 - ppm', 'VOC - ppb'
]
if (Mode == 'terminal'):
print(tableprint.header(header, width=12))
elif (Mode == 'pipe'):
print(header)
elif (Mode == 'file'):
file = open(outfile, 'a+')
file.write(','.join(header) + "\n")
file.close()
while True:
for waveplus in waveplus_devices:
tries = 0
while tries <= MaxRetries:
try:
tries = tries + 1
connected = waveplus.connect()
if not connected:
def on_training_started(engine, outputfun=sys.stdout.write, header=[]):
outputfun("<== Training Started ==>")
for line in tp.header(header, style="grid").split("\n"):
outputfun(line)
def main(config='config/train.yaml', **kwargs):
"""Trains a model on the given features and vocab.
:config: A training configuration. Note that all parameters in the config can also be manually adjusted with --ARG VALUE
:returns: None
"""
config_parameters = parse_config_or_kwargs(config, **kwargs)
outputdir = os.path.join(
config_parameters['outputpath'],
config_parameters['model'],
datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S-%f'))
try:
os.makedirs(outputdir)
except IOError:
pass
logger = genlogger(outputdir, 'train.log')
logger.info("Storing data at: {}".format(outputdir))
logger.info("<== Passed Arguments ==>")
# Print arguments into logs
for line in pformat(config_parameters).split('\n'):
logger.info(line)
# seed setting
seed=config_parameters['seed'] # 1~5
np.random.seed(seed)
torch.manual_seed(seed)
kaldi_string = parsecopyfeats(
config_parameters['features'], **config_parameters['feature_args'])
scaler = getattr(pre, config_parameters['scaler'])(**config_parameters['scaler_args'])
logger.info("<== Estimating Scaler ({}) ==>".format(scaler.__class__.__name__))
inputdim = -1
for kid, feat in kaldi_io.read_mat_ark(kaldi_string):
scaler.partial_fit(feat)
inputdim = feat.shape[-1]
assert inputdim > 0, "Reading inputstream failed"
logger.info(
"Features: {} Input dimension: {}".format(
config_parameters['features'],
inputdim))
logger.info("<== Labels ==>")
# Can be label, DAT, DADA ... default is 'label'
target_label_name = config_parameters.get('label_type', 'label')
if target_label_name == 'label':
label_df = pd.read_csv(config_parameters['labels'], sep=' ', names=['speaker', 'filename', 'physical', 'system', 'label'])
else: # 'DAT' or 'DADA'
label_df = pd.read_csv(config_parameters['labels'], sep=' ', names=['speaker', 'filename', 'physical', 'system', 'label', 'domain'])
label_encoder = pre.LabelEncoder()
if target_label_name == 'label':
label_encoder.fit(label_df[target_label_name].values.astype(str))
# Labelencoder needs an iterable to work, so just put a list around it and fetch again the 0-th element ( just the encoded string )
label_df['class_encoded'] = label_df[target_label_name].apply(lambda x: label_encoder.transform([x])[0])
train_labels = label_df[['filename', 'class_encoded']].set_index('filename').loc[:, 'class_encoded'].to_dict()
else: # 'DAT' or 'DADA'
label_encoder_sub = pre.LabelEncoder()
label_encoder.fit(label_df['label'].values.astype(str))
label_df['lab_encoded'] = label_df['label'].apply(lambda x: label_encoder.transform([x])[0])
label_encoder_sub.fit(label_df['domain'].values.astype(str))
label_df['domain_encoded'] = label_df['domain'].apply(lambda x: label_encoder_sub.transform([x])[0])
train_labels = label_df[['filename', 'lab_encoded', 'domain_encoded']].set_index('filename').to_dict('index')
train_labels = {k:np.array(list(v.values())) for k, v in train_labels.items()}
# outdomain
outdomain = config_parameters['outdomain']
outdomain_label = label_encoder_sub.transform([outdomain])[0]
logger.info("Outdomain: {}, Outdomain label: {}".format(outdomain, outdomain_label))
if target_label_name == 'label':
train_dataloader, cv_dataloader = create_dataloader_train_cv(kaldi_string, train_labels, transform=scaler.transform, target_label_name=target_label_name, **config_parameters['dataloader_args'])
else: #'DAT' or 'DADA'
outdomain_train_dataloader, indomain_train_dataloader, cv_dataloader = create_dataloader_train_cv(kaldi_string, train_labels, transform=scaler.transform, target_label_name=target_label_name, outdomain_label=outdomain_label, **config_parameters['dataloader_args'])
if target_label_name == 'label':
model = getattr(models, config_parameters['model'])(inputdim=inputdim, outputdim=len(label_encoder.classes_), **config_parameters['model_args'])
else: # 'DAT' or 'DADA'
model = getattr(models, config_parameters['model'])(inputdim=inputdim, outputdim1=len(label_encoder.classes_), outputdim2=len(label_encoder_sub.classes_), **config_parameters['model_args'])
logger.info("<== Model ==>")
for line in pformat(model).split('\n'):
logger.info(line)
optimizer = getattr(torch.optim, config_parameters['optimizer'])(model.parameters(), **config_parameters['optimizer_args'])
scheduler = getattr(torch.optim.lr_scheduler, config_parameters['scheduler'])(optimizer, **config_parameters['scheduler_args'])
criterion = getattr(loss, config_parameters['loss'])(**config_parameters['loss_args'])
trainedmodelpath = os.path.join(outputdir, 'model.th')
model = model.to(device)
criterion_improved = criterion_improver(config_parameters['improvecriterion'])
header = [
'Epoch',
'Lr',
'Loss(T)',
'Loss(CV)',
"Acc(T)",
"Acc(CV)",
]
for line in tp.header(header, style='grid').split('\n'):
logger.info(line)
for epoch in range(1, config_parameters['epochs']+1):
if target_label_name == 'label':
train_loss, train_acc = runepoch(train_dataloader, None, model, criterion, target_label_name, optimizer, dotrain=True, epoch=epoch)
else: # 'DAT' or 'DADA'
train_loss, train_acc = runepoch(outdomain_train_dataloader, indomain_train_dataloader, model, criterion, target_label_name, optimizer, dotrain=True, epoch=epoch)
cv_loss, cv_acc = runepoch(cv_dataloader, None, model, criterion, target_label_name, dotrain=False, epoch=epoch)
logger.info(
tp.row(
(epoch,) + (optimizer.param_groups[0]['lr'],) +
(str(train_loss), str(cv_loss), str(train_acc), str(cv_acc)),
style='grid'))
epoch_meanloss = cv_loss[0] if type(cv_loss)==tuple else cv_loss
if epoch % config_parameters['saveinterval'] == 0:
torch.save({'model': model,
'scaler': scaler,
'encoder': label_encoder,
'config': config_parameters},
os.path.join(outputdir, 'model_{}.th'.format(epoch)))
# ReduceOnPlateau needs a value to work
schedarg = epoch_meanloss if scheduler.__class__.__name__ == 'ReduceLROnPlateau' else None
scheduler.step(schedarg)
if criterion_improved(epoch_meanloss):
torch.save({'model': model,
'scaler': scaler,
'encoder': label_encoder,
'config': config_parameters},
trainedmodelpath)
if optimizer.param_groups[0]['lr'] < 1e-7:
break
logger.info(tp.bottom(len(header), style='grid'))
logger.info("Results are in: {}".format(outputdir))
def on_epoch_begin(self, epoch, logs={}):
tp.banner(f"Epoch {epoch}")
print(tp.header(['iter', 'loss']))
def train(model, experiment, monitor, num_epochs, augment=False):
"""Train the given network against the given data
Parameters
----------
model : keras.models.Model or glms.GLM
A GLM or Keras Model object
experiment : experiments.Experiment
An Experiment object
monitor : io.Monitor
Saves the model parameters and plots of performance progress
num_epochs : int
Number of epochs to train for
reduce_lr_every : int
How often to reduce the learning rate
reduce_rate : float
A fraction (constant) to multiply the learning rate by
"""
assert isinstance(model, (Model, GLM)), "'model' must be a GLM or Keras model"
# initialize training iteration
iteration = 0
train_start = time()
# loop over epochs
try:
for epoch in range(num_epochs):
tp.banner('Epoch #{} of {}'.format(epoch + 1, num_epochs))
print(tp.header(["Iteration", "Loss", "Runtime"]), flush=True)
# loop over data batches for this epoch
for X, y in experiment.train(shuffle=True):
# update on save_every, assuming it is positive
if (monitor is not None) and (iteration % monitor.save_every == 0):
# performs validation, updates performance plots, saves results to dropbox
monitor.save(epoch, iteration, X, y, model.predict)
# train on the batch
tstart = time()
loss = model.train_on_batch({'stim':X, 'loss':y})[0]
elapsed_time = time() - tstart
# update
iteration += 1
print(tp.row([iteration, float(loss), tp.humantime(elapsed_time)]), flush=True)
print(tp.bottom(3))
except KeyboardInterrupt:
print('\nCleaning up')
# allows the monitor to perform any post-training visualization
if monitor is not None:
elapsed_time = time() - train_start
monitor.cleanup(iteration, elapsed_time)
tp.banner('Training complete!')
def check_grad(f_df, xref, stepsize=1e-6, tol=1e-6, width=15, style='round', out=sys.stdout):
"""
Compares the numerical gradient to the analytic gradient
Parameters
----------
f_df : function
The analytic objective and gradient function to check
x0 : array_like
Parameter values to check the gradient at
stepsize : float, optional
Stepsize for the numerical gradient. Too big and this will poorly estimate the gradient.
Too small and you will run into precision issues (default: 1e-6)
tol : float, optional
Tolerance to use when coloring correct/incorrect gradients (default: 1e-5)
width : int, optional
Width of the table columns (default: 15)
style : string, optional
Style of the printed table, see tableprint for a list of styles (default: 'round')
"""
CORRECT = u'\x1b[32m\N{CHECK MARK}\x1b[0m'
INCORRECT = u'\x1b[31m\N{BALLOT X}\x1b[0m'
obj, grad = wrap(f_df, xref, size=0)
x0 = destruct(xref)
df = grad(x0)
# header
out.write(tp.header(["Numerical", "Analytic", "Error"], width=width, style=style) + "\n")
out.flush()
# helper function to parse a number
def parse_error(number):
# colors
failure = "\033[91m"
passing = "\033[92m"
warning = "\033[93m"
end = "\033[0m"
base = "{}{:0.3e}{}"
# correct
if error < 0.1 * tol:
return base.format(passing, error, end)
# warning
elif error < tol:
return base.format(warning, error, end)
# failure
else:
return base.format(failure, error, end)
# check each dimension
num_errors = 0
for j in range(x0.size):
# take a small step in one dimension
dx = np.zeros(x0.size)
dx[j] = stepsize
# compute the centered difference formula
df_approx = (obj(x0 + dx) - obj(x0 - dx)) / (2 * stepsize)
df_analytic = df[j]
# absolute error
abs_error = np.linalg.norm(df_approx - df_analytic)
# relative error
error = abs_error if np.allclose(abs_error, 0) else abs_error / \
(np.linalg.norm(df_analytic) + np.linalg.norm(df_approx))
num_errors += error >= tol
errstr = CORRECT if error < tol else INCORRECT
out.write(tp.row([df_approx, df_analytic, parse_error(error) + ' ' + errstr],
width=width, style=style) + "\n")
out.flush()
out.write(tp.bottom(3, width=width, style=style) + "\n")
return num_errors
def main(features: str, vocab_file: str,
config='config/trainconfig.yaml', **kwargs):
"""Trains a model on the given features and vocab.
:features: str: Input features. Needs to be kaldi formatted file
:vocab_file:str: Vocabulary generated by using build_vocab.py
:config: A training configuration. Note that all parameters in the config can also be manually adjusted with --ARG=VALUE
:returns: None
"""
config_parameters = parse_config_or_kwargs(config, **kwargs)
outputdir = os.path.join(
config_parameters['encodermodel'] + '_' +
config_parameters['decodermodel'],
datetime.datetime.now().strftime('%Y-%m-%d_%H-%M'))
try:
os.makedirs(outputdir)
except IOError:
pass
logger = genlogger(outputdir, 'train.log')
logger.info("Storing data at: {}".format(outputdir))
logger.info("<== Passed Arguments ==>")
# Print arguments into logs
for line in pformat(config_parameters).split('\n'):
logger.info(line)
kaldi_string = parsecopyfeats(
features, **config_parameters['feature_args'])
scaler = getattr(
pre, config_parameters['scaler'])(
**config_parameters['scaler_args'])
inputdim = -1
logger.info(
"<== Estimating Scaler ({}) ==>".format(
scaler.__class__.__name__))
for kid, feat in kaldi_io.read_mat_ark(kaldi_string):
scaler.partial_fit(feat)
inputdim = feat.shape[-1]
assert inputdim > 0, "Reading inputstream failed"
vocabulary = torch.load(vocab_file)
vocab_size = len(vocabulary)
logger.info(
"Features: {} Input dimension: {} Vocab Size: {}".format(
features, inputdim, vocab_size))
if 'load_pretrained' in config_parameters and config_parameters['load_pretrained']:
encodermodeldump = torch.load(
config_parameters['load_pretrained'],
map_location=lambda storage, loc: storage)
pretrainedmodel = encodermodeldump['encodermodel']
encodermodel = models.PreTrainedCNN(
inputdim=inputdim, pretrained_model=pretrainedmodel, **
config_parameters['encodermodel_args'])
else:
encodermodel = getattr(
models, config_parameters['encodermodel'])(
inputdim=inputdim, **config_parameters['encodermodel_args'])
decodermodel = getattr(
models, config_parameters['decodermodel'])(
vocab_size=vocab_size, **config_parameters['decodermodel_args'])
logger.info("<== EncoderModel ==>")
for line in pformat(encodermodel).split('\n'):
logger.info(line)
logger.info("<== DecoderModel ==>")
for line in pformat(decodermodel).split('\n'):
logger.info(line)
params = list(encodermodel.parameters()) + list(decodermodel.parameters())
train_dataloader, cv_dataloader = create_dataloader_train_cv(
kaldi_string,
config_parameters['captions_file'],
vocab_file,
transform=scaler.transform,
**config_parameters['dataloader_args'])
optimizer = getattr(
torch.optim, config_parameters['optimizer'])(
params,
**config_parameters['optimizer_args'])
scheduler = getattr(
torch.optim.lr_scheduler,
config_parameters['scheduler'])(
optimizer,
**config_parameters['scheduler_args'])
criterion = torch.nn.CrossEntropyLoss()
trainedmodelpath = os.path.join(outputdir, 'model.th')
encodermodel = encodermodel.to(device)
decodermodel = decodermodel.to(device)
criterion_improved = criterion_improver(
config_parameters['improvecriterion'])
for line in tp.header(
['Epoch', 'MeanLoss(T)', 'StdLoss(T)', 'Loss(CV)', 'StdLoss(CV)',
"Acc(T)", "Acc(CV)", "Forcing?"],
style='grid').split('\n'):
logger.info(line)
teacher_forcing_ratio = config_parameters['teacher_forcing_ratio']
for epoch in range(1, config_parameters['epochs']+1):
use_teacher_forcing = True if random.random() < teacher_forcing_ratio else False
train_loss_mean_std, train_acc = trainepoch(
train_dataloader, encodermodel, decodermodel, criterion, optimizer,
vocabulary, use_teacher_forcing)
cv_loss_mean_std, cv_acc = sample_cv(
cv_dataloader, encodermodel, decodermodel, criterion)
logger.info(
tp.row(
(epoch,) + train_loss_mean_std + cv_loss_mean_std +
(train_acc, cv_acc, use_teacher_forcing),
style='grid'))
epoch_meanloss = cv_loss_mean_std[0]
if epoch % config_parameters['saveinterval'] == 0:
torch.save({'encodermodel': encodermodel,
'decodermodel': decodermodel, 'scaler': scaler,
'config': config_parameters},
os.path.join(outputdir, 'model_{}.th'.format(epoch)))
# ReduceOnPlateau needs a value to work
schedarg = epoch_meanloss if scheduler.__class__.__name__ == 'ReduceLROnPlateau' else None
scheduler.step(schedarg)
if criterion_improved(epoch_meanloss):
torch.save({'encodermodel': encodermodel,
'decodermodel': decodermodel, 'scaler': scaler,
'config': config_parameters},
trainedmodelpath)
else:
dump = torch.load(trainedmodelpath)
encodermodel.load_state_dict(dump['encodermodel'].state_dict())
decodermodel.load_state_dict(dump['decodermodel'].state_dict())
if optimizer.param_groups[0]['lr'] < 1e-6:
break
logger.info(tp.bottom(8, style='grid'))
# Sample results
from sample import sample
sample(
data_path=features,
encoder_path=trainedmodelpath,
vocab_path=vocab_file,
output=os.path.join(
outputdir,
'output_word.txt'))
try:
#---- Connect to device ----#
wave = Wave(SerialNumber)
wave.scan()
if (Mode == 'terminal'):
print "\nPress ctrl-C to exit program\n"
print "Device serial number: %s" % (SerialNumber)
header = [
'Datetime', 'Humidity', 'Temperature', 'Radon ST avg', 'Radon LT avg'
]
if (Mode == 'terminal'):
print tableprint.header(header, width=20)
elif (Mode == 'pipe'):
print header
while True:
wave.connect()
# read current values
date_time = wave.read(SENSOR_IDX_DATETIME)
humidity = wave.read(SENSOR_IDX_HUMIDITY)
temperature = wave.read(SENSOR_IDX_TEMPERATURE)
radon_st_avg = wave.read(SENSOR_IDX_RADON_ST_AVG)
radon_lt_avg = wave.read(SENSOR_IDX_RADON_LT_AVG)
data = [date_time, humidity, temperature, radon_st_avg, radon_lt_avg]
# Print data
waveplus = WavePlus(SerialNumber)
ledController = LedController(LedMode)
if (Mode == 'terminal'):
print("\nPress ctrl+C to exit program\n")
print(f"Device serial number: {SerialNumber}")
header = [
'Date, time', 'Humidity', 'Radon ST avg', 'Radon LT avg',
'Temperature', 'Pressure', 'CO2 level', 'VOC level'
]
COLUMN_WIDTH = 12
if (Mode == 'terminal'):
print(tableprint.header(header, width=COLUMN_WIDTH))
elif (Mode == 'pipe'):
print(header)
MAX_FAILURES = 5
failure_count = 0
while True:
ledController.OnCommsStart()
try:
waveplus.connect()
failure_count = 0
except:
ledController.OnCommsEnd()
if failure_count < MAX_FAILURES:
def check_grad(f_df,
xref,
stepsize=1e-6,
tol=1e-6,
width=15,
style='round',
out=sys.stdout):
"""
Compares the numerical gradient to the analytic gradient
Parameters
----------
f_df : function
The analytic objective and gradient function to check
x0 : array_like
Parameter values to check the gradient at
stepsize : float, optional
Stepsize for the numerical gradient. Too big and this will poorly estimate the gradient.
Too small and you will run into precision issues (default: 1e-6)
tol : float, optional
Tolerance to use when coloring correct/incorrect gradients (default: 1e-5)
width : int, optional
Width of the table columns (default: 15)
style : string, optional
Style of the printed table, see tableprint for a list of styles (default: 'round')
"""
CORRECT = u'\x1b[32m\N{CHECK MARK}\x1b[0m'
INCORRECT = u'\x1b[31m\N{BALLOT X}\x1b[0m'
obj, grad = wrap(f_df, xref, size=0)
x0 = destruct(xref)
df = grad(x0)
# header
out.write(
tp.header(["Numerical", "Analytic", "Error"], width=width, style=style)
+ "\n")
out.flush()
# helper function to parse a number
def parse_error(number):
# colors
failure = "\033[91m"
passing = "\033[92m"
warning = "\033[93m"
end = "\033[0m"
base = "{}{:0.3e}{}"
# correct
if error < 0.1 * tol:
return base.format(passing, error, end)
# warning
elif error < tol:
return base.format(warning, error, end)
# failure
else:
return base.format(failure, error, end)
# check each dimension
num_errors = 0
for j in range(x0.size):
# take a small step in one dimension
dx = np.zeros(x0.size)
dx[j] = stepsize
# compute the centered difference formula
df_approx = (obj(x0 + dx) - obj(x0 - dx)) / (2 * stepsize)
df_analytic = df[j]
# absolute error
abs_error = np.linalg.norm(df_approx - df_analytic)
# relative error
error = abs_error if np.allclose(abs_error, 0) else abs_error / \
(np.linalg.norm(df_analytic) + np.linalg.norm(df_approx))
num_errors += error >= tol
errstr = CORRECT if error < tol else INCORRECT
out.write(
tp.row([df_approx, df_analytic,
parse_error(error) + ' ' + errstr],
width=width,
style=style) + "\n")
out.flush()
out.write(tp.bottom(3, width=width, style=style) + "\n")
return num_errors
# shift = self.mean * (now - self.last_computed_sys_time)
# tti[0] = tti[0] + shift
# tti[1] = tti[1] + shift
# tti[2] = tti[1] + shift
tti.append(now)
return tti
samples = NTPMonitor(5, sys.argv[1], 3, 300, .999999999)
width = 22
print(
tableprint.header([
'Last Offset (us)', 'Last Delay', 'Mean Offset (us)', 'RMS Offset',
'Offset St. Dev. (us)', 'Interval Width (us)', 'Interval Lower',
'Interval Upper'
],
style='clean',
width=width))
print(
tableprint.row([
samples.last.offset * 1e+6,
np.average([x.offset for x in samples.data[-30:]]) * 1e6,
samples.mean * 1e+6, samples.rms_offset * 1e6, samples.stdev * 1e+6,
samples.tt_interval()[3] * 1e6,
samples.tt_interval()[0],
samples.tt_interval()[1]
],
width=width,
style='clean',
try:
#---- Initialize ----#
waveplus = WavePlus(SerialNumber)
if (Mode == 'terminal'):
print "\nPress ctrl+C to exit program\n"
print "Device serial number: %s" % (SerialNumber)
header = [
'Humidity', 'Radon ST avg', 'Radon LT avg', 'Temperature', 'Pressure',
'CO2 level', 'VOC level'
]
if (Mode == 'terminal'):
print tableprint.header(header, width=12)
elif (Mode == 'pipe'):
print header
while True:
waveplus.connect()
# read values
sensors = waveplus.read()
# extract
humidity = str(sensors.getValue(SENSOR_IDX_HUMIDITY)) + " " + str(
sensors.getUnit(SENSOR_IDX_HUMIDITY))
radon_st_avg = str(
sensors.getValue(SENSOR_IDX_RADON_SHORT_TERM_AVG)) + " " + str(
def start(self):
print('\n'.join((self.hr,
tp.header(self.column_names, self.width),
self.hr)), flush=True)
def table_header(self, header, *args, **kwargs):
self.write(tableprint.header(header, *args, **kwargs))
def main(config='config/ReLU/0Pool/crnn_maxpool.yaml', **kwargs):
"""Trains a model on the given features and vocab.
:features: str: Input features. Needs to be kaldi formatted file
:config: A training configuration. Note that all parameters in the config can also be manually adjusted with --ARG=VALUE
:returns: None
"""
config_parameters = parse_config_or_kwargs(config, **kwargs)
outputdir = os.path.join(
config_parameters['outputpath'],
config_parameters['model'],
datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%f'))
try:
os.makedirs(outputdir)
except IOError:
pass
logger = genlogger(outputdir, 'train.log')
logger.info("Storing data at: {}".format(outputdir))
logger.info("<== Passed Arguments ==>")
# Print arguments into logs
for line in pformat(config_parameters).split('\n'):
logger.info(line)
kaldi_string = parsecopyfeats(
config_parameters['features'], **config_parameters['feature_args'])
scaler = getattr(
pre, config_parameters['scaler'])(
**config_parameters['scaler_args'])
inputdim = -1
logger.info(
"<== Estimating Scaler ({}) ==>".format(
scaler.__class__.__name__))
for kid, feat in kaldi_io.read_mat_ark(kaldi_string):
scaler.partial_fit(feat)
inputdim = feat.shape[-1]
assert inputdim > 0, "Reading inputstream failed"
logger.info(
"Features: {} Input dimension: {}".format(
config_parameters['features'],
inputdim))
logger.info("<== Labels ==>")
label_df = pd.read_csv(config_parameters['labels'], sep='\t')
label_df.event_labels = label_df.event_labels.str.split(',')
label_df = label_df.set_index('filename')
uniquelabels = list(np.unique(
[item
for row in label_df.event_labels.values
for item in row]))
many_hot_encoder = ManyHotEncoder(
label_list=uniquelabels,
time_resolution=1
)
label_df['manyhot'] = label_df['event_labels'].apply(
lambda x: many_hot_encoder.encode(x, 1).data.flatten())
utt_labels = label_df.loc[:, 'manyhot'].to_dict()
train_dataloader, cv_dataloader = create_dataloader_train_cv(
kaldi_string,
utt_labels,
transform=scaler.transform,
**config_parameters['dataloader_args'])
model = getattr(
models,
config_parameters['model'])(
inputdim=inputdim,
output_size=len(uniquelabels),
**config_parameters['model_args'])
logger.info("<== Model ==>")
for line in pformat(model).split('\n'):
logger.info(line)
optimizer = getattr(
torch.optim, config_parameters['optimizer'])(
model.parameters(),
**config_parameters['optimizer_args'])
scheduler = getattr(
torch.optim.lr_scheduler,
config_parameters['scheduler'])(
optimizer,
**config_parameters['scheduler_args'])
criterion = getattr(losses, config_parameters['loss'])(
**config_parameters['loss_args'])
trainedmodelpath = os.path.join(outputdir, 'model.th')
model = model.to(device)
criterion_improved = criterion_improver(
config_parameters['improvecriterion'])
header = [
'Epoch',
'UttLoss(T)',
'UttLoss(CV)',
"UttAcc(T)",
"UttAcc(CV)",
"mAUC(CV)"]
for line in tp.header(
header,
style='grid').split('\n'):
logger.info(line)
poolingfunction_name = config_parameters['poolingfunction']
pooling_function = parse_poolingfunction(poolingfunction_name)
for epoch in range(1, config_parameters['epochs']+1):
train_utt_loss_mean_std, train_utt_acc, train_auc_utt = runepoch(
train_dataloader, model, criterion, optimizer, dotrain=True, poolfun=pooling_function)
cv_utt_loss_mean_std, cv_utt_acc, cv_auc_utt = runepoch(
cv_dataloader, model, criterion, dotrain=False, poolfun=pooling_function)
logger.info(
tp.row(
(epoch,) +
(train_utt_loss_mean_std[0],
cv_utt_loss_mean_std[0],
train_utt_acc, cv_utt_acc, cv_auc_utt),
style='grid'))
epoch_meanloss = cv_utt_loss_mean_std[0]
if epoch % config_parameters['saveinterval'] == 0:
torch.save({'model': model,
'scaler': scaler,
'encoder': many_hot_encoder,
'config': config_parameters},
os.path.join(outputdir, 'model_{}.th'.format(epoch)))
# ReduceOnPlateau needs a value to work
schedarg = epoch_meanloss if scheduler.__class__.__name__ == 'ReduceLROnPlateau' else None
scheduler.step(schedarg)
if criterion_improved(epoch_meanloss):
torch.save({'model': model,
'scaler': scaler,
'encoder': many_hot_encoder,
'config': config_parameters},
trainedmodelpath)
if optimizer.param_groups[0]['lr'] < 1e-7:
break
logger.info(tp.bottom(len(header), style='grid'))
logger.info("Results are in: {}".format(outputdir))
return outputdir
