Merge pull request #12374 from chenwhql/py_calc_memory

Add memory usage estimate API 

paddle/fluid/API.spec

@@ -336,6 +336,7 @@ paddle.fluid.contrib.BeamSearchDecoder.decode ArgSpec(args=['self'], varargs=Non
 paddle.fluid.contrib.BeamSearchDecoder.early_stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.contrib.BeamSearchDecoder.read_array ArgSpec(args=['self', 'init', 'is_ids', 'is_scores'], varargs=None, keywords=None, defaults=(False, False))
 paddle.fluid.contrib.BeamSearchDecoder.update_array ArgSpec(args=['self', 'array', 'value'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.contrib.memory_usage ArgSpec(args=['program', 'batch_size'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.transpiler.DistributeTranspiler.create_splited_vars ArgSpec(args=['self', 'source_var', 'block', 'tag'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)

python/paddle/fluid/contrib/__init__.py

@@ -14,5 +14,7 @@
 
 import decoder
 from decoder import *
+import memory_usage_calc
+from memory_usage_calc import *
 
-__all__ = decoder.__all__
+__all__ = decoder.__all__ + memory_usage_calc.__all__

python/paddle/fluid/contrib/memory_usage_calc.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+#
+# 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.
+"""
+This module privides a memory usage calculate function for user.
+The purpose of this API is to allow users to estimate memory usage of
+a program under a special batch size, then user can set appropriate 
+batch size to fully utilize a GPU. 
+
+This API is still under active development and may change drastically.
+"""
+
+from .. import core
+from ..framework import Program, Variable
+
+__all__ = ['memory_usage']
+
+dtype_to_size = {
+    core.VarDesc.VarType.FP16: 2,
+    core.VarDesc.VarType.FP32: 4,
+    core.VarDesc.VarType.FP64: 8,
+    core.VarDesc.VarType.INT16: 2,
+    core.VarDesc.VarType.INT32: 4,
+    core.VarDesc.VarType.INT64: 8,
+    core.VarDesc.VarType.BOOL: 1,
+    core.VarDesc.VarType.UINT8: 1,
+}
+
+DEBUG = False
+
+
+def memory_usage(program, batch_size):
+    """
+    Get the estimate memory usage of program with input batch size.
+
+    Args:
+        program(Program): The current Program.
+        batch_size(int): The current input data batch_size.  
+    
+    Returns:
+        min_total_memory(float): the estimate memory usage lower bound.
+        max_total_memory(float): the estimate memory usage upper bound.
+        unit_str(string): the unit of estimate usage result.
+    
+    Examples:
+        
+        >>> import paddle.fluid as fluid
+        >>> lower_usage, upper_usage, unit = fluid.contrib.memory_usage(
+                fluid.default_main_program(), batch_size=10)
+        >>> print "memory usage is about %.3f - %.3f %s" % \
+                (lower_usage, upper_usage, unit)
+
+    """
+
+    # Parameters check
+    if not isinstance(program, Program):
+        raise TypeError(
+            "Calculating Memory Usage requires Program as its Parameter."
+            "But you passed in %s" % (type(prgram)))
+    if batch_size <= 0:
+        raise ValueError("The batch size need to be positive.")
+
+    # Get the var_name list of first block and calculate
+    total_memory = 0.0
+    for var in program.global_block().vars.itervalues():
+        data_count = 1
+        for x in var.shape:
+            if x == -1:
+                data_count *= batch_size
+            else:
+                data_count *= x
+        var_memory = data_count * dtype_to_size[var.dtype]
+        if DEBUG:
+            print "%s memory usage: %d" % (var.name, var_memory)
+        total_memory += var_memory
+    if DEBUG:
+        print "total memory usage: %.2f" % (total_memory)
+
+    # Convert appropriate unit
+    unit_str = "B"
+    if total_memory > 1024:
+        total_memory /= 1024
+        unit_str = "KB"
+        if total_memory > 1024:
+            total_memory /= 1024
+            unit_str = "MB"
+
+    # Append extra memory consumption (5% - 10%)
+    min_total_memory = total_memory * 1.05
+    max_total_memory = total_memory * 1.1
+
+    return min_total_memory, max_total_memory, unit_str

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

+#   Copyright (c) 2018 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.
+
+from __future__ import print_function
+import paddle
+import paddle.fluid as fluid
+import contextlib
+import unittest
+
+
+def train_simulator(test_batch_size=10):
+    if test_batch_size <= 0:
+        raise ValueError("batch_size should be a positive integeral value, "
+                         "but got batch_size={}".format(test_batch_size))
+
+    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(cost)
+
+    sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
+    sgd_optimizer.minimize(avg_cost)
+
+    # Calculate memory usage in current network config 
+    lower_usage, upper_usage, unit = fluid.contrib.memory_usage(
+        fluid.default_main_program(), batch_size=test_batch_size)
+
+    print("memory usage is about %.3f - %.3f %s" %
+          (lower_usage, upper_usage, unit))
+
+
+class TestMemoryUsage(unittest.TestCase):
+    def test_with_unit_B(self):
+        with self.program_scope_guard():
+            train_simulator()
+
+    def test_with_unit_KB(self):
+        with self.program_scope_guard():
+            train_simulator(test_batch_size=1000)
+
+    def test_with_unit_MB(self):
+        with self.program_scope_guard():
+            train_simulator(test_batch_size=100000)
+
+    @contextlib.contextmanager
+    def program_scope_guard(self):
+        prog = fluid.Program()
+        startup_prog = fluid.Program()
+        scope = fluid.core.Scope()
+        with fluid.scope_guard(scope):
+            with fluid.program_guard(prog, startup_prog):
+                yield
+
+
+if __name__ == '__main__':
+    unittest.main()