Merge pull request #15926 from dzhwinter/test/add_ir_mem_opt_tests

add ir memory optimize test base

python/paddle/fluid/tests/unittests/ir_memory_optimize_net_base.py

@@ -0,0 +1,150 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+
+import six
+import unittest
+import time
+import math
+import multiprocessing
+import numpy as np
+
+import paddle
+import paddle.fluid.core as core
+import paddle.fluid as fluid
+from paddle.fluid import compiler
+
+# open eager delete mode
+os.environ['FLAGS_eager_delete_tensor_gb'] = '0.0'
+os.environ['FLAGS_fast_eager_deletion_mode'] = 'true'
+os.environ['CPU_NUM'] = '2'
+
+
+class BuildIrMemOptBase(unittest.TestCase):
+    def check_network_convergence(self,
+                                  network,
+                                  use_cuda=True,
+                                  memory_opt=True,
+                                  use_ir_memory_optimize=True,
+                                  enable_inplace=True,
+                                  iter=5):
+        if use_cuda and not core.is_compiled_with_cuda():
+            print('Skip use_cuda=True because Paddle is not compiled with cuda')
+            return
+
+        if os.name == 'nt':
+            print(
+                'Skip use_parallel_executor=True because Paddle comes without parallel support on windows'
+            )
+            return
+        fluid.default_startup_program().random_seed = 100
+        fluid.default_main_program().random_seed = 100
+        batch_size = 32
+        batch_size *= fluid.core.get_cuda_device_count() if use_cuda else int(
+            os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+
+        # build network
+        word_dict = paddle.dataset.imdb.word_dict()
+        train_reader = paddle.batch(
+            paddle.dataset.imdb.train(word_dict), batch_size=batch_size)
+
+        data = fluid.layers.data(
+            name="words", shape=[1], dtype="int64", lod_level=1)
+
+        label = fluid.layers.data(name="label", shape=[1], dtype="int64")
+
+        cost = network(data, label, len(word_dict))
+        optimizer = fluid.optimizer.Adam(learning_rate=0.001)
+        optimizer.minimize(cost)
+        if memory_opt:
+            fluid.memory_optimize(fluid.default_main_program())
+
+        # execution
+        place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
+        feeder = fluid.DataFeeder(feed_list=[data, label], place=place)
+        reader = feeder.decorate_reader(train_reader, multi_devices=True)
+        exe = fluid.Executor(place)
+        exe.run(fluid.default_startup_program())
+
+        train_cp = compiler.CompiledProgram(fluid.default_main_program())
+        train_cp = train_cp.with_data_parallel(loss_name=cost.name)
+        fetch_list = [cost.name]
+
+        begin = time.time()
+        first_loss, last_loss = None, None
+        step_id = 0
+        custom_iter = getattr(self, "iter", None)
+        if not custom_iter == None:
+            iter = custom_iter
+        for data in reader():
+            ret = exe.run(train_cp, feed=data, fetch_list=fetch_list)
+            print(ret)
+            step_id += 1
+            if step_id == 1:
+                first_loss = ret[0]
+            if step_id == iter:
+                last_loss = ret[0]
+                break
+        end = time.time()
+
+        print("%.4f Instance per second" % (
+            (batch_size * iter) / (end - begin)))
+
+        print(first_loss, last_loss)
+        avg_last_loss_val = np.array(last_loss).mean()
+        avg_first_loss_val = np.array(first_loss).mean()
+        if math.isnan(float(avg_last_loss_val)) or math.isnan(
+                float(avg_first_loss_val)):
+            sys.exit("got NaN loss, training failed.")
+
+        return first_loss, last_loss
+
+
+class TestIrMemOptBase(BuildIrMemOptBase):
+    def setUp(self):
+        self.network = None
+
+    def test_network(self):
+        if self.network is None or not core.is_compiled_with_cuda():
+            return
+
+        baseline_first_loss, baseline_last_loss = None, None
+        for use_cuda in [True]:
+            for use_python_mem_opt in [True, False]:
+                print(
+                    'network: {}, use_cuda: {}, use_python_mem_opt: {}, use_ir_mem_opt : {}'.
+                    format(self.network.__name__, use_cuda, use_python_mem_opt,
+                           not use_python_mem_opt))
+                with fluid.program_guard(fluid.Program(), fluid.Program()):
+                    with fluid.scope_guard(core.Scope()):
+                        if use_cuda is True and use_python_mem_opt is True:
+                            baseline_first_loss, baseline_last_loss = self.check_network_convergence(
+                                self.network,
+                                use_cuda=use_cuda,
+                                memory_opt=use_python_mem_opt)
+                        else:
+                            cur_first_loss, cur_last_loss = self.check_network_convergence(
+                                self.network,
+                                use_cuda=use_cuda,
+                                memory_opt=use_python_mem_opt)
+
+                            self.assertAlmostEquals(
+                                np.mean(baseline_last_loss),
+                                np.mean(cur_last_loss),
+                                delta=1e-2)
+                            self.assertAlmostEquals(
+                                np.mean(baseline_first_loss),
+                                np.mean(cur_first_loss),
+                                delta=1e-2)

python/paddle/fluid/tests/unittests/test_eager_deletion_dynamic_rnn_base.py

@@ -56,6 +56,8 @@ def train(network, use_cuda, use_parallel_executor, batch_size=32, pass_num=2):
         train_reader, multi_devices=use_parallel_executor)
 
     exe = fluid.Executor(place)
+    fluid.default_startup_program().random_seed = 1
+    fluid.default_main_program().random_seed = 1
     exe.run(fluid.default_startup_program())
 
     train_cp = compiler.CompiledProgram(fluid.default_main_program())

python/paddle/fluid/tests/unittests/test_ir_memory_optimize_nlp.py

@@ -0,0 +1,55 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# nlp model stack of op operate on lod. It's a classical test case in optimize pass.
+
+from __future__ import print_function
+
+import paddle.fluid as fluid
+import unittest
+from ir_memory_optimize_net_base import TestIrMemOptBase
+
+
+def lstm_net(data,
+             label,
+             dict_dim,
+             emb_dim=128,
+             hid_dim=128,
+             hid_dim2=96,
+             class_dim=2,
+             emb_lr=30.0):
+    emb = fluid.layers.embedding(
+        input=data,
+        size=[dict_dim, emb_dim],
+        param_attr=fluid.ParamAttr(learning_rate=emb_lr))
+    fc0 = fluid.layers.fc(input=emb, size=hid_dim * 4)
+
+    lstm_h, c = fluid.layers.dynamic_lstm(
+        input=fc0, size=hid_dim * 4, is_reverse=False)
+    lstm_max = fluid.layers.sequence_pool(input=lstm_h, pool_type='max')
+    lstm_max_tanh = fluid.layers.tanh(lstm_max)
+    fc1 = fluid.layers.fc(input=lstm_max_tanh, size=hid_dim2, act='tanh')
+    prediction = fluid.layers.fc(input=fc1, size=class_dim, act='softmax')
+    cost = fluid.layers.cross_entropy(input=prediction, label=label)
+    avg_cost = fluid.layers.mean(x=cost)
+    return avg_cost
+
+
+class TestIrMemOptRNN(TestIrMemOptBase):
+    def setUp(self):
+        self.network = lstm_net
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/test_ir_memory_optimize_transformer.py

@@ -28,9 +28,6 @@ os.environ[
 from test_parallel_executor_transformer import transformer, ModelHyperParams, transformer_model, transformer, prepare_batch_input
 from parallel_executor_test_base import TestParallelExecutorBase
 
-# disable temporarily because of timeout.
-sys.exit(0)
-
 
 # NOTE(dzhwinter): test diferent strategy colisions.
 # open the eager delete tensor strategy by default.