def test_get_sign(self):
dim = 100
act = 10
gen = Generator(dim, act)
signs = [str(i) for i in range(10)]
sign_index = TrieSignIndex(gen, vocabulary=signs)
for s in signs:
self.assertTrue(sign_index.contains(s))
id = sign_index.get_id(s)
self.assertTrue(sign_index.contains_id(id))
s2 = sign_index.get_sign(id)
self.assertEqual(s,s2)\
#get sign for an id that doesn't exist
id = 86
s = sign_index.get_sign(id)
self.assertEqual(s,None)
self.assertFalse(sign_index.contains_id(id))
self.assertEqual(len(sign_index.sign_trie),len(signs))
self.assertTrue(sign_index.contains_id(len(signs)-1))
self.assertFalse(sign_index.contains_id(len(signs)))
def test_contains(self):
dim = 100
act = 10
gen = Generator(dim, act)
sign_index = TrieSignIndex(generator=gen)
sign_index.add("0")
self.assertTrue(sign_index.contains("0"))
self.assertFalse(sign_index.contains("1"))
sign_index.remove("0")
self.assertFalse(sign_index.contains("0"))
def test_size(self):
gen = Generator(100, 10)
sign_index = TrieSignIndex(generator=gen)
# adding elements should increase size
self.assertEqual(len(sign_index), 0)
sign_index.add("0")
self.assertEqual(len(sign_index), 1)
# duplicated elements are not added
sign_index.add("0")
self.assertEqual(len(sign_index), 1)
sign_index.add("1")
self.assertEqual(len(sign_index), 2)
# removing elements should reduce size
size_before = len(sign_index)
sign_index.remove("0")
size_after = len(sign_index)
self.assertEqual(size_after, size_before - 1)
def test_save(self):
dim = 100
act = 10
gen = Generator(dim, act)
signs = [str(i) for i in range(10)]
sign_index = TrieSignIndex(gen, vocabulary=signs)
filename = "index.hdf5"
directory = os.path.dirname(os.path.abspath(__file__))
output_file = directory+"/"+filename
self.assertFalse(os.path.exists(output_file))
try:
sign_index.save(output_file)
self.assertTrue(os.path.exists(output_file))
h5file = h5py.File(output_file,'r')
h5signs = h5file["signs"]
h5ri = h5file["ri"]
self.assertEqual(len(h5signs),len(signs))
print(h5ri[0])
print(h5ri.attrs["k"])
print(h5ri.attrs["s"])
print(h5ri.attrs["state"].tostring())
h5file.close()
except:
raise
finally:
if os.path.exists(output_file):
os.remove(output_file)
self.assertFalse(os.path.exists(output_file))
def test_get_ri(self):
dim = 100
act = 10
gen = Generator(dim, act)
sign_index = TrieSignIndex(gen)
sign_index.add("0")
self.assertTrue(sign_index.contains("0"))
ri0 = sign_index.get_ri("0")
self.assertIsInstance(ri0, RandomIndex)
self.assertEqual(ri0.dim, dim)
def test_load(self):
""" The ids should be the same when the index is loaded back up
"""
dim = 100
act = 10
gen = Generator(dim, act)
signs1 = [str(i) for i in range(1000)]
index1 = TrieSignIndex(gen, vocabulary=signs1)
filename = "index.hdf5"
directory = os.path.dirname(os.path.abspath(__file__))
index_file = directory + "/" + filename
self.assertFalse(os.path.exists(index_file))
try:
index1.save(index_file)
self.assertTrue(os.path.exists(index_file))
index2 = TrieSignIndex.load(index_file)
self.assertEqual(len(index2),len(index1))
for sign in signs1:
self.assertTrue(index1.contains(sign))
self.assertTrue(index2.contains(sign))
id1 = index1.get_id(sign)
id2 = index2.get_id(sign)
self.assertEqual(id1,id2)
ri1 = index1.get_ri(sign).to_vector()
ri2 = index2.get_ri(sign).to_vector()
np.testing.assert_array_equal(ri1,ri2)
except:
raise
finally:
if os.path.exists(index_file):
os.remove(index_file)
self.assertFalse(os.path.exists(index_file))
# iterates over lines but loads them as chunks
#n_rows = 100000
#sentences = chunk_it(corpus_dataset,n_rows=n_rows, chunk_size=20000)
n_rows = len(corpus_dataset)
sentences = chunk_it(corpus_dataset, chunk_size=100000)
pipeline = WaCKyPipe(datagen=sentences)
# ======================================================================================
# Load Vocabulary
# ======================================================================================
vocab_file = data_dir + "wacky_vocab_6M_spacy.hdf5"
vocab_hdf5 = h5py.File(vocab_file, 'r')
ri_gen = Generator(dim=k, num_active=s)
print("Loading Vocabulary...")
sign_index = TrieSignIndex(ri_gen, list(vocab_hdf5["vocabulary"][:]), pregen_indexes=False)
if subsampling:
freq = TrieSignIndex.map_frequencies(list(vocab_hdf5["vocabulary"][:]),
list(vocab_hdf5["frequencies"][:]),
sign_index)
total_freq = np.sum(vocab_hdf5["frequencies"])
print("done")
# ======================================================================================
# Neural Random Projections Model
# ======================================================================================
pos_labels = Input(n_units=k, name="yp")
neg_labels = Input(n_units=k, name="yn")
def run(**kwargs):
arg_dict.from_dict(kwargs)
args = arg_dict.to_namespace()
# ======================================================================================
# Load Corpus & Vocab
# ======================================================================================
corpus = PTBReader(path=args.corpus, mark_eos=args.mark_eos)
corpus_stats = h5py.File(os.path.join(args.corpus, "ptb_stats.hdf5"),
mode='r')
ri_generator = Generator(dim=args.k_dim,
num_active=args.s_active,
symmetric=True)
# vocab = marisa_trie.Trie(corpus_stats["vocabulary"])
index = TrieSignIndex(generator=ri_generator,
vocabulary=corpus_stats["vocabulary"],
pregen_indexes=True)
# for i in range(1000):
# w = index.get_sign(i)
# ri: RandomIndex = index.get_ri(w)
# print(w)
# print(ri)
# print(ri)
# print(index.get_id(w))
# pre-gen indices for vocab, we could do this iteratively ... same thing
# ris = [ri_generator.generate() for _ in range(len(vocab))]
# print(vocab.keys())
# index = TrieSignIndex(generator=ri_generator,vocabulary=vocab)
# TODO could create the NRP model with NCE only and input with random indices could be passed to the model
# dynamically, also for inference and evaluation, we could either work with a dynamic encoding process
# or give it the current ri tensor with all the known ris if we know there are no OOV words (words that might not
# have been seen during training.
# table with random indices for all known symbols
# ri_tensor = RandomIndexTensor.from_ri_list(ris, k=args.k_dim, s=args.s_active)
def corpus_pipeline(corpus_stream,
n_gram_size=args.ngram_size,
epochs=1,
batch_size=args.batch_size,
shuffle=args.shuffle,
flatten=False):
""" Corpus Processing Pipeline.
Transforms the corpus reader -a stream of sentences or words- into a stream of n-gram batches.
Args:
n_gram_size: the size of the n-gram window
corpus_stream: the stream of sentences of words
epochs: number of epochs we want to iterate over this corpus
batch_size: batch size for the n-gram batch
shuffle: if true, shuffles the n-grams according to a buffer size
flatten: if true sliding windows are applied over a stream of words rather than within each sentence
(n-grams can cross sentence boundaries)
"""
if flatten:
word_it = flatten_it(corpus_stream)
n_grams = window_it(word_it, n_gram_size)
else:
sentence_n_grams = (window_it(sentence, n_gram_size)
for sentence in corpus_stream)
n_grams = flatten_it(sentence_n_grams)
# at this point this is an n_gram iterator
# n_grams = ([vocab[w] for w in ngram] for ngram in n_grams)
n_grams = ([index.get_id(w) for w in ngram] for ngram in n_grams)
if epochs > 1:
n_grams = repeat_it(n_grams, epochs)
if shuffle:
n_grams = shuffle_it(n_grams, args.shuffle_buffer_size)
n_grams = batch_it(n_grams, size=batch_size, padding=False)
return n_grams
# print("counting dataset samples...")
training_len = sum(1 for _ in corpus_pipeline(
corpus.training_set(), batch_size=1, epochs=1, shuffle=False))
validation_len = None
test_len = None
if args.eval_progress:
validation_len = sum(1 for _ in corpus_pipeline(
corpus.validation_set(), batch_size=1, epochs=1, shuffle=False))
test_len = sum(1 for _ in corpus_pipeline(
corpus.test_set(), batch_size=1, epochs=1, shuffle=False))
# print("done")
# print("dset len ", training_len)
# ======================================================================================
# Load Params, Prepare results assets
# ======================================================================================
# Experiment parameter summary
res_param_filename = os.path.join(
args.out_dir, "params_{id}_{run}.csv".format(id=args.id, run=args.run))
with open(res_param_filename, "w") as param_file:
writer = csv.DictWriter(f=param_file, fieldnames=arg_dict.keys())
writer.writeheader()
writer.writerow(arg_dict)
param_file.flush()
# make dir for model checkpoints
if args.save_model:
model_ckpt_dir = os.path.join(
args.out_dir, "model_{id}_{run}".format(id=args.id, run=args.run))
os.makedirs(model_ckpt_dir, exist_ok=True)
model_path = os.path.join(
model_ckpt_dir, "nnlm_{id}_{run}.ckpt".format(id=args.id,
run=args.run))
# start perplexity file
ppl_header = ["id", "run", "epoch", "step", "lr", "dataset", "perplexity"]
ppl_fname = os.path.join(
args.out_dir, "perplexity_{id}_{run}.csv".format(id=args.id,
run=args.run))
ppl_file = open(ppl_fname, "w")
ppl_writer = csv.DictWriter(f=ppl_file, fieldnames=ppl_header)
ppl_writer.writeheader()
# ======================================================================================
# MODEL
# ======================================================================================
# Configure weight initializers based on activation functions
if args.h_act == "relu":
h_act = tx.relu
h_init = tx.he_normal_init()
elif args.h_act == "tanh":
h_act = tx.tanh
h_init = tx.glorot_uniform()
elif args.h_act == "elu":
h_act = tx.elu
h_init = tx.he_normal_init()
elif args.h_act == "selu":
h_act = tf.nn.selu
h_init = tx.glorot_uniform()
# Configure embedding and logit weight initializers
if args.embed_init == "normal":
embed_init = tx.random_normal(mean=0., stddev=args.embed_init_val)
elif args.embed_init == "uniform":
embed_init = tx.random_uniform(minval=-args.embed_init_val,
maxval=args.embed_init_val)
if args.logit_init == "normal":
logit_init = tx.random_normal(mean=0., stddev=args.logit_init_val)
elif args.logit_init == "uniform":
logit_init = tx.random_uniform(minval=-args.logit_init_val,
maxval=args.logit_init_val)
f_init = None
if args.use_f_predict:
if args.f_init == "normal":
f_init = tx.random_normal(mean=0., stddev=args.f_init_val)
elif args.f_init == "uniform":
f_init = tx.random_uniform(minval=-args.f_init_val,
maxval=args.f_init_val)
model = NRP(ctx_size=args.ngram_size - 1,
sign_index=index,
k_dim=args.k_dim,
s_active=args.s_active,
embed_dim=args.embed_dim,
h_dim=args.h_dim,
embed_init=embed_init,
logit_init=logit_init,
num_h=args.num_h,
h_activation=h_act,
h_init=h_init,
use_dropout=args.dropout,
embed_dropout=args.embed_dropout,
keep_prob=args.keep_prob,
l2_loss=args.l2_loss,
l2_loss_coef=args.l2_loss_coef,
f_init=f_init,
embed_share=args.embed_share,
logit_bias=args.logit_bias,
use_nce=args.nce,
nce_samples=args.nce_samples,
nce_noise_amount=0.04)
model_runner = tx.ModelRunner(model)
# Input params can be changed during training by setting their value
# lr_param = tx.InputParam(init_value=args.lr)
lr_param = tensorx.train.EvalStepDecayParam(
value=args.lr,
improvement_threshold=args.eval_threshold,
less_is_better=True,
decay_rate=args.lr_decay_rate,
decay_threshold=args.lr_decay_threshold)
if args.optimizer == "sgd":
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=lr_param.tensor)
elif args.optimizer == "adam":
optimizer = tf.train.AdamOptimizer(learning_rate=lr_param.tensor,
beta1=args.optimizer_beta1,
beta2=args.optimizer_beta2,
epsilon=args.optimizer_epsilon)
elif args.optimizer == "ams":
optimizer = tx.AMSGrad(learning_rate=lr_param.tensor,
beta1=args.optimizer_beta1,
beta2=args.optimizer_beta2,
epsilon=args.optimizer_epsilon)
def clip_grad_global(grads):
grads, _ = tf.clip_by_global_norm(grads, 12)
return grads
def clip_grad_local(grad):
return tf.clip_by_norm(grad, args.clip_value)
if args.clip_grads:
if args.clip_local:
clip_fn = clip_grad_local
else:
clip_fn = clip_grad_global
if args.clip_grads:
model_runner.config_optimizer(optimizer,
optimizer_params=lr_param,
gradient_op=clip_fn,
global_gradient_op=not args.clip_local)
else:
model_runner.config_optimizer(optimizer, optimizer_params=lr_param)
# ======================================================================================
# EVALUATION
# ======================================================================================
def eval_model(runner,
dataset_it,
len_dataset=None,
display_progress=False):
if display_progress:
pb = tqdm(total=len_dataset, ncols=60, position=1)
batches_processed = 0
sum_loss = 0
for batch in dataset_it:
batch = np.array(batch, dtype=np.int64)
ctx = batch[:, :-1]
target = batch[:, -1:]
mean_loss = runner.eval(ctx, target)
sum_loss += mean_loss
if display_progress:
pb.update(args.batch_size)
batches_processed += 1
if display_progress:
pb.close()
return np.exp(sum_loss / batches_processed)
def evaluation(runner: tx.ModelRunner,
progress_bar,
cur_epoch,
step,
display_progress=False):
ppl_validation = eval_model(
runner,
corpus_pipeline(corpus.validation_set(), epochs=1, shuffle=False),
validation_len, display_progress)
res_row = {
"id": args.id,
"run": args.run,
"epoch": cur_epoch,
"step": step,
"lr": lr_param.value,
"dataset": "validation",
"perplexity": ppl_validation
}
ppl_writer.writerow(res_row)
if args.eval_test:
# pb.write("[Eval Test Set]")
ppl_test = eval_model(
runner,
corpus_pipeline(corpus.test_set(), epochs=1, shuffle=False),
test_len, display_progress)
res_row = {
"id": args.id,
"run": args.run,
"epoch": cur_epoch,
"step": step,
"lr": lr_param.value,
"dataset": "test",
"perplexity": ppl_test
}
ppl_writer.writerow(res_row)
ppl_file.flush()
if args.eval_test:
progress_bar.set_postfix({"test PPL ": ppl_test})
# pb.write("valid. ppl = {}".format(ppl_validation))
return ppl_validation
# ======================================================================================
# TRAINING LOOP
# ======================================================================================
# print("Starting TensorFlow Session")
# preparing evaluation steps
# I use ceil because I make sure we have padded batches at the end
epoch_step = 0
global_step = 0
current_epoch = 0
patience = 0
cfg = tf.ConfigProto()
cfg.gpu_options.allow_growth = True
sess = tf.Session(config=cfg)
model_runner.set_session(sess)
model_runner.init_vars()
progress = tqdm(total=training_len * args.epochs,
position=args.pid + 1,
disable=not args.display_progress)
training_data = corpus_pipeline(corpus.training_set(),
batch_size=args.batch_size,
epochs=args.epochs,
shuffle=args.shuffle)
evaluations = []
try:
for ngram_batch in training_data:
epoch = progress.n // training_len + 1
# Start New Epoch
if epoch != current_epoch:
current_epoch = epoch
epoch_step = 0
if args.display_progress:
progress.set_postfix({"epoch": current_epoch})
# ================================================
# EVALUATION
# ================================================
if epoch_step == 0:
current_eval = evaluation(model_runner,
progress,
epoch,
global_step,
display_progress=args.eval_progress)
evaluations.append(current_eval)
lr_param.update(current_eval)
# print(lr_param.eval_history)
# print("improvement ", lr_param.eval_improvement())
if global_step > 0:
if args.early_stop and epoch > 1:
if lr_param.eval_improvement(
) < lr_param.improvement_threshold:
if patience >= 3:
break
patience += 1
else:
patience = 0
# ================================================
# TRAIN MODEL
# ================================================
ngram_batch = np.array(ngram_batch, dtype=np.int64)
ctx_ids = ngram_batch[:, :-1]
word_ids = ngram_batch[:, -1:]
model_runner.train(ctx_ids, word_ids)
progress.update(args.batch_size)
epoch_step += 1
global_step += 1
# if not early stop, evaluate last state of the model
if not args.early_stop or patience < 3:
current_eval = evaluation(model_runner, progress, epoch,
epoch_step)
evaluations.append(current_eval)
ppl_file.close()
if args.save_model:
model_runner.save_model(model_name=model_path,
step=global_step,
write_state=False)
model_runner.close_session()
progress.close()
tf.reset_default_graph()
# return the best validation evaluation
return min(evaluations)
except Exception as e:
traceback.print_exc()
os.remove(ppl_file.name)
os.remove(param_file.name)
raise e
# n_rows = len(corpus_dataset)
# ======================================================================================
# Load Vocabulary
# ======================================================================================
if args.lemmas:
vocab_file = data_dir + "bnc_vocab_lemma.hdf5"
else:
vocab_file = data_dir + "bnc_vocab.hdf5"
vocab_hdf5 = h5py.File(vocab_file, 'r')
ri_gen = Generator(dim=k, num_active=s)
print("Loading Vocabulary...")
index = TrieSignIndex(ri_gen,
list(vocab_hdf5["vocabulary"][:]),
pregen_indexes=False)
if subsampling:
freq = TrieSignIndex.map_frequencies(list(vocab_hdf5["vocabulary"][:]),
list(vocab_hdf5["frequencies"][:]),
index)
total_freq = np.sum(vocab_hdf5["frequencies"])
print("done")
# ======================================================================================
# Neural Random Projections Model
# ======================================================================================
print("vocab loaded")
print(t1 - t0)
top10w = list(vocabulary[0:10])
top10f = list(frequencies[0:10])
top10ids = [trie.get(top10w[i]) for i in range(10)]
top10w_trie = [trie.restore_key(i) for i in top10ids]
print(top10w)
print(top10f)
print(top10w_trie)
ri_gen = Generator(dim=1000, num_active=10)
t0 = time.time()
sign_index = TrieSignIndex(ri_gen, list(vocabulary[:]))
t1 = time.time()
print(t1 - t0)
print(top10ids)
top10w_index = [sign_index.get_sign(i) for i in top10ids]
print(top10w_index)
#test load top ten
print("=============================================")
index = TrieSignIndex(generator=ri_gen, vocabulary=top10w)
print(top10w)
top10ids = [index.get_id(w) for w in top10w]
print(top10ids)
freq = TrieSignIndex.map_frequencies(top10w, top10f, index)
top10freq = [freq[i] for i in top10ids]
result = pool.map(func=text_to_ri, iterable=args)
pool.close()
if __name__ == '__main__':
# model parameters
max_sentences = 10000
# corpus and output assets
home = os.getenv("HOME")
corpus_file = home + "/data/gold_standards/wacky_1M.hdf5"
output_vectors = home + "/data/results/wacky_ri_1M.hdf5"
# load sign index
print("loading vocabulary")
vocab_file = home + "/data/results/wacky_vocab_1M.hdf5"
h5v = h5py.File(vocab_file, 'r')
vocabulary = h5v["vocabulary"]
frequencies = h5v["frequencies"]
ri_gen = Generator(dim=1000, num_active=10)
sign_index = TrieSignIndex(generator=ri_gen,
vocabulary=list(vocabulary[()]),
pregen_indexes=True)
print("done")
# index_filename = None
parallel_ri(corpus_file, max_sentences, window_size=3, n_processes=16)
from deepsign.rp.index import TrieSignIndex as Index
from deepsign.data.corpora.toefl import TOEFLReader
import seaborn as sns
# model dir
home = os.getenv("HOME")
data_dir = home + "/data/gold_standards/"
result_dir = home + "/data/results/"
model_dir = result_dir + "nrp/300d_reg_embeddings/"
model_file = model_dir + "model_bnc"
embeddings_file = model_dir + "embeddings.npy"
index_file = model_dir + "index.hdf5"
# load index
print("loading word index")
index = Index.load(index_file)
print(len(index))
# load toefl
questions_file = data_dir + "toefl/questions.csv"
answers_file = data_dir + "toefl/answers.csv"
toefl = TOEFLReader(questions_file=questions_file, answers_file=answers_file)
# words in toelf and not in index
toefl_remove = set(w for w in toefl.words if not index.contains(w))
for (i, question) in enumerate(toefl.questions):
qw = question[0]
aw = question[1]
# print(question)
answer = toefl.answer(i)
def test_nce_nrp(self):
vocab_size = 1000
k = 500
s = 8
embed_size = 128
nce_samples = 10
noise_ratio = 0.1
use_nce = True
vocab = [str(i) for i in range(vocab_size)]
generator = Generator(k, s)
sign_index = TrieSignIndex(generator,
vocabulary=vocab,
pregen_indexes=True)
ris = [
sign_index.get_ri(sign_index.get_sign(i))
for i in range(len(sign_index))
]
# ris = [generator.generate() for _ in range(vocab_size)]
ri_tensor = ris_to_sp_tensor_value(ri_seq=ris,
dim=k,
all_positive=False)
ri_tensor_input = tx.SparseInput(n_units=k, value=ri_tensor)
if use_nce:
label_inputs = tx.SparseInput(k, name="target_random_indices")
else:
label_inputs = [
tx.Input(1, dtype=tf.int64, name="ids"),
tx.InputParam(dtype=tf.int32,
value=vocab_size,
name="vocab_size")
]
eval_label_inputs = [
tx.Input(1, dtype=tf.int64, name="ids_eval"),
tx.InputParam(dtype=tf.int32, value=vocab_size, name="vocab_size")
]
model = NRP(
run_inputs=tx.SparseInput(n_units=k, name="random_index_inputs"),
label_inputs=label_inputs,
eval_label_input=eval_label_inputs,
ctx_size=2,
# vocab_size=vocab_size,
k_dim=k,
ri_tensor_input=ri_tensor_input, # current dictionary state
embed_dim=embed_size,
h_dim=128,
num_h=1,
h_activation=tx.relu,
use_dropout=True,
embed_dropout=True,
keep_prob=0.70,
use_nce=use_nce,
nce_samples=nce_samples,
nce_noise_amount=noise_ratio,
noise_input=tx.SparseInput(k, name="noise"))
tf.summary.histogram("embeddings", model.embeddings.weights)
for h in model.h_layers:
tf.summary.histogram("h", h.linear.weights)
# model.eval_tensors.append(model.train_loss_tensors[0])
runner = tx.ModelRunner(model)
runner.set_log_dir("/tmp")
runner.log_graph()
options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
# options = None
runner.set_session(runtime_stats=True, run_options=options)
# options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
# runner.config_optimizer(tf.train.GradientDescentOptimizer(learning_rate=0.005))#,
# SGD with 0.025
# lr = tx.InputParam(init_value=0.0002)
lr = tx.InputParam(value=0.025)
# runner.config_optimizer(tf.train.AdamOptimizer(learning_rate=lr.tensor, beta1=0.9), params=lr,
runner.config_optimizer(
tf.train.GradientDescentOptimizer(learning_rate=lr.tensor),
optimizer_params=lr,
global_gradient_op=False,
# gradient_op=lambda grad: tf.clip_by_global_norm(grad, 10.0)[0])
gradient_op=lambda grad: tf.clip_by_norm(grad, 1.0))
data = np.array([[0, 2], [5, 7], [9, 8], [3, 4], [1, 9], [12, 8]])
labels = np.array([[32], [56], [12], [2], [5], [23]])
ppl_curve = []
n = 256
batch_size = 128
dataset = np.column_stack((data, labels))
# print(dataset)
dataset = views.repeat_it([dataset], n)
dataset = views.flatten_it(dataset)
# shuffle 5 at a time
dataset = views.shuffle_it(dataset, 6)
dataset = views.batch_it(dataset, batch_size)
# print(np.array(list(dataset)))
# d = list(views.take_it(1, views.shuffle_it(d, 4)))[0]
data_stream = dataset
for data_stream in tqdm(data_stream, total=n * 5 / batch_size):
sample = np.array(data_stream)
ctx = sample[:, :-1]
ctx.flatten()
ctx = ctx.flatten()
ctx_ris = [sign_index.get_ri(sign_index.get_sign(i)) for i in ctx]
ctx_ris = ris_to_sp_tensor_value(
ctx_ris,
dim=sign_index.feature_dim(),
all_positive=not sign_index.generator.symmetric)
lbl_ids = sample[:, -1:]
lbl = lbl_ids.flatten()
if use_nce:
lbl_ris = [
sign_index.get_ri(sign_index.get_sign(i)) for i in lbl
]
lbl_ris = ris_to_sp_tensor_value(
lbl_ris,
dim=sign_index.feature_dim(),
all_positive=not sign_index.generator.symmetric)
noise = generate_noise(k_dim=k,
batch_size=lbl_ris.dense_shape[0] *
nce_samples,
ratio=noise_ratio)
runner.train(ctx_ris, [lbl_ris, noise],
output_loss=True,
write_summaries=True)
else:
runner.train(model_input_data=ctx_ris,
loss_input_data=lbl_ids,
output_loss=True,
write_summaries=True)
runner.close_session()