def main(argv):
parser = argparse.ArgumentParser(description=' Batch Movie Encoder. Converts from one container to another with specified codec for all files in the specified Destination path.',
epilog='If the codec specified is the same as the codec in the source container only a copy will not be done and no transcoding to speed up the process.')
parser.add_argument('-p',help='Destination path',required=True)
parser.add_argument('-i',choices=Encoder.SUPPORTEDINPUTCONTAINER,
help='The input movie container format. Generally the same as the file extension. Eg .avi',required=True)
parser.add_argument('-o',choices=Encoder.SUPPORTEDTARGETCONTAINER,
help='The output movie container format.',required=True)
parser.add_argument('-c',choices=Encoder.SUPPORTEDTARGETVIDEOFORMATS,
help='The video codec to encode with.',required=False,default='H264')
parser.add_argument('-r',help='Recurse subdirectories.')
parser.add_argument('-d',help='Delete source file after conversion completion.')
EncoderOptions= parser.parse_args()
try:
movieEncoder=Encoder(EncoderOptions.p,EncoderOptions.i,EncoderOptions.o,EncoderOptions.c)
movieEncoder.queueFiles(helper.getFilesExt(EncoderOptions.p,EncoderOptions.i))
movieEncoder.encodeBatch()
except Exception as e:
print e
def encode(self):
raw_content = open(self.args.input, 'r').read()
encoder = Encoder(raw_content)
encoded_content, reversed_mappings = encoder.encode()
with open(self.args.output, 'wb') as f:
f.write(encoded_content)
with open(self.args.mappings, 'w') as f:
json_mappings = json.dumps(reversed_mappings, separators=(',', ':'))
f.write(json_mappings)
def translate(**kwargs):
## logging configuration
log_lvl = log_lvls.get(kwargs.get('log_lvl', '10'))
logging.config.fileConfig(os.path.join(pwd, "logging.conf"))
logging.getLogger().setLevel(log_lvl)
prg = kwargs.get('prg', None)
out_dir = kwargs.get('out_dir', res_dir)
sk = kwargs.get('sketch', True)
fs = kwargs.get('fs', False)
cgen = kwargs.get('custom_gen', False)
cntr = kwargs.get('cntr', False)
skv = kwargs.get('skv', 0)
lib = kwargs.get('lib', True)
codegen_jar = os.path.join(root_dir, "codegen", "lib", "codegen.jar")
logging.info('parsing {}'.format(prg))
prg_ast = parse(prg,lib=lib)
util.add_object(prg_ast)
encoder = Encoder(prg_ast, out_dir, fs)
logging.info('encoding to Sketch')
encoder.to_sk()
# Sketch options
opts = kwargs.get('opts', [])
# print counter examples
if cntr: opts.extend(['-V3', '--debug-cex'])
if skv != 0: opts.extend(['-V{}'.format(skv)])
# place to keep sketch's temporary files
opts.extend(["--fe-tempdir", out_dir])
opts.append("--fe-keep-tmp")
# custom codegen
if cgen: opts.extend(["--fe-custom-codegen", codegen_jar])
# run Sketch
output_path = os.path.join(out_dir, "output", "{}.txt".format(encoder.demo_name))
if sk:
if os.path.exists(output_path): os.remove(output_path)
sketch.set_default_option(opts)
logging.info('sk_dir: {}, output_path: {}'.format(encoder.sk_dir, output_path))
_, r = sketch.run(encoder.sk_dir, output_path)
# if sketch fails, halt the process here
if not r: return 1
elif not prg:
jskparser.error("need to pass in some file")
return 0
def test(polys, ber, message=None, print_flag=False):
if message == None:
f = open('message.txt', 'rb') # holds a 1000-bit binary string
message = f.read().strip()
e = Encoder(polys, message, ber, print_flag=print_flag)
encoded = e.encoded
sent = e.send()
t = Trellis(polys, sent)
decoded = t.decode_message()
if print_flag:
print 'Message: %s' % message
print 'Decoded: %s' % decoded
return hamming_distance(message, decoded)
def main():
enc = Encoder(pin_clk=12, pin_dt=14, clicks=4, accel=5, max_val=127)
osc = Client(OSC_SERVER, OSC_PORT)
oldval = 0
try:
while True:
val = enc.value
if oldval != val:
oldval = val
osc.send(OSC_TOPIC, ('m', (0, 0xB0, MIDI_CC, val)))
enc.cur_accel = max(0, enc.cur_accel - enc.accel)
sleep_ms(UPDATE_DELAY)
except Exception as exc:
enc.close()
print(exc)
def encode_decode(self, k):
print "\nTesting encoding and then decoding with k = %s" % k
md5 = hashlib.md5()
with FileChunker(k, SYMBOLSIZE, DEFAULT_FILE) as chunker:
chunk = chunker.chunk()
while chunk:
padding = chunk.padding
symbols = [(i, chunk[i]) for i in xrange(k)]
encoder = Encoder(k, symbols)
symbols = []
# Start at k/2 and produce 1.25k more symbols to get a mix
# of parity and source symbols
for i in xrange(k * 2):
symbols.append(encoder.next())
encoder = None
decoder = Decoder(k)
for tup in symbols:
decoder.append(tup)
decoder.decode()
decoded = bytearray()
for i in xrange(k):
esi, s = decoder.next()
decoded += s.tostring()
decoder = None
if padding:
padding = 0 - padding
print "Removing padding", padding, "bytes"
decoded = decoded[:padding]
md5.update(decoded)
# Continue on to the next chunk
chunk = chunker.chunk()
print "Original digest:", self.original_digest
print "Decoded digest:", md5.hexdigest()
return self.original_digest == md5.hexdigest()
def startStopRecording_(self, notification):
if self.recording:
self.recorder.join()
# Create timelapse after recording?
if encode:
self.encoder = Encoder(
self.image_dir, self.encoder_output_basedir)
self.encoder.start()
else:
self.recorder = Recorder(self.image_dir, screenshot_interval)
self.recorder.start()
self.recording = not self.recording
self.setStatus()
def __init__(self):
from encoder import Encoder
self.encoder = Encoder(3.1196)
# the transitions
self.m = defaultdict(lambda: defaultdict(lambda: Automaton.m_inf))
# self.m = defaultdict(dict)
# emissions for the states
self.emissions = {}
self.m_emittors = defaultdict(set)
# how edge values can be coded
self.quantizer = None
def __init__(self, posX, posY, posTheta):
self.w = 15
self.h = 10
self.heading = 0
self.posTheta = posTheta
self.posX = posX
self.posY = posY
self.targetTheta = 0
self.logicalTheta = 0
self.logicalX = 0
self.logicalY = 0
self.Rw = 9
self.Tr = 240
self.D = 9
self.dT1 = 0
self.dT2 = 0
self.T1 = 0
self.T2 = 0
self.counter = 0
self.leftEncoder = Encoder()
self.rightEncoder = Encoder()
self.leftRangeSensor = RangeSensor(self, self.h / 2 + 1, 0, -1)
self.frontRangeSensor = RangeSensor(self, self.w / 2 + 1, 1, 0)
self.rightRangeSensor = RangeSensor(self, self.h / 2 + 1, 0, 1)
self.statsWidget = RobotStatsWidget(self)
def __init__( self,
device = '/dev/ttyUSB0',
set_id = True,
device_id = "01",
debug = True,
):
self.debug = debug
self.ser = serial.Serial(
port = device,
baudrate = 9600,
parity = serial.PARITY_NONE,
stopbits = serial.STOPBITS_ONE,
bytesize = serial.EIGHTBITS,
xonxoff = True,
)
self.encoder = Encoder(1)
if set_id:
self._set_device_id(device_id)
def main():
use_cuda = args.use_cuda
half_precision = args.half_precision
print("Cuda set to {} | Cuda availability: {}".format(
use_cuda, torch.cuda.is_available()))
experiment = "vae_latent3"
logger = SummaryWriter(log_dir='./logs', comment=experiment)
train_data = UnlabeledContact(
data='/home/ygx/data/fspeptide/fs_peptide.npy')
print('Number of samples: {}'.format(len(train_data)))
trainloader = DataLoader(train_data, batch_size=args.batch_size)
# Contact matrices are 21x21
input_size = 441
encoder = Encoder(input_size=input_size, latent_size=3)
decoder = Decoder(latent_size=3, output_size=input_size)
#vae = VAE(encoder, decoder, use_cuda=use_cuda)
vae = VAE(encoder, decoder)
#criterion = nn.BCELoss()
if use_cuda:
encoder = encoder.cuda()
decoder = decoder.cuda()
vae = vae.cuda()
#criterion = criterion.cuda()
if half_precision:
encoder = encoder.half()
decoder = decoder.half()
vae = vae.half()
optimizer = optim.SGD(vae.parameters(), lr=0.001)
losses = AverageMeter()
epoch_loss = 0
total_loss = 0
for epoch in range(100):
for batch_idx, data in enumerate(trainloader):
inputs = data['cont_matrix']
inputs = inputs.resize_(args.batch_size, 1, 21, 21)
inputs = inputs.float()
if use_cuda:
inputs = inputs.cuda()
if half_precision:
inputs = inputs.half()
inputs = Variable(inputs)
# Compute output
optimizer.zero_grad()
#dec = vae(inputs)
recon_batch, mu, logvar = vae(inputs)
# Measure the loss
loss = entropy_kl_loss(recon_batch, inputs, mu, logvar,
args.batch_size, input_size)
#kl = kl_loss(vae.z_mean, vae.z_sigma)
#loss = criterion(dec, inputs) #+ kl # Adding KL is caussing loss > 1
losses.update(loss.data[0], inputs.size(0))
# Compute the gradient
loss.backward()
optimizer.step()
epoch_loss += loss.data[0]
# Logging
# Adding graph is a lot of overhead
#logger.add_graph_onnx(vae)
# log loss values every iteration
logger.add_scalar('data/(train)loss_val', losses.val,
batch_idx + 1)
logger.add_scalar('data/(train)loss_avg', losses.avg,
batch_idx + 1)
# log the layers and layers gradient histogram and distributions
#for tag, value in vae.named_parameters():
# tag = tag.replace('.', '/')
# logger.add_histogram('model/(train)' + tag, to_numpy(value), batch_idx + 1)
# logger.add_histogram('model/(train)' + tag + '/grad', to_numpy(value.grad), batch_idx + 1)
# log the outputs of the autoencoder
logger.add_image('model/(train)output',
make_grid(recon_batch.data), batch_idx + 1)
#logger.add_image('model/(train)output', make_grid(dec.data), batch_idx + 1)
if batch_idx % args.log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(trainloader.dataset),
100. * batch_idx / len(trainloader), loss.data[0]))
#if epoch < 10:
# Get latent encoding
#latent_array = encoder(inputs).data[0].cpu().numpy()
#filename = 'latent_epoch' + str(epoch)
#np.save('./latent_saves/kl_bce_latent3/' + filename, latent_array)
# Get reconstructed image
#reconstructed_array = vae(inputs).data[0].cpu().numpy().reshape(21, 21)
#recon_filename = 'reconstructed_epoch' + str(epoch)
#np.save('./reconstruct_saves/kl_bce_latent3/' + recon_filename, reconstructed_array)
if epoch % 10 == 0:
torch.save(vae.state_dict(), args.save_path + 'epoch' + str(epoch))
#latent_array = encoder(inputs).data[0].cpu().numpy()
#filename = 'latent_epoch' + str(epoch)
#np.save('./latent_saves/kl_bce_latent3/' + filename, latent_array)
reconstructed_array, _, _ = vae(
inputs).data[0].cpu().float().numpy().reshape(21, 21)
recon_filename = 'reconstructed_epoch' + str(epoch)
np.save('./reconstruct_saves/kl_bce_latent3/' + recon_filename,
reconstructed_array)
r_val = [train_set, valid_set, test_set]
return r_val
if __name__ == '__main__':
batch_size = 10
n_input = 3
n_hidden = 50
learning_rate = 0.01
rng = numpy.random.RandomState(12321)
word_dict = Dict(n_input, rng)
datasets = load_data(sys.argv[1], word_dict)
train_set, valid_set, test_set = datasets
n_train_batches = [train_set[b][0].get_value(borrow=True).shape[-1] // batch_size for b in train_set.keys()]
n_valid_batches = [valid_set[b][0].get_value(borrow=True).shape[-1] // batch_size for b in valid_set.keys()]
n_test_batches = [test_set[b][0].get_value(borrow=True).shape[-1] // batch_size for b in test_set.keys()]
encoder = Encoder(rng, n_hidden=n_hidden, n_input=n_input)
# Holds indices for the batch
index = T.lscalar()
# Emb_dim x sequence_len X batch_size
x = T.tensor3('x')
# Batch_size x 1
y = T.ivector('y')
# sequence_len x batch_size
m = T.matrix('m')
get_hidden_states = theano.function(
inputs=[index],
outputs=encoder.compute_hidden_states_no_output(x),
givens=[
(x, test_set[10][0][index * batch_size: (index + 1) * batch_size]),
class Transmitter:
"""
Sends content to the led ticker device.
- transmission is blocking
"""
def __init__( self,
device = '/dev/ttyUSB0',
set_id = True,
device_id = "01",
debug = True,
):
self.debug = debug
self.ser = serial.Serial(
port = device,
baudrate = 9600,
parity = serial.PARITY_NONE,
stopbits = serial.STOPBITS_ONE,
bytesize = serial.EIGHTBITS,
xonxoff = True,
)
self.encoder = Encoder(1)
if set_id:
self._set_device_id(device_id)
def add_message(self, message, page):
""" adds a message to the specified page """
self._send_receive(commands.get_message_cmd(message, page = page))
def set_schedule(self, pages):
""" sets the schedule (order) of the pages """
self._send_receive(commands.get_schedule_cmd(pages))
def clear_screen(self):
""" delete all pages """
self.add_message(" ", 0)
self.set_schedule([0])
def delete_pages(self):
""" delete all pages """
self._send('<D*>')
def end(self):
""" terminates the connection to the led ticker """
self.ser.close()
def _send_receive(self, command):
self.ser.flushInput()
self._send(command)
self._receive_response();
def _send(self, command):
""" sends a command to the device """
data = self.encoder.encode(command)
self._debug_log(data)
self.ser.write(self.encoder.encode(command))
def _receive_response(self):
""" receives a ACK/NACK response
ACK/NACK is sent as ascii, since its not the same length ether
4 or 3 character are read to prevent blocking
"""
response = self.ser.read(1)
if response == 'N':
chars_to_read = 3
else:
chars_to_read = 2
response += self.ser.read(chars_to_read)
self._debug_log('response: ' + response)
return response
def _debug_log(self, message):
if self.debug:
print message
def _set_device_id(self, device_id):
# set the device id
self.ser.flushInput()
self.ser.write('<ID><01><E>' + device_id)
self._debug_log('response: ' + self.ser.read(2))
class Automaton(object):
""" Classic Moore-automaton class with
@m: transitions per states
@emissions: emission per states
@m_emittors: states per emitted letter"""
eps = 1e-7
m_inf = float("-inf")
def __init__(self):
from encoder import Encoder
self.encoder = Encoder(3.1196)
# the transitions
self.m = defaultdict(lambda: defaultdict(lambda: Automaton.m_inf))
# self.m = defaultdict(dict)
# emissions for the states
self.emissions = {}
self.m_emittors = defaultdict(set)
# how edge values can be coded
self.quantizer = None
@staticmethod
def read_transitions(filename):
tr = {}
f = open(filename)
for l in f:
(state1, state2, probstr) = l.strip().split()
if state1 not in tr:
tr[state1] = {}
prob = float(probstr)
if not (prob < 0.0):
raise ValueError("invalid probabilities in {0}, ".format(filename) + "only logprobs are accepted.")
tr[state1][state2] = prob
f.close()
return tr
@staticmethod
def create_from_dump(f):
""" Reads automaton dump from @f"""
automaton = Automaton()
# create states and emissions
for line in f:
l = line.strip().split()
if len(l) == 4:
s1, _, s2, weight = l
s2 = s2.strip(":")
weight = float(weight)
# check this with 1e-10 instead of 0.0 because of floating
# point precision error
if weight > 1e-10:
raise ValueError("Only logprogs are accepted in dumps")
automaton.m[s1][s2] = weight
elif len(l) == 2:
state = l[0].rstrip(":")
emission = eval(l[1])
automaton.emissions[state] = emission
automaton.m_emittors[emission].add(state)
for state in automaton.m.iterkeys():
automaton.check_state_sum(state)
automaton.finalize()
return automaton
@staticmethod
def _create_automaton_from_alphabet(alphabet):
""" Creates states of the automaton given by @alphabet
@alphabet is a dict from letters to the number of states that emits
that letter
"""
automaton = Automaton()
# create states and emissions
is_degenerate = True
for letter in alphabet:
for index in xrange(alphabet[letter]):
state = "".join(letter) + "_" + str(index)
automaton.emissions[state] = letter
automaton.m_emittors[letter].add(state)
if is_degenerate and not Automaton.is_epsilon_state(state):
# found at least one emitting state
is_degenerate = False
if is_degenerate:
raise Exception("Automaton has no emittors")
return automaton
@staticmethod
def create_uniform_automaton(alphabet, initial_transitions=None):
"""Creates an automaton with alphabet and uniform transition probabilities.
If initial_transitions is given (dict of (state, dict of (state, probability),
returned by read_transitions) the remaining probability mass will be
divided up among the uninitialized transitions in an uniform manner.
"""
automaton = Automaton._create_automaton_from_alphabet(alphabet)
states = automaton.emissions.keys()
states.append("^")
states.append("$")
if initial_transitions:
for s in initial_transitions.keys():
if s not in states:
raise Exception("invalid state name in initial " + "transitions given by option -I")
for s2 in initial_transitions[s]:
if s2 not in states:
raise Exception("invalid state name in initial " + "transitions given by option -I")
# calculate uniform transition distributions
for s1 in states:
if s1 == "$":
continue
init_total = 0.0
states_initialized = set()
if initial_transitions and s1 in initial_transitions:
for s2 in initial_transitions[s1]:
if s2 not in states:
raise Exception("invalid state name in init: %s" % s2)
prob = initial_transitions[s1][s2]
automaton.m[s1][s2] = prob
init_total += math.exp(prob)
states_initialized.add(s2)
# TODO refactor this
if init_total > 1.0000001:
sys.stderr.write("state: {0}, init total: {1}\n".format(s1, init_total))
raise Exception("Too much probability for init_total")
# divide up remaining mass into equal parts
valid_next_states = set([s2 for s2 in states if Automaton.is_valid_transition(s1, s2)])
if valid_next_states == states_initialized:
continue
u = (1.0 - init_total) / (len(valid_next_states) - len(states_initialized))
for s2 in valid_next_states - states_initialized:
try:
automaton.m[s1][s2] = math.log(u)
except ValueError:
automaton.m[s1][s2] = Automaton.m_inf
automaton.finalize()
return automaton
@staticmethod
def create_from_corpus(corpus):
""" Creates an automaton from a corpus, where @corpus is a dict from
items (str or tuple) to counts"""
automaton = Automaton()
alphabet = set()
total = float(sum(corpus.itervalues()))
for item, cnt in corpus.iteritems():
item = ("^",) + item + ("$",)
for i in range(len(item) - 1):
alphabet.add(item[i])
alphabet.add(item[i + 1])
if item[i + 1] in automaton.m[item[i]]:
automaton.m[item[i]][item[i + 1]] += cnt / total
else:
automaton.m[item[i]][item[i + 1]] = cnt / total
# changing to log probs and normalize
for state1, outs in automaton.m.iteritems():
for state2 in outs.iterkeys():
outs[state2] = math.log(outs[state2])
automaton.normalize_state(state1)
for l in alphabet:
automaton.emissions[l] = l
automaton.m_emittors[l].add(l)
automaton.finalize()
return automaton
@staticmethod
def is_epsilon_state(state):
return state.startswith("EPSILON_")
@staticmethod
def is_valid_transition(state1, state2):
# subsequent non emitting states are not allowed
# the only exception is '^' -> '$'
if (state1, state2) == ("^", "$"):
return True
return (
not (Automaton.nonemitting(state1) and Automaton.nonemitting(state2))
and not state2 == "^"
and not state1 == "$"
)
@staticmethod
def nonemitting(state):
return state == "^" or state == "$" or Automaton.is_epsilon_state(state)
def finalize(self):
for state1, transitions in self.m.iteritems():
self.m[state1] = dict(transitions)
self.m = dict(self.m)
self.m_emittors = dict(self.m_emittors)
def emittors(self, letter):
return self.m_emittors[letter]
def update_probability_of_string_in_state(self, string, state, memo):
"""The probability of the event that the automaton emits
'string' and finishes the quest at 'state'.
'state' did not emit yet:
It will emit the next symbol following 'string'.
"""
total = Automaton.m_inf
# compute real emissions
for previousState in self.emissions:
previousState_i = self.state_indices[previousState]
# stop if no transitions to state
if not state in self.m[previousState]:
continue
state_emit = self.emissions[previousState]
state_emit_l = len(state_emit)
if state_emit == string[-state_emit_l:]:
head = string[:-state_emit_l]
soFar = Automaton.m_inf
if head in memo and memo[head][previousState_i] is not None:
soFar = memo[head][previousState_i]
soFar += self.m[previousState][state]
total = max(soFar, total)
# check the case of epsilon emission
for epsilonState in self.m.keys():
epsilonState_i = self.state_indices[epsilonState]
if epsilonState in self.emissions:
continue
# if the automaton is not complete, avoid KeyError:
if not state in self.m[epsilonState]:
continue
if not string in memo or memo[string][epsilonState_i] is None:
continue
soFar = memo[string][epsilonState_i]
soFar += self.m[epsilonState][state]
total = max(soFar, total)
if string not in memo:
memo[string] = [None] * len(self.state_indices)
memo[string][self.state_indices[state]] = total
def update_probability_of_string(self, string, memo):
"""Probability that the automaton emits this string"""
states = set(self.m.keys())
states.add("$")
states.remove("^")
# first compute the epsilon states probs because of the
# memoization dependency
for state in sorted(states, key=lambda x: not Automaton.is_epsilon_state(x)):
self.update_probability_of_string_in_state(string, state, memo)
def probability_of_strings(self, strings):
"""
Expects a list of strings.
Outputs a map from those strings to probabilities.
"""
topsorted = closure_and_top_sort(strings)
# remove empty string
topsorted = topsorted[1:]
memo = self.init_memo()
output = {}
for string in topsorted:
self.update_probability_of_string(string, memo)
output[string] = memo[string][self.state_indices["$"]]
return output
def init_memo(self):
# to save memory if memo is huge, inner dicts in memo are actually
# lists with state indices
states = set(self.m.keys())
states.add("$")
self.state_indices = dict([(s, i) for i, s in enumerate(states)])
memo = {(): [None] * len(states)}
epsilon_reachables = set(["^"])
while True:
targets = set()
for state in epsilon_reachables:
state_i = self.state_indices[state]
for target in self.m[state]:
target_i = self.state_indices[target]
if target in epsilon_reachables:
continue
if Automaton.is_epsilon_state(target):
targets.add(target)
# start is not memoized
so_far = Automaton.m_inf
if memo[()][target_i] is not None:
so_far = memo[()][target_i]
prob_this_way = self.m[state][target]
if state != "^":
prob_this_way += memo[()][state_i]
memo[()][target_i] = max(so_far, prob_this_way)
epsilon_reachables |= targets
if len(targets) == 0:
break
return memo
@staticmethod
def kullback(p1, p2):
if p1 == 0.0:
return 0.0
return p1 * math.log(p1 / p2)
@staticmethod
def squarerr(p1, p2):
return (p1 - p2) ** 2
@staticmethod
def l1err(p1, p2):
return abs(p1 - p2)
def distance_from_corpus(self, corpus, distfp, reverse=False, distances={}):
distance = 0.0
probs = self.probability_of_strings(list(corpus.keys()))
for item, corpus_p in corpus.iteritems():
if corpus_p > 0.0:
modeled_p = math.exp(probs[item])
if modeled_p == 0.0:
modeled_p = 1e-50
dist = distfp(corpus_p, modeled_p) if not reverse else distfp(modeled_p, corpus_p)
distance += dist
distances[item] = dist
return distance
def round_and_normalize_state(self, state):
if self.quantizer:
self.round_transitions(self.m[state])
self.normalize_state(state)
def round_transitions(self, edges):
for state, weight in edges.iteritems():
edges[state] = self.quantizer.representer(weight)
def normalize_state(self, state):
edges = self.m[state]
total_log = math.log(sum(math.exp(v) for v in edges.values()))
for s2 in edges.keys():
edges[s2] -= total_log
def round_and_normalize(self):
for state in self.m.iterkeys():
self.round_and_normalize_state(state)
def smooth(self):
"""Smooth zero transition probabilities"""
eps = math.log(Automaton.eps)
for state, edges in self.m.iteritems():
for other_state in edges:
old_val = edges.get(other_state, Automaton.m_inf)
if old_val < eps:
edges[other_state] = eps
# normalize the transitions
self.normalize_state(state)
def boost_edge(self, edge, factor):
"""Adds @factor logprob to @edge"""
s1, s2 = edge
self.m[s1][s2] += factor
self.round_and_normalize_state(s1)
self.check_state_sum(s1)
def check_state_sum(self, state):
edges = self.m[state]
s_sum = sum([math.exp(log_prob) for log_prob in edges.values()])
if abs(1.0 - s_sum) < 1e-3:
return
else:
raise Exception("transitions from state {0} ".format(state) + "don't sum to 1, but {0}".format(s_sum))
def dump(self, f):
if self.quantizer is not None:
emit_bits, trans_bits = self.encoder.automaton_bits(self)
total_bits = emit_bits + trans_bits
f.write(
"total bits: {0} ({1} transition bits, {2} emission bits)\n".format(total_bits, emit_bits, trans_bits)
)
states = sorted(self.m.keys())
for s1 in states:
for s2 in states + ["$"]:
if s2 in self.m[s1]:
f.write("{0} -> {1}: {2}\n".format(s1, s2, self.m[s1][s2]))
for s1, em in self.emissions.iteritems():
f.write("{0}: {1}\n".format(s1, repr(em).replace(" ", "")))
def split_state(self, state, new_state, ratio):
hub_in = "EPSILON_{0}_{1}_in".format(state, new_state)
hub_out = "EPSILON_{0}_{1}_out".format(state, new_state)
self.m[hub_in] = {state + "_0": math.log(1 - ratio), new_state + "_0": math.log(ratio)}
self.m[hub_out] = {}
self.emissions[new_state + "_0"] = (new_state,)
self.m_emittors[(new_state,)] = set([new_state + "_0"])
for s1, trs in self.m.items():
if s1 in (hub_in, hub_out):
continue
for s2, p in trs.items():
if s2.startswith(state):
self.m[s1][hub_in] = p
self.m[s1][s2] = float("-inf")
for s2, p in self.m[state + "_0"].items():
self.m[hub_out][s2] = p
self.m[state + "_0"] = {hub_out: 0.0}
self.m[new_state + "_0"] = {hub_out: 0.0}
def language(self):
generated_mass = 0.0
emits = set(self.emissions.itervalues())
memo = self.init_memo()
prev_mass = -1.0
while abs(generated_mass - prev_mass) >= 1e-4 and 1.0 - generated_mass > 0.01:
prev_mass = generated_mass
for word in memo.keys():
for emit in emits:
new_word = word + emit
self.update_probability_of_string(new_word, memo)
# filter small probs
memo = dict([(k, [(None if (lp is None or lp < -100) else lp) for lp in l]) for k, l in memo.iteritems()])
# filter small prob words
memo = dict([(k, l) for k, l in memo.iteritems() if sum(filter(lambda x: x is not None, l)) > -200])
# compute generated mass
generated_mass = sum(
[
math.exp(prob_list[self.state_indices["$"]])
for s, prob_list in memo.iteritems()
if (s != () and prob_list[self.state_indices["$"]] is not None)
]
)
# compute hq - only debug
# hq = sum([-probs[self.state_indices["$"]] * math.exp(probs[self.state_indices["$"]]) for probs in memo.itervalues()])
for k in memo.keys():
if memo[k][self.state_indices["$"]] is None:
del memo[k]
return memo
class Timelapse(NSObject):
""" Creates a timelapse video """
def applicationDidFinishLaunching_(self, notification):
self.check_dependencies()
# Initialize recording
self.recording = start_recording
# Set correct output paths
self.recorder_output_basedir = os.path.join(dir_base, dir_pictures,
dir_app)
self.encoder_output_basedir = os.path.join(dir_base, dir_movies)
self.image_dir = self.create_dir(self.recorder_output_basedir)
# Create a reference to the statusbar (menubar)
self.statusbar = NSStatusBar.systemStatusBar()
# Create item in statusbar
self.statusitem = self.statusbar.statusItemWithLength_(
NSVariableStatusItemLength)
self.statusitem.setHighlightMode_(1) # Highlight upon clicking
# Create a simple menu and bind it to the status item
self.menu = self.createMenu()
self.statusitem.setMenu_(self.menu)
# Load icons and show them in the statusbar
self.loadIcons()
self.setStatus()
def loadIcons(self):
self.icon_recording = NSImage.alloc().initWithContentsOfFile_(
os.path.join("timelapse", dir_resources, image_recording))
self.icon_idle = NSImage.alloc().initWithContentsOfFile_(
os.path.join("timelapse", dir_resources, image_idle))
def setStatus(self):
""" Sets the image and menu text according to recording status """
if self.recording:
self.statusitem.setImage_(self.icon_recording)
self.recordButton.setTitle_(text_recorder_running)
self.statusitem.setToolTip_(tooltip_running)
else:
self.statusitem.setImage_(self.icon_idle)
self.recordButton.setTitle_(text_recorder_idle)
self.statusitem.setToolTip_(tooltip_idle)
def createMenu(self):
""" Status bar menu """
menu = NSMenu.alloc().init()
# Bind record event
self.recordButton = NSMenuItem.alloc(
).initWithTitle_action_keyEquivalent_(text_recorder_idle,
'startStopRecording:', '')
menu.addItem_(self.recordButton)
# Quit event
menuitem = NSMenuItem.alloc().initWithTitle_action_keyEquivalent_(
'Quit', 'terminate:', '')
menu.addItem_(menuitem)
return menu
def startStopRecording_(self, notification):
if self.recording:
self.recorder.join()
# Create timelapse after recording?
if encode:
self.encoder = Encoder(self.image_dir,
self.encoder_output_basedir)
self.encoder.start()
else:
self.recorder = Recorder(self.image_dir, screenshot_interval)
self.recorder.start()
self.recording = not self.recording
self.setStatus()
@objc.python_method
def create_dir(self, base_dir):
""" Creates a specified directory and the path to it if necessary """
if create_session_subdir:
# Create a new subdirectory
output_dir = os.path.join(base_dir, self.get_sub_dir(base_dir))
else:
# Don't create a subdirectory. Use base directory for output
output_dir = base_dir
# Create path if it doesn't exist
try:
print(output_dir)
os.makedirs(output_dir)
except OSError as e:
print("Error while creating directory:", e)
exit()
return output_dir
@objc.python_method
def get_sub_dir(self, base_dir):
""" Returns the next nonexistend subdirectory to base_dir """
subdir_base = os.path.join(base_dir, subdir_suffix)
# Check if we can use subdir without any session id
subdir = subdir_base
# Use a session id only if subdir already exists
session_number = 0
while os.path.exists(subdir):
# We can't use subdir. Create directory with session id
session_number += 1
subdir = subdir_base + "-" + str(session_number)
return subdir
def check_dependencies(self):
try:
subprocess.check_call(['ffmpeg'])
except subprocess.CalledProcessError:
print("ffmpeg command was found")
pass # ffmpeg is found, but returns non-zero exit as expected
# This is a quick and dirty check; it leaves some spurious output
# for the user to puzzle over.
except OSError:
print(not_found_msg)
def main():
global device
device = next(chromecast)
enc = Encoder(12, 13, clicks=2, reverse=True)
np = volume.NeoPixelRing(4, device, machine.Pin(15), 16)
button = machine.Pin(5, machine.Pin.IN)
cast = connect2device(np)
current_vol = cast.get_volume
print('Connected to:', cast_name[device], device, 'current vol:', current_vol)
enc.set_val(current_vol)
last_enc_val = current_vol
last_change_tick = time.ticks_ms()
np.change_device(device, current_vol)
while True:
val = enc.value
if last_enc_val != val:
print(val)
np.set_vol(val)
last_enc_val = val
last_change_tick = time.ticks_ms()
#CHANGING VOLUME
if (time.ticks_diff(time.ticks_ms(), last_change_tick) > 200) and (last_enc_val != current_vol):
cast.set_volume(val)
current_vol = cast.get_volume
print('current volume:', current_vol)
#SLEEP AFTER DELAY
if (time.ticks_diff(time.ticks_ms(), last_change_tick) > 10000): #10 sec
cast.disconnect()
np.turn_off()
print("SLEEP")
esp.deepsleep()
#CHANGING CHROMECAST WITH ENCODER BUTTON
if button.value():
print('BUTTON PRESSED')
b_start = time.ticks_ms()
while button.value():
if (time.ticks_diff(time.ticks_ms(), b_start) > 2000):
print('STOPPING PLAYBACK')
np.stop()
cast.stop_playback()
time.sleep_ms(1500)
np.set_vol(current_vol)
last_change_tick = time.ticks_ms()
break
if time.ticks_diff(time.ticks_ms(), b_start) < 2000:
cast.disconnect()
prev_device = device
device = next(chromecast)
if device is not prev_device:
cast = connect2device(np)
current_vol = cast.get_volume
enc.set_val(current_vol)
np.change_device(device, current_vol)
print('switched to:', cast_name[device], device, 'current vol:', current_vol)
last_change_tick = time.ticks_ms()
time.sleep_ms(100)
def __init__(self, model_pre_path):
self.encoder = Encoder(model_pre_path)
self.decoder = Decoder(model_pre_path)
class Timelapse(NSObject):
""" Creates a timelapse video """
def applicationDidFinishLaunching_(self, notification):
self.check_dependencies()
# Initialize recording
self.recording = start_recording
# Set correct output paths
self.recorder_output_basedir = os.path.join(
dir_base, dir_pictures, dir_app)
self.encoder_output_basedir = os.path.join(dir_base, dir_movies)
self.image_dir = self.create_dir(self.recorder_output_basedir)
# Create a reference to the statusbar (menubar)
self.statusbar = NSStatusBar.systemStatusBar()
# Create item in statusbar
self.statusitem = self.statusbar.statusItemWithLength_(
NSVariableStatusItemLength)
self.statusitem.setHighlightMode_(1) # Highlight upon clicking
# Create a simple menu and bind it to the status item
self.menu = self.createMenu()
self.statusitem.setMenu_(self.menu)
# Load icons and show them in the statusbar
self.loadIcons()
self.setStatus()
def loadIcons(self):
self.icon_recording = NSImage.alloc().initWithContentsOfFile_(
os.path.join("timelapse", dir_resources, image_recording))
self.icon_idle = NSImage.alloc().initWithContentsOfFile_(
os.path.join("timelapse", dir_resources, image_idle))
def setStatus(self):
""" Sets the image and menu text according to recording status """
if self.recording:
self.statusitem.setImage_(self.icon_recording)
self.recordButton.setTitle_(text_recorder_running)
self.statusitem.setToolTip_(tooltip_running)
else:
self.statusitem.setImage_(self.icon_idle)
self.recordButton.setTitle_(text_recorder_idle)
self.statusitem.setToolTip_(tooltip_idle)
def createMenu(self):
""" Status bar menu """
menu = NSMenu.alloc().init()
# Bind record event
self.recordButton = NSMenuItem.alloc().initWithTitle_action_keyEquivalent_(
text_recorder_idle, 'startStopRecording:', '')
menu.addItem_(self.recordButton)
# Quit event
menuitem = NSMenuItem.alloc().initWithTitle_action_keyEquivalent_(
'Quit', 'terminate:', '')
menu.addItem_(menuitem)
return menu
def startStopRecording_(self, notification):
if self.recording:
self.recorder.join()
# Create timelapse after recording?
if encode:
self.encoder = Encoder(
self.image_dir, self.encoder_output_basedir)
self.encoder.start()
else:
self.recorder = Recorder(self.image_dir, screenshot_interval)
self.recorder.start()
self.recording = not self.recording
self.setStatus()
@objc.python_method
def create_dir(self, base_dir):
""" Creates a specified directory and the path to it if necessary """
if create_session_subdir:
# Create a new subdirectory
output_dir = os.path.join(base_dir, self.get_sub_dir(base_dir))
else:
# Don't create a subdirectory. Use base directory for output
output_dir = base_dir
# Create path if it doesn't exist
try:
print(output_dir)
os.makedirs(output_dir)
except OSError as e:
print("Error while creating directory:", e)
exit()
return output_dir
@objc.python_method
def get_sub_dir(self, base_dir):
""" Returns the next nonexistend subdirectory to base_dir """
subdir_base = os.path.join(base_dir, subdir_suffix)
# Check if we can use subdir without any session id
subdir = subdir_base
# Use a session id only if subdir already exists
session_number = 0
while os.path.exists(subdir):
# We can't use subdir. Create directory with session id
session_number += 1
subdir = subdir_base + "-" + str(session_number)
return subdir
def check_dependencies(self):
try:
subprocess.check_call(['ffmpeg'])
except subprocess.CalledProcessError:
print("ffmpeg command was found")
pass # ffmpeg is found, but returns non-zero exit as expected
# This is a quick and dirty check; it leaves some spurious output
# for the user to puzzle over.
except OSError:
print(not_found_msg)
def train_decoder(train_set, opt, learning_r, encoder=None, epoch=500, batch_size=32, pre_trained_path='',
test_set=None,
name="Validation", tensorboard=False, img_tag=""):
data_loader = make_data_loader(data_to_loader=train_set, batch_size=batch_size)
best_test_loss = 999999999
if encoder is None:
encoder = Encoder()
decoder = Decoder()
if pre_trained_path != '':
encoder = torch.load(pre_trained_path)["model"]
if use_gpu:
decoder = decoder.cuda()
encoder = encoder.cuda()
criterion = nn.MSELoss()
params = list(decoder.parameters()) + list(encoder.parameters())
if opt == "ADAM":
optimizer = optim.Adam(params, lr=learning_r)
else:
optimizer = optim.SGD(params, lr=learning_r)
for epoch in range(epoch): # loop over the dataset multiple times
print("Epoch: ", epoch)
for i, data in enumerate(data_loader, 0):
# get the inputs
inputs, labels = data
if use_gpu:
inputs = inputs.cuda()
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
hidden = encoder(inputs)
outputs = decoder(hidden)
loss = criterion(outputs, inputs)
loss.backward()
optimizer.step()
train_loss = test_decoder(encoder, decoder, vad_set=train_set, name="Training", show_log=True, img_tag=img_tag,
tensorboard=tensorboard, epoch=epoch)
if test_set is not None:
test_loss = test_decoder(encoder, decoder, vad_set=test_set, name=name, show_log=True, img_tag=img_tag,
tensorboard=tensorboard, epoch=epoch)
if tensorboard and name is not "Validation":
writer_test.add_scalar('MSE_loss', test_loss, epoch)
if test_loss < best_test_loss:
best_test_loss = test_loss
if tensorboard and name is not "Validation":
writer_train.add_scalar('MSE_loss', train_loss, epoch)
return encoder, decoder, best_test_loss
from machine import Pin, I2C, Timer from board import * from bno055 import BNO055 # IMU from drv8833 import DRV8833 # your implementation from motor import PIDMotor # your implementation from encoder import Encoder # your implementation, don't forget clear_count from balance import Balance import gc # for garbage collection methods # Setup motors ########## Check Pin Numbers! ########## # Change pin numbers here to match yours or rewire your robot leftEnc = Encoder(34, 39, 2) leftM = DRV8833(19, 16) rightEnc = Encoder(36, 4, 1) rightM = DRV8833(17, 21) ########## Check Pin Numbers! ########## pidL = PIDMotor(leftM, leftEnc) pidR = PIDMotor(rightM, rightEnc) # setup IMU i2c = I2C(0, sda=23, scl=22, freq=13000) imu = BNO055(i2c) # status LED led = Pin(LED, mode=Pin.OUT)
def __init__(self,
hyperparams,
is_training,
inputs,
input_lengths,
mel_targets=None,
linear_targets=None,
audio_lengths=None):
self.encoder = Encoder(hyperparams, is_training, inputs, input_lengths)
self.decoder = Decoder(hyperparams, is_training,
self.encoder.encoder_outputs, mel_targets)
if is_training:
with tf.variable_scope('loss'):
mel_loss = tf.abs(mel_targets - self.decoder.mel_outputs)
l1 = tf.abs(linear_targets - self.decoder.linear_outputs)
self.linear_loss = tf.reduce_mean(l1)
self.mel_loss = tf.reduce_mean(mel_loss)
self.loss = self.linear_loss + self.mel_loss
with tf.variable_scope('optimizer'):
self.global_step = tf.get_variable(
"global_step",
shape=[],
trainable=False,
initializer=tf.zeros_initializer,
dtype=tf.int32)
step = tf.cast(self.global_step + 1, dtype=tf.float32)
self.learning_rate = hyperparams.initial_learning_rate * \
tf.train.exponential_decay(1., step, 3000, 0.95)
optimizer = tf.train.AdamOptimizer(self.learning_rate,
hyperparams.adam_beta1,
hyperparams.adam_beta2)
self.gradients, variables = zip(
*optimizer.compute_gradients(self.loss))
clipped_gradients, _ = tf.clip_by_global_norm(
self.gradients, 1.0)
# Memo the total length of audio this model was trained on
self.total_length = tf.get_variable(
'total_train_length', [],
initializer=tf.zeros_initializer,
dtype=tf.float32)
update_total_length = tf.assign_add(
self.total_length, tf.reduce_sum(audio_lengths))
# Add dependency on UPDATE_OPS; otherwise batchnorm won't work correctly. See:
# https://github.com/tensorflow/tensorflow/issues/1122
with tf.control_dependencies(
tf.get_collection(tf.GraphKeys.UPDATE_OPS) +
[update_total_length]):
self.optimize = optimizer.apply_gradients(
zip(clipped_gradients, variables),
global_step=self.global_step)
# Logging
self.training_summary = tf.summary.merge([
tf.summary.scalar("total_loss", self.loss),
tf.summary.scalar("mel_loss", self.mel_loss),
tf.summary.scalar("linear_loss", self.linear_loss),
tf.summary.scalar("total_length_hours",
self.total_length / 3600)
])
self.validation_summary = tf.summary.merge([
tf.summary.scalar("validation_total_loss", self.loss),
tf.summary.scalar("validation_mel_loss", self.mel_loss),
tf.summary.scalar("validation_linear_loss",
self.linear_loss),
tf.summary.image(
"v_alignment_matrix",
tf.expand_dims(self.decoder.alignments, 3))
])
class DataIterator:
"""数据集迭代类"""
def __init__(self,
model_conf: ModelConfig,
mode: RunMode,
ran_captcha=None):
"""
:param model_conf: 工程配置
:param mode: 运行模式（区分：训练/验证）
"""
self.model_conf = model_conf
self.mode = mode
self.path_map = {
RunMode.Trains:
self.model_conf.trains_path[DatasetType.TFRecords],
RunMode.Validation:
self.model_conf.validation_path[DatasetType.TFRecords]
}
self.batch_map = {
RunMode.Trains: self.model_conf.batch_size,
RunMode.Validation: self.model_conf.validation_batch_size
}
self.data_dir = self.path_map[mode]
self.next_element = None
self.image_path = []
self.label_list = []
self._label_list = []
self._size = 0
self.encoder = Encoder(self.model_conf, self.mode)
self.ran_captcha = ran_captcha
@staticmethod
def parse_example(serial_example):
features = tf.io.parse_single_example(
serial_example,
features={
'label': tf.io.FixedLenFeature([], tf.string),
'input': tf.io.FixedLenFeature([], tf.string),
})
_input = tf.cast(features['input'], tf.string)
_label = tf.cast(features['label'], tf.string)
return _input, _label
@staticmethod
def total_sample(file_name):
sample_nums = 0
for _ in tf.compat.v1.python_io.tf_record_iterator(file_name):
sample_nums += 1
return sample_nums
def read_sample_from_tfrecords(self, path):
"""
从TFRecords中读取样本
:param path: TFRecords文件路径
:return:
"""
if isinstance(path, list):
for p in path:
self._size += self.total_sample(p)
else:
self._size = self.total_sample(path)
min_after_dequeue = 1000
batch = self.batch_map[self.mode]
if self.model_conf.da_random_captcha['Enable']:
batch = random.randint(int(batch / 3 * 2), batch)
dataset_train = tf.data.TFRecordDataset(filenames=path,
num_parallel_reads=20).map(
self.parse_example)
dataset_train = dataset_train.shuffle(
min_after_dequeue,
reshuffle_each_iteration=True).prefetch(128).batch(
batch, drop_remainder=True).repeat()
iterator = tf.compat.v1.data.make_one_shot_iterator(dataset_train)
self.next_element = iterator.get_next()
@property
def size(self):
"""样本数"""
return self._size
@property
def labels(self):
"""标签"""
return self.label_list
@staticmethod
def to_sparse(input_batch, label_batch):
"""密集输入转稀疏"""
batch_inputs = input_batch
batch_labels = utils.sparse.sparse_tuple_from_sequences(label_batch)
return batch_inputs, batch_labels
def generate_captcha(self, num) -> (list, list):
_images = []
_labels = []
for i in range(num):
try:
image, labels, font_type = self.ran_captcha.create()
_images.append(image)
_labels.append(''.join(labels).encode())
except Exception as e:
print(e)
pass
return _images, _labels
def generate_batch_by_tfrecords(self, session):
"""根据TFRecords生成当前批次，输入为当前TensorFlow会话，输出为稀疏型X和Y"""
# print(session.graph)
batch = self.batch_map[self.mode]
_input, _label = session.run(self.next_element)
if self.model_conf.da_random_captcha['Enable']:
remain_batch = batch - len(_label)
extra_input, extra_label = self.generate_captcha(remain_batch)
_input = np.concatenate((_input, extra_input), axis=0)
_label = np.concatenate((_label, extra_label), axis=0)
input_batch = []
label_batch = []
for index, (i1, i2) in enumerate(zip(_input, _label)):
try:
label_array = self.encoder.text(i2)
if self.model_conf.model_field == ModelField.Image:
input_array = self.encoder.image(i1)
else:
input_array = self.encoder.text(i1)
if input_array is None:
tf.compat.v1.logging.warn(
"{}, Cannot identify image file labeled: {}, ignored.".
format(input_array, label_array))
continue
if isinstance(input_array, str):
tf.compat.v1.logging.warn(
"{}, \nInput errors labeled: {} [{}], ignored.".format(
input_array, i1, label_array))
continue
if isinstance(label_array, dict):
# tf.logging.warn("The sample label {} contains invalid charset: {}.".format(
# label_array['label'], label_array['char']
# ))
continue
if input_array.shape[-1] != self.model_conf.image_channel:
# pass
tf.compat.v1.logging.warn(
"{}, \nInput shape: {}, ignored.".format(
self.model_conf.image_channel,
input_array.shape[-1]))
continue
label_len_correct = len(
label_array) != self.model_conf.max_label_num
using_cross_entropy = self.model_conf.loss_func == LossFunction.CrossEntropy
if label_len_correct and using_cross_entropy and not self.model_conf.auto_padding:
tf.compat.v1.logging.warn(
"The number of labels must be fixed when using cross entropy, label: {}, "
"the number of tags is incorrect, ignored.".format(i2))
continue
if len(
label_array
) > self.model_conf.max_label_num and using_cross_entropy:
tf.compat.v1.logging.warn(
"The number of label[{}] exceeds the maximum number of labels, ignored.{}"
.format(i2, label_array))
continue
input_batch.append(input_array)
label_batch.append(label_array)
except OSError:
random_suffix = hashlib.md5(i1).hexdigest()
file_format = EXCEPT_FORMAT_MAP[self.model_conf.model_field]
with open(file="oserror_{}.{}".format(random_suffix,
file_format),
mode="wb") as f:
f.write(i1)
tf.compat.v1.logging.warn("OSError [{}]".format(i2))
continue
# 如果图片尺寸不固定则padding当前批次，使用最大的宽度作为序列最大长度
if self.model_conf.model_field == ModelField.Image and self.model_conf.resize[
0] == -1:
input_batch = tf.keras.preprocessing.sequence.pad_sequences(
sequences=input_batch,
maxlen=None,
dtype='float32',
padding='post',
truncating='post',
value=0)
self.label_list = label_batch
return self.to_sparse(input_batch, self.label_list)
class VariationalDynamicalEncoder(object):
def __init__(self,
params,
dim_x,
dim_z,
dim_recon,
p_x='bernoulli',
q_z='gaussian_marg',
p_z='gaussian_marg',
l2_loss=1e-6):
self.params, self.dim_x, self.dim_z, self.dim_recon = params, dim_x, dim_z, dim_recon
N, M, T, D, n_z = self.params.n_batch, self.params.batch_size, self.params.n_time_steps, self.params.d_inputs, self.params.n_z
Tau_pred, C = self.params.pred_seq_len, self.params.n_classes
self.distributions = {'p_x': p_x, 'q_z': q_z, 'p_z': p_z}
self.l2_loss = l2_loss
''' Create Graph '''
self.G = tf.Graph()
with self.G.as_default():
self.x = tf.placeholder(tf.float32, [None, T, D]) # train M, T, D
self.x_con = tf.placeholder(tf.float32, [None, T, D])
self.y_one_hot = tf.placeholder(tf.float32, [None, Tau_pred, C])
self.encoder = Encoder(params=self.params)
self.decoder = Decoder(params=self.params)
self.predict = Predict_trend(params=self.params)
self._objective()
self.session = tf.Session(graph=self.G)
self.saver = tf.train.Saver()
def _gen_sample(self, mu, log_sigma_sq):
epsilon = tf.random_normal((tf.shape(mu)), 0, 1)
sample = tf.add(mu, tf.multiply(tf.exp(0.5 * log_sigma_sq), epsilon))
return sample
def _generate_zx(self, x, reuse=False):
with tf.variable_scope('encoder', reuse=reuse):
z_mu, z_lsgms = self.encoder.output(x, reuse=reuse)
z_sample = self._gen_sample(z_mu, z_lsgms)
return z_sample, z_mu, z_lsgms
def _generate_xz(self, z, reuse=False):
with tf.variable_scope('decoder', reuse=reuse):
x_hat = self.decoder.output(z, reuse=reuse)
return x_hat
def _generate_yz(self, z, reuse=False):
with tf.variable_scope('decoder', reuse=reuse):
y_hat = self.predict.output(z, reuse=reuse)
return y_hat
def _objective(self):
############
''' Cost '''
############
self.z_sample, self.z_mu, self.z_lsgms = self._generate_zx(self.x)
self.x_hat = self._generate_xz(self.z_sample)
self.z_tau, _, _ = self._generate_zx(self.x_hat, reuse=True)
# if self.distributions['p_z'] == 'gaussian_marg':
# prior_z = tf.reduce_sum(utils.tf_gaussian_marg(self.z_mu, self.z_lsgms), 1)
#
# if self.distributions['q_z'] == 'gaussian_marg':
# post_z = tf.reduce_sum(utils.tf_gaussian_ent(self.z_lsgms), 1)
#
# if self.distributions['p_x'] == 'bernoulli':
# self.log_lik = - tf.reduce_sum(utils.tf_binary_xentropy(self.x, self.x_hat), 1)
l2 = tf.add_n([tf.nn.l2_loss(v) for v in tf.trainable_variables()])
latent_cost = -0.5 * tf.reduce_sum(
1 + self.z_lsgms - tf.square(self.z_mu) - tf.exp(self.z_lsgms),
axis=1)
latent_loss = tf.reduce_mean(latent_cost)
z_mean, z_var = tf.nn.moments(self.z_sample, axes=[0], keep_dims=True)
z_tau_mean, z_tau_var = tf.nn.moments(self.z_tau,
axes=[0],
keep_dims=True)
num = tf.reduce_mean(tf.multiply((self.z_sample - z_mean),
(self.z_tau - z_tau_mean)),
axis=[0, 1])
den = tf.reduce_mean(tf.multiply(z_var, tf.transpose(z_tau_var)))
self.y_pred = self._generate_yz(self.z_sample) # _, T, C
# # TODO format code into metric
# eval_metric_ops = {
# "accuracy": tf.metrics.accuracy(labels=self.y_one_hot, predictions=self.predictions["classes"])
# }
#
# self.predictions = {
# "classes": tf.argmax(input=self.y_pred, axis=2),
# "class_target": tf.argmax(input=self.y_one_hot, axis=2),
# "probabilities": tf.nn.softmax(self.y_pred, name="softmax")
# }
_, self.accuracy = tf.metrics.accuracy(
labels=tf.argmax(self.y_one_hot, 2),
predictions=tf.argmax(self.y_pred, 2))
# classify
self.predict_loss = tf.losses.softmax_cross_entropy(
onehot_labels=self.y_one_hot, logits=self.y_pred)
self.corr_loss = -num / (den + 1e-6)
self.mse_loss = tf.losses.mean_squared_error(labels=self.x_con,
predictions=self.x_hat)
# self.cost = tf.reduce_mean(post_z - prior_z) + self.corr_loss + self.mse_loss + self.l2_loss * l2
self.cost = latent_loss + self.mse_loss + self.l2_loss * l2 + self.predict_loss
##################
''' Evaluation '''
##################
self.z_sample_eval, _, _ = self._generate_zx(self.x, reuse=True)
self.x_hat_eval = self._generate_xz(self.z_sample_eval, reuse=True)
self.eval_log_lik = -tf.reduce_mean(
tf.reduce_sum(utils.tf_binary_xentropy(self.x, self.x_hat_eval),
1))
def train(self,
x,
x_con,
x_valid,
x_con_valid,
epochs,
num_batches,
print_every=1,
learning_rate=3e-4,
beta1=0.9,
beta2=0.999,
seed=31415,
stop_iter=100,
y_one_hot=None,
save_path=None,
load_path=None):
self.num_examples = x.shape[0]
self.num_batches = num_batches
assert self.num_examples % self.num_batches == 0, '#Examples % #Batches != 0'
self.batch_size = self.num_examples // self.num_batches
''' Session and Summary '''
self.save_path = save_path
if save_path is None:
self.save_ckpt_path = 'checkpoints/model_VAE_{}-{}_{}.cpkt'.format(
learning_rate, self.batch_size, time.time())
else:
self.save_ckpt_path = save_path + 'model_VAE_{}-{}_{}.cpkt'.format(
learning_rate, self.batch_size, time.time())
np.random.seed(seed)
tf.set_random_seed(seed)
with self.G.as_default():
self.optimizer_origin = tf.train.AdamOptimizer(
learning_rate=learning_rate, beta1=beta1, beta2=beta2)
self.optimizer = tf.contrib.estimator.clip_gradients_by_norm(
self.optimizer_origin, clip_norm=1.0)
self.train_op = self.optimizer.minimize(self.cost)
self.init_g = tf.global_variables_initializer()
self.init_l = tf.local_variables_initializer()
self._test_vars = None
with self.session as sess:
# sess.run(self.init)
sess.run(self.init_g)
sess.run(self.init_l)
if load_path == 'default':
full_path = tf.train.latest_checkpoint(self.save_path)
print('restore model from {}'.format(full_path))
self.saver.restore(sess, full_path)
elif load_path is not None:
full_path = tf.train.latest_checkpoint(load_path)
print('restore model from {}'.format(full_path))
self.saver.restore(sess, full_path)
best_eval_log_lik = -np.inf
best_mse = np.inf
stop_counter = 0
for epoch in range(epochs):
# TODO create shuffle Data
# X_train_shuffle, X_tau_train_shuffle, y_classify_train_shuffle = shuffle_data(X_train, X_tau_train, y_classify_train)
''' Training '''
training_cost, accuracy = 0, 0
for x_batch, x_con_batch, y_one_hot_batch in utils.feed_numpy(
self.batch_size, x, x_con, y=y_one_hot):
training_result = sess.run(
[self.train_op, self.cost, self.accuracy],
feed_dict={
self.x: x_batch,
self.x_con: x_con_batch,
self.y_one_hot: y_one_hot_batch
})
training_cost += training_result[1]
accuracy += training_result[2]
training_cost, accuracy = training_cost / self.num_batches, accuracy / self.num_batches
''' Evaluation '''
stop_counter += 1
if epoch % print_every == 0:
# test_vars = tf.get_collection(bookkeeper.GraphKeys.TEST_VARIABLES)
# if test_vars:
# if test_vars != self._test_vars:
# self._test_vars = list(test_vars)
# self._test_var_init_op = tf.initialize_variables(test_vars)
# self._test_var_init_op.run()
mse = sess.run(self.mse_loss,
feed_dict={
self.x: x,
self.x_con: x_con
})
# corr_loss = self.corr_loss.eval(feed_dict={self.x: x, self.x_con: x_con})
if mse < best_mse:
best_eval_log_lik = mse
self.saver.save(sess, self.save_ckpt_path)
stop_counter = 0
utils.print_metrics(epoch + 1,
['Training', ' cost', training_cost],
['Accuracy', 'train', accuracy],
['MSE ', ' train', mse])
if stop_counter >= stop_iter:
print(
'Stopping No change in validation log-likelihood for {} iterations'
.format(stop_iter))
print('Best validation log-likelihood: {}'.format(
best_eval_log_lik))
print('Model saved in {}'.format(self.save_path))
break
def encode(self, x, sample=False):
if sample:
return self.session.run([self.z_sample, self.z_mu, self.z_lsgms],
feed_dict={self.x: x})
else:
return self.session.run([self.z_mu, self.z_lsgms],
feed_dict={self.x: x})
def decode(self, z):
return self.session.run([self.x_hat], feed_dict={self.z_sample: z})
def initialize(self, train_data):
"""Initializes model parameters from pre-defined hyperparameters and other hyperparameters
that are computed based on statistics over the training data."""
nl_lengths = []
code_lengths = []
nl_token_counter = Counter()
code_token_counter = Counter()
for ex in train_data:
trg_sequence = [START
] + ex.span_minimal_diff_comment_tokens + [END]
nl_token_counter.update(trg_sequence)
nl_lengths.append(len(trg_sequence))
old_nl_sequence = ex.old_comment_tokens
nl_token_counter.update(old_nl_sequence)
nl_lengths.append(len(old_nl_sequence))
code_sequence = ex.span_diff_code_tokens
code_token_counter.update(code_sequence)
code_lengths.append(len(code_sequence))
self.max_nl_length = int(
np.percentile(np.asarray(sorted(nl_lengths)), LENGTH_CUTOFF_PCT))
self.max_code_length = int(
np.percentile(np.asarray(sorted(code_lengths)), LENGTH_CUTOFF_PCT))
self.max_vocab_extension = self.max_nl_length + self.max_code_length
nl_counts = np.asarray(sorted(nl_token_counter.values()))
nl_threshold = int(np.percentile(nl_counts, VOCAB_CUTOFF_PCT)) + 1
code_counts = np.asarray(sorted(code_token_counter.values()))
code_threshold = int(np.percentile(nl_counts, VOCAB_CUTOFF_PCT)) + 1
self.embedding_store = EmbeddingStore(nl_threshold, NL_EMBEDDING_SIZE,
nl_token_counter, code_threshold,
CODE_EMBEDDING_SIZE,
code_token_counter, DROPOUT_RATE,
True)
self.code_encoder = Encoder(CODE_EMBEDDING_SIZE, HIDDEN_SIZE,
NUM_LAYERS, DROPOUT_RATE)
self.nl_encoder = Encoder(NL_EMBEDDING_SIZE, HIDDEN_SIZE, NUM_LAYERS,
DROPOUT_RATE)
self.decoder = UpdateDecoder(NL_EMBEDDING_SIZE, DECODER_HIDDEN_SIZE,
2 * HIDDEN_SIZE, self.embedding_store,
NL_EMBEDDING_SIZE, DROPOUT_RATE)
self.encoder_final_to_decoder_initial = nn.Parameter(
torch.randn(2 * NUM_ENCODERS * HIDDEN_SIZE,
DECODER_HIDDEN_SIZE,
dtype=torch.float,
requires_grad=True))
self.code_features_to_embedding = nn.Linear(CODE_EMBEDDING_SIZE +
NUM_CODE_FEATURES,
CODE_EMBEDDING_SIZE,
bias=False)
self.nl_features_to_embedding = nn.Linear(NL_EMBEDDING_SIZE +
NUM_NL_FEATURES,
NL_EMBEDDING_SIZE,
bias=False)
self.optimizer = torch.optim.Adam(self.parameters(), lr=LR)
class CommentUpdateModel(nn.Module):
"""Edit model which learns to map a sequence of code edits to a sequence of comment edits and then applies the edits to the
old comment in order to produce an updated comment."""
def __init__(self, model_path):
super(CommentUpdateModel, self).__init__()
self.model_path = model_path
self.torch_device_name = 'cpu'
def initialize(self, train_data):
"""Initializes model parameters from pre-defined hyperparameters and other hyperparameters
that are computed based on statistics over the training data."""
nl_lengths = []
code_lengths = []
nl_token_counter = Counter()
code_token_counter = Counter()
for ex in train_data:
trg_sequence = [START
] + ex.span_minimal_diff_comment_tokens + [END]
nl_token_counter.update(trg_sequence)
nl_lengths.append(len(trg_sequence))
old_nl_sequence = ex.old_comment_tokens
nl_token_counter.update(old_nl_sequence)
nl_lengths.append(len(old_nl_sequence))
code_sequence = ex.span_diff_code_tokens
code_token_counter.update(code_sequence)
code_lengths.append(len(code_sequence))
self.max_nl_length = int(
np.percentile(np.asarray(sorted(nl_lengths)), LENGTH_CUTOFF_PCT))
self.max_code_length = int(
np.percentile(np.asarray(sorted(code_lengths)), LENGTH_CUTOFF_PCT))
self.max_vocab_extension = self.max_nl_length + self.max_code_length
nl_counts = np.asarray(sorted(nl_token_counter.values()))
nl_threshold = int(np.percentile(nl_counts, VOCAB_CUTOFF_PCT)) + 1
code_counts = np.asarray(sorted(code_token_counter.values()))
code_threshold = int(np.percentile(nl_counts, VOCAB_CUTOFF_PCT)) + 1
self.embedding_store = EmbeddingStore(nl_threshold, NL_EMBEDDING_SIZE,
nl_token_counter, code_threshold,
CODE_EMBEDDING_SIZE,
code_token_counter, DROPOUT_RATE,
True)
self.code_encoder = Encoder(CODE_EMBEDDING_SIZE, HIDDEN_SIZE,
NUM_LAYERS, DROPOUT_RATE)
self.nl_encoder = Encoder(NL_EMBEDDING_SIZE, HIDDEN_SIZE, NUM_LAYERS,
DROPOUT_RATE)
self.decoder = UpdateDecoder(NL_EMBEDDING_SIZE, DECODER_HIDDEN_SIZE,
2 * HIDDEN_SIZE, self.embedding_store,
NL_EMBEDDING_SIZE, DROPOUT_RATE)
self.encoder_final_to_decoder_initial = nn.Parameter(
torch.randn(2 * NUM_ENCODERS * HIDDEN_SIZE,
DECODER_HIDDEN_SIZE,
dtype=torch.float,
requires_grad=True))
self.code_features_to_embedding = nn.Linear(CODE_EMBEDDING_SIZE +
NUM_CODE_FEATURES,
CODE_EMBEDDING_SIZE,
bias=False)
self.nl_features_to_embedding = nn.Linear(NL_EMBEDDING_SIZE +
NUM_NL_FEATURES,
NL_EMBEDDING_SIZE,
bias=False)
self.optimizer = torch.optim.Adam(self.parameters(), lr=LR)
def get_batches(self, dataset, shuffle=False):
"""Divides the dataset into batches based on pre-defined BATCH_SIZE hyperparameter.
Each batch is tensorized so that it can be directly passed into the network."""
batches = []
if shuffle:
random.shuffle(dataset)
curr_idx = 0
while curr_idx < len(dataset):
start_idx = curr_idx
end_idx = min(start_idx + BATCH_SIZE, len(dataset))
code_token_ids = []
code_lengths = []
old_nl_token_ids = []
old_nl_lengths = []
trg_token_ids = []
trg_extended_token_ids = []
trg_lengths = []
invalid_copy_positions = []
inp_str_reps = []
inp_ids = []
code_features = []
nl_features = []
for i in range(start_idx, end_idx):
code_sequence = dataset[i].span_diff_code_tokens
code_sequence_ids = self.embedding_store.get_padded_code_ids(
code_sequence, self.max_code_length)
code_length = min(len(code_sequence), self.max_code_length)
code_token_ids.append(code_sequence_ids)
code_lengths.append(code_length)
old_nl_sequence = dataset[i].old_comment_tokens
old_nl_length = min(len(old_nl_sequence), self.max_nl_length)
old_nl_sequence_ids = self.embedding_store.get_padded_nl_ids(
old_nl_sequence, self.max_nl_length)
old_nl_token_ids.append(old_nl_sequence_ids)
old_nl_lengths.append(old_nl_length)
ex_inp_str_reps = []
ex_inp_ids = []
extra_counter = len(self.embedding_store.nl_vocabulary)
max_limit = len(self.embedding_store.nl_vocabulary
) + self.max_vocab_extension
out_ids = set()
copy_inputs = code_sequence[:
code_length] + old_nl_sequence[:
old_nl_length]
for c in copy_inputs:
nl_id = self.embedding_store.get_nl_id(c)
if self.embedding_store.is_nl_unk(
nl_id) and extra_counter < max_limit:
if c in ex_inp_str_reps:
nl_id = ex_inp_ids[ex_inp_str_reps.index(c)]
else:
nl_id = extra_counter
extra_counter += 1
out_ids.add(nl_id)
ex_inp_str_reps.append(c)
ex_inp_ids.append(nl_id)
trg_sequence = trg_sequence = [
START
] + dataset[i].span_minimal_diff_comment_tokens + [END]
trg_sequence_ids = self.embedding_store.get_padded_nl_ids(
trg_sequence, self.max_nl_length)
trg_extended_sequence_ids = self.embedding_store.get_extended_padded_nl_ids(
trg_sequence, self.max_nl_length, ex_inp_ids,
ex_inp_str_reps)
trg_token_ids.append(trg_sequence_ids)
trg_extended_token_ids.append(trg_extended_sequence_ids)
trg_lengths.append(min(len(trg_sequence), self.max_nl_length))
inp_str_reps.append(ex_inp_str_reps)
inp_ids.append(ex_inp_ids)
invalid_copy_positions.append(
get_invalid_copy_locations(ex_inp_str_reps,
self.max_vocab_extension,
trg_sequence,
self.max_nl_length))
code_features.append(
get_code_features(code_sequence, dataset[i],
self.max_code_length))
nl_features.append(
get_nl_features(old_nl_sequence, dataset[i],
self.max_nl_length))
batches.append(
UpdateBatchData(
torch.tensor(code_token_ids,
dtype=torch.int64,
device=self.get_device()),
torch.tensor(code_lengths,
dtype=torch.int64,
device=self.get_device()),
torch.tensor(old_nl_token_ids,
dtype=torch.int64,
device=self.get_device()),
torch.tensor(old_nl_lengths,
dtype=torch.int64,
device=self.get_device()),
torch.tensor(trg_token_ids,
dtype=torch.int64,
device=self.get_device()),
torch.tensor(trg_extended_token_ids,
dtype=torch.int64,
device=self.get_device()),
torch.tensor(trg_lengths,
dtype=torch.int64,
device=self.get_device()),
torch.tensor(invalid_copy_positions,
dtype=torch.uint8,
device=self.get_device()), inp_str_reps,
inp_ids,
torch.tensor(code_features,
dtype=torch.float32,
device=self.get_device()),
torch.tensor(nl_features,
dtype=torch.float32,
device=self.get_device())))
curr_idx = end_idx
return batches
def get_encoder_output(self, batch_data):
"""Gets hidden states, final state, and a length masks corresponding to each encoder."""
code_embedded_tokens = self.code_features_to_embedding(
torch.cat([
self.embedding_store.get_code_embeddings(batch_data.code_ids),
batch_data.code_features
],
dim=-1))
code_hidden_states, code_final_state = self.code_encoder.forward(
code_embedded_tokens, batch_data.code_lengths, self.get_device())
old_nl_embedded_tokens = self.nl_features_to_embedding(
torch.cat([
self.embedding_store.get_nl_embeddings(batch_data.old_nl_ids),
batch_data.nl_features
],
dim=-1))
old_nl_hidden_states, old_nl_final_state = self.nl_encoder.forward(
old_nl_embedded_tokens, batch_data.old_nl_lengths,
self.get_device())
encoder_hidden_states, input_lengths = merge_encoder_outputs(
code_hidden_states, batch_data.code_lengths, old_nl_hidden_states,
batch_data.old_nl_lengths, self.get_device())
encoder_final_state = torch.einsum(
'bd,dh->bh',
torch.cat([code_final_state, old_nl_final_state], dim=-1),
self.encoder_final_to_decoder_initial)
mask = (torch.arange(encoder_hidden_states.shape[1],
device=self.get_device()).view(1, -1) >=
input_lengths.view(-1, 1)).unsqueeze(1)
code_masks = (torch.arange(code_hidden_states.shape[1],
device=self.get_device()).view(1, -1) >=
batch_data.code_lengths.view(-1, 1)).unsqueeze(1)
old_nl_masks = (torch.arange(old_nl_hidden_states.shape[1],
device=self.get_device()).view(1, -1) >=
batch_data.old_nl_lengths.view(-1, 1)).unsqueeze(1)
return encoder_hidden_states, encoder_final_state, mask, code_hidden_states, old_nl_hidden_states, code_masks, old_nl_masks
def forward(self, batch_data):
"""Computes the loss against the gold sequences corresponding to the examples in the batch. NOTE: teacher-forcing."""
encoder_hidden_states, initial_state, inp_length_mask, code_hidden_states, old_nl_hidden_states, code_masks, old_nl_masks = self.get_encoder_output(
batch_data)
decoder_input_embeddings = self.embedding_store.get_nl_embeddings(
batch_data.trg_nl_ids)[:, :-1]
decoder_states, decoder_final_state, generation_logprobs, copy_logprobs = self.decoder.forward(
initial_state, decoder_input_embeddings, encoder_hidden_states,
code_hidden_states, old_nl_hidden_states, inp_length_mask,
code_masks, old_nl_masks)
gold_generation_ids = batch_data.trg_nl_ids[:, 1:].unsqueeze(-1)
gold_generation_logprobs = torch.gather(
input=generation_logprobs, dim=-1,
index=gold_generation_ids).squeeze(-1)
copy_logprobs = copy_logprobs.masked_fill(
batch_data.invalid_copy_positions[:, 1:, :encoder_hidden_states.
shape[1]], float('-inf'))
gold_copy_logprobs = copy_logprobs.logsumexp(dim=-1)
gold_logprobs = torch.logsumexp(torch.cat([
gold_generation_logprobs.unsqueeze(-1),
gold_copy_logprobs.unsqueeze(-1)
],
dim=-1),
dim=-1)
gold_logprobs = gold_logprobs.masked_fill(
torch.arange(batch_data.trg_nl_ids[:, 1:].shape[-1],
device=self.get_device()).unsqueeze(0) >=
batch_data.trg_nl_lengths.unsqueeze(-1) - 1, 0)
likelihood_by_example = gold_logprobs.sum(dim=-1)
# Normalizing by length. Seems to help
likelihood_by_example = likelihood_by_example / (
batch_data.trg_nl_lengths - 1).float()
return -(likelihood_by_example).mean()
def beam_decode(self, batch_data):
"""Performs beam search on the decoder to get candidate predictions for every example in the batch."""
encoder_hidden_states, initial_state, inp_length_mask, code_hidden_states, old_nl_hidden_states, code_masks, old_nl_masks = self.get_encoder_output(
batch_data)
predictions, scores = self.decoder.beam_decode(
initial_state, encoder_hidden_states, code_hidden_states,
old_nl_hidden_states, inp_length_mask, self.max_nl_length,
batch_data, code_masks, old_nl_masks, self.get_device())
decoded_output = []
batch_size = initial_state.shape[0]
for i in range(batch_size):
beam_output = []
for j in range(len(predictions[i])):
token_ids = predictions[i][j]
tokens = self.embedding_store.get_nl_tokens(
token_ids, batch_data.input_ids[i],
batch_data.input_str_reps[i])
beam_output.append((tokens, scores[i][j]))
decoded_output.append(beam_output)
return decoded_output
def get_device(self):
"""Returns the proper device."""
if self.torch_device_name == 'gpu':
return torch.device('cuda')
else:
return torch.device('cpu')
def run_gradient_step(self, batch_data):
"""Performs gradient step."""
self.optimizer.zero_grad()
loss = self.forward(batch_data)
loss.backward()
self.optimizer.step()
return float(loss.cpu())
def run_train(self, train_data, valid_data):
"""Runs training over the entire training set across several epochs. Following each epoch,
loss on the validation data is computed. If the validation loss has improved, save the model.
Early-stopping is employed to stop training if validation hasn't improved for a certain number
of epochs."""
valid_batches = self.get_batches(valid_data)
train_batches = self.get_batches(train_data, shuffle=True)
best_loss = float('inf')
patience_tally = 0
for epoch in range(MAX_EPOCHS):
if patience_tally > PATIENCE:
print('Terminating')
break
self.train()
random.shuffle(train_batches)
train_loss = 0
for batch_data in train_batches:
train_loss += self.run_gradient_step(batch_data)
self.eval()
validation_loss = 0
with torch.no_grad():
for batch_data in valid_batches:
validation_loss += float(self.forward(batch_data).cpu())
validation_loss = validation_loss / len(valid_batches)
if validation_loss <= best_loss:
torch.save(self, self.model_path)
saved = True
best_loss = validation_loss
patience_tally = 0
else:
saved = False
patience_tally += 1
print('Epoch: {}'.format(epoch))
print('Training loss: {}'.format(train_loss / len(train_batches)))
print('Validation loss: {}'.format(validation_loss))
if saved:
print('Saved')
print('-----------------------------------')
def get_likelihood_scores(self, comment_generation_model,
formatted_beam_predictions, test_example):
"""Computes the generation likelihood score for each beam prediction based on the pre-trained
comment generation model."""
batch_examples = []
for j in range(len(formatted_beam_predictions)):
batch_examples.append(
Example(test_example.id, test_example.old_comment,
test_example.old_comment_tokens,
' '.join(formatted_beam_predictions[j]),
formatted_beam_predictions[j], test_example.old_code,
test_example.old_code_tokens, test_example.new_code,
test_example.new_code_tokens))
batch_data = comment_generation_model.get_batches(batch_examples)[0]
return np.asarray(
comment_generation_model.compute_generation_likelihood(batch_data))
def get_generation_model(self):
"""Loads the pre-trained comment generation model needed for re-ranking.
NOTE: the path is hard-coded here so may need to be modified."""
comment_generation_model = torch.load(FULL_GENERATION_MODEL_PATH)
comment_generation_model.torch_device_name = 'cpu'
comment_generation_model.cpu()
for c in comment_generation_model.children():
c.cpu()
comment_generation_model.eval()
return comment_generation_model
def run_evaluation(self, test_data, rerank):
"""Predicts updated comments for all comments in the test set and computes evaluation metrics."""
self.eval()
test_batches = self.get_batches(test_data)
test_predictions = []
generation_predictions = []
gold_strs = []
pred_strs = []
src_strs = []
references = []
pred_instances = []
with torch.no_grad():
for b_idx, batch_data in enumerate(test_batches):
test_predictions.extend(self.beam_decode(batch_data))
if not rerank:
test_predictions = [pred[0][0] for pred in test_predictions]
else:
comment_generation_model = self.get_generation_model()
with torch.no_grad():
generation_test_batches = comment_generation_model.get_batches(
test_data)
for gen_batch_data in generation_test_batches:
generation_predictions.extend(
comment_generation_model.greedy_decode(gen_batch_data))
reranked_predictions = []
for i in range(len(test_predictions)):
formatted_beam_predictions = []
model_scores = np.zeros(len(test_predictions[i]),
dtype=np.float)
generated = generation_predictions[i]
old_comment_tokens = test_data[i].old_comment_tokens
for b, (b_pred, b_score) in enumerate(test_predictions[i]):
b_pred_str = diff_utils.format_minimal_diff_spans(
test_data[i].old_comment_tokens, b_pred)
formatted_beam_predictions.append(b_pred_str.split(' '))
model_scores[b] = b_score
likelihood_scores = self.get_likelihood_scores(
comment_generation_model, formatted_beam_predictions,
test_data[i])
old_meteor_scores = compute_sentence_meteor(
[[old_comment_tokens]
for _ in range(len(formatted_beam_predictions))],
formatted_beam_predictions)
rerank_scores = [
(model_scores[j] * MODEL_LAMBDA) +
(likelihood_scores[j] * LIKELIHOOD_LAMBDA) +
(old_meteor_scores[j] * OLD_METEOR_LAMBDA)
for j in range(len(formatted_beam_predictions))
]
sorted_indices = np.argsort(-np.asarray(rerank_scores))
reranked_predictions.append(
test_predictions[i][sorted_indices[0]][0])
test_predictions = reranked_predictions
for i in range(len(test_predictions)):
pred_str = diff_utils.format_minimal_diff_spans(
test_data[i].old_comment_tokens, test_predictions[i])
prediction = pred_str.split()
gold_str = test_data[i].new_comment
gold_strs.append(gold_str)
pred_strs.append(pred_str)
src_strs.append(test_data[i].old_comment)
references.append([test_data[i].new_comment_tokens])
pred_instances.append(prediction)
print('Old comment: {}'.format(test_data[i].old_comment))
print('Gold comment: {}'.format(gold_str))
print('Predicted comment: {}'.format(pred_str))
print('Raw prediction: {}\n'.format(' '.join(test_predictions[i])))
try:
print('Old code:\n{}\n'.format(get_old_code(test_data[i])))
except:
print('Failed to print old code\n')
print('New code:\n{}\n'.format(get_new_code(test_data[i])))
print('----------------------------')
if rerank:
prediction_file = '{}_rerank.txt'.format(
self.model_path.split('.')[0])
else:
prediction_file = '{}.txt'.format(self.model_path.split('.')[0])
write_predictions(pred_strs, prediction_file)
write_predictions(src_strs, 'src.txt')
write_predictions(gold_strs, 'ref.txt')
predicted_accuracy = compute_accuracy(gold_strs, pred_strs)
predicted_bleu = compute_bleu(references, pred_instances)
predicted_meteor = compute_meteor(references, pred_instances)
predicted_sari = compute_sari(test_data, pred_instances)
predicted_gleu = compute_gleu(test_data, 'src.txt', 'ref.txt',
prediction_file)
print('Predicted Accuracy: {}'.format(predicted_accuracy))
print('Predicted BLEU: {}'.format(predicted_bleu))
print('Predicted Meteor: {}'.format(predicted_meteor))
print('Predicted SARI: {}'.format(predicted_sari))
print('Predicted GLEU: {}\n'.format(predicted_gleu))
def __init__(self, args):
self._log_step = args.log_step
self._batch_size = args.batch_size
self._image_size = args.image_size
self._latent_dim = args.latent_dim
self._coeff_gan = args.coeff_gan
self._coeff_vae = args.coeff_vae
self._coeff_reconstruct = args.coeff_reconstruct
self._coeff_latent = args.coeff_latent
self._coeff_kl = args.coeff_kl
self._norm = 'instance' if args.instance_normalization else 'batch'
self._use_resnet = args.use_resnet
self._augment_size = self._image_size + (30 if self._image_size == 256
else 15)
self._image_shape = [self._image_size, self._image_size, 3]
self.is_train = tf.placeholder(tf.bool, name='is_train')
self.lr = tf.placeholder(tf.float32, name='lr')
self.global_step = tf.train.get_or_create_global_step(graph=None)
image_a = self.image_a = \
tf.placeholder(tf.float32, [self._batch_size] + self._image_shape, name='image_a')
image_b = self.image_b = \
tf.placeholder(tf.float32, [self._batch_size] + self._image_shape, name='image_b')
z = self.z = \
tf.placeholder(tf.float32, [self._batch_size, self._latent_dim], name='z')
# Data augmentation
seed = random.randint(0, 2**31 - 1)
def augment_image(image):
image = tf.image.resize_images(
image, [self._augment_size, self._augment_size])
image = tf.random_crop(image,
[self._batch_size] + self._image_shape,
seed=seed)
image = tf.map_fn(
lambda x: tf.image.random_flip_left_right(x, seed), image)
return image
image_a = tf.cond(self.is_train, lambda: augment_image(image_a),
lambda: image_a)
image_b = tf.cond(self.is_train, lambda: augment_image(image_b),
lambda: image_b)
# Generator
G = Generator('G',
is_train=self.is_train,
norm=self._norm,
image_size=self._image_size)
# Discriminator
D = Discriminator('D',
is_train=self.is_train,
norm=self._norm,
activation='leaky',
image_size=self._image_size)
# Encoder
E = Encoder('E',
is_train=self.is_train,
norm=self._norm,
activation='relu',
image_size=self._image_size,
latent_dim=self._latent_dim,
use_resnet=self._use_resnet)
# conditional VAE-GAN: B -> z -> B'
z_encoded, z_encoded_mu, z_encoded_log_sigma = E(image_b)
image_ab_encoded = G(image_a, z_encoded)
# conditional Latent Regressor-GAN: z -> B' -> z'
image_ab = self.image_ab = G(image_a, z)
z_recon, z_recon_mu, z_recon_log_sigma = E(image_ab)
# Discriminate real/fake images
D_real = D(image_b)
D_fake = D(image_ab)
D_fake_encoded = D(image_ab_encoded)
loss_vae_gan = (tf.reduce_mean(tf.squared_difference(D_real, 0.9)) +
tf.reduce_mean(tf.square(D_fake_encoded)))
loss_image_cycle = tf.reduce_mean(tf.abs(image_b - image_ab_encoded))
loss_gan = (tf.reduce_mean(tf.squared_difference(D_real, 0.9)) +
tf.reduce_mean(tf.square(D_fake)))
loss_latent_cycle = tf.reduce_mean(tf.abs(z - z_recon))
loss_kl = -0.5 * tf.reduce_mean(1 + 2 * z_encoded_log_sigma -
z_encoded_mu**2 -
tf.exp(2 * z_encoded_log_sigma))
loss = self._coeff_vae * loss_vae_gan - self._coeff_reconstruct * loss_image_cycle + \
self._coeff_gan * loss_gan - self._coeff_latent * loss_latent_cycle - \
self._coeff_kl * loss_kl
# Optimizer
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
self.optimizer_D = tf.train.AdamOptimizer(learning_rate=self.lr, beta1=0.5) \
.minimize(loss, var_list=D.var_list, global_step=self.global_step)
self.optimizer_G = tf.train.AdamOptimizer(learning_rate=self.lr, beta1=0.5) \
.minimize(-loss, var_list=G.var_list)
self.optimizer_E = tf.train.AdamOptimizer(learning_rate=self.lr, beta1=0.5) \
.minimize(-loss, var_list=E.var_list)
# Summaries
self.loss_vae_gan = loss_vae_gan
self.loss_image_cycle = loss_image_cycle
self.loss_latent_cycle = loss_latent_cycle
self.loss_gan = loss_gan
self.loss_kl = loss_kl
self.loss = loss
tf.summary.scalar('loss/vae_gan', loss_vae_gan)
tf.summary.scalar('loss/image_cycle', loss_image_cycle)
tf.summary.scalar('loss/latent_cycle', loss_latent_cycle)
tf.summary.scalar('loss/gan', loss_gan)
tf.summary.scalar('loss/kl', loss_kl)
tf.summary.scalar('loss/total', loss)
tf.summary.scalar('model/D_real', tf.reduce_mean(D_real))
tf.summary.scalar('model/D_fake', tf.reduce_mean(D_fake))
tf.summary.scalar('model/D_fake_encoded',
tf.reduce_mean(D_fake_encoded))
tf.summary.scalar('model/lr', self.lr)
tf.summary.image('image/A', image_a[0:1])
tf.summary.image('image/B', image_b[0:1])
tf.summary.image('image/A-B', image_ab[0:1])
tf.summary.image('image/A-B_encoded', image_ab_encoded[0:1])
self.summary_op = tf.summary.merge_all()
out_channels //= channel_shrinkage_factor
i = num_conv_volumes
layers.append(('deconv%d' % i, nn.ConvTranspose2d(in_channels, 3, *conv_spec),))
layers.append(('tanh', nn.Tanh()))
self._net = nn.Sequential(OrderedDict(layers))
def forward(self, x):
assert x.size()[1] == self._num_in_channels
return self._net(x)
if __name__ == '__main__':
import torch
from encoder import Encoder
input_image_size = 64
expected_image_shape = (3, input_image_size, input_image_size)
enc = Encoder(input_image_size)
dec = Decoder(enc.num_out_channels)
x = torch.autograd.Variable(torch.rand(1, *expected_image_shape))
z = enc.forward(x)
x_ = dec.forward(z)
print(dec)
print(x_.size())
assert x_.size()[-3:] == expected_image_shape
# prepare the training data for the NetworkManager
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
x_train = x_train.astype('float32') / 255.
x_test = x_test.astype('float32') / 255.
y_train = to_categorical(y_train, 10)
y_test = to_categorical(y_test, 10)
dataset = [x_train, y_train, x_test,
y_test] # pack the dataset for the NetworkManager
with policy_sess.as_default():
# create the Encoder and build the internal policy network
controller = Encoder(policy_sess,
state_space,
B=B,
K=K_,
reg_param=REGULARIZATION,
controller_cells=CONTROLLER_CELLS,
restore_controller=RESTORE_CONTROLLER)
# create the Network Manager
manager = NetworkManager(dataset, epochs=MAX_EPOCHS, batchsize=BATCHSIZE)
print()
# train for number of trails
for trial in range(B):
with policy_sess.as_default():
K.set_session(policy_sess)
if trial == 0:
k = None
ON = True
OFF = False
# Entry
ADDRESS = 0
USERNAME = 1
if __name__ == '__main__':
# Components
ld = Led(red_pin=4, green_pin=17)
ld.setup()
motor = Motor(step_pin=23, dir_pin=24, power_pin=25)
motor.setup()
encoder = Encoder(clk_pin=22, dir_pin=27)
encoder.setup()
db = Database(host, "philosoraptor", "explosion", "doorman")
db.setup()
rfid = Rfid()
rfid.setup()
# Main Loop
while True:
ld.setstate(ld.BOTH)
addr = rfid.getaddr()
if addr:
#print "Address: %s" % addr
found, entry = db.checkaddr(addr)
def train_encoder(train_set, hidden_num, opt, learning_r, epoch=500, batch_size=32, pre_trained_path='', test_set=None,
name="Validation", tensorboard=False):
data_loader = make_data_loader(data_to_loader=train_set, batch_size=batch_size)
best_test_loss = 9999999999999
best_test_accuracy = 0
best_train_accuracy = 0
best_train_loss = 9999999999
accuracy_array = []
train_accuracy_array = []
is_pretrained = False
net = Encoder()
predictor = Predictor(hidden_num=hidden_num)
if pre_trained_path != '':
net = torch.load(pre_trained_path)["model"]
is_pretrained = True
if use_gpu:
net = net.cuda()
predictor = predictor.cuda()
criterion = nn.CrossEntropyLoss()
params = list(net.parameters()) + list(predictor.parameters())
if opt == 'ADAM':
optimizer = optim.Adam(params, lr=learning_r)
else:
optimizer = optim.SGD(params, lr=learning_r)
for epoch in range(epoch): # loop over the dataset multiple times
print("Epoch: ", epoch)
for i, data in enumerate(data_loader, 0):
# get the inputs
inputs, labels = data
if use_gpu:
inputs = inputs.cuda()
labels = labels.cuda()
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
outputs = net(inputs)
outputs = predictor(outputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
train_loss, accuracy = test_encoder(net, predictor, train_set, name="Training", show_log=True)
train_accuracy_array.append(accuracy)
if train_loss < best_train_loss:
best_train_loss = train_loss
best_train_accuracy = accuracy
if test_set is not None:
test_loss, accuracy = test_encoder(net, predictor, test_set, name=name, show_log=True)
accuracy_array.append(accuracy)
if test_loss < best_test_loss:
best_test_loss = test_loss
best_test_accuracy = accuracy
if is_pretrained:
name = "is_pretrained"
else:
name = "scratch"
if tensorboard:
writer_train.add_scalar('loss_'+name, best_train_loss, epoch)
writer_test.add_scalar('losss_'+name, best_test_loss, epoch)
writer_train.add_scalar('top-1_accuracys_'+name, best_train_accuracy, epoch)
writer_test.add_scalar('top-1_accuracys_'+name, best_test_accuracy, epoch)
if len(accuracy_array) < 3:
writer_train.add_scalar('top-3_accuracys_'+name, 0, epoch)
writer_test.add_scalar('top-3_accuracys_'+name, 0, epoch)
else:
writer_train.add_scalar('top-3_accuracys_'+name, sorted(train_accuracy_array)[-3], epoch)
writer_test.add_scalar('top-3_accuracys_'+name, sorted(accuracy_array)[-3], epoch)
return net, predictor, best_test_loss, best_test_accuracy, accuracy_array
class Robot():
"""
Paul v0.0:
posX si posY reprezinta coordonatele de pe plansa.
logicX si logicY sunt coordonatele relativ la pozitia de start a
robotului; sunt coordonate deduse de robot.
"""
MaxSpeed = 24
def __init__(self, posX, posY, posTheta):
self.w = 15
self.h = 10
self.heading = 0
self.posTheta = posTheta
self.posX = posX
self.posY = posY
self.targetTheta = 0
self.logicalTheta = 0
self.logicalX = 0
self.logicalY = 0
self.Rw = 9
self.Tr = 240
self.D = 9
self.dT1 = 0
self.dT2 = 0
self.T1 = 0
self.T2 = 0
self.counter = 0
self.leftEncoder = Encoder()
self.rightEncoder = Encoder()
self.leftRangeSensor = RangeSensor(self, self.h / 2 + 1, 0, -1)
self.frontRangeSensor = RangeSensor(self, self.w / 2 + 1, 1, 0)
self.rightRangeSensor = RangeSensor(self, self.h / 2 + 1, 0, 1)
self.statsWidget = RobotStatsWidget(self)
def setOrientation(self, posTheta):
self.posTheta = posTheta
self.statsWidget.setOrientation(self.posTheta)
def setTargetDirection(self, targetTheta):
self.targetTheta = targetTheta
if self.targetTheta > 2 * math.pi:
self.targetTheta = self.targetTheta % (2 * math.pi)
elif self.targetTheta < 0:
self.targetTheta = self.targetTheta % (2 * math.pi)
#if self.targetTheta > 2 * math.pi:
# self.target = 2 * math.pi
# print('targetTheta: %f' % self.targetTheta)
#elif self.targetTheta < 0:
# self.targetTheta = 0
# print('targetTheta: %f' % self.targetTheta)
self.statsWidget.setTargetDirection(self.targetTheta)
def move(self):
self.leftEncoder.addTicks(self.dT1)
self.rightEncoder.addTicks(self.dT2)
#print('dT1: ' + str(self.dT1) + ', dT2: ' + str(self.dT2))
b = 9
if (self.dT2 - self.dT1) < 0.00001:
dHeading = 0
dRealX = self.dT1 * math.cos(self.heading)
dRealY = self.dT1 * math.sin(self.heading)
elif (self.dT2 - self.dT1) > 24:
dHeading = (self.dT2 - self.dT1) / b
dRealX = 0
dRealY = 0
else:
R = b / 2 * (self.dT2 + self.dT1) / (self.dT2 - self.dT1)
dHeading = (self.dT2 - self.dT1) / b
dRealX = R * (math.sin(dHeading + self.heading) - math.sin(self.heading))
dRealY = R * (math.cos(self.heading) - math.cos(dHeading + self.heading))
#TODO: This is wrong
dHeading2 = 2 * math.pi * (self.Rw / self.D) * \
(self.dT1 - self.dT2) / self.Tr
#TODO: This is wrong
dRealX2 = self.Rw * math.cos(self.heading) * \
(self.dT1 + self.dT2) * math.pi / self.Tr
dRealY2 = self.Rw * math.sin(self.heading) * \
(self.dT1 + self.dT2) * math.pi / self.Tr
#print('dHeading: ' + str(dHeading2) + ', dRealX2: ' + str(dRealX2) + ', dRealY2: ' + str(dRealY2))
#print('dRealTheta: ' + str(dHeading) + ', dRealX: ' + str(dRealX) + ', dRealY: ' + str(dRealY))
if self.checkCollision(self.posX + dRealX2, self.posY + dRealY2) == False:
self.heading = self.heading + dHeading2
self.posX = self.posX + dRealX2
self.posY = self.posY + dRealY2
self.setOrientation(self.posTheta)
self.setTargetDirection(self.targetTheta - dHeading2)
#print('heading: ' + str(self.heading) + ', posX: ' + str(self.posX) + ', posY: ' + str(self.posY))
self.counter += 1
if self.counter == 5:
self.nextStep()
self.counter = 0
def draw(self, painter):
painter.setPen(QtGui.QColor(0xffffff))
painter.setBrush(QtGui.QColor(0x00CC66))
#rect = QtCore.QRect(0, 0, self.w, self.h)
#original = QtGui.QPolygon(rect, True)
#painter.drawPolygon(original)
coords = []
coords.append(QtCore.QPoint(0, 0))
coords.append(QtCore.QPoint(self.w, 0))
coords.append(QtCore.QPoint(self.w, self.h))
coords.append(QtCore.QPoint(0, self.h))
original = QtGui.QPolygon(coords)
original.translate(-self.w/2, -self.h/2)
transform = QtGui.QTransform().rotateRadians(self.posTheta)
#painter.fillRect(self.posX, self.posX, self.w, self.h, QtGui.QColor(0x0ff000))
rotated = transform.map(original)
#original.translate(self.posX, self.posY)
rotated.translate(self.posX, self.posY)
#QtCore.qDebug(str(rotated.point(0)))
#QtCore.qDebug(str(rotated.point(1)))
#QtCore.qDebug(str(rotated.point(2)))
#QtCore.qDebug(str(rotated.point(3)))
#painter.drawPolygon(original)
painter.setBrush(QtGui.QColor(0x0066CC))
painter.drawPolygon(rotated)
self.leftRangeSensor.draw(painter, QtGui.QColor(0x00FF00))
self.frontRangeSensor.draw(painter, QtGui.QColor(0xFF0000))
self.rightRangeSensor.draw(painter, QtGui.QColor(0x0000FF))
def nextStep(self):
leftDist = self.leftRangeSensor.getDistance()
frontDist = self.frontRangeSensor.getDistance()
rightDist = self.rightRangeSensor.getDistance()
self.statsWidget.setLeftRangeSensorDistance(leftDist)
self.statsWidget.setFrontRangeSensorDistance(frontDist)
self.statsWidget.setRightRangeSensorDistance(rightDist)
dHeading2 = 2 * math.pi * (self.Rw / self.D) * \
(self.dT1 - self.dT2) / self.Tr
# dT = dT1 - dT2
dT = self.targetTheta * self.D * self.Tr / (2 * math.pi * self.Rw)
if dT < 5:
self.setLeftMotorSpeed(24)
self.setRightMotorSpeed(24)
else:
self.setLeftMotorSpeed(dT)
self.setRightMotorSpeed(-dT)
self.T1 = 0
self.T2 = 0
def increaseLeftMotorSpeed(self, percent):
speed = self.dT1 + (percent / 100.0) * Robot.MaxSpeed
#print('speed: %f, dT1: %f, percent: %f, MaxSpeed: %f' % (speed, self.dT1, percent, Robot.MaxSpeed))
self.setLeftMotorSpeed(speed)
def increaseRightMotorSpeed(self, percent):
speed = self.dT2 + (percent / 100.0) * Robot.MaxSpeed
#print('speed: %f, dT2: %f, percent: %f, MaxSpeed: %f' % (speed, self.dT2, percent, Robot.MaxSpeed))
self.setRightMotorSpeed(speed)
def setLeftMotorSpeed(self, speed):
self.dT1 = speed
if speed < - Robot.MaxSpeed:
speed = - Robot.MaxSpeed
elif speed > Robot.MaxSpeed:
speed = Robot.MaxSpeed
self.statsWidget.setLeftMotorSpeed(speed)
#print('dT1: %f dT2: %f' % (self.dT1, self.dT2))
def setRightMotorSpeed(self, speed):
self.dT2 = speed
if speed < - Robot.MaxSpeed:
speed = - Robot.MaxSpeed
elif speed > Robot.MaxSpeed:
speed = Robot.MaxSpeed
self.statsWidget.setRightMotorSpeed(speed)
#print('dT1: %f dT2: %f' % (self.dT1, self.dT2))
'''
O verificare de coliziune rudimentara
'''
def checkCollision(self, newPosX, newPosY):
r = max(self.w, self.h) / 2
x1 = newPosX - r - 1
y1 = newPosY - r - 1
x2 = newPosX + r + 1
y2 = newPosX + r + 1
if x1 < 0 or y1 < 0 or\
x2 >= ImageMap.image.width() or\
y2 >= ImageMap.image.height():
return True
if ImageMap.image.pixel(x1, y1) != 0xFFFFFFFF:
return True
if ImageMap.image.pixel(x1, y2) != 0xFFFFFFFF:
return True
if ImageMap.image.pixel(x2, y1) != 0xFFFFFFFF:
return True
if ImageMap.image.pixel(x2, y2) != 0xFFFFFFFF:
return True
return False
def getStatsWidget(self):
return self.statsWidget
class CommentGenerationModel(nn.Module):
"""Simple Seq2Seq w/ attention + copy model for pure comment generation (i.e. generating a comment given a method)."""
def __init__(self, model_path):
super(CommentGenerationModel, self).__init__()
self.model_path = model_path
self.torch_device_name = 'cpu'
def initialize(self, train_data, embedding_store=None):
"""Initializes model parameters from pre-defined hyperparameters and other hyperparameters
that are computed based on statistics over the training data."""
nl_lengths = []
code_lengths = []
nl_token_counter = Counter()
code_token_counter = Counter()
for ex in train_data:
trg_sequence = [START] + ex.new_comment_tokens + [END]
nl_token_counter.update(trg_sequence)
nl_lengths.append(len(trg_sequence))
code_sequence = ex.new_code_tokens
code_token_counter.update(code_sequence)
code_lengths.append(len(code_sequence))
self.max_nl_length = int(
np.percentile(np.asarray(sorted(nl_lengths)), LENGTH_CUTOFF_PCT))
self.max_code_length = int(
np.percentile(np.asarray(sorted(code_lengths)), LENGTH_CUTOFF_PCT))
nl_counts = np.asarray(sorted(nl_token_counter.values()))
nl_threshold = int(np.percentile(nl_counts, VOCAB_CUTOFF_PCT)) + 1
code_counts = np.asarray(sorted(code_token_counter.values()))
code_threshold = int(np.percentile(nl_counts, VOCAB_CUTOFF_PCT)) + 1
if embedding_store is None:
self.embedding_store = EmbeddingStore(
nl_threshold, NL_EMBEDDING_SIZE, nl_token_counter,
code_threshold, CODE_EMBEDDING_SIZE, code_token_counter,
DROPOUT_RATE, 0, 0, 0, False)
else:
self.embedding_store = embedding_store
self.code_encoder = Encoder(CODE_EMBEDDING_SIZE, HIDDEN_SIZE,
NUM_LAYERS, DROPOUT_RATE)
self.decoder = GenerationDecoder(NL_EMBEDDING_SIZE,
DECODER_HIDDEN_SIZE, 2 * HIDDEN_SIZE,
self.embedding_store,
NL_EMBEDDING_SIZE, DROPOUT_RATE)
self.optimizer = torch.optim.Adam(self.parameters(), lr=LR)
def get_batches(self, dataset, shuffle=False):
"""Divides the dataset into batches based on pre-defined BATCH_SIZE hyperparameter.
Each batch is tensorized so that it can be directly passed into the network."""
batches = []
if shuffle:
random.shuffle(dataset)
curr_idx = 0
while curr_idx < len(dataset):
start_idx = curr_idx
end_idx = min(start_idx + BATCH_SIZE, len(dataset))
code_token_ids = []
code_lengths = []
trg_token_ids = []
trg_extended_token_ids = []
trg_lengths = []
invalid_copy_positions = []
inp_str_reps = []
inp_ids = []
for i in range(start_idx, end_idx):
code_sequence = dataset[i].new_code_tokens
code_sequence_ids = self.embedding_store.get_padded_code_ids(
code_sequence, self.max_code_length)
code_length = min(len(code_sequence), self.max_code_length)
code_token_ids.append(code_sequence_ids)
code_lengths.append(code_length)
ex_inp_str_reps = []
ex_inp_ids = []
extra_counter = len(self.embedding_store.nl_vocabulary)
max_limit = len(
self.embedding_store.nl_vocabulary) + self.max_code_length
out_ids = set()
for c in code_sequence[:code_length]:
nl_id = self.embedding_store.get_nl_id(c)
if self.embedding_store.is_nl_unk(
nl_id) and extra_counter < max_limit:
if c in ex_inp_str_reps:
nl_id = ex_inp_ids[ex_inp_str_reps.index(c)]
else:
nl_id = extra_counter
extra_counter += 1
out_ids.add(nl_id)
ex_inp_str_reps.append(c)
ex_inp_ids.append(nl_id)
trg_sequence = [START] + dataset[i].new_comment_tokens + [END]
trg_sequence_ids = self.embedding_store.get_padded_nl_ids(
trg_sequence, self.max_nl_length)
trg_extended_sequence_ids = self.embedding_store.get_extended_padded_nl_ids(
trg_sequence, self.max_nl_length, ex_inp_ids,
ex_inp_str_reps)
trg_token_ids.append(trg_sequence_ids)
trg_extended_token_ids.append(trg_extended_sequence_ids)
trg_lengths.append(min(len(trg_sequence), self.max_nl_length))
inp_str_reps.append(ex_inp_str_reps)
inp_ids.append(ex_inp_ids)
invalid_copy_positions.append(
get_invalid_copy_locations(ex_inp_str_reps,
self.max_code_length,
trg_sequence,
self.max_nl_length))
batches.append(
GenerationBatchData(
torch.tensor(code_token_ids,
dtype=torch.int64,
device=self.get_device()),
torch.tensor(code_lengths,
dtype=torch.int64,
device=self.get_device()),
torch.tensor(trg_token_ids,
dtype=torch.int64,
device=self.get_device()),
torch.tensor(trg_extended_token_ids,
dtype=torch.int64,
device=self.get_device()),
torch.tensor(trg_lengths,
dtype=torch.int64,
device=self.get_device()),
torch.tensor(invalid_copy_positions,
dtype=torch.uint8,
device=self.get_device()), inp_str_reps,
inp_ids))
curr_idx = end_idx
return batches
def get_encoder_output(self, batch_data):
"""Gets hidden states, final state, and a length-mask from the encoder."""
code_embedded_tokens = self.embedding_store.get_code_embeddings(
batch_data.code_ids)
code_hidden_states, code_final_state = self.code_encoder.forward(
code_embedded_tokens, batch_data.code_lengths, self.get_device())
mask = (torch.arange(code_hidden_states.shape[1],
device=self.get_device()).view(1, -1) >=
batch_data.code_lengths.view(-1, 1)).unsqueeze(1)
return code_hidden_states, code_final_state, mask
def forward(self, batch_data):
"""Computes the loss against the gold sequences corresponding to the examples in the batch. NOTE: teacher-forcing."""
encoder_hidden_states, initial_state, inp_length_mask = self.get_encoder_output(
batch_data)
decoder_input_embeddings = self.embedding_store.get_nl_embeddings(
batch_data.trg_nl_ids)[:, :-1]
decoder_states, decoder_final_state, generation_logprobs, copy_logprobs = self.decoder.forward(
initial_state, decoder_input_embeddings, encoder_hidden_states,
inp_length_mask)
gold_generation_ids = batch_data.trg_nl_ids[:, 1:].unsqueeze(-1)
gold_generation_logprobs = torch.gather(
input=generation_logprobs, dim=-1,
index=gold_generation_ids).squeeze(-1)
copy_logprobs = copy_logprobs.masked_fill(
batch_data.invalid_copy_positions[:, 1:, :encoder_hidden_states.
shape[1]], float('-inf'))
gold_copy_logprobs = copy_logprobs.logsumexp(dim=-1)
gold_logprobs = torch.logsumexp(torch.cat([
gold_generation_logprobs.unsqueeze(-1),
gold_copy_logprobs.unsqueeze(-1)
],
dim=-1),
dim=-1)
gold_logprobs = gold_logprobs.masked_fill(
torch.arange(batch_data.trg_nl_ids[:, 1:].shape[-1],
device=self.get_device()).unsqueeze(0) >=
batch_data.trg_nl_lengths.unsqueeze(-1) - 1, 0)
likelihood_by_example = gold_logprobs.sum(dim=-1)
# Normalizing by length. Seems to help
likelihood_by_example = likelihood_by_example / (
batch_data.trg_nl_lengths - 1).float()
return -(likelihood_by_example).mean()
def compute_generation_likelihood(self, batch_data):
"""This is not used by the generation model but rather the comment update model for re-ranking. It computes P(Comment|Method)."""
with torch.no_grad():
encoder_hidden_states, initial_state, inp_length_mask = self.get_encoder_output(
batch_data)
decoder_input_embeddings = self.embedding_store.get_nl_embeddings(
batch_data.trg_nl_ids)[:, :-1]
decoder_states, decoder_final_state, generation_logprobs, copy_logprobs = self.decoder.forward(
initial_state, decoder_input_embeddings, encoder_hidden_states,
inp_length_mask)
gold_generation_ids = batch_data.trg_nl_ids[:, 1:].unsqueeze(-1)
gold_generation_logprobs = torch.gather(
input=generation_logprobs, dim=-1,
index=gold_generation_ids).squeeze(-1)
copy_logprobs = copy_logprobs.masked_fill(
batch_data.
invalid_copy_positions[:, 1:, :encoder_hidden_states.shape[1]],
float('-inf'))
gold_copy_logprobs = copy_logprobs.logsumexp(dim=-1)
gold_logprobs = torch.logsumexp(torch.cat([
gold_generation_logprobs.unsqueeze(-1),
gold_copy_logprobs.unsqueeze(-1)
],
dim=-1),
dim=-1)
gold_logprobs = gold_logprobs.masked_fill(
torch.arange(batch_data.trg_nl_ids[:, 1:].shape[-1],
device=self.get_device()).unsqueeze(0) >=
batch_data.trg_nl_lengths.unsqueeze(-1) - 1, 0)
return torch.exp(
gold_logprobs.sum(dim=-1) /
(batch_data.trg_nl_lengths - 1).float())
def greedy_decode(self, batch_data):
"""Predicts a comment for every method in the batch in a greedy manner."""
encoder_hidden_states, initial_state, inp_length_mask = self.get_encoder_output(
batch_data)
predictions, scores = self.decoder.greedy_decode(
initial_state, encoder_hidden_states, inp_length_mask,
self.max_nl_length, batch_data, self.get_device())
batch_size = initial_state.shape[0]
decoded_tokens = []
for i in range(batch_size):
token_ids = predictions[i]
tokens = self.embedding_store.get_nl_tokens(
token_ids, batch_data.input_ids[i],
batch_data.input_str_reps[i])
decoded_tokens.append(tokens)
return decoded_tokens
def get_device(self):
"""Returns the proper device."""
if self.torch_device_name == 'gpu':
return torch.device('cuda')
else:
return torch.device('cpu')
def run_gradient_step(self, batch_data):
"""Performs gradient step."""
self.optimizer.zero_grad()
loss = self.forward(batch_data)
loss.backward()
self.optimizer.step()
return float(loss.cpu())
def run_train(self, train_data, valid_data):
"""Runs training over the entire training set across several epochs. Following each epoch,
loss on the validation data is computed. If the validation loss has improved, save the model.
Early-stopping is employed to stop training if validation hasn't improved for a certain number
of epochs."""
valid_batches = self.get_batches(valid_data)
train_batches = self.get_batches(train_data, shuffle=True)
best_loss = float('inf')
patience_tally = 0
for epoch in range(MAX_EPOCHS):
if patience_tally > PATIENCE:
print('Terminating')
break
self.train()
random.shuffle(train_batches)
train_loss = 0
for batch_data in train_batches:
train_loss += self.run_gradient_step(batch_data)
self.eval()
validation_loss = 0
with torch.no_grad():
for batch_data in valid_batches:
validation_loss += float(self.forward(batch_data).cpu())
validation_loss = validation_loss / len(valid_batches)
if validation_loss <= best_loss:
torch.save(self, self.model_path)
saved = True
best_loss = validation_loss
patience_tally = 0
else:
saved = False
patience_tally += 1
print('Epoch: {}'.format(epoch))
print('Training loss: {}'.format(train_loss / len(train_batches)))
print('Validation loss: {}'.format(validation_loss))
if saved:
print('Saved')
print('-----------------------------------')
sys.stdout.flush()
def run_evaluation(self, test_data):
"""Generates predicted comments for all comments in the test set and computes evaluation metrics."""
self.eval()
test_batches = self.get_batches(test_data)
test_predictions = []
# Lists of string predictions
gold_strs = []
pred_strs = []
# Lists of tokenized predictions
references = []
pred_instances = []
with torch.no_grad():
for b_idx, batch_data in enumerate(test_batches):
print('Evaluating {}'.format(b_idx))
sys.stdout.flush()
test_predictions.extend(self.greedy_decode(batch_data))
for i in range(len(test_predictions)):
prediction = test_predictions[i]
gold_str = test_data[i].new_comment
pred_str = ' '.join(prediction)
gold_strs.append(gold_str)
pred_strs.append(pred_str)
references.append([test_data[i].new_comment_tokens])
pred_instances.append(prediction)
print('Gold comment: {}'.format(gold_str))
print('Predicted comment: {}'.format(pred_str))
print('----------------------------')
predicted_accuracy = compute_accuracy(gold_strs, pred_strs)
predicted_bleu = compute_bleu(references, pred_instances)
predicted_meteor = compute_meteor(references, pred_instances)
print('Predicted Accuracy: {}'.format(predicted_accuracy))
print('Predicted BLEU: {}'.format(predicted_bleu))
print('Predicted Meteor: {}\n'.format(predicted_meteor))
from decoder import Decoder
from base import BaseModel
from utils.general import Config
@click.option('--data', default="configs/data_small.json",
help='Path to data json config')
@click.option('--vocab', default="configs/vocab_small.json",
help='Path to vocab json config')
@click.option('--training', default="configs/training_small.json",
help='Path to training json config')
@click.option('--model', default="configs/model.json",
help='Path to model json config')
# @click.option('--output', default="results/small/",
# help='Dir for results and model weights')
_config_args = [
'configs/data_small.json',
'configs/vocab_small.json',
'configs/training_small.json',
'configs/model.json',
# 'results/small/',
]
_config = Config(_config_args)
encoder = Encoder(_config)
decoder = Decoder(_config, _vocab.n_tok, _vocab.id_end)
dataloader = read_dataset('../pic', img_size, batch_size)
version = input('result version:')
c_loss_weight = 0.3
RF_loss_weight = 0.7
generator_loss_weight = 0.7
path = './result_' + version + '/arg_' + version + '.txt'
f = open(path, 'a+')
arg='epoch='+str(epoch)+'\n'+'batch_size='+str(batch_size)+'\n'+'img_size='+str(img_size)+'\n'+\
'c_size='+str(c_size)+'\n'+'z_size='+str(z_size)+'\n'+'RF_loss_weight=generator_loss_weight='+str(RF_loss_weight)+'\n'+'c_loss_weight='+str(c_loss_weight)+'\n'
f.write(arg + '\n')
f.close()
unloader = transforms.ToPILImage()
encoder = Encoder(c_size, z_size)
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
m.weight.data.normal_(0.0, 0.02)
elif classname.find('BatchNorm') != -1:
m.weight.data.normal_(1.0, 0.02)
m.bias.data.fill_(0)
g = Generator(c_size + z_size)
g.apply(weights_init)
d = Discriminator()
import sys
from encoder import Encoder
f=open('stats.out')
lines=f.readlines()
f.close()
dist=[]
for x in range(len(lines)):
p=float(lines[x].strip())
dist.append((p, x))
enc=Encoder(dist)
enc.encode(sys.argv[1], sys.argv[2])
enc.decode(sys.argv[2], sys.argv[3])
def __init__(self, vocab_size):
super(EncoderDecoder, self).__init__()
self.embedding_layer = nn.Embedding(vocab_size, EMBEDDING_DIM)
self.encoder = Encoder(vocab_size, self.embedding_layer)
self.decoder = AttentionDecoder(vocab_size, self.embedding_layer)
