def _process_gradients(
self, edl_embedding_gradients, indexed_grads, grads, request_version
):
if not self._use_async:
# grads of ElasticDL Embedding layer
for k, v in edl_embedding_gradients.items():
if k in self._edl_embedding_gradients:
self._edl_embedding_gradients[k] = merge_indexed_slices(
self._edl_embedding_gradients[k], v
)
else:
self._edl_embedding_gradients[k] = v
# grads of Keras Embedding layer
for k, v in indexed_grads.items():
if k not in self._gradient_sum_indexed:
self._gradient_sum_indexed[k] = v
else:
grads_s = self._gradient_sum_indexed[k]
self._gradient_sum_indexed[k] = merge_indexed_slices(
grads_s, v
)
# other grads
for k, v in grads.items():
if not self._use_async and k in self._gradient_sum:
self._gradient_sum[k] = self._gradient_sum[k] + v
else:
self._gradient_sum[k] = v
self._grad_n += 1
need_to_update_model = self._use_async
if not self._use_async and self._grad_n >= self._grad_to_wait:
need_to_update_model = True
# get gradient average for sync SGD
for k in self._gradient_sum:
self._gradient_sum[k] = (
self._gradient_sum[k] / self._grad_to_wait
)
edl_embedding_gradients = self._edl_embedding_gradients
indexed_grads = self._gradient_sum_indexed
grads = self._gradient_sum
if need_to_update_model:
self._update_optimizer(request_version)
self._update_model(grads, indexed_grads, edl_embedding_gradients)
def report_gradient_to_ps(self, grads):
self._timing.start_record_time("report_gradient")
reqs = [
elasticdl_pb2.PushGradientsRequest() for i in range(self._ps_num)
]
ps_grads = {}
non_embed_vars_n = len(self._non_embed_vars)
for g, v in zip(
grads[:non_embed_vars_n], self._non_embed_vars.values()
):
ps_id = self._var_to_ps[v.name]
if ps_id not in ps_grads:
ps_grads[ps_id] = {v.name: g}
else:
if v.name not in ps_grads[ps_id]:
ps_grads[ps_id][v.name] = g
else:
if isinstance(g, tf.IndexedSlices):
ps_grads[ps_id][v.name] = merge_indexed_slices(
ps_grads[ps_id][v.name], g
)
else:
ps_grads[ps_id][v.name] += g
for ps_id, pair in ps_grads.items():
for name, g in pair.items():
if isinstance(g, tf.IndexedSlices):
v, i = deduplicate_indexed_slices(g.values, g.indices)
ps_grads[ps_id][name] = tf.IndexedSlices(v, i)
for ps_id in ps_grads:
req = reqs[ps_id]
for name, g in ps_grads[ps_id].items():
# Keras embedding layer has a dense parameter,
# but an indexed slices type gradient
if isinstance(g, tf.IndexedSlices):
serialize_indexed_slices(
Tensor(None, g.values.numpy(), g.indices.numpy()),
req.gradients.embedding_tables[name],
)
else:
serialize_ndarray(
g.numpy(), req.gradients.dense_parameters[name]
)
edl_embedding_name_values = self._collect_edl_embedding_name_values()
if edl_embedding_name_values:
edl_embedding_grads = grads[non_embed_vars_n:]
bet_number = 0
for name, embedding_and_ids in edl_embedding_name_values:
bet_number += len(embedding_and_ids)
if len(edl_embedding_grads) != bet_number:
raise ValueError(
"elasticdl.layers.embedding related gradient number %d "
"does not match the number of its output tensor %d."
% (len(edl_embedding_grads), bet_number)
)
grad_accum_iter = 0
for name, embedding_and_ids in edl_embedding_name_values:
g_values = None
g_indices = None
for _, ids in embedding_and_ids:
grad = edl_embedding_grads[grad_accum_iter]
grad_accum_iter += 1
# ElasticDL embedding layer with Sparse Gradients
if isinstance(grad, tf.IndexedSlices):
grad = grad.values
if g_values is not None:
g_values = tf.concat([g_values, grad], axis=0)
g_indices = tf.concat([g_indices, ids], axis=0)
else:
g_values = grad
g_indices = ids
# Sum up the values of the duplicated indices in the
# gradients. It can reduce the gradient payload of the
# dense embedding.
g_values, g_indices = deduplicate_indexed_slices(
values=g_values, indices=g_indices
)
results = scatter_embedding_vector(
g_values.numpy(), g_indices.numpy(), self._ps_num
)
for ps_id in results:
req = reqs[ps_id]
gv, gi = results[ps_id]
serialize_indexed_slices(
Tensor(None, gv, gi),
req.gradients.embedding_tables[name],
)
report_futures = []
for ps_id in range(self._ps_num):
req = reqs[ps_id]
req.gradients.version = self._model_versions_from_ps[ps_id]
req.learning_rate = K.get_value(self._model.optimizer.lr)
report_future = self._ps_stubs[ps_id].push_gradients.future(req)
report_futures.append(report_future)
accepted = False
max_version = -1
for report_future in report_futures:
res = report_future.result()
if res.accepted:
accepted = True
if res.version > max_version:
max_version = res.version
self._timing.end_record_time("report_gradient")
return accepted, max_version
def push_gradients(self, request, _):
res = elasticdl_pb2.PushGradientsResponse()
if self._use_async:
grad_vars = []
for name, pb in request.gradients.dense_parameters.items():
grad = pb_to_ndarray(pb)
self._parameters.check_grad(Tensor(name, grad, None))
grad = tf.constant(grad)
var = self._parameters.get_non_embedding_param(name)
grad_vars.append((grad, var))
for name, pb in request.gradients.embedding_tables.items():
grad = pb_to_indexed_slices(pb)
self._parameters.check_grad(
Tensor(name, grad.values, grad.indices))
if name in self._parameters.non_embedding_params:
var = self._parameters.get_non_embedding_param(name)
grad_vars.append((grad, var))
else:
grad_vars.append((grad, name))
learning_rate = request.learning_rate
# TODO: if request.learning_rate == 0.0, modulate learning_rate
# in self._optimizer with staleness
if self._lr_staleness_modulation and learning_rate > 0.0:
staleness = max(
1, self._parameters.version - request.gradients.version)
# Modulate learning rate by staleness
learning_rate /= staleness
self._set_optimizer_learning_rate(learning_rate)
self._optimizer.apply_gradients(grad_vars)
with self._version_lock:
self._parameters.version += 1
self._save_params_to_checkpoint_if_needed()
version = self._parameters.version
self._report_version_if_needed(version)
res.accepted = True
res.version = self._parameters.version
return res
else:
if (request.gradients.version <
self._parameters.version - self._sync_version_tolerance):
res.accepted = False
res.version = self._parameters.version
return res
with self._lock:
for name, pb in request.gradients.dense_parameters.items():
grad = pb_to_ndarray(pb)
self._parameters.check_grad(Tensor(name, grad, None))
if name in self._grads_buffer:
self._grads_buffer[name] = (self._grads_buffer[name] +
grad)
else:
self._grads_buffer[name] = grad
for name, pb in request.gradients.embedding_tables.items():
grad = pb_to_indexed_slices(pb)
self._parameters.check_grad(
Tensor(name, grad.values, grad.indices))
if name in self._grads_buffer:
self._grads_buffer[name] = merge_indexed_slices(
self._grads_buffer[name], grad)
else:
self._grads_buffer[name] = grad
self._grads_n += 1
res.accepted = True
updated_version = False
version = self._parameters.version
if self._grads_n == self._grads_to_wait:
grad_vars = []
for name, grad in self._grads_buffer.items():
# Dense gradients are averaged,
# while sparse gradients are summed
if not isinstance(grad, tf.IndexedSlices):
grad = grad / self._grads_to_wait
grad = tf.constant(grad)
var = self._parameters.get_non_embedding_param(name)
if var is None:
grad_vars.append((grad, name))
else:
grad_vars.append((grad, var))
self._set_optimizer_learning_rate(request.learning_rate)
self._optimizer.apply_gradients(grad_vars)
self._grads_n = 0
self._grads_buffer.clear()
self._parameters.version += 1
self._save_params_to_checkpoint_if_needed()
version = self._parameters.version
updated_version = True
if updated_version:
self._report_version_if_needed(version)
res.version = version
return res
def push_gradients(
self, grads, edl_grads, learning_rate, model_versions,
):
"""
Push gradients to PS. There two kinds of gradients:
- gradients of normal layers
- sparse gradients of ElasticDL embedding layers
"""
reqs = [
elasticdl_pb2.PushGradientsRequest() for i in range(self.ps_num)
]
ps_grads = {}
# 1. handle grads
for grad in grads:
ps_id = self.parameter_to_ps[grad.name]
if ps_id not in ps_grads:
ps_grads[ps_id] = {grad.name: grad}
else:
if grad.name not in ps_grads[ps_id]:
ps_grads[ps_id][grad.name] = grad
else:
if grad.indices is not None:
ps_grads[ps_id][grad.name] = merge_indexed_slices(
ps_grads[ps_id][grad.name], grad
)
else:
ps_grads[ps_id][grad.name].values += grad.values
for ps_id, pair in ps_grads.items():
for name, grad in pair.items():
if grad.indices is not None:
v, i = deduplicate_indexed_slices(
grad.values, grad.indices
)
ps_grads[ps_id][name] = Tensor(None, v, i)
for ps_id in ps_grads:
req = reqs[ps_id]
for name, grad in ps_grads[ps_id].items():
# Keras embedding layer has a dense parameter,
# but an indexed slices type gradient
if grad.indices is not None:
serialize_indexed_slices(
Tensor(None, grad.values, grad.indices),
req.gradients.embedding_tables[name],
)
else:
serialize_ndarray(
grad.values, req.gradients.dense_parameters[name]
)
# 2. handle sparse grads of elasticdl embedding layers
groups = {}
for grad in edl_grads:
if grad.name not in groups:
groups[grad.name] = grad
else:
groups[grad.name] = merge_indexed_slices(
groups[grad.name], grad
)
# Sum up the values of the duplicated indices in the
# gradients. It can reduce the gradient payload of the
# dense embedding.
for name, grad in groups.items():
v, i = deduplicate_indexed_slices(grad.values, grad.indices)
groups[name] = Tensor(None, v, i)
results = scatter_embedding_vector(
groups[name].values, groups[name].indices, self.ps_num
)
for ps_id in results:
req = reqs[ps_id]
gv, gi = results[ps_id]
serialize_indexed_slices(
Tensor(None, gv, gi), req.gradients.embedding_tables[name],
)
# 3. push gradients to PS
report_futures = []
for ps_id in range(self.ps_num):
req = reqs[ps_id]
req.gradients.version = model_versions[ps_id]
req.learning_rate = learning_rate
report_future = self.ps_stubs[ps_id].push_gradients.future(req)
report_futures.append(report_future)
accepted = False
max_version = -1
for report_future in report_futures:
res = report_future.result()
if res.accepted:
accepted = True
if res.version > max_version:
max_version = res.version
return accepted, max_version