def second_pass(self, file_lines):
memory_address = self.MEM_START_ADDR
for line in file_lines:
parser = Parser(instruction=line)
encoder = Encoder(instruction_type=parser.instruction_type)
if parser.instruction_type == InstructionType.c_instruction:
hack_line = encoder.encode(dest=parser.dest,
comp=parser.comp,
jump=parser.jump)
elif parser.instruction_type == InstructionType.a_instruction:
try:
integer_address = int(parser.address)
except ValueError:
if self.symbol_table.get(parser.address) is None:
self.symbol_table[parser.address] = memory_address
memory_address += 1
integer_address = self.symbol_table.get(parser.address)
hack_line = encoder.encode(address=integer_address)
else:
continue
self.hack_file.write(hack_line + '\r\n')
def _build(self, Log, Scorer):
if Log == None:
Log = Logger()
if Scorer == None:
Scorer = Score()
self.Log = Log
self.Scorer = Scorer
self.Encoder = Encoder()
self.GPT = GPT2LanguageModel(model_name=self.model)
def _build(self, mod, Log):
''' Builds application using variables provided by user '''
if Log == None:
Log = Log()
if (self._seed < 1):
random.seed(time.time())
self._seed = random.random()
self.Log = Log
self.Scorer = Score(mod, self.Log)
self.Encoder = Encoder(seed=self._seed, probability=self._probability)
self.GPT = GPT2LanguageModel(model_name=self.model)
def __init__(self, params, embedding_matrix):
super(RVAE_dilated, self).__init__()
self.params = params
self.word_embeddings = nn.Embedding(params.word_vocab_size, params.word_embed_size)
self.word_embeddings.weight = Parameter(t.from_numpy(embedding_matrix).float(),
requires_grad=False)
self.encoder = Encoder(self.params)
self.context_to_mu = nn.Linear(self.params.encoder_rnn_size * 2, self.params.latent_variable_size)
self.context_to_logvar = nn.Linear(self.params.encoder_rnn_size * 2, self.params.latent_variable_size)
self.decoder = Decoder(self.params)
def load_model_from_package(cls, package):
encoder = Encoder(package['d_input'],
package['n_layers_enc'],
package['n_head'],
package['d_k'],
package['d_v'],
package['d_model'],
package['d_inner'],
dropout=package['dropout'],
pe_maxlen=package['pe_maxlen'])
decoder = Decoder(
package['sos_id'],
package['eos_id'],
package['vocab_size'],
package['d_word_vec'],
package['n_layers_dec'],
package['n_head'],
package['d_k'],
package['d_v'],
package['d_model'],
package['d_inner'],
dropout=package['dropout'],
tgt_emb_prj_weight_sharing=package['tgt_emb_prj_weight_sharing'],
pe_maxlen=package['pe_maxlen'],
)
model = cls(encoder, decoder)
model.load_state_dict(package['state_dict'])
LFR_m, LFR_n = package['LFR_m'], package['LFR_n']
return model, LFR_m, LFR_n
def main():
# Load a dictionary of Michael's quotes to their season and episode
print("Attempting to load quotes from file")
quotes = load_quotes()
if quotes is None:
print("Scraping the web for new quotes")
quotes = scrape()
print("Creating sentence encoder")
encoder = Encoder()
print("Attempting to load quote embeddings from file")
quote_embeddings = load_quote_embeddings()
if quote_embeddings is None:
print("Generating new quote embeddings")
quote_embeddings = generate_quote_embeddings(encoder, quotes)
print("Saving new quote embeddings to {0}".format(embeddings_file))
save_pickle(quote_embeddings, embeddings_file)
print("Creating predictor")
predictor = Predictor(encoder, quote_embeddings)
while True:
input_sentence = query_input()
prediction = predictor.predict_output(input_sentence)
output_quote = prediction[0]
output_season = prediction[1]['season']
output_episode = prediction[1]['episode']
print("Michael says: \"{0}\" in season {1}, episode {2}".format(
output_quote, output_season, output_episode))
def make_model(src_vocab,
tgt_vocab,
N=6,
d_model=512,
d_ff=2048,
h=8,
dropout=0.1):
"Helper: Construct a model from hyperparameters."
c = copy.deepcopy
attn = MultiHeadedAttention(h, d_model)
ff = PositionwiseFeedForward(d_model, d_ff, dropout)
position = PositionalEncoding(d_model, dropout)
model = EncoderDecoder(
Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N),
Decoder(DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout), N),
nn.Sequential(Embeddings(d_model, src_vocab), c(position)),
nn.Sequential(Embeddings(d_model, tgt_vocab), c(position)),
Generator(d_model, tgt_vocab))
# This was important from their code.
# Initialize parameters with Glorot / fan_avg.
for p in model.parameters():
if p.dim() > 1:
nn.init.xavier_uniform(p)
return model
def train_model(dpath, ppath, epoch, version):
if dpath.endswith(".csv"):
d = pd.read_csv(dpath)
else:
raise ValueError("data format is not supported")
pipe = joblib.load(ppath)
encoder = Encoder(pipe)
x = encoder.encode(d.iloc[:, 1:-1])
m = create_model(
[
x.shape[1],
]
)
m.fit(x, d.iloc[:, -1], batch_size=1000, epochs=epoch)
m.save(f"model/{version}")
def run_ensemble(train_df):
encoder = None
encoder = Encoder(train_df)
encoder.transform(train_df)
estimators = []
scores = []
labels = []
nums = list(range(1, 5, 1)) + list(range(5, 60, 5)) + list(
range(60, 100, 10)) + list(range(100, 500, 50))
for n in nums:
lr = modelDict["GBM"](n_estimators=n)
n_train_df = pd.get_dummies(train_df)
train_score, val_score = lr.train(n_train_df)
scores += [train_score, val_score]
estimators += [n, n]
labels += ['train', 'val']
return scores, labels, estimators
def run_tests():
e = Encoder()
d = Decoder(coding_polynomial=e.coding_polynomial,
k=e.k,
t=e.r,
gf_index=e.gf.index)
message = "zaqwsxcderfvbgtyhnmjuik,ol.p;/zaqwsxedcrfvtgbyhnujmzaqwsxcderf"
codeword = e.encode(message)
decoded_message = d.decode(codeword, 'basic')
for i in range(2, 28):
codeword.elements[i] = codeword.elements[i].multiplicative_inversion()
print('27 errors occurred...')
print(codeword)
decoded_message = d.decode(codeword, 'basic')
print('Decoded message: ' + decoded_message[:len(message)])
def test_integration(model_service, xy):
p = Path(curdir / "saved_model")
assert "http://127.0.0.1:8501" == model_service
fname = curdir / ".tmp.joblib"
train_sk_pipe(fname, xy[0])
assert os.path.exists(fname)
pipe = load(fname)
encoder = Encoder(pipe)
matrixs = encoder.encode(xy[0][:100]).tolist()
res = requests.post(
model_service + "/v1/models/tp_pred:predict",
data=json.dumps({"instances": matrixs}),
)
assert len(res.json()["predictions"]) == 100
def run_example_program():
e = Encoder()
d = Decoder(coding_polynomial=e.coding_polynomial,
k=e.k,
t=e.r,
gf_index=e.gf.index)
message = "zaqwsxcderfvbgtyhnmjuik,ol.p;/zaqwsxedcrf"
print('Message: ' + message)
codeword = e.encode(message)
print('Codeword: ' + str(codeword))
decoded_message = d.decode(codeword, 'basic')
print('Decoded message: ' + decoded_message[:len(message)])
for i in range(1, 28):
codeword.elements[i] = codeword.elements[i].multiplicative_inversion()
print('27 errors occurred...')
print(codeword)
decoded_message = d.decode(codeword, 'basic')
print('Decoded message: ' + decoded_message[:len(message)])
def test_encoder_fix_errors(ii, k, test_type, message):
e = Encoder()
d = Decoder(coding_polynomial=e.coding_polynomial, k=e.k, t=e.r, gf_index=e.gf.index)
encoded_message = e.encode(message)
if test_type == 'multiple':
random_indexes = random.sample(range(0, len(encoded_message)), k)
else:
random_start = random.randint(0, len(encoded_message)-k-1)
random_indexes = [i for i in range(random_start, random_start + k)]
# print("{}): {}".format(k, random_indexes))
for i in random_indexes:
encoded_message.elements[i] = encoded_message.elements[i].multiplicative_inversion()
try:
start = time.time()
decoded_message = d.decode(encoded_message, 'basic')
stop = time.time()
passed.write("{}, {}, {}, {}, {}\n".format(k, test_type, message, random_indexes, stop-start))
except CannotDetectErrorException as c:
failed.write("{}, {}, {}, {}\n".format(k, test_type, message, random_indexes))
assert False
assert message in decoded_message
def run(train_df, test_df):
encoder = None
encoder = Encoder(train_df)
lr = modelDict["GBM"](need_scale=False)
encoder.transform(train_df)
n_train_df = pd.get_dummies(train_df)
lr.train(n_train_df)
encoder.transform(test_df)
n_test_df = pd.get_dummies(test_df)
y = lr.test(n_test_df)
save(test_df, y, encoder)
def __init__(self):
self.leftFEncoder = Encoder()
self.leftMEncoder = Encoder()
self.leftREncoder = Encoder()
self.rightFEncoder = Encoder()
self.rightMEncoder = Encoder()
self.rightREncoder = Encoder()
self.pose = Pose()
self.lastTime = 0
class odometry_node:
def __init__(self):
self.leftFEncoder = Encoder()
self.leftMEncoder = Encoder()
self.leftREncoder = Encoder()
self.rightFEncoder = Encoder()
self.rightMEncoder = Encoder()
self.rightREncoder = Encoder()
self.pose = Pose()
self.lastTime = 0
def setTime(self, newTime):
self.lastTime = newTime
def main(self):
self.odomPub = rospy.Publisher('odometry/filtered',
Odometry,
queue_size=10)
self.tfPub = TransformBroadcaster()
rospy.init_node('odometry_node')
self.nodeName = rospy.get_name()
rospy.loginfo("{0} started".format(self.nodeName))
self.ticksPerMeter = int(rospy.get_param('~ticks_per_meter', 780))
self.wheelSeparation = float(rospy.get_param('~wheel_separation', 0.7))
self.rate = float(rospy.get_param('~rate', 10.0))
self.baseFrameID = rospy.get_param('~base_frame_id', 'base_footprint')
self.odomFrameID = rospy.get_param('~odom_frame_id', 'odom')
self.encoderMin = int(rospy.get_param('~encoder_min', -32768))
self.encoderMax = int(rospy.get_param('~encoder_max', 32767))
self.setTime(rospy.get_time())
rate = rospy.Rate(self.rate)
rospy.Subscriber("encoder_counts",
EncoderCounts,
callback=self.callback)
rospy.spin()
def callback(self, data):
self.calculate_pose(data)
def calculate_pose(self, data):
lc = data.left_wheel_counts #front to back
rc = data.right_wheel_counts #same here yo
#update left encoders
self.leftFEncoder.update(lc[0])
self.leftMEncoder.update(lc[1])
self.leftREncoder.update(lc[2])
#update right encoders
self.rightFEncoder.update(rc[0])
self.rightMEncoder.update(rc[1])
self.rightREncoder.update(rc[2])
#get Travels
leftFTravel = self.leftFEncoder.getDelta() / self.ticksPerMeter
leftMTravel = self.leftMEncoder.getDelta() / self.ticksPerMeter
leftRTravel = self.leftREncoder.getDelta() / self.ticksPerMeter
rightFTravel = (self.rightFEncoder.getDelta() / self.ticksPerMeter)
rightMTravel = (self.rightMEncoder.getDelta() / self.ticksPerMeter)
rightRTravel = (self.rightREncoder.getDelta() / self.ticksPerMeter)
rospy.loginfo(leftFTravel)
rospy.loginfo(rightFTravel)
#time stuff
newTime = rospy.get_time()
deltaTime = newTime - self.lastTime
self.setTime(newTime)
#middle travel
aveLT = (leftFTravel + leftMTravel + leftRTravel) / 3
aveRT = (rightFTravel + rightMTravel + rightRTravel) / 3
rospy.loginfo(aveLT)
rospy.loginfo(aveRT)
midTravel = (aveRT + (-1 * (aveLT))) / 2
midTheta = (aveRT - (-1 * (aveLT))) / self.wheelSeparation
if -1 * aveRT == aveLT:
deltaX = aveLT * cos(self.pose.theta)
deltaY = aveLT * sin(self.pose.theta)
else:
radius = midTravel / midTheta
# Find the instantaneous center of curvature (ICC) offset.
iccDx = radius * sin(self.pose.theta)
iccDy = -radius * cos(self.pose.theta)
deltaX = cos(midTheta) * iccDx + sin(midTheta) * iccDy - iccDx
deltaY = sin(midTheta) * iccDx + cos(midTheta) * iccDy - iccDy
#set poses
self.pose.x += deltaX
self.pose.y += deltaY
self.pose.theta = (self.pose.theta + midTheta) % (2 * pi)
self.pose.xVel = midTravel / deltaTime if deltaTime > 0 else 0.
self.pose.yVel = 0
self.pose.thetaVel = midTheta / deltaTime if deltaTime > 0 else 0.
self.lastTime = newTime
now = rospy.get_rostime()
q = Quaternion()
q.x = 0
q.y = 0
q.z = sin(self.pose.theta / 2)
q.w = cos(self.pose.theta / 2)
self.tfPub.sendTransform((self.pose.x, self.pose.y, 0),
(q.x, q.y, q.z, q.w), now, self.baseFrameID,
self.odomFrameID)
odom = Odometry()
odom.header.stamp = now
odom.header.frame_id = self.odomFrameID
odom.child_frame_id = self.baseFrameID
odom.pose.pose.position.x = self.pose.x
odom.pose.pose.position.y = self.pose.y
odom.pose.pose.position.z = 0
odom.pose.pose.orientation = q
odom.pose.covariance[0] = 1e-3
odom.pose.covariance[7] = 1e-3
odom.pose.covariance[14] = 1e-3
odom.pose.covariance[21] = 100000
odom.pose.covariance[28] = 100000
odom.pose.covariance[35] = 1e-3
odom.twist.twist.linear.x = self.pose.xVel
odom.twist.twist.linear.y = 0
odom.twist.twist.angular.z = self.pose.thetaVel
odom.twist.covariance[0] = 1e-3
odom.twist.covariance[7] = 1e-3
odom.twist.covariance[14] = 100000
odom.twist.covariance[21] = 100000
odom.twist.covariance[28] = 100000
odom.twist.covariance[35] = 1e-3
self.odomPub.publish(odom)
def __getitem__(self, item):
return getattr(self, item)
def test_minimal(self):
num_words, num_tags, num_chars = 10, 10, 100
encoder = Encoder(num_words, num_tags, num_chars=num_chars)
assert encoder.num_tags == num_tags
assert isinstance(encoder.word_embedding, nn.Embedding)
def encoder():
e = Encoder()
return e
def main(args):
# load dictionary and generate char_list, sos_id, eos_id
char_list, sos_id, eos_id = process_dict(args.dict)
vocab_size = len(char_list)
tr_dataset = AudioDataset('train', args.batch_size)
cv_dataset = AudioDataset('dev', args.batch_size)
tr_loader = AudioDataLoader(tr_dataset,
batch_size=1,
num_workers=args.num_workers,
shuffle=args.shuffle,
feature_dim=args.feature_dim,
char_list=char_list,
path_list=tr_dataset.path_lst,
label_list=tr_dataset.han_lst,
LFR_m=args.LFR_m,
LFR_n=args.LFR_n)
cv_loader = AudioDataLoader(cv_dataset,
batch_size=1,
num_workers=args.num_workers,
feature_dim=args.feature_dim,
char_list=char_list,
path_list=cv_dataset.path_lst,
label_list=cv_dataset.han_lst,
LFR_m=args.LFR_m,
LFR_n=args.LFR_n)
data = {'tr_loader': tr_loader, 'cv_loader': cv_loader}
encoder = Encoder(args.d_input * args.LFR_m,
args.d_low_dim,
args.n_layers_enc,
args.n_head,
args.d_k,
args.d_v,
args.d_model,
args.d_inner,
dropout=args.dropout,
pe_maxlen=args.pe_maxlen)
decoder = Decoder(
sos_id,
eos_id,
vocab_size,
args.d_word_vec,
args.n_layers_dec,
args.n_head,
args.d_k,
args.d_v,
args.d_model,
args.d_inner,
dropout=args.dropout,
tgt_emb_prj_weight_sharing=args.tgt_emb_prj_weight_sharing,
pe_maxlen=args.pe_maxlen)
model = Transformer(encoder, decoder)
print(model)
model.cuda()
# optimizer
optimizier = TransformerOptimizer(
torch.optim.Adam(model.parameters(), betas=(0.9, 0.98), eps=1e-09),
args.init_lr, args.d_model, args.warmup_steps)
# solver
solver = Solver(data, model, optimizier, args)
solver.train()
word_map['<unk>'] = len(word_map) + 1
word_map['<start>'] = len(word_map) + 1
word_map['<end>'] = len(word_map) + 1
word_map['<pad>'] = 0
checkpoint = '../input/image-copy-2/checkpoint_copy.pt'
decoder = DecoderWithAttention(embed_dim=emb_dim,
decoder_dim=decoder_dim,
vocab_size=len(word_map),
dropout=dropout)
decoder_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, decoder.parameters()),
lr=decoder_lr)
encoder = Encoder()
# Move to GPU, if available
decoder = decoder.to(device)
encoder = encoder.to(device)
decoder.eval()
encoder.eval()
from scipy.misc import imread, imresize
if checkpoint is not None:
checkpoint = torch.load(checkpoint)
decoder.load_state_dict(checkpoint['decoder_state_dict'])
decoder_optimizer.load_state_dict(checkpoint['decoder_optimizer_dict'])
from termcolor import colored
from src.encoder import Encoder
from src.universal_function import universalFunction
def present():
print(colored('For following instructions:', 'green'))
print(open('instructions.txt', "r").read())
print()
print(colored('We have this LL(1) grammar:', 'green'))
print(open('grammar.txt', "r").read())
print()
print(colored('Parse table of this grammar is:', 'green'))
print(open('parse-table.txt', "r").read())
if __name__ == "__main__":
present()
# filePath = 'data/in/book-example.txt'
filePath = 'data/in/test.txt'
fileString = open(filePath, "r").read()
print(colored('input file:', 'green'))
encoder = Encoder(fileString)
instructions = encoder.encodeLines()
programCode = encoder.calcuateProgramCode()
inputValues = encoder.getInputVaules()
input_for_universal_program = inputValues + [programCode]
universalFunction(input_for_universal_program, instructions)
class pyReadability():
'''
Main class for the applicaiton. Loads the data, encodes it and runs scoring algortim.
'''
def __init__(self,
model,
interact,
topK,
seed,
mod,
probability,
Log=None):
''' Take in variables and starts the program. See readme for inputs'''
self.model = model
self.topK = topK
self.interact = interact
self._seed = seed
self._probability = probability
self._score = -1
self._build(mod, Log)
def _build(self, mod, Log):
''' Builds application using variables provided by user '''
if Log == None:
Log = Log()
if (self._seed < 1):
random.seed(time.time())
self._seed = random.random()
self.Log = Log
self.Scorer = Score(mod, self.Log)
self.Encoder = Encoder(seed=self._seed, probability=self._probability)
self.GPT = GPT2LanguageModel(model_name=self.model)
def _run(self, text):
''' Runs GTP and gets results '''
logits = self.GPT.predict(text, "")
probabilities = torch.nn.functional.softmax(logits)
best_indices = logits.topk(self.topK)[1]
self.best_words = [self.GPT[idx.item()] for idx in best_indices]
self.best_probabilities = probabilities[best_indices].tolist()
def _getWords(self):
''' returns Top-K Words from GPT-2 '''
return self.best_words
def _getPropability(self):
''' returns top-k Propabilities from GPT-2 '''
return [round(p * 100, 2) for p in self.best_probabilities]
def _process(self, text, guess):
''' scores inputted text and logs it '''
self._run(text)
outputLst = self._output()
self.Log.Trace(("Answer List : {}".format(outputLst)))
score = self.Scorer.score(outputLst, guess)
self.Log.Trace(score)
self.Log.Info("Score of \'{}\': {}".format(score[0], score[1]))
def start(self, text=""):
'''
starts program
text = Text to be inputted
'''
if text == "" and not self.interact:
raise EnvironmentError(
"Please input valid text or use the --interact flag")
if text != "":
encoded = self.Encoder.encode(text=text)
for item in encoded:
if item[0] == '':
continue
self._process(item[0], item[1])
# Code for Manual Input, meant for debugging not for production use
else:
while self.interact:
text = self.Log.Input("Input Text >> ")
if text == "":
self.Log.Info("Please provide a valid input")
continue
if text == "#?":
self.Log.Info(
"Available Commands: \n#?: Shows available commands\n#end: Ends Execution"
)
continue
if text == "#end":
self.Log.Info("Ending Program")
break
guess = self.Log.Input("What will the next word be >> ")
self._process(text, guess)
self._score = self.Scorer.calcScore()
self.Log.Info("Normalized Score: {} | Unnormalized Score: {}".format(
self.getNormScore(), self.getUnNormScore()))
def getNormScore(self):
''' returns the normalized score '''
return self._score[0]
def getUnNormScore(self):
''' returns the unormalized score '''
return self._score[1]
def getSeed(self):
''' returns the seed used '''
return self._seed
def getEncoder(self):
''' returns the encoder object '''
return self.Encoder
def _output(self):
''' returns top-k words and propabilities '''
return [(self._getWords()[i], self._getPropability()[i])
for i in range(self.topK)]
class Guesser():
def __init__(self,
model="gpt2",
interact=False,
score=False,
topK=10,
Log=None,
Scorer=None):
self.model = model
self.topK = topK
self._build(Log, Scorer)
def _build(self, Log, Scorer):
if Log == None:
Log = Logger()
if Scorer == None:
Scorer = Score()
self.Log = Log
self.Scorer = Scorer
self.Encoder = Encoder()
self.GPT = GPT2LanguageModel(model_name=self.model)
def _getBestWords(self, text):
''' Creates finds best words and calculates their propbablity of occuring '''
logits = self.GPT.predict(text, "")
best_logits, best_indices = logits.topk(self.topK)
# converts best indicie list into a list of words
best_words = [self.GPT[idx.item()] for idx in best_indices]
# calculates probabilities
probabilities = torch.nn.functional.softmax(logits)
# creates a list of probabilites based on best_indicies. This is a parallel array to best_words
best_probabilities = self._getPropability(
probabilities[best_indices].tolist())
# returns a list of tuples. each tuple contains the world in position 0 and the probability in position 1
return [(best_words[i], best_probabilities[i])
for i in range(self.topK)]
def _getPropability(self, probabilities):
''' returns top-k Propabilities from GPT-2 '''
return [round(p * 100, 2) for p in probabilities]
def _run(self, text, guess):
''' scores inputted text and logs it '''
# gives a list of the best words that GPT predicts
guessList = self._getBestWords(text)
self.Log.Info(("Answer List : {}".format(guessList)))
# scores guess word with the list of preditions
score = self.Scorer.score(guessList, guess)
self.Log.Info(score)
def start(self, text=""):
'''
Starts The Program
text: Text to be fed into GPT. If left blank initiate interactive mode
'''
if text != "":
encoded_text = self.Encoder.encode(text=text)
for text, guess in encoded_text[0]:
if text == '':
continue
self._run(text, guess)
else:
while True:
text = self.Log.Input("Input Text >> ")
if text == "":
self.Log.Info("Please provide a valid input")
continue
if text == "#?":
self.Log.Info(
"Available Commands: \n#?: Shows available commands\n#end: Ends Execution"
)
continue
if text == "#end":
self.Log.Info("Ending Program")
break
guess = self.Log.Input("What will the next word be >> ")
self._run(text, guess)
self.Log.Info("Score: {}".format(self.Scorer.calcScore()))
def save(data: pd.DataFrame, y, encoder: Encoder, save_path=out_path):
data['charges'] = y
if encoder: encoder.rev_transform(data)
data.to_csv(save_path, sep=',', index=False)
class RVAE_dilated(nn.Module):
def __init__(self, params, embedding_matrix):
super(RVAE_dilated, self).__init__()
self.params = params
self.word_embeddings = nn.Embedding(params.word_vocab_size, params.word_embed_size)
self.word_embeddings.weight = Parameter(t.from_numpy(embedding_matrix).float(),
requires_grad=False)
self.encoder = Encoder(self.params)
self.context_to_mu = nn.Linear(self.params.encoder_rnn_size * 2, self.params.latent_variable_size)
self.context_to_logvar = nn.Linear(self.params.encoder_rnn_size * 2, self.params.latent_variable_size)
self.decoder = Decoder(self.params)
def forward(self, drop_prob,
encoder_input_tuple=None,
use_cuda=False,
z=None,
inference=False):
"""
:param encoder_input_tuple: An tensor with shape of [batch_size, seq_len] of Long type
:param drop_prob: probability of an element of decoder input to be zeroed in sense of dropout
:param use_cuda: whether to use gpu
:param z: context if sampling is performing
:return: unnormalized logits of sentence words distribution probabilities
with shape of [batch_size, seq_len, word_vocab_size]
kld loss estimation
"""
encoder_input, lengths = encoder_input_tuple
batch_size = encoder_input.size()[0]
encoder_input = Variable(encoder_input, volatile=inference).cuda() if use_cuda else Variable(encoder_input, volatile=inference)
encoder_input_matrix = self.word_embeddings(encoder_input)
if z is None:
''' Get context from encoder and sample z ~ N(mu, std)
'''
packed_seq = pack_padded_sequence(encoder_input_matrix, lengths, batch_first=True)
context = self.encoder.forward(packed_seq, batch_size)
mu = self.context_to_mu(context)
logvar = self.context_to_logvar(context)
std = t.exp(0.5 * logvar)
z = Variable(t.randn([batch_size, self.params.latent_variable_size]))
if use_cuda:
z = z.cuda()
z = z * std + mu
kld = (-0.5 * t.sum(logvar - t.pow(mu, 2) - t.exp(logvar) + 1, 1)).mean()
else:
kld = None
decoder_input = encoder_input_matrix
out = self.decoder.forward(decoder_input, z, drop_prob)
return encoder_input, out, kld
def learnable_parameters(self):
# word_embedding is constant parameter thus it must be dropped from list of parameters for optimizer
return [p for p in self.parameters() if p.requires_grad]
def trainer(self, optimizer, use_cuda, dropout):
def train_one_batch(data_tuple):
target, logits, kld = self.forward(drop_prob=dropout,
encoder_input_tuple=data_tuple,
use_cuda=use_cuda,
z=None)
batch_size = target.data.size()[0]
sequence_length = target.data.size()[1]
logits = logits.view(-1, self.params.word_vocab_size)
target = target.view(-1)
cross_entropy = F.cross_entropy(logits, target)
loss = sequence_length * cross_entropy + kld
optimizer.zero_grad()
loss.backward()
optimizer.step()
return kld, loss
return train_one_batch
def validater(self, use_cuda, dropout):
def validate(data_tuple):
target, logits, kld = self.forward(drop_prob=dropout,
encoder_input_tuple=data_tuple,
use_cuda=use_cuda,
z=None)
logits = logits.view(-1, self.params.word_vocab_size)
target = target.view(-1)
cross_entropy = F.cross_entropy(logits, target)
return np.exp2(cross_entropy.data.cpu().numpy()[0])
return validate
:return:
"""
parser = ArgumentParser()
parser.add_argument("path", help="path to .csv data file")
return parser.parse_args()
if __name__ == '__main__':
args = args()
DATA_PATH = args.path
seed = 42
np.random.seed(seed)
harvester = DataHarvester(DATA_PATH)
harvester.read_file()
harvester.cut_lines()
encoder = Encoder(harvester.read_data)
encoder.encode_data()
encoder.encode_label()
X = encoder.encoded
Y = encoder.encoded_label
model_builder = ModelBuilder(encoder.num_of_label_classes, encoder.num_of_data_classes)
estimator = KerasClassifier(build_fn=model_builder, epochs=20, batch_size=5, verbose=5)
kfold = KFold(n_splits=30, shuffle=True, random_state=seed)
results = cross_val_score(estimator, X, Y, cv=kfold)
print("Baseline: %.2f%% (%.2f%%)" % (results.mean() * 100, results.std() * 100))
def __init__(self, hidden_encoder_size, z_dim, hidden_decoder_size, output_size, rnn_type, device):
super(GrammarVAE, self).__init__()
self.encoder = Encoder(hidden_encoder_size, z_dim)
self.decoder = Decoder(z_dim, hidden_decoder_size, output_size, device, rnn_type)
self.device = device