add memory optimization transpiler demo (#7443)

* add memory optimization transpiler demo

* add memory benchmark compile option

* add gflags instead of macro

* refine code

paddle/framework/executor.cc

@@ -23,6 +23,7 @@ limitations under the License. */
 #include "paddle/framework/op_registry.h"
 #include "paddle/platform/place.h"
 
+DECLARE_bool(do_memory_benchmark);
 DEFINE_bool(check_nan_inf, false,
             "Checking whether operator produce NAN/INF or not. It will be "
             "extremely slow so please use this flag wisely.");
@@ -117,6 +118,10 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
     VLOG(3) << op->DebugStringEx(local_scope);
     op->Run(*local_scope, place_);
+    if (FLAGS_do_memory_benchmark) {
+      VLOG(2) << "Memory used after operator " + op->Type() + " running: "
+              << memory::memory_usage(place_);
+    }
     if (FLAGS_check_nan_inf) {
       for (auto& vname : op->OutputVars(true)) {
         auto* var = local_scope->FindVar(vname);
@@ -130,6 +135,12 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
   if (create_vars && create_local_scope) {
     scope->DeleteScope(local_scope);
   }
+  if (FLAGS_do_memory_benchmark) {
+    VLOG(2) << "-------------------------------------------------------";
+    VLOG(2) << "Memory used after deleting local scope: "
+            << memory::memory_usage(place_);
+    VLOG(2) << "-------------------------------------------------------";
+  }
 }
 
 }  // namespace framework

paddle/framework/scope.cc

@@ -20,6 +20,10 @@ limitations under the License. */
 #include "paddle/framework/threadpool.h"
 #include "paddle/string/printf.h"
 
+DEFINE_bool(do_memory_benchmark, false,
+            "Doing memory benchmark. It will make deleting scope synchronized, "
+            "and add some memory usage logs");
+
 namespace paddle {
 namespace framework {
 
@@ -88,8 +92,12 @@ void Scope::DeleteScope(Scope* scope) {
   auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
   PADDLE_ENFORCE(it != this->kids_.end(), "Cannot find %p as kid scope", scope);
   this->kids_.erase(it);
-  // Make delete async.
-  Async([scope] { delete scope; });
+  // When making memory benchmark on Fluid, we have to delete scope sync.
+  if (FLAGS_do_memory_benchmark) {
+    delete scope;
+  } else {
+    Async([scope] { delete scope; });
+  }
 }
 
 void Scope::Rename(const std::string& origin_name,

python/paddle/v2/fluid/__init__.py

@@ -86,7 +86,9 @@ def __bootstrap__():
 
     os.environ['OMP_NUM_THREADS'] = str(num_threads)
 
-    read_env_flags = ['use_pinned_memory', 'check_nan_inf']
+    read_env_flags = [
+        'use_pinned_memory', 'check_nan_inf', 'do_memory_benchmark'
+    ]
     if core.is_compile_gpu():
         read_env_flags += ['fraction_of_gpu_memory_to_use', 'op_sync']
     core.init_gflags([sys.argv[0]] +

python/paddle/v2/fluid/tests/CMakeLists.txt

@@ -6,3 +6,4 @@ endforeach()
 
 add_subdirectory(book)
 add_subdirectory(book_distribute)
+add_subdirectory(book_memory_optimization)

python/paddle/v2/fluid/tests/book_memory_optimization/CMakeLists.txt

@@ -0,0 +1,11 @@
+file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
+string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
+
+list(REMOVE_ITEM TEST_OPS test_memopt_image_classification_train)
+py_test(test_memopt_image_classification_train_resnet SRCS test_memopt_image_classification_train.py ARGS resnet)
+py_test(test_memopt_image_classification_train_vgg SRCS test_memopt_image_classification_train.py ARGS vgg)
+
+# default test
+foreach(src ${TEST_OPS})
+    py_test(${src} SRCS ${src}.py)
+endforeach()

python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py

@@ -0,0 +1,44 @@
+import numpy as np
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+
+x = fluid.layers.data(name='x', shape=[13], dtype='float32')
+
+y_predict = fluid.layers.fc(input=x, size=1, act=None)
+
+y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+
+cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+avg_cost = fluid.layers.mean(x=cost)
+
+sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1)
+sgd_optimizer.minimize(avg_cost)
+
+# memopt_program = fluid.default_main_program()
+memopt_program = fluid.memory_optimize(fluid.default_main_program())
+
+BATCH_SIZE = 200
+
+train_reader = paddle.batch(
+    paddle.reader.shuffle(
+        paddle.dataset.uci_housing.train(), buf_size=500),
+    batch_size=BATCH_SIZE)
+
+place = fluid.CPUPlace()
+feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
+exe = fluid.Executor(place)
+
+exe.run(fluid.default_startup_program())
+
+PASS_NUM = 100
+for pass_id in range(PASS_NUM):
+    fluid.io.save_persistables(exe, "./fit_a_line.model/")
+    fluid.io.load_persistables(exe, "./fit_a_line.model/")
+    for data in train_reader():
+        avg_loss_value, = exe.run(memopt_program,
+                                  feed=feeder.feed(data),
+                                  fetch_list=[avg_cost])
+
+        if avg_loss_value[0] < 10.0:
+            exit(0)  # if avg cost less than 10.0, we think our code is good.
+exit(1)

python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py

@@ -0,0 +1,133 @@
+from __future__ import print_function
+
+import sys
+
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+
+
+def resnet_cifar10(input, depth=32):
+    def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'):
+        tmp = fluid.layers.conv2d(
+            input=input,
+            filter_size=filter_size,
+            num_filters=ch_out,
+            stride=stride,
+            padding=padding,
+            act=None,
+            bias_attr=False)
+        return fluid.layers.batch_norm(input=tmp, act=act)
+
+    def shortcut(input, ch_in, ch_out, stride):
+        if ch_in != ch_out:
+            return conv_bn_layer(input, ch_out, 1, stride, 0, None)
+        else:
+            return input
+
+    def basicblock(input, ch_in, ch_out, stride):
+        tmp = conv_bn_layer(input, ch_out, 3, stride, 1)
+        tmp = conv_bn_layer(tmp, ch_out, 3, 1, 1, act=None)
+        short = shortcut(input, ch_in, ch_out, stride)
+        return fluid.layers.elementwise_add(x=tmp, y=short, act='relu')
+
+    def layer_warp(block_func, input, ch_in, ch_out, count, stride):
+        tmp = block_func(input, ch_in, ch_out, stride)
+        for i in range(1, count):
+            tmp = block_func(tmp, ch_out, ch_out, 1)
+        return tmp
+
+    assert (depth - 2) % 6 == 0
+    n = (depth - 2) / 6
+    conv1 = conv_bn_layer(
+        input=input, ch_out=16, filter_size=3, stride=1, padding=1)
+    res1 = layer_warp(basicblock, conv1, 16, 16, n, 1)
+    res2 = layer_warp(basicblock, res1, 16, 32, n, 2)
+    res3 = layer_warp(basicblock, res2, 32, 64, n, 2)
+    pool = fluid.layers.pool2d(
+        input=res3, pool_size=8, pool_type='avg', pool_stride=1)
+    return pool
+
+
+def vgg16_bn_drop(input):
+    def conv_block(input, num_filter, groups, dropouts):
+        return fluid.nets.img_conv_group(
+            input=input,
+            pool_size=2,
+            pool_stride=2,
+            conv_num_filter=[num_filter] * groups,
+            conv_filter_size=3,
+            conv_act='relu',
+            conv_with_batchnorm=True,
+            conv_batchnorm_drop_rate=dropouts,
+            pool_type='max')
+
+    conv1 = conv_block(input, 64, 2, [0.3, 0])
+    conv2 = conv_block(conv1, 128, 2, [0.4, 0])
+    conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0])
+    conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0])
+    conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0])
+
+    drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5)
+    fc1 = fluid.layers.fc(input=drop, size=512, act=None)
+    bn = fluid.layers.batch_norm(input=fc1, act='relu')
+    drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5)
+    fc2 = fluid.layers.fc(input=drop2, size=512, act=None)
+    return fc2
+
+
+classdim = 10
+data_shape = [3, 32, 32]
+
+images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
+label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+
+net_type = "vgg"
+if len(sys.argv) >= 2:
+    net_type = sys.argv[1]
+
+if net_type == "vgg":
+    print("train vgg net")
+    net = vgg16_bn_drop(images)
+elif net_type == "resnet":
+    print("train resnet")
+    net = resnet_cifar10(images, 32)
+else:
+    raise ValueError("%s network is not supported" % net_type)
+
+predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
+cost = fluid.layers.cross_entropy(input=predict, label=label)
+avg_cost = fluid.layers.mean(x=cost)
+
+optimizer = fluid.optimizer.Adam(learning_rate=0.001)
+opts = optimizer.minimize(avg_cost)
+
+accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
+
+# memopt_program = fluid.default_main_program()
+memopt_program = fluid.memory_optimize(fluid.default_main_program())
+
+BATCH_SIZE = 128
+PASS_NUM = 1
+
+train_reader = paddle.batch(
+    paddle.reader.shuffle(
+        paddle.dataset.cifar.train10(), buf_size=128 * 10),
+    batch_size=BATCH_SIZE)
+
+place = fluid.CPUPlace()
+exe = fluid.Executor(place)
+feeder = fluid.DataFeeder(place=place, feed_list=[images, label])
+exe.run(fluid.default_startup_program())
+
+for pass_id in range(PASS_NUM):
+    accuracy.reset(exe)
+    for data in train_reader():
+        loss, acc = exe.run(memopt_program,
+                            feed=feeder.feed(data),
+                            fetch_list=[avg_cost] + accuracy.metrics)
+        pass_acc = accuracy.eval(exe)
+        print("loss:" + str(loss) + " acc:" + str(acc) + " pass_acc:" + str(
+            pass_acc))
+        # this model is slow, so if we can train two mini batch, we think it works properly.
+        exit(0)
+exit(1)