From 71c2b296eb6537439917781a6b38f271b3eba9a9 Mon Sep 17 00:00:00 2001
From: typhoonzero <typhoonzero1986@gmail.com>
Date: Fri, 20 Oct 2017 22:03:51 +0800
Subject: [PATCH] update

---
 python/paddle/v2/parameters.py |  4 ++++
 python/paddle/v2/topology.py   | 27 +++++++++++++++++++++++++++
 python/paddle/v2/trainer.py    |  5 +++++
 3 files changed, 36 insertions(+)

diff --git a/python/paddle/v2/parameters.py b/python/paddle/v2/parameters.py
index 4cfd91882..d0b5ff12f 100644
--- a/python/paddle/v2/parameters.py
+++ b/python/paddle/v2/parameters.py
@@ -101,6 +101,10 @@ class Parameters(object):
 
         self.__param_conf__[param_conf.name] = param_conf
 
+    def update_param_conf(self, model_config):
+        for p in model_config.parameters:
+            self.__param_conf__[p.name] = p
+
     def keys(self):
         """
         keys are the names of each parameter.
diff --git a/python/paddle/v2/topology.py b/python/paddle/v2/topology.py
index 2db66be25..8dbe944ae 100644
--- a/python/paddle/v2/topology.py
+++ b/python/paddle/v2/topology.py
@@ -19,6 +19,7 @@ import paddle.trainer_config_helpers as conf_helps
 import layer as v2_layer
 import config_base
 import cPickle
+from paddle.trainer import config_parser as cp
 
 __all__ = ['Topology']
 
@@ -50,6 +51,32 @@ class Topology(object):
 
         assert isinstance(self.__model_config__, ModelConfig)
 
+    def update_from_default(self):
+        # HACK(typhoonzero): update ParameterConfig(proto) in case of optimizers
+        # are defined after layers, or between layers.
+        # Must be called from trainer.__init__()
+        for parameter in self.__model_config__.parameters:
+            print "####", parameter.decay_rate, cp.g_default_decay_rate
+            if parameter.momentum == 0.0 and cp.g_default_momentum:
+                parameter.momentum = cp.g_default_momentum
+            if parameter.decay_rate == 0.0 and cp.g_default_decay_rate:
+                parameter.decay_rate = cp.g_default_decay_rate
+            if parameter.initial_mean == 0.0:
+                parameter.initial_mean = cp.g_default_initial_mean
+            if parameter.initial_std == 0.01:
+                parameter.initial_std = cp.g_default_initial_std
+            if parameter.initial_strategy == 0:
+                parameter.initial_strategy = cp.g_default_initial_strategy
+            if parameter.initial_smart == False:
+                parameter.initial_smart = cp.g_default_initial_smart
+            if parameter.num_batches_regularization == 1 and cp.g_default_num_batches_regularization:
+                parameter.num_batches_regularization = cp.g_default_num_batches_regularization
+            if parameter.gradient_clipping_threshold == 0.0 and cp.g_default_gradient_clipping_threshold:
+                parameter.gradient_clipping_threshold = cp.g_default_gradient_clipping_threshold
+            if parameter.device == -1 and cp.g_default_device:
+                parameter.device = cp.g_default_device
+            # FIXME(typhoonzero): ignored: update_hooks, g_default_compact_func
+
     def use_sparse_updater(self):
         """
         check if any parameter require to use sparse_update
diff --git a/python/paddle/v2/trainer.py b/python/paddle/v2/trainer.py
index 076e75593..d937d182b 100644
--- a/python/paddle/v2/trainer.py
+++ b/python/paddle/v2/trainer.py
@@ -64,6 +64,11 @@ class SGD(object):
                             "paddle.v2.optimizer.Optimizer")
         import py_paddle.swig_paddle as api
         topology = Topology(cost, extra_layers=extra_layers)
+        # HACK(typhoonzero): update ParameterConfig(proto) in case of optimizers
+        # are defined after layers, or between layers.
+        topology.update_from_default()
+        parameters.update_param_conf(topology.proto())
+
         self.__optimizer__ = update_equation
         self.__topology__ = topology
         self.__parameters__ = parameters
-- 
GitLab