Merge pull request #1 from guru4elephant/for_pslib

For pslib

Author: guru4elephant

python/paddle/fluid/distributed/__init__.py

@@ -0,0 +1,32 @@
+from .node import DownpourServer
+from .node import DownpourWorker
+from ..backward import append_backward
+import ps_pb2 as pslib
+from paddle.fluid.distribute_lookup_table import find_distributed_lookup_table
+from google.protobuf import text_format
+
+class DownpourSGD(object):
+    def __init__(self, learning_rate=0.001, window=1):
+        # todo(guru4elephant): if optimizer is not None, will warning here
+        self.learning_rate_ = learning_rate
+        self.window_ = window
+
+    def minimize(self, loss, startup_program=None,
+                 parameter_list=None, no_grad_set=None,
+                 prefetch_slots=None, prefetch_slots_emb=None):
+        params_grads = sorted(append_backward(loss), key=lambda x:x[0].name)
+        table_name = find_distributed_lookup_table(loss.block.program)
+        server = DownpourServer()
+        worker = DownpourWorker(self.window_)
+        server.add_sparse_table(0, learning_rate,
+                                prefetch_slots, prefetch_slots_emb)
+        server.add_dense_table(1, learning_rate, params, grads)
+        worker.add_sparse_table(0, learning_rate,
+                                prefetch_slots, prefetch_slots_emb)
+        worker.add_dense_table(1, learning_rate, params, grads)
+        ps_param = pslib.PSParameter()
+        ps_param.server_param.CopyFrom(server.get_desc())
+        #ps_param.worker_param.CopyFrom(worker.get_desc())
+        worker_skipped_ops = ["lookup_table", "lookup_table_grad"]
+        ps_param_str = text_format.MessageToString(ps_param)
+        return [ps_param_str, worker_skipped_ops]

python/paddle/fluid/distributed/downpour.py

@@ -0,0 +1,20 @@
+from mpi4py import MPI
+
+class MPIHelper(object):
+    def __init__(self):
+        self.comm = MPI.COMM_WORLD
+
+    def get_rank(self):
+        return self.comm.Get_rank()
+
+    def get_size(self):
+        return self.comm.Get_size()
+
+    def get_ip(self):
+        import socket
+        local_ip = socket.gethostbyname(socket.gethostname())
+        return local_ip
+
+    def get_hostname(self):
+        import socket
+        return socket.gethostname()

python/paddle/fluid/distributed/helper.py

@@ -0,0 +1,73 @@
+import ps_pb2 as pslib
+
+class Server(object):
+    def __init__(self):
+        pass
+
+
+class Worker(object):
+    def __init__(self):
+        pass
+
+
+class DownpourServer(Server):
+    def __init__(self):
+        #self.server_ = pslib.ServerParameter().downpour_server_param
+        self.server_ = pslib.ServerParameter()
+
+    def add_sparse_table(self, table_id, learning_rate,
+                         slot_key, slot_value_var, slot_grad_var):
+        #table = self.server_.downpour_table_param.add()
+        table = self.server_.downpour_server_param.downpour_table_param.add()
+        table.table_id = table_id
+        table.type = PS_SPARSE_TABLE
+        table.accessor.accessor_class = "DownpourFeatureValueAccessor"
+        table.accessor.dense_sgd_param.adam.learning_rate = learning_rate
+        table.accessor.fea_dim = slot_value_var[0].shape[1]
+
+    def add_dense_table(self, table_id, learning_rate, 
+                        param_var, grad_var):
+        #table = self.server_.downpour_table_param.add()
+        table = self.server_.downpour_server_param.downpour_table_param.add()
+        table.table_id = table_id
+        table.type = PS_DENSE_TABLE
+        table.accessor.accessor_class = "DownpourDenseValueAccessor"
+        table.accessor.sparse_sgd_param.learning_rate = learning_rate
+        table.accessor.fea_dim = 1
+        #table.accessor.fea_dim = reduce(lambda x, y: x.shape, 1 for x in param_var)
+
+    def get_desc(self):
+        return self.server_
+
+
+class DownpourWorker(Worker):
+    def __init__(self, window):
+        self.window = window
+        #self.worker_ = pslib.WorkerParameter().downpour_worker_param
+        #self.worker_ = pslib.WorkerParameter()
+        self.worker_ = pslib.DownpourTrainerParameter()
+        #self.worker_.pull_dense_per_batch = window
+        #self.worker_.push_dense_per_batch = window
+        #self.worker_.downpour_worker_param.pull_dense_per_batch = window
+        #self.worker_.downpour_worker_param.push_dense_per_batch = window
+        self.worker_.pull_dense_per_batch = window
+        self.worker_.push_dense_per_batch = window
+        print(self.worker_)
+
+    def add_sparse_table(self, table_id, 
+                         slot_keys, slot_value_vars, slot_grad_vars):
+        #table = self.worker_.sparse_table.add()
+        table = self.worker_.downpour_worker_param.sparse_table.add()
+        table.table_id = table_id
+        table.slot.extend(slot_keys)
+        self.worker_.extend([grad.name for grad in slot_grad_vars])
+
+    def add_dense_table(self, table_id, param_vars, grad_vars):
+        #table = self.worker_.dense_table.add()
+        table = self.worker_.downpour_worker_param.dense_table.add()
+        table.table_id = table_id
+        table.dense_variable_name.extend([p.name for p in param_vars])
+        table.dense_gradient_variable_name.extend([g.name for g in grad_vars])
+
+    def get_desc(self):
+        return self.worker_