def loss(self, model, paths, total_paths, alpha=1.0):
start, next_report, loss = time.time(), 5.0, 0.0
num_nodes = 0
for job_no, job in enumerate(
chunkize_serial(prepare_sentences(model, paths), 250)):
batch_loss = np.zeros(1, dtype=np.float32)
batch_work = np.zeros(model.layer1_size, dtype=np.float32)
batch_node = sum([
loss_o2(model.node_embedding,
model.context_embedding,
path,
self.negative,
self.window_size,
model.table,
alpha,
model.layer1_size,
batch_work,
py_loss=batch_loss) for path in job if path is not None
])
num_nodes += batch_node
loss += batch_loss[0]
elapsed = time.time() - start
if elapsed >= next_report:
log.debug("PROGRESS: at %.2f%% path, %.0f paths/s" %
(100.0 * num_nodes / total_paths,
num_nodes / elapsed if elapsed else 0.0))
# log.debug("loss: {}".format(loss))
next_report = elapsed + 1.0 # don't flood the log, wait at least a second between progress reports
# def worker_loss(job, next_report):
# """Train the model, lifting lists of paths from the jobs queue."""
#
# py_work = np.zeros(model.layer1_size, dtype=np.float32)
# job_nodes = sum([loss_o2(model.node_embedding, model.context_embedding, path, self.negative,
# self.window_size, model.table, alpha, model.layer1_size,
# py_work, py_loss=loss) for path in job]) # execute the sgd
# num_nodes[0] += job_nodes
# elapsed = time.time() - start
#
# if elapsed >= next_report:
# print("PROGRESS: at %.2f%% path, %.0f paths/s" % (
# 100.0 * num_nodes[0] / total_paths, num_nodes[0] / elapsed if elapsed else 0.0))
# next_report = elapsed + 1.0 # don't flood the log, wait at least a second between progress reports
# print(loss)
# return next_report
#
# for job_no, job in enumerate(chunkize_serial(prepare_sentences(model, paths), 250)):
# next_report = worker_loss(job, next_report)
log.info(num_nodes)
log.info(loss)
return loss
def loss(self, model, edges):
ret_loss = 0
for edge in prepare_sentences(model, edges):
assert len(
edge) == 2, "edges have to be done by 2 nodes :{}".format(edge)
edge_loss = np.log(
sigmoid(
np.dot(model.node_embedding[edge[1].index],
model.node_embedding[edge[0].index].T)))
assert edge_loss <= 0, "malformed loss"
ret_loss -= edge_loss
return ret_loss
def learn_second(network,
lr,
model,
examples_files,
total_example,
alpha=1.0,
batch_size=20):
"""
Helper function used to optimize O2
:param network: network model to optimize
:param lr: learning rate
:param model: model containing the shared data
:param examples_files: list of files containing the examples
:param total_example: total example for training
:param alpha: trade-off param
:param batch_size: size of the batch
:return: loss value
"""
num_batch = 0
log.info("compute o2")
optimizer = SGD(network.parameters(), lr)
log.debug("read example file: {}".format("\t".join(examples_files)))
loss_val = 0
if alpha <= 0:
return loss_val
for batch in emb_utils.batch_generator(emb_utils.prepare_sentences(
model, graph_utils.combine_example_files_iter(examples_files),
network.transfer_fn(model.vocab)),
batch_size,
long_tensor=LongTensor):
input, output = batch
loss = (alpha * network.forward(
input, output, negative_sampling_fn=model.negative_sample))
loss_val += loss.data[0]
optimizer.zero_grad()
loss.backward()
optimizer.step()
num_batch += 1
if (num_batch) % 10000 == 0:
log.info("community embedding batches completed: {}".format(
num_batch / (total_example / batch_size)))
log.debug("O2 loss: {}".format(loss_val))
return loss_val
def learn_community(network, lr, model, nodes, beta=1.0, batch_size=20):
"""
Helper function used to optimize O3
:param network: model to optimize
:param lr: learning rate
:param model: model containing the shared data
:param nodes: nodes on which execute the learning
:param beta: trade-off value
:param batch_size: size of the batch
:return: loss value
"""
num_batch = 0
log.info("compute o3")
optimizer = SGD(network.parameters(), lr)
loss_val = 0
if beta <= 0.:
return loss_val
for batch in emb_utils.batch_generator(emb_utils.prepare_sentences(
model, nodes, network.transfer_fn()),
batch_size,
long_tensor=LongTensor):
input, output = batch
loss = network.forward(input, model)
loss.data *= (beta / model.k)
loss_val += loss.data[0]
optimizer.zero_grad()
loss.backward()
optimizer.step()
num_batch += 1
if (num_batch) % 10000 == 0:
log.info("community embedding batches completed: {}".format(
num_batch / (total_example / batch_size)))
log.debug("O3 loss: {}".format(loss_val))
return loss_val
def learn_first(network, lr, model, edges, num_iter=1, batch_size=20):
"""
Helper function used to optimize O1
:param network: neural network to train
:param lr: learning rate
:param model: model containing the shared data
:param edges: numpy list of edges used for training
:param num_iter: iteration number over the edges
:param batch_size: size of the batch
:return: loss value
"""
log.info("computing o1")
optimizer = SGD(network.parameters(), lr)
num_batch = 0
total_batch = (edges.shape[0] * num_iter) / batch_size
loss_val = 0
for batch in emb_utils.batch_generator(emb_utils.prepare_sentences(
model, edges, network.transfer_fn(model.vocab)),
batch_size,
long_tensor=LongTensor):
input, output = batch
loss = network.forward(input,
output,
negative_sampling_fn=model.negative_sample)
loss_val += loss.data[0]
optimizer.zero_grad()
loss.backward()
optimizer.step()
num_batch += 1
if (num_batch) % 10000 == 0:
log.info("community embedding batches completed: {}".format(
num_batch / total_batch))
log.debug("O1 loss: {}".format(loss_val))
return loss_val
def learn_second(network, lr, model, examples_files, alpha=1.0):
"""
Helper function used to optimize O1 and O3
:param loss: loss to optimize
:param lr: learning rate
:param model: deprecated_model used to compute the batches and the negative sampling
:param examples_files: list of files containing the examples
:param num_iter: iteration number over the edges
:return:
"""
log.info("compute o2")
optimizer = SGD(network.parameters(), lr)
log.debug("read example file: {}".format("\t".join(examples_files)))
for batch in emb_utils.batch_generator(
emb_utils.prepare_sentences(
model, graph_utils.combine_example_files_iter(examples_files),
network.transfer_fn(model.vocab)), 20):
input, output = batch
loss = (alpha * network.forward(
input, output, negative_sampling_fn=model.negative_sample))
optimizer.zero_grad()
loss.backward()
optimizer.step()
def learn_first(network, lr, model, edges, num_iter=1):
"""
Helper function used to optimize O1 and O3
:param network: neural network to train
:param lr: learning rate
:param model: deprecated_model used to compute the batches and the negative sampling
:param edges: numpy list of edges used for training
:param num_iter: iteration number over the edges
:return:
"""
log.info("computing o1")
optimizer = SGD(network.parameters(), lr)
for batch in emb_utils.batch_generator(
emb_utils.prepare_sentences(
model, emb_utils.RepeatCorpusNTimes(edges, n=num_iter),
network.transfer_fn(model.vocab)), 20):
input, output = batch
loss = network.forward(input,
output,
negative_sampling_fn=model.negative_sample)
optimizer.zero_grad()
loss.backward()
optimizer.step()
def train(self, model, edges, chunksize=150, iter=1):
"""
Update the model's neural weights from a sequence of paths (can be a once-only generator stream).
"""
assert model.node_embedding.dtype == np.float32
log.info("O1 training model with %i workers on %i vocabulary and %i features and 'negative sampling'=%s" %
(self.workers, len(model.vocab), model.layer1_size, self.negative))
if not model.vocab:
raise RuntimeError("you must first build vocabulary before training the model")
edges = RepeatCorpusNTimes(edges, iter)
total_node = edges.corpus.shape[0] * edges.corpus.shape[1] * edges.n
log.debug('total edges: %d' % total_node)
start, next_report, word_count = time.time(), [5.0], [0]
#int(sum(v.count * v.sample_probability for v in self.vocab.values()))
jobs = Queue(maxsize=2*self.workers) # buffer ahead only a limited number of jobs.. this is the reason we can't simply use ThreadPool :(
lock = threading.Lock()
def worker_train():
"""Train the model, lifting lists of paths from the jobs queue."""
while True:
job = jobs.get(block=True)
if job is None: # data finished, exit
jobs.task_done()
# print('thread %s break' % threading.current_thread().name)
break
py_work = np.zeros(model.layer1_size, dtype=np.float32)
job_words = sum(train_o1(model.node_embedding, edge, self.lr, self.negative, model.table,
py_size=model.layer1_size, py_work=py_work) for edge in job if edge is not None)
jobs.task_done()
lock.acquire(timeout=30)
try:
word_count[0] += job_words
elapsed = time.time() - start
if elapsed >= next_report[0]:
log.info("PROGRESS: at %.2f%% words\tword_computed %d\talpha %.05f\t %.0f words/s" %
(100.0 * word_count[0] / total_node, word_count[0], self.lr, word_count[0] / elapsed if elapsed else 0.0))
next_report[0] = elapsed + 5.0 # don't flood the log, wait at least a second between progress reports
finally:
lock.release()
workers = [threading.Thread(target=worker_train, name='thread_'+str(i)) for i in range(self.workers)]
for thread in workers:
thread.daemon = True # make interrupting the process with ctrl+c easier
thread.start()
# convert input strings to Vocab objects (eliding OOV/downsampled words), and start filling the jobs queue
for job_no, job in enumerate(chunkize_serial(prepare_sentences(model, edges), chunksize)):
jobs.put(job)
for _ in range(self.workers):
jobs.put(None) # give the workers heads up that they can finish -- no more work!
for thread in workers:
thread.join()
elapsed = time.time() - start
log.info("training on %i words took %.1fs, %.0f words/s" %
(word_count[0], elapsed, word_count[0]/ elapsed if elapsed else 0.0))
def train(self,
model,
paths,
total_nodes,
alpha=1.0,
node_count=0,
chunksize=150):
"""
Update the model's neural weights from a sequence of paths (can be a once-only generator stream).
:param model: model containing the shared data
:param paths: generator of the paths
:param total_nodes: total number of nodes in the path
:param alpha: trade-off parameter
:param node_count: init of the number of nodes
:param chunksize: size of the batch
:return:
"""
assert model.node_embedding.dtype == np.float32
assert model.context_embedding.dtype == np.float32
log.info(
"O3 CONTEXT training model with %i workers on %i vocabulary and %i features, using \t'negative sampling'=%s\t'windows'=%s"
% (self.workers, len(model.vocab), model.layer1_size,
self.negative, self.window_size))
if alpha <= 0.:
return
if not model.vocab:
raise RuntimeError(
"you must first build vocabulary before training the model")
start, next_report = time.time(), [1.0]
if total_nodes is None:
raise AttributeError('need the number of node')
node_count = [0]
jobs = Queue(
maxsize=2 * self.workers
) # buffer ahead only a limited number of jobs.. this is the reason we can't simply use ThreadPool :(
lock = threading.Lock(
) # for shared state (=number of nodes trained so far, log reports...)
def worker_train():
"""Train the model, lifting lists of paths from the jobs queue."""
py_work = np.zeros(model.layer1_size, dtype=np.float32)
while True:
job = jobs.get()
if job is None: # data finished, exit
break
lr = max(self.min_lr,
self.lr * (1 - 1.0 * node_count[0] / total_nodes))
job_nodes = sum(
train_o2(model.node_embedding,
model.context_embedding,
path,
lr,
self.negative,
self.window_size,
model.table,
py_alpha=alpha,
py_size=model.layer1_size,
py_work=py_work)
for path in job) #execute the sgd
with lock:
node_count[0] += job_nodes
elapsed = time.time() - start
if elapsed >= next_report[0]:
log.info(
"PROGRESS: at %.2f%% nodes, lr %.05f, %.0f nodes/s"
% (100.0 * node_count[0] / total_nodes, lr,
node_count[0] / elapsed if elapsed else 0.0))
next_report[
0] = elapsed + 1.0 # don't flood the log, wait at least a second between progress reports
workers = [
threading.Thread(target=worker_train) for _ in range(self.workers)
]
for thread in workers:
thread.daemon = True # make interrupting the process with ctrl+c easier
thread.start()
# convert input strings to Vocab objects (eliding OOV/downsampled nodes), and start filling the jobs queue
for job_no, job in enumerate(
chunkize_serial(prepare_sentences(model, paths), chunksize)):
jobs.put(job)
log.debug(
"reached the end of input; waiting to finish %i outstanding jobs" %
jobs.qsize())
for _ in range(self.workers):
jobs.put(
None
) # give the workers heads up that they can finish -- no more work!
for thread in workers:
thread.join()
elapsed = time.time() - start
log.info("training on %i nodes took %.1fs, %.0f nodes/s" %
(node_count[0], elapsed,
node_count[0] / elapsed if elapsed else 0.0))
