def main(config_file, run_type='train', checkpoint=''):
# pylint: disable=no-member
config = Config(config_file)
print(config)
if run_type == 'train':
from trainer import Trainer
trainer = Trainer(dataset_dir=config.dataset_dir,
log_dir=config.log_dir,
generator_channels=config.generator_channels,
discriminator_channels=config.discriminator_channels,
nz=config.nz,
style_depth=config.style_depth,
lrs=config.lrs,
betas=config.betas,
eps=config.eps,
phase_iter=config.phase_iter,
batch_size=config.batch_size,
n_cpu=config.n_cpu,
opt_level=config.opt_level)
trainer.run(log_iter=config.log_iter, checkpoint=checkpoint)
elif run_type == 'inference':
from inferencer import Inferencer
inferencer = Inferencer(
generator_channels=config.generator_channels,
nz=config.nz,
style_depth=config.style_depth,
)
inferencer.inference(n=8)
else:
raise NotImplementedError
def _initialize(self, corpus, vocab, prior_tree, number_of_topics, alpha_alpha):
Inferencer._initialize(self, vocab, prior_tree, number_of_topics, alpha_alpha);
# initialize the documents, key by the document path, value by a list of non-stop and tokenized words, with duplication.
self._corpus = corpus;
self._parsed_corpus = self.parse_data();
# initialize the size of the collection, i.e., total number of documents.
self._number_of_documents = len(self._parsed_corpus);
'''
# initialize a D-by-K matrix gamma, valued at N_d/K
#self._gamma = numpy.zeros((self._number_of_documents, self._number_of_topics)) + self._alpha_alpha[numpy.newaxis, :] + 1.0 * self._number_of_types / self._number_of_topics;
# initialize a V-by-K matrix beta, valued at 1/V, subject to the sum over every row is 1
#self._eta = numpy.random.gamma(100., 1. / 100., (self._number_of_topics, self._number_of_types));
#self._E_log_eta = compute_dirichlet_expectation(self._eta);
'''
# initialize the size of the vocabulary, i.e. total number of distinct tokens.
# self._number_of_terms = len(self._type_to_index)
# initialize a D-by-K matrix gamma, valued at N_d/K
# self._gamma = numpy.zeros((self._number_of_documents, self._number_of_topics)) + self._alpha_alpha + 1.0 * self._number_of_paths / self._number_of_topics;
# self._gamma = numpy.tile(self._alpha_alpha + 1.0 * self._number_of_terms / self._number_of_topics, (self._number_of_documents, 1));
self._gamma = self._alpha_alpha + 2.0 * self._number_of_paths / self._number_of_topics * numpy.random.random((self._number_of_documents, self._number_of_topics));
# initialize a _E_log_beta variable, indexed by node, valued by a K-by-C matrix, where C stands for the number of children of that node
self._var_beta = numpy.random.gamma(100., 1. / 100., (self._number_of_topics, self._number_of_edges));
# for edge_index in self._index_to_edge:
# self._var_beta[:, [edge_index]] += numpy.sum(phi_sufficient_statistics[:, self._paths_through_edge[edge_index]], axis=1)[:, numpy.newaxis];
'''
def _initialize(self,
corpus,
vocab,
number_of_topics,
alpha_alpha,
alpha_beta,
alpha_eta=0,
alpha_sigma_square=1.0
):
Inferencer._initialize(self, vocab, number_of_topics, alpha_alpha, alpha_beta);
self._parsed_corpus, self._responses = self.parse_data(corpus);
# define the total number of document
self._number_of_documents = len(self._parsed_corpus);
# initialize a D-by-K matrix gamma, valued at N_d/K
self._gamma = numpy.zeros((self._number_of_documents, self._number_of_topics)) + self._alpha_alpha[numpy.newaxis, :] + 1.0 * self._number_of_types / self._number_of_topics;
# self._gamma = numpy.random.gamma(100., 1./100, (self._number_of_documents, self._number_of_topics))
# initialize a V-by-K matrix _eta, valued at 1/V, subject to the sum over every row is 1
self._beta = numpy.random.gamma(100., 1. / 100., (self._number_of_topics, self._number_of_types));
# self._beta /= numpy.sum(self._beta, 1)[:, numpy.newaxis]
# self._E_log_eta = compute_dirichlet_expectation(self._beta);
self._eta = numpy.zeros((1, self._number_of_topics)) + alpha_eta
self._sigma_square = alpha_sigma_square
def test(config):
_config_test(config)
de2idx, idx2de = load_de_vocab()
en2idx, idx2en = load_en_vocab()
model = ConvSeq2Seq(config)
graph_handler = GraphHandler(config)
inferencer = Inferencer(config, model)
sess = tf.Session()
graph_handler.initialize(sess)
global_step = 0
refs = []
hypotheses = []
with codecs.open(os.path.join(config.eval_dir, config.model_name), "w", "utf-8") as fout:
for i, batch in tqdm(enumerate(get_batch_for_test())):
preds = inferencer.run(sess, batch)
sources = batch['source']
targets = batch['target']
for source, target, pred in zip(sources, targets, preds):
got = " ".join(idx2en[idx] for idx in pred).split("</S>")[0].strip()
fout.write("- source: " + source +"\n")
fout.write("- expected: " + target + "\n")
fout.write("- got: " + got + "\n\n")
fout.flush()
ref = target.split()
hypothesis = got.split()
if len(ref) > 3 and len(hypothesis) > 3:
refs.append([ref])
hypotheses.append(hypothesis)
score = corpus_bleu(refs, hypotheses)
fout.write("Bleu Score = " + str(100*score))
def __init__(self,
hash_oov_words=False,
number_of_samples=10,
burn_in_sweeps=5):
Inferencer.__init__(self, hash_oov_words)
self._number_of_samples = number_of_samples
self._burn_in_sweeps = burn_in_sweeps
def __init__(self,
hash_oov_words=False,
maximum_gamma_update_iteration=50,
minimum_mean_change_threshold=1e-3):
Inferencer.__init__(self, hash_oov_words)
self._maximum_gamma_update_iteration = maximum_gamma_update_iteration
self._minimum_mean_change_threshold = minimum_mean_change_threshold
def __init__(self,
hyper_parameter_optimize_interval=1,
#hyper_parameter_iteration=100,
#hyper_parameter_decay_factor=0.9,
#hyper_parameter_maximum_decay=10,
#hyper_parameter_converge_threshold=1e-6,
):
Inferencer.__init__(self, hyper_parameter_optimize_interval);
def __init__(self,
hash_oov_words=False,
maximum_gamma_update_iteration=50,
minimum_mean_change_threshold=1e-3
):
Inferencer.__init__(self, hash_oov_words);
self._maximum_gamma_update_iteration = maximum_gamma_update_iteration;
self._minimum_mean_change_threshold = minimum_mean_change_threshold;
def __init__(self,
hyper_parameter_optimize_interval=1,
symmetric_alpha_alpha=True,
symmetric_alpha_beta=True,
):
Inferencer.__init__(self, hyper_parameter_optimize_interval);
self._symmetric_alpha_alpha = symmetric_alpha_alpha
self._symmetric_alpha_beta = symmetric_alpha_beta
def __init__(self,
hash_oov_words=False,
number_of_samples=10,
burn_in_sweeps=5
):
Inferencer.__init__(self, hash_oov_words);
self._number_of_samples = number_of_samples;
self._burn_in_sweeps = burn_in_sweeps;
def __init__(
self,
hyper_parameter_optimize_interval=1,
#hyper_parameter_iteration=100,
#hyper_parameter_decay_factor=0.9,
#hyper_parameter_maximum_decay=10,
#hyper_parameter_converge_threshold=1e-6,
):
Inferencer.__init__(self, hyper_parameter_optimize_interval)
def __init__(
self,
hyper_parameter_optimize_interval=10,
symmetric_alpha_alpha=True,
symmetric_alpha_beta=True,
):
Inferencer.__init__(self, hyper_parameter_optimize_interval)
self._symmetric_alpha_alpha = symmetric_alpha_alpha
self._symmetric_alpha_beta = symmetric_alpha_beta
def __init__(self,
hyper_parameter_optimize_interval=10,
symmetric_alpha_alpha=True,
symmetric_alpha_beta=True,
#local_parameter_iteration=1,
):
Inferencer.__init__(self, hyper_parameter_optimize_interval);
self._symmetric_alpha_alpha=symmetric_alpha_alpha
self._symmetric_alpha_beta=symmetric_alpha_beta
def __init__(self,
hyper_parameter_optimize_interval=1,
symmetric_alpha_alpha=True,
symmetric_alpha_beta=True,
#scipy_optimization_method="BFGS",
scipy_optimization_method="L-BFGS-B",
#scipy_optimization_method = "CG"
):
Inferencer.__init__(self, hyper_parameter_optimize_interval);
self._symmetric_alpha_alpha = symmetric_alpha_alpha
self._symmetric_alpha_beta = symmetric_alpha_beta
self._scipy_optimization_method = scipy_optimization_method
def _initialize(self, corpus, vocab, number_of_topics, alpha_alpha, alpha_beta):
Inferencer._initialize(self, vocab, number_of_topics, alpha_alpha, alpha_beta);
#self._corpus = corpus;
self._parsed_corpus = self.parse_data(corpus);
# define the total number of document
self._number_of_documents = len(self._parsed_corpus[0]);
# initialize a D-by-K matrix gamma, valued at N_d/K
self._gamma = numpy.zeros((self._number_of_documents, self._number_of_topics)) + self._alpha_alpha[numpy.newaxis, :] + 1.0 * self._number_of_types / self._number_of_topics;
# initialize a V-by-K matrix beta, valued at 1/V, subject to the sum over every row is 1
self._eta = numpy.random.gamma(100., 1. / 100., (self._number_of_topics, self._number_of_types));
def _initialize(self, corpus, vocab, number_of_topics, alpha_alpha, alpha_beta):
Inferencer._initialize(self, vocab, number_of_topics, alpha_alpha, alpha_beta);
self._corpus = corpus;
self._parsed_corpus = self.parse_data();
# define the total number of document
self._number_of_documents = len(self._parsed_corpus[0]);
# initialize a D-by-K matrix gamma, valued at N_d/K
self._gamma = numpy.zeros((self._number_of_documents, self._number_of_topics)) + self._alpha_alpha[numpy.newaxis, :] + 1.0 * self._number_of_types / self._number_of_topics;
# initialize a V-by-K matrix beta, valued at 1/V, subject to the sum over every row is 1
self._eta = numpy.random.gamma(100., 1. / 100., (self._number_of_topics, self._number_of_types));
def __init__(self,
hyper_parameter_optimize_interval=1,
):
'''
update_hyper_parameter=True,
alpha_update_decay_factor=0.9,
alpha_maximum_decay=10,
alpha_converge_threshold=0.000001,
alpha_maximum_iteration=100,
model_likelihood_threshold=0.00001,
gamma_converge_threshold=0.000001,
gamma_maximum_iteration=20
'''
Inferencer.__init__(self, hyper_parameter_optimize_interval);
def main():
flags(sys.argv)
model_params = Inferencer_Params(image_size=256,
model_path=flags.model_dir)
model_inferencer = Inferencer(model_params)
pipeline_params = Pipeline_Inferencer_Params(data_dir_x=os.path.join(
flags.dataset, 'test_X'),
data_dir_y=None)
pipeline = Pipeline_Inferencer(inferencer=model_inferencer,
params=pipeline_params,
pre_processing=None)
count = 0
first_pass = True
while first_pass or img_out is not None:
if first_pass:
first_pass = False
if not os.path.exists(flags.outdir):
os.makedirs(flags.outdir)
img_out = pipeline.run()
if img_out is not None:
filename = get_filename(count, 'mask_')
imageio.imwrite(os.path.join(flags.outdir, filename), img_out)
print(' [*] save file ' + filename)
count += 1
def _initialize(self, corpus, vocab, number_of_topics, alpha_alpha, alpha_beta):
Inferencer._initialize(self, vocab, number_of_topics, alpha_alpha, alpha_beta);
self._parsed_corpus = self.parse_data(corpus);
# define the total number of document
self._number_of_documents = len(self._parsed_corpus);
# define the counts over different topics for all documents, first indexed by doc_id id, the indexed by topic id
self._n_dk = numpy.zeros((self._number_of_documents, self._number_of_topics));
# define the counts over words for all topics, first indexed by topic id, then indexed by token id
self._n_kv = numpy.zeros((self._number_of_topics, self._number_of_types));
self._n_k = numpy.zeros(self._number_of_topics);
# define the topic assignment for every word in every document, first indexed by doc_id id, then indexed by word word_pos
self._k_dn = {};
self.random_initialize();
def _initialize(self, corpus_en, voc_en, corpus_cn, voc_cn,
number_of_topics_ge, alpha_alpha, alpha_beta, lam):
Inferencer._initialize(self,
voc_en,
voc_cn,
number_of_topics_ge,
alpha_alpha,
alpha_beta,
lam=0.5)
self._corpus_en = corpus_en
self._corpus_cn = corpus_cn
self._trans_en_cn = np.zeros(
(self._number_of_types_cn, self._number_of_types_en))
self._trans_cn_en = np.zeros(
(self._number_of_types_en, self._number_of_types_cn))
self.parse_data()
# define the total number of document
self._number_of_documents_en = len(self._word_idss_en)
self._number_of_documents_cn = len(self._word_idss_cn)
self._number_of_documents = self._number_of_documents_en + self._number_of_documents_cn
self._n_dk_en = np.zeros(
(self._number_of_documents_en, self._number_of_topics))
self._n_dk_cn = np.zeros(
(self._number_of_documents_cn, self._number_of_topics))
self._n_kv_en = np.zeros(
(self._number_of_topics, self._number_of_types_en))
self._n_kv_cn = np.zeros(
(self._number_of_topics, self._number_of_types_cn))
# define the topic assignment for every word in every document, first indexed by doc_id id, then indexed by word word_pos
self._n_k_en = np.zeros(self._number_of_topics)
self._n_k_cn = np.zeros(self._number_of_topics)
self._k_dn_en = {}
self._k_dn_cn = {}
self.psi_en = np.zeros(
(self._number_of_topics, self._number_of_types_en))
self.psi_cn = np.zeros(
(self._number_of_topics, self._number_of_types_cn))
self.phi_en = np.zeros(
(self._number_of_topics, self._number_of_types_en))
self.phi_cn = np.zeros(
(self._number_of_topics, self._number_of_types_cn))
self.random_initialize()
def _initialize(self, corpus, vocab, number_of_topics, alpha_alpha, alpha_beta):
Inferencer._initialize(self, vocab, number_of_topics, alpha_alpha, alpha_beta);
self._corpus = corpus;
self.parse_data();
# define the total number of document
self._number_of_documents = len(self._word_idss);
# define the counts over different topics for all documents, first indexed by doc_id id, the indexed by topic id
self._n_dk = numpy.zeros((self._number_of_documents, self._number_of_topics));
# define the counts over words for all topics, first indexed by topic id, then indexed by token id
self._n_kv = numpy.zeros((self._number_of_topics, self._number_of_types));
self._n_k = numpy.zeros(self._number_of_topics);
# define the topic assignment for every word in every document, first indexed by doc_id id, then indexed by word word_pos
self._k_dn = {};
self.random_initialize();
def __init__(self,
update_hyper_parameter=True,
alpha_update_decay_factor=0.9,
alpha_maximum_decay=10,
alpha_converge_threshold=0.000001,
alpha_maximum_iteration=100,
model_likelihood_threshold=0.00001,
gamma_converge_threshold=0.000001,
gamma_maximum_iteration=20
):
Inferencer.__init__(self, update_hyper_parameter, alpha_update_decay_factor, alpha_maximum_decay, alpha_converge_threshold, alpha_maximum_iteration, model_likelihood_threshold);
#self._alpha_update_decay_factor = alpha_update_decay_factor;
#self._alpha_maximum_decay = alpha_maximum_decay;
#self._alpha_converge_threshold = alpha_converge_threshold;
#self._alpha_maximum_iteration = alpha_maximum_iteration;
self._gamma_maximum_iteration = gamma_maximum_iteration;
self._gamma_converge_threshold = gamma_converge_threshold;
def __init__(self,
update_hyper_parameter=True,
alpha_update_decay_factor=0.9,
alpha_maximum_decay=10,
alpha_converge_threshold=0.000001,
alpha_maximum_iteration=100,
model_likelihood_threshold=0.00001,
number_of_samples=10,
burn_in_samples=5
):
Inferencer.__init__(self, update_hyper_parameter, alpha_update_decay_factor, alpha_maximum_decay, alpha_converge_threshold, alpha_maximum_iteration, model_likelihood_threshold);
#self._alpha_update_decay_factor = alpha_update_decay_factor;
#self._alpha_maximum_decay = alpha_maximum_decay;
#self._alpha_converge_threshold = alpha_converge_threshold;
#self._alpha_maximum_iteration = alpha_maximum_iteration;
self._number_of_samples = number_of_samples;
self._burn_in_samples = burn_in_samples;
def _initialize(self, corpus, vocab, labels, alpha_alpha, alpha_beta):
Inferencer._initialize(self, vocab, labels, alpha_alpha, alpha_beta)
self._number_of_topics = len(self._label_to_index)
self._parsed_corpus, self._parsed_labels = self.parse_data(corpus)
'''
# define the total number of document
self._number_of_documents = len(self._parsed_corpus);
# define the counts over different topics for all documents, first indexed by doc_id id, the indexed by topic id
self._n_dk = numpy.zeros((self._number_of_documents, self._number_of_topics), dtype=int)
# define the counts over words for all topics, first indexed by topic id, then indexed by token id
self._n_kv = numpy.zeros((self._number_of_topics, self._number_of_types), dtype=int)
self._n_k = numpy.zeros(self._number_of_topics, dtype=int)
# define the topic assignment for every word in every document, first indexed by doc_id id, then indexed by word word_pos
self._k_dn = {};
'''
# self._number_of_documents, self._k_dn, self._n_dk, self._n_k, self._n_kv = self.random_initialize();
self._k_dn, self._n_dk, self._n_k, self._n_kv = self.random_initialize(
)
def main():
A1 = fset.LeftSkewTrapezoid('A1', (0, 0), (1, 1), (4, 0))
A2 = fset.Triangle('A2', (2, 0), (5, 1), (8, 0))
A3 = fset.RightSkewTrapezoid('A3', (6, 0), (8, 1), (10, 0))
rule1 = 'IF x IS A1 THEN y IS B1'
rule2 = 'IF x IS A2 THEN y IS B2'
rule3 = 'IF x IS A3 THEN y IS B3'
fsets = [A1, A2, A3]
rules = [rule1, rule2, rule3]
inferencer = Inferencer()
inferencer.add_fsets(fsets)
inferencer.add_rules(rules)
input_test = {'x': 7}
print inferencer.evaluate(input_test)
""" load data """
train_loader, val_loader, test_loader, m_dh = prepare_data(
TRAIN_PATH, VAL_PATH, TEST_PATH, DH_PATH, LOAD_FROM_DUMP, BATCH_SIZE)
""" model setup """
INPUT_DIM, OUTPUT_DIM = len(m_dh.de_vocab), len(m_dh.en_vocab)
enc = Encoder(INPUT_DIM, ENC_EMB_DIM, ENC_HID_DIM, DEC_HID_DIM, 0)
attn = Attention(ENC_HID_DIM, DEC_HID_DIM, ATTN_DIM)
dec = Decoder(OUTPUT_DIM, DEC_EMB_DIM, ENC_HID_DIM, DEC_HID_DIM, 0, attn)
model = Seq2Seq(enc, dec)
""" load model """
state_dict = torch.load('ckpts/best.pt')
model.load_state_dict(state_dict['model_state'])
en_infer = Inferencer(m_dh.en_vocab)
de_infer = Inferencer(m_dh.de_vocab)
criterion = torch.nn.CrossEntropyLoss(ignore_index=1)
src, trg = next(iter(test_loader))
trg_text = en_infer.decode(trg)
with open('validate_sample/target.txt', 'w') as f:
f.writelines(trg_text)
print(trg_text)
inversion = Inversion(model,
MAX_LEN,
INPUT_DIM,
criterion,
ENTROPY_S,
TRAIN_PATH, VAL_PATH, TEST_PATH, DH_PATH, LOAD_FROM_DUMP, 3)
""" model setup """
INPUT_DIM, OUTPUT_DIM = len(m_dh.de_vocab), len(m_dh.en_vocab)
enc = Encoder(INPUT_DIM, ENC_EMB_DIM, ENC_HID_DIM, DEC_HID_DIM, ENC_DROPOUT)
attn = Attention(ENC_HID_DIM, DEC_HID_DIM, ATTN_DIM)
dec = Decoder(OUTPUT_DIM, DEC_EMB_DIM, ENC_HID_DIM, DEC_HID_DIM, DEC_DROPOUT,
attn)
model = Seq2Seq(enc, dec)
""" load model """
state_dict = torch.load('ckpts/best.pt')
model.load_state_dict(state_dict['model_state'])
model.eval()
en_infer = Inferencer(m_dh.en_vocab)
src, trg = next(iter(test_loader))
""" ______________ """
import matplotlib.pyplot as plt
import numpy
def plot_head_map(mma, target_labels, source_labels):
fig, ax = plt.subplots()
heatmap = ax.pcolor(mma, cmap=plt.cm.Blues)
# put the major ticks at the middle of each cell
ax.set_xticks(numpy.arange(mma.shape[1]) + 0.5,
minor=False) # mma.shape[1] = target seq 길이
ax.set_yticks(numpy.arange(mma.shape[0]) + 0.5,
minor=False) # mma.shape[0] = input seq 길이
def main():
input_directory = sys.argv[1];
output_directory = sys.argv[2];
if len(sys.argv)==4 and len(sys.argv[3])>0:
log_beta_path = sys.argv[3]
else:
log_beta_path = None;
from inferencer import Inferencer;
lda_inferencer = Inferencer();
lda_inferencer.load_params(os.path.join(output_directory, "current-params"));
lda_inferencer.load_tree(os.path.join(output_directory, "current-tree"));
lda_inferencer.format_output(input_directory);
if lda_inferencer._update_hyper_parameter:
lda_inferencer.dump_parameters(os.path.join(output_directory, "current-params"));
lda_inferencer.dump_E_log_beta(os.path.join(output_directory, "current-E-log-beta"));
if log_beta_path!=None:
lda_inferencer.export_E_log_beta(log_beta_path);
#if not hybrid_mode:
#lda_inferencer.dump_gamma(os.path.join(output_directory, "current-gamma"));
lda_inferencer.export_gamma(os.path.join(output_directory, "gamma"));
def main(TYPE="train", W_PATH=None):
print("TYPE", TYPE)
print("W_PATH", W_PATH)
engine = Trainer(W_PATH) if TYPE == "train" else Inferencer(W_PATH)
engine.run()
def main(
metadata_path,
source_dir,
target_dir,
output_dir,
root,
batch_size,
reload,
reload_dir,
):
"""Main function"""
# import Inferencer module
inferencer = Inferencer(root)
device = inferencer.device
sample_rate = inferencer.sample_rate
print(f"[INFO]: Inferencer is loaded from {root}.")
metadata = json.load(open(metadata_path))
print(f"[INFO]: Metadata list is loaded from {metadata_path}.")
output_dir = Path(output_dir) / Path(root).stem / \
f"{metadata['source_corpus']}2{metadata['target_corpus']}"
output_dir.mkdir(parents=True, exist_ok=True)
if reload:
metadata, conv_mels = reload_from_numpy(device, metadata, reload_dir)
else:
metadata, conv_mels = conversion(inferencer, device, root, metadata,
source_dir, target_dir, output_dir)
waveforms = []
max_memory_use = conv_mels[0].size(0) * batch_size
with torch.no_grad():
pbar = tqdm(total=metadata["n_samples"])
left = 0
while (left < metadata["n_samples"]):
batch_size = max_memory_use // conv_mels[left].size(0) - 1
right = left + min(batch_size, metadata["n_samples"] - left)
waveforms.extend(
inferencer.spectrogram2waveform(conv_mels[left:right]))
pbar.update(batch_size)
left += batch_size
pbar.close()
for pair, waveform in tqdm(zip(metadata["pairs"], waveforms)):
waveform = waveform.detach().cpu().numpy()
prefix = Path(pair["src_utt"]).stem
postfix = Path(pair["tgt_utts"][0]).stem
file_path = output_dir / f"{prefix}_to_{postfix}.wav"
pair["converted"] = f"{prefix}_to_{postfix}.wav"
if Path(root).stem == "BLOW":
wavfile.write(file_path, sample_rate, waveform)
else:
sf.write(file_path, waveform, sample_rate)
metadata_output_path = output_dir / "metadata.json"
json.dump(metadata, metadata_output_path.open("w"), indent=2)
