From 22b9b6662b215b663ce2cebdf7624ea1212bb9c1 Mon Sep 17 00:00:00 2001
From: Yu Yang <yuyang18@baidu.com>
Date: Thu, 15 Dec 2016 21:01:47 +0800
Subject: [PATCH] Add unittest to coverage SgdThreadUpdater's enableBufType

---
 paddle/trainer/ThreadParameterUpdater.cpp     |  3 +++
 paddle/trainer/tests/fake_file_list.list      |  1 +
 .../tests/simple_sparse_neural_network.py     | 23 +++++++++++++++++++
 .../tests/simple_sparse_neural_network_dp.py  | 21 +++++++++++++++++
 paddle/trainer/tests/test_TrainerOnePass.cpp  |  9 +++++++-
 5 files changed, 56 insertions(+), 1 deletion(-)
 create mode 100644 paddle/trainer/tests/fake_file_list.list
 create mode 100644 paddle/trainer/tests/simple_sparse_neural_network.py
 create mode 100644 paddle/trainer/tests/simple_sparse_neural_network_dp.py

diff --git a/paddle/trainer/ThreadParameterUpdater.cpp b/paddle/trainer/ThreadParameterUpdater.cpp
index 9caa92a4d75..049022b1f10 100644
--- a/paddle/trainer/ThreadParameterUpdater.cpp
+++ b/paddle/trainer/ThreadParameterUpdater.cpp
@@ -55,6 +55,9 @@ void SgdThreadUpdater::init(std::vector<ParameterPtr>& parameters) {
       // not create parameter buf for PARAMETER_GRADIENT for sparse update in
       // Parameter::enableType(). But gradient parameter buf is still used
       // in SgdThreadUpdater. We need to explicitly create it.
+      //
+      // The AverageOptimizer::restore/apply method will use PARAMETER_GRADIENT
+      // as a temp buffer.
       para->enableBufType(PARAMETER_GRADIENT);
     }
   }
diff --git a/paddle/trainer/tests/fake_file_list.list b/paddle/trainer/tests/fake_file_list.list
new file mode 100644
index 00000000000..f27ceed277f
--- /dev/null
+++ b/paddle/trainer/tests/fake_file_list.list
@@ -0,0 +1 @@
+do_not_matter.txt
diff --git a/paddle/trainer/tests/simple_sparse_neural_network.py b/paddle/trainer/tests/simple_sparse_neural_network.py
new file mode 100644
index 00000000000..9604e1b9b45
--- /dev/null
+++ b/paddle/trainer/tests/simple_sparse_neural_network.py
@@ -0,0 +1,23 @@
+from paddle.trainer_config_helpers import *
+
+settings(batch_size=128, learning_method=AdaGradOptimizer(), learning_rate=1e-4)
+
+file_list = 'trainer/tests/fake_file_list.list'
+
+define_py_data_sources2(
+    train_list=file_list,
+    test_list=file_list,
+    module="simple_sparse_neural_network_dp",
+    obj="process")
+
+embedding = embedding_layer(
+    input=data_layer(
+        name="word_ids", size=65536),
+    size=128,
+    param_attr=ParamAttr(sparse_update=True))
+prediction = fc_layer(input=embedding, size=10, act=SoftmaxActivation())
+
+outputs(
+    classification_cost(
+        input=prediction, label=data_layer(
+            name='label', size=10)))
diff --git a/paddle/trainer/tests/simple_sparse_neural_network_dp.py b/paddle/trainer/tests/simple_sparse_neural_network_dp.py
new file mode 100644
index 00000000000..8bfd1f37e71
--- /dev/null
+++ b/paddle/trainer/tests/simple_sparse_neural_network_dp.py
@@ -0,0 +1,21 @@
+from paddle.trainer.PyDataProvider2 import provider, integer_sequence, integer_value
+import random
+
+
+def init_hook(settings, is_train, **kwargs):
+    settings.is_train = is_train
+
+
+@provider(
+    input_types={'word_ids': integer_value(65536),
+                 'label': integer_value(10)},
+    min_pool_size=0,
+    init_hook=init_hook)
+def process(settings, filename):
+    if settings.is_train:
+        data_size = 2**20
+    else:
+        data_size = 2**10
+
+    for _ in xrange(data_size):
+        yield random.randint(0, 65535), random.randint(0, 9)
diff --git a/paddle/trainer/tests/test_TrainerOnePass.cpp b/paddle/trainer/tests/test_TrainerOnePass.cpp
index ee21008aec5..4d0174f784a 100644
--- a/paddle/trainer/tests/test_TrainerOnePass.cpp
+++ b/paddle/trainer/tests/test_TrainerOnePass.cpp
@@ -27,6 +27,9 @@ static const string& configFile1 = "trainer/tests/sample_trainer_config.conf";
 static const string& configFile2 =
     "trainer/tests/sample_trainer_config_parallel.conf";
 
+static const string& configFileSimpleSparse =
+    "trainer/tests/simple_sparse_neural_network.py";
+
 DECLARE_bool(use_gpu);
 DECLARE_string(config);
 DECLARE_int32(gpu_id);
@@ -298,11 +301,15 @@ TEST(checkRemoteUpdater, cpuDeltaTrainerOldUpdater) {
   checkRemoteParameterUpdaterTest(configFile1, false, false, 1, true, 10);
 }
 
+TEST(SgdThreadUpdater, simpleSparseNN) {
+  trainerOnePassTest(configFileSimpleSparse, false, false, 1, 0.5, true);
+}
+
 int main(int argc, char** argv) {
+  testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);
   initPython(argc, argv);
   gNumDevices = hl_get_device_count();
-  testing::InitGoogleTest(&argc, argv);
 
   FLAGS_num_passes = 1;          // train one pass
   FLAGS_saving_period = 100000;  // do not save parameteres
-- 
GitLab