def __init__(self,
input_dim,
output_dim,
model_dim,
n_head,
key_dim,
value_dim,
hidden_dim,
n_layers,
pad_idx,
max_seq_len=200,
dropout=0.1):
super(Transformer, self).__init__()
self.pad_idx = pad_idx
self.encoder = encoder.Encoder(input_dim=input_dim,
model_dim=model_dim,
n_head=n_head,
key_dim=key_dim,
value_dim=value_dim,
n_layers=n_layers,
hidden_dim=hidden_dim,
max_seq_len=max_seq_len,
dropout=dropout)
self.decoder = decoder.Decoder(input_dim=input_dim,
model_dim=model_dim,
n_head=n_head,
key_dim=key_dim,
value_dim=value_dim,
n_layers=n_layers,
hidden_dim=hidden_dim,
max_seq_len=max_seq_len,
dropout=dropout)
self.linear_output = nn.Linear(in_features=model_dim,
out_features=output_dim)
def run(self, FSAngle, FSVelocity):
car = vehicle.Car(constants.CAR_POS_X, constants.CAR_POS_Y,
constants.CAR_ANGLE)
iteration = 0
dec = decoder.Decoder(FSAngle, FSVelocity, self.car, False)
while not self.exit:
dt = self.clock.get_time() / 1000
for event in pygame.event.get():
if event.type == pygame.QUIT:
self.exit = True
ds, drot = dec.get_movement_params()
fuzzy_text = dec.fuzzy_text
self.car.update(dt, ds, drot)
current_pixel_color = self.draw_screen(fuzzy_text)
iteration = iteration + 1
if self.car.is_idle(iteration) or self.car.is_collided(
current_pixel_color):
break
self.clock.tick(self.ticks)
pygame.quit()
return vehicle.distance(self.car.center_position().x,
self.car.center_position().y, constants.GOAL.x,
constants.GOAL.y)
def _create_blocks(self):
#create the environment-agent interface (decoder)
self.decoder = decoder.Decoder(self, -1)
#create the "hippocampus" / state tracer
self.action_buffer = OrNode(self.network, (self.n_actions, 1), -1)
self.hippocampus = hippocampus.Hippocampus(self, -1)
#create the "cortex" / reward estimator
#create the action encoder
if self.n_replicates > 1:
self.cortex = cortex.MultiCortex(self,
-1,
noisy=self.noisy,
n_replicates=self.n_replicates,
dynrange=self.dynrange)
self.encoder = encoder.MultiEncoder(self, -1)
else:
self.cortex = cortex.Cortex(self, -1, noisy=self.noisy)
self.encoder = encoder.Encoder(self, -1)
#create stubs for SNIP
self.stubs['state'] = self.network.createInputStubGroup(
size=self.n_states)
self.stubs['action'] = self.network.createInputStubGroup(
size=self.n_actions)
self.stubs['reward'] = self.network.createInputStubGroup(size=1)
self.stubs['punishment'] = self.network.createInputStubGroup(size=1)
self.stubs['draw'] = self.network.createInputStubGroup(size=1)
def setUp(self):
self.reg = cpu.Registers()
self.mem = memory.Memory()
self.alu = cpu.ArithmeticLogicUnit()
self.decoder = decoder.Decoder(self.reg, self.mem, self.alu)
self.clock = cpu.Clock(self.reg, self.decoder)
def test_beam_search(self):
alphabet = 'ab'
probabilities = np.array([[0.4, 0, 0.6], [0.4, 0, 0.6]])
expected = 'a'
bs_decoder = decoder.Decoder()
actual = bs_decoder.beam_search(probabilities, alphabet)
self.assertEqual(actual, expected)
def __init__(self, latent_spaces, batch_size):
super(ANVAE, self).__init__()
self.batch_size = batch_size
self.latent_spaces = 3
self.level_sizes = [1, 1, 1]
self.input_s = [32, 32, 1]
self.latent_channels = 20
self.h_dim = 1000
self.encoder = encoder.Encoder(self.latent_spaces, self.input_s)
self.decoder = decoder.Decoder(self.encoder(
tf.zeros([self.batch_size, 32, 32, 1]), False),
latent_channels=self.latent_channels,
level_sizes=self.level_sizes)
self.discriminator = discriminator.Discriminator(
self.latent_spaces, self.input_s, self.h_dim)
self.lr_ae = .0001
self.lr_dc = .0001
self.lr_gen = .0001
self.ae_optimizer = tf.keras.optimizers.Adamax(self.lr_ae, clipnorm=2)
self.gen_optimizer = tf.keras.optimizers.Adamax(self.lr_gen,
clipnorm=2)
self.dc_optimizer = tf.keras.optimizers.Adamax(self.lr_dc, clipnorm=2)
self.ae_loss_weight = 1.
self.gen_loss_weight = 6.
self.dc_loss_weight = 6.
self.lastEncVars = []
self.lastDecVars = []
self.lastDiscVars = []
self.debugCount = 0
self.counter = 1
self.log_writer = tf.summary.create_file_writer(logdir='./tf_summary')
self.step_count = 0
self.conv_layers = []
self.sr_u = {}
self.sr_v = {}
self.num_power_iter = 4
for layer in self.encoder.layers:
if isinstance(layer, tf.keras.layers.Conv2D) or isinstance(
layer, tf.keras.layers.DepthwiseConv2D):
self.conv_layers.append(layer)
for layer in self.decoder.layers:
if isinstance(layer, tf.keras.layers.Conv2D) or isinstance(
layer, tf.keras.layers.DepthwiseConv2D):
self.conv_layers.append(layer)
for layer in self.discriminator.layers:
if isinstance(layer, tf.keras.layers.Conv2D) or isinstance(
layer, tf.keras.layers.DepthwiseConv2D):
self.conv_layers.append(layer)
def decode(self):
self.stage_one_enable(False)
try:
self.decoder = decoder.Decoder(self.lang.text())
self.lines = self.decoder.get_strings().splitlines()
except FileNotFoundError as e:
QMessageBox.critical(self, "Error", str(e), QMessageBox.Ok,
QMessageBox.Ok)
self.stage_one_enable(True)
return
except IndexError:
QMessageBox.critical(self, "Error",
"Could not find a lotr.str in the file",
QMessageBox.Ok, QMessageBox.Ok)
self.decoder.file.close()
name = ""
for line in self.lines:
if not line.lower().startswith(("controlbar", '"', "end")):
continue
if line.startswith("CONTROLBAR"):
name = line.split(":")[1]
elif line.startswith('"') and name is not None:
self.buttons[name] = line[1:-1]
elif line.lower().startswith("end"):
name = None
QMessageBox.information(
self, "Done",
"Done decoding, you may now begin editing shortcuts below",
QMessageBox.Ok, QMessageBox.Ok)
self.stage_two_enable(True)
self.search_box.setCompleter(QCompleter(self.buttons.keys()))
def main():
alpha_list = ['B', 'P', 'F', 'D']
digit_list = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
crater_list = ['1', '2', '3', '4', '5']
connect = database.Database()
connect.connect()
# connect.create_database()
connect.create_table()
test = decoder.Decoder([['B'], ['P'], alpha_list, alpha_list, ['0'],
crater_list, digit_list, digit_list, digit_list,
digit_list, digit_list])
barcode_template = decoder.generate_template(test)
while True:
collected_data = decoder.DataEntry()
collected_data.barcode_input()
decoder.compare_data(barcode_template, collected_data.data)
print("----------------Barcode Descrpition---------------\n")
print("Barcode to be validated: " + str(collected_data.data))
print("--------------------------------------------------")
barcode_config = decoder.BarcodeDescription(collected_data.data)
barcode_config.output_description()
print("--------------------------------------------------")
array = barcode_config.export_barcode()
connect.insert_record(array)
connect.end_connection()
def __init__(self, num_iterations, H, R, G):
self.num_iterations = num_iterations # number codewords to pass through transmitter and decoder when computing bit error rate
self.H = H # parity matrix
self.R = R # decoding matrix
self.G = G # code generator matrix
self.Transmitter = transmitter.Transmitter(G)
self.Decoder = decoder.Decoder(10, H, R)
def testStringFromCTC(self):
"""Tests that the decoder can decode sequences including multi-codes.
"""
# - f - a r - m(1/2)m -junk sp b a r - n -
ctc_labels = [9, 6, 9, 1, 3, 9, 4, 9, 5, 5, 9, 5, 0, 2, 1, 3, 9, 4, 9]
decode = decoder.Decoder(filename=_testdata('charset_size_10.txt'))
text = decode.StringFromCTC(ctc_labels, merge_dups=True, null_label=9)
self.assertEqual(text, 'farm barn')
def _create_blocks(self):
#create the environment-agent interface (decoder)
self.decoder = decoder.Decoder(self, 0)
#create the "hippocampus" / state tracer
self.hippocampus = hippocampus.Hippocampus(self, 1)
#create the "cortex" / reward estimator
self.cortex = cortex.Cortex(self, 2)
#create the action encoder
self.encoder = encoder.Encoder(self, 3)
def main(unused_argv):
vocab = dataset.Vocab(FLAGS.vocab_path, 200000)
# Check for presence of required special tokens.
assert vocab.tokenToId(dataset.PAD_TOKEN) > 0
assert vocab.tokenToId(dataset.UNKNOWN_TOKEN) > 0
assert vocab.tokenToId(dataset.SENTENCE_START) > 0
assert vocab.tokenToId(dataset.SENTENCE_END) > 0
assert vocab.tokenToId(dataset.WORD_BEGIN) > 0
assert vocab.tokenToId(dataset.WORD_CONTINUE) > 0
assert vocab.tokenToId(dataset.WORD_END) > 0
params = selector.parameters(
mode=FLAGS.mode, # train, eval, decode
min_lr=0.01, # min learning rate.
lr=0.1, # learning rate
batch_size=1,
c_timesteps=600, # context length
q_timesteps=30, # question length
min_input_len=2, # discard context, question < than this words
hidden_size=200, # for rnn cell and embedding
emb_size=200, # If 0, don't use embedding
max_decode_steps=4,
maxout_size=32,
max_grad_norm=2)
batcher = batch_reader.Generator(FLAGS.data_path,
vocab,
params,
FLAGS.context_key,
FLAGS.question_key,
FLAGS.answer_key,
FLAGS.max_context_sentences,
FLAGS.max_question_sentences,
bucketing=FLAGS.use_bucketing,
truncate_input=FLAGS.truncate_input)
tf.set_random_seed(FLAGS.random_seed)
if params.mode == 'train':
model = selector.Model(params,
len(vocab),
num_cpus=FLAGS.num_cpus,
num_gpus=FLAGS.num_gpus)
_train(model, batcher)
elif params.mode == 'eval':
model = selector.Model(params,
len(vocab),
num_cpus=FLAGS.num_cpus,
num_gpus=FLAGS.num_gpus)
_eval(model, batcher)
elif params.mode == 'decode':
model = selector.Model(params,
len(vocab),
num_cpus=FLAGS.num_cpus,
num_gpus=FLAGS.num_gpus)
machine = decoder.Decoder(model, batcher, params, vocab)
machine.loop()
def main():
files = gen_filenames()
if not len(files):
print_help()
return
for i, f in enumerate(files):
ddd = decoder.Decoder()
result = ddd.tryWholeFile(f)
print "".join(result)
def setup_computer(self, start_ip):
"""Build the computer."""
self.reg = cpu.Registers()
self.mem = memory.Memory()
self.alu = cpu.ArithmeticLogicUnit()
self.decoder_obj = decoder.Decoder(self.reg, self.mem, self.alu)
self.clock = cpu.Clock(self.reg, self.decoder_obj)
self.reg.ip = start_ip
def __init__(self, dataset_params, encoder_params, decoder_params):
self.X = tf.placeholder(tf.float32,
shape=(None, dataset_params['n_features']),
name="X")
self._encoder = enc.Encoder(inputs=self.X, **encoder_params)
z = self._encoder.get_latent_representation()
self._decoder = dec.Decoder(inputs=z, **decoder_params)
self._Xhat = self._decoder.get_outputs()
def analyze_generate():
a = load_user_data(conf_path)
dbSession = model.startSession(a)
dec = decoder.Decoder()
q = dbSession.query(model.Tweet)
tq = q.filter(model.Tweet.isAnalyze < 2)[:10000]
for t in tq:
print(t.user, parse_text(t.text))
print(dec.decode(parse_text(t.text)))
def __init__(self):
self.bv = None
self.av = None
self.tag = None
self.err_tag = []
self.stat_dict = {}
self.activity_dict = {}
self.comments = []
self.decoder = decoder.Decoder()
self.sum_table_3 = 0
self.sum_table_4 = 0
def setup_computer():
"""Build the computer."""
reg = cpu.Registers()
mem = memory.Memory()
alu = cpu.ArithmeticLogicUnit()
decoder_obj = decoder.Decoder(reg, mem, alu)
clock = cpu.Clock(reg, decoder_obj)
reg.ip = ADDR(0x20)
return mem, clock
def main():
test_data_file = sys.argv[1]
model_file = sys.argv[2]
sys_file = sys.argv[3]
m = model.Model(model_file)
y_true = []
y_pred = []
with open(sys_file, "w") as wfp:
wfp.write('%%%%% test data:\n')
with open(test_data_file, "r") as rfp:
for index, line in enumerate(tqdm(rfp)):
line = line.strip()
d = decoder.Decoder(m)
d.read_instance(line, index)
d.fill_class_prob()
d.fill_final_prob()
y_pred.append(d.find_predicted_class_label())
y_true.append(line.split()[0])
wfp.write(d.report_sys_string() + '\n')
sorted_list = sorted(m.feature_weight_dictionary.items(),
key=lambda x: x[0])
label_list = []
for element in sorted_list:
label_list.append(element[0])
cm = pd.DataFrame(confusion_matrix(y_true, y_pred, labels=label_list),
index=label_list,
columns=label_list)
print("Confusion matrix for the testing data:")
print("row is the truth, column is the system output\n")
print(cm.to_string() + "\n")
print("accuracy=%s\n\n" % (accuracy_score(y_true, y_pred)))
def main(_):
# create foler for generated images
if not os.path.exists(FLAGS.out_dir):
os.makedirs(FLAGS.out_dir)
# Import data
datas = filter(lambda x: x.endswith('jpg'), os.listdir(FLAGS.data_dir))
# data placeholders
X = tf.placeholder(tf.float32, shape=[None, 4096]) # 64*64 = 4096
# initializes encoder and decoder
_encoder = encoder.Encoder(pkeep=0.75)
_decoder = decoder.Decoder(pkeep=0.75)
# encodes and decodes
X_fake = _decoder.decode(_encoder.encode(X))
# loss : quadratic error
loss = tf.reduce_mean(tf.pow(X - X_fake, 2))
# invoke the optimizer
solver = tf.train.AdamOptimizer(learning_rate).minimize(loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# train
for it in range(training_iter):
for start, end in zip(range(0, len(datas), batch_size),
range(batch_size, len(datas), batch_size)):
# next input batch
batch_image_files = datas[start:end]
batch_images = map(
lambda x: utils.crop_resize(os.path.join(
FLAGS.data_dir, x)), batch_image_files)
batch_images = np.array(batch_images).astype(np.float32)
# keep the last batch for tesing
if it == training_iter - 1:
# run session
_, loss_curr = sess.run([solver, loss],
feed_dict={X: batch_images})
if it % display_step == 0:
print('Iter: {}'.format(it))
print('cost: {:.4}'.format(loss_curr))
else:
# test on 10 test images
test = sess.run(X_fake, feed_dict={X: batch_images[:10]})
utils.save(test, batch_images[:10])
def test_no_conditioning(self):
dec = decoder.Decoder(self.wemb,
self.num_layers,
self.dropout,
use_attention=False,
bidirectional_encoder=False)
dec.init_state(batch_size=self.batch_size, encoder_final=None)
dec_outs, attns = dec(self.tgt, memory_bank=None, memory_lengths=None)
self.assertEqual(list(dec_outs.size()),
[self.T_tgt, self.batch_size, self.dim])
def show_img(self):
self.dec = deco.Decoder(self.controller.data, self.controller.cal)
self.dec.image_at('634ghz', attr='s_db')
plt.savefig('current.jpg')
image = Image.open("current.jpg")
photo = ImageTk.PhotoImage(image)
self.chart_img.pack_forget()
self.chart_img = Label(self.chart, image=photo)
self.chart_img.image = photo
self.chart_img.pack(side=TOP, fill=BOTH, expand=True)
plt.clf()
def Inference(train_dir,
model_str,
infer_data,
decoder_file,
num_lines,
graph_def_file=None,
reader=None):
"""Restores a model from a checkpoint and evaluates it.
Args:
train_dir: Directory to find checkpoints.
model_str: Network specification string.
infer_data: Inference data file pattern.
decoder_file: File to read to decode the labels.
num_lines: Number of lines in infer_data
graph_def_file: File to write graph definition to for freezing.
reader: Function that returns an actual reader to read Examples from input
files. If None, uses tf.TFRecordReader().
Returns:
(char error rate, word recall error rate, sequence error rate) as percent.
Raises:
ValueError: If unimplemented feature is used.
"""
decode = None
ocr_result = ''
if decoder_file:
decode = decoder.Decoder(decoder_file)
# Run inference
with tf.Graph().as_default():
model = InitNetwork(infer_data, model_str, 'eval', reader=reader)
sess = tf.Session('')
if graph_def_file is not None:
# Write the eval version of the graph to a file for freezing.
if not tf.gfile.Exists(graph_def_file):
with tf.gfile.FastGFile(graph_def_file, 'w') as f:
f.write(
sess.graph.as_graph_def(
add_shapes=True).SerializeToString())
ckpt = tf.train.get_checkpoint_state(train_dir)
if ckpt and ckpt.model_checkpoint_path:
step = model.Restore(ckpt.model_checkpoint_path, sess)
if decode:
ocr_result = decode.SoftmaxInfer(sess, model, num_lines)
else:
raise ValueError(
'Non-softmax decoder evaluation not implemented!')
return ocr_result
def model_init(self):
if self.device == torch.device('cpu'):
if not os.path.exists(self.epath):
#input_size,embedding_size,hidden_size
self.encode_network = encoder.Encoder(self.cp.S_len,
self.embedding_size,
self.hidden_dim)
if os.path.exists(self.epath):
self.encode_network = torch.load(self.epath,
map_location='cpu')
if not os.path.exists(self.dpath):
#input_dim,embedding_dim,hidden_dim
self.decode_network = decoder.Decoder(self.cp.T_len,
self.embedding_size,
self.hidden_dim)
if os.path.exists(self.dpath):
self.decode_network = torch.load(self.dpath,
map_location='cpu')
else:
if not os.path.exists(self.epath):
#input_size,embedding_size,hidden_size
self.encode_network = encoder.Encoder(
self.cp.S_len, self.embedding_size,
self.hidden_dim).to(self.device)
if os.path.exists(self.epath):
self.encode_network = torch.load(self.epath).to(self.device)
if not os.path.exists(self.dpath):
#input_dim,embedding_dim,hidden_dim
self.decode_network = decoder.Decoder(
self.cp.T_len, self.embedding_size,
self.hidden_dim).to(self.device)
if os.path.exists(self.dpath):
self.decode_network = torch.load(self.dpath).to(self.device)
def create_train_model(hparams):
# get src/tgt vocabulary table
graph = tf.Graph()
with graph.as_default() as graph:
src_vocab_table, tgt_vocab_table = vocab_table_util.get_vocab_table(
hparams.src_vocab_file, hparams.tgt_vocab_file)
reversed_tgt_vocab_table = lookup_ops.index_to_string_table_from_file(
hparams.tgt_vocab_file, default_value=vocab_table_util.UNK)
reversed_src_vocab_table = lookup_ops.index_to_string_table_from_file(
hparams.src_vocab_file, default_value=vocab_table_util.UNK)
with tf.variable_scope("NMTModel",
initializer=tf.truncated_normal_initializer(
stddev=0.01)) as nmtmodel_scope:
with tf.variable_scope("train_iterator"):
src_dataset_file = "%s.%s" % (hparams.train_prefix,
hparams.src)
tgt_dataset_file = "%s.%s" % (hparams.train_prefix,
hparams.tgt)
iterator = iterator_utils.get_nmt_iterator(
src_dataset_file, tgt_dataset_file, src_vocab_table,
tgt_vocab_table, hparams.batch_size, hparams.eos,
hparams.sos, hparams.source_reverse, hparams.random_seed)
with tf.variable_scope("shared_encoder") as encoder_scope:
encoder = en.Encoder(hparams,
tf.contrib.learn.ModeKeys.TRAIN,
dtype=tf.float32,
scope=encoder_scope)
with tf.variable_scope("shared_decoder") as decoder_scope:
decoder = de.Decoder(hparams,
tf.contrib.learn.ModeKeys.TRAIN,
dtype=tf.float32,
scope=decoder_scope)
nmt_model = mdl.NMTModel(hparams, src_vocab_table, tgt_vocab_table,
encoder, decoder, iterator,
tf.contrib.learn.ModeKeys.TRAIN,
reversed_tgt_vocab_table,
reversed_src_vocab_table)
saver = tf.train.Saver(tf.global_variables())
return TrainModel(graph=graph,
model=nmt_model,
encoder=encoder,
decoder=decoder,
iterator=iterator,
saver=saver)
def __init__(self, dataset_params, encoder_params, channel_params,
decoder_params, decision_params):
self.X = tf.compat.v1.placeholder(tf.float32,
shape=(None, 13,
dataset_params['n_features']),
name="X")
self.Noise = tf.compat.v1.placeholder(tf.float32,
shape=(),
name="Noise")
# encoder -----Transmitter Model
self._encoder = enc.Encoder(inputs=self.X, **encoder_params)
self._z = self._encoder.get_latent_representation(
) # shape (?, 1, 104)
# Stochastic Channel Model
#print("self._z", self._z.shape)
self._channel = sc.Channel(inputs=self._z, **channel_params)
self._channelout = self._channel.get_ChannelOuput()
# AWGN noise layer -----Channel Model(Part)
w = noi.gaussian_noise_layer(input_layer=self._channelout,
std=self.Noise)
#print("w shape is: ", w.shape)
w = tf.reshape(w, [-1, 476])
#print("w before complex is: ", w.shape)
# convert to complex number and then slice
w = RToC.get_c(w)
#print("w after complex is: ", w.shape)
# Slicer
self.slice_output = tf.slice(w, [0, k1], [-1, k2 - k1 + 1])
#print("first slice is: ", self.slice_output.shape)
# conver it back to real number
self.slice_output = CToR.get_r(self.slice_output)
#print("self.slice_output shape is: ", self.slice_output.shape)
# decoder layer -input shape is (batch_size, 352)--Real
self._decoder = dec.Decoder(inputs=self.slice_output, **decoder_params)
u = self._decoder.get_outputs()
# decision -----Receiver Model
self._decision = decis.Decision(inputs=u, **decision_params)
self._Xhat = self._decision.get_decision()
def main(_):
# create foler for generated images
if not os.path.exists(FLAGS.out_dir):
os.makedirs(FLAGS.out_dir)
# Import data
mnist = input_data.read_data_sets(FLAGS.data_dir, one_hot=True)
# data placeholders
X = tf.placeholder(tf.float32, shape=[None, 784])
# initializes encoder and decoder
_encoder = encoder.Encoder(pkeep=0.75)
_decoder = decoder.Decoder(pkeep=0.75)
# encodes and decodes
X_fake = _decoder.decode(_encoder.encode(X))
# loss : quadratic error
loss = tf.reduce_mean(tf.pow(X - X_fake, 2))
# invoke the optimizer
solver = tf.train.AdamOptimizer(learning_rate).minimize(loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
nb_batch = int(mnist.train.num_examples / batch_size)
# train
for it in range(training_iter):
for i in range(nb_batch):
# next input batch
X_mb, _ = mnist.train.next_batch(batch_size)
# run session
_, loss_curr = sess.run([solver, loss], feed_dict={X: X_mb})
if it % display_step == 0:
print('Iter: {}'.format(it))
print('cost: {:.4}'.format(loss_curr))
# test on 10 test images
test = sess.run(X_fake, feed_dict={X: mnist.test.images[:10]})
save(test, mnist.test.images)
def create_infer_model(hparams):
graph = tf.Graph()
with graph.as_default() as graph:
src_vocab_table, tgt_vocab_table = vocab_table_util.get_vocab_table(
hparams.src_vocab_file, hparams.tgt_vocab_file)
reversed_tgt_vocab_table = lookup_ops.index_to_string_table_from_file(
hparams.tgt_vocab_file, default_value=vocab_table_util.UNK)
reversed_src_vocab_table = lookup_ops.index_to_string_table_from_file(
hparams.src_vocab_file, default_value=vocab_table_util.UNK)
with tf.variable_scope("NMTModel") as nmtmodel_scope:
with tf.variable_scope("infer_iterator"):
src_dataset_file = "%s.%s" % (hparams.dev_prefix, hparams.src)
iterator = iterator_utils.get_nmt_infer_iterator(
src_dataset_file, src_vocab_table, hparams.batch_size,
hparams.source_reverse, hparams.eos)
with tf.variable_scope("shared_encoder") as encoder_scope:
encoder = en.Encoder(hparams,
tf.contrib.learn.ModeKeys.INFER,
dtype=tf.float32,
scope=encoder_scope)
with tf.variable_scope("shared_decoder") as decoder_scope:
decoder = de.Decoder(hparams,
tf.contrib.learn.ModeKeys.INFER,
dtype=tf.float32,
scope=decoder_scope)
nmt_model = mdl.NMTModel(
hparams,
src_vocab_table,
tgt_vocab_table,
encoder,
decoder,
iterator,
tf.contrib.learn.ModeKeys.INFER,
reversed_tgt_vocab_table=reversed_tgt_vocab_table,
reversed_src_vocab_table=reversed_src_vocab_table)
saver = tf.train.Saver(tf.global_variables())
return InferModel(graph=graph,
model=nmt_model,
encoder=encoder,
decoder=decoder,
iterator=iterator,
saver=saver)
def __init__(self, latent_spaces, batch_size):
super(ANVAE, self).__init__()
self.batch_size = batch_size
self.latent_spaces = 3
self.level_sizes = [1, 1, 1]
self.input_s = [32, 32, 1]
self.latent_channels = 20
self.encoder = encoder.Encoder(self.latent_spaces, self.input_s)
self.decoder = decoder.Decoder(self.encoder(
tf.zeros([self.batch_size, 32, 32, 1]), False),
latent_channels=self.latent_channels,
level_sizes=self.level_sizes)
inputs, disc_l, outputs = self.disc_function()
self.discriminator = tf.keras.Model(inputs=[inputs],
outputs=[outputs, disc_l])
self.lr_ae = .0001
self.lr_disc = .0001
self.recon_loss_div = 1
self.latent_loss_div = 1
self.sig_mult = 10
self.enc_optimizer = tf.keras.optimizers.Adam(self.lr_ae,
0.5,
epsilon=.1,
clipvalue=0.5,
clipnorm=1)
self.dec_optimizer = tf.keras.optimizers.Adam(self.lr_ae,
0.5,
clipvalue=0.5,
clipnorm=1)
self.disc_optimizer = tf.keras.optimizers.Adam(self.lr_disc,
0.5,
clipvalue=0.5,
clipnorm=1)
self.lastEncVars = []
self.lastDecVars = []
self.lastDiscVars = []
self.debugCount = 0
def evaluate(FSAngle, FSVelocity, road_matrix, memory):
"""
Runs a single simulation, movement params are calculated based on the fuzzy systems FSAngle and FSVelocity
"""
car = vehicle.Car(constants.CAR_POS_X, constants.CAR_POS_Y,
constants.CAR_ANGLE)
iteration = 0
past_pos = car.center_position()
dec = decoder.Decoder(FSAngle, FSVelocity, car)
dt = TIME_STEP
total_distance = 0
punishment = 0
left_right = 0
while iteration <= MAX_ITERATIONS:
car.left_sensor_input, car.front_sensor_input, car.right_sensor_input = get_sensors(
car, road_matrix, memory)
ds, drot = dec.get_movement_params()
car.update(dt, ds, drot)
iteration += 1
total_distance += ds
left_right += abs(
float(car.left_sensor_input) - float(car.right_sensor_input))
if iteration % 100 == 0:
past_x, past_y = past_pos
curr_x, curr_y = car.center_position()
if vehicle.distance(past_x, past_y, curr_x, curr_y) < MIN_DISTANCE:
break
else:
past_pos = car.center_position()
if car.is_idle(iteration) or car.is_collided2(road_matrix):
punishment = 150
break
return left_right / iteration + punishment