diff --git a/doc/fluid/api/gen_doc.sh b/doc/fluid/api/gen_doc.sh old mode 100755 new mode 100644 diff --git a/paddle/fluid/framework/lod_tensor.h b/paddle/fluid/framework/lod_tensor.h index dee505fee0dccd8d60bb290a8bec4df243e504a2..4f130d265900483ec7a7c541f2610d17a352913f 100644 --- a/paddle/fluid/framework/lod_tensor.h +++ b/paddle/fluid/framework/lod_tensor.h @@ -142,6 +142,7 @@ class LoDTensor : public Tensor { return (lod_)[level].size() - 1; } + // Split LoDTensor and copy to each place specified in places. std::vector SplitLoDTensor( const std::vector places) const; diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index 17885143247f0e0db8f12931e3c3412e7114ef3d..7be93fa6002ae93c3e1b75c8f7fe5ca5f40b271f 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -150,13 +150,30 @@ void ParallelExecutor::BCastParamsToGPUs( #endif } -void ParallelExecutor::Run(const std::vector &fetch_tensors, - const std::string &fetched_var_name) { +void ParallelExecutor::Run( + const std::vector &fetch_tensors, + const std::string &fetched_var_name, + const std::unordered_map &feed_tensors) { platform::RecordBlock b(0); + SplitTensorToPlaces(feed_tensors); auto fetch_data = member_->executor_->Run(fetch_tensors); *member_->global_scope_->Var(fetched_var_name)->GetMutable() = fetch_data; } +void ParallelExecutor::SplitTensorToPlaces( + const std::unordered_map &feed_tensors) { + for (auto it : feed_tensors) { + auto lod_tensors = it.second.SplitLoDTensor(member_->places_); + for (size_t j = 0; j < member_->places_.size(); ++j) { + // TODO(panxy0718): Do I need to delete this var? + member_->local_scopes_[j] + ->Var(it.first) + ->GetMutable() + ->ShareDataWith(lod_tensors[j]); + } + } +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/parallel_executor.h b/paddle/fluid/framework/parallel_executor.h index 964b476234e622cae934d41bc3793bc3114a5f1a..c7c58b2b808383621a6d492f9188b0d36bfa6858 100644 --- a/paddle/fluid/framework/parallel_executor.h +++ b/paddle/fluid/framework/parallel_executor.h @@ -42,9 +42,13 @@ class ParallelExecutor { bool allow_op_delay); void Run(const std::vector& fetch_tensors, - const std::string& fetched_var_name = "fetched_var"); + const std::string& fetched_var_name, + const std::unordered_map& feed_tensors); private: + void SplitTensorToPlaces( + const std::unordered_map& feed_tensors); + ParallelExecutorPrivate* member_; void BCastParamsToGPUs(const ProgramDesc& startup_program) const; diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py index a2c830b3c943b114f3024f23f73f78bf87e1da34..4153049c05a71fe827abfefd68b31b1d095b5782 100644 --- a/python/paddle/fluid/parallel_executor.py +++ b/python/paddle/fluid/parallel_executor.py @@ -26,25 +26,29 @@ class ParallelExecutor(object): use_cuda, num_threads=None, allow_op_delay=False): - places = [] + self._places = [] + self._act_places = [] if use_cuda: for i in xrange(core.get_cuda_device_count()): p = core.Place() - p.set_place(core.CUDAPlace(i)) - places.append(p) + self._act_places.append(core.CUDAPlace(i)) + p.set_place(self._act_places[-1]) + self._places.append(p) else: for i in xrange(multiprocessing.cpu_count()): p = core.Place() - p.set_place(core.CPUPlace()) - places.append(p) + self._act_places.append(core.CPUPlace(i)) + p.set_place(self._act_places[-1]) + self._places.append(p) + assert self._places, "no place for execution" if num_threads is None: if use_cuda: # Experiments on se-resnext shows that too many threads hurt # performance. Worth tunning for other models in the future. - num_threads = len(places) + num_threads = len(self._places) else: - min(len(places) * 2, multiprocessing.cpu_count()) + min(len(self._places) * 2, multiprocessing.cpu_count()) startup = framework.default_startup_program() main = framework.default_main_program() @@ -53,7 +57,7 @@ class ParallelExecutor(object): self.executor = core.ParallelExecutor( num_threads, True if use_cuda else False, # use_event - places, + self._places, set([ p.name for p in main.global_block().iter_parameters() if not p.stop_gradient @@ -65,8 +69,22 @@ class ParallelExecutor(object): allow_op_delay) self.scope = scope - def run(self, fetch_list): + def run(self, fetch_list, feed_dict={}): + """ + :param fetch_list: A list of variable names that will be fetched. + :param feed_dict: A dict mapping for feed variable name to LoDTensor + or numpy array. + :return: fetched value list. + """ + feed_tensor_dict = {} + for i, feed_name in enumerate(feed_dict): + feed_tensor = feed_dict[feed_name] + if not isinstance(feed_tensor, core.LoDTensor): + feed_tensor = core.LoDTensor() + feed_tensor.set(feed_dict[feed_name], self._act_places[0]) + feed_tensor_dict[feed_name] = feed_tensor + fetch_var_name = '@FETCHED_VAR_NAME@' - self.executor.run(fetch_list, fetch_var_name) + self.executor.run(fetch_list, fetch_var_name, feed_tensor_dict) arr = self.scope.find_var(fetch_var_name).get_lod_tensor_array() return [arr[i] for i in range(len(arr))] diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor.py b/python/paddle/fluid/tests/unittests/test_parallel_executor.py index a79e4b3e183eaef06be27a724893799923e84ac1..0cbef82e33c306e2b89ea4f3d66d48da3840ccbd 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor.py @@ -21,13 +21,17 @@ import paddle.dataset.mnist as mnist import paddle.dataset.wmt16 as wmt16 -def simple_fc_net(): - reader = fluid.layers.open_recordio_file( - filename='./mnist.recordio', - shapes=[[-1, 784], [-1, 1]], - lod_levels=[0, 0], - dtypes=['float32', 'int64']) - img, label = fluid.layers.read_file(reader) +def simple_fc_net(use_feed): + if use_feed: + img = fluid.layers.data(name='image', shape=[784], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + else: + reader = fluid.layers.open_recordio_file( + filename='./mnist.recordio', + shapes=[[-1, 784], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64']) + img, label = fluid.layers.read_file(reader) hidden = img for _ in xrange(4): hidden = fluid.layers.fc( @@ -42,13 +46,18 @@ def simple_fc_net(): return loss -def fc_with_batchnorm(): - reader = fluid.layers.open_recordio_file( - filename='./mnist.recordio', - shapes=[[-1, 784], [-1, 1]], - lod_levels=[0, 0], - dtypes=['float32', 'int64']) - img, label = fluid.layers.read_file(reader) +def fc_with_batchnorm(use_feed): + if use_feed: + img = fluid.layers.data(name='image', shape=[784], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + else: + reader = fluid.layers.open_recordio_file( + filename='./mnist.recordio', + shapes=[[-1, 784], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64']) + img, label = fluid.layers.read_file(reader) + hidden = img for _ in xrange(1): hidden = fluid.layers.fc( @@ -135,7 +144,9 @@ def bottleneck_block(input, num_filters, stride, cardinality, reduction_ratio): return fluid.layers.elementwise_add(x=short, y=scale, act='relu') -def SE_ResNeXt152Small(batch_size=2): +def SE_ResNeXt152Small(batch_size=2, use_feed=False): + assert not use_feed, "SE_ResNeXt doesn't support feed yet" + img = fluid.layers.fill_constant( shape=[batch_size, 3, 224, 224], dtype='float32', value=0.0) label = fluid.layers.fill_constant( @@ -185,30 +196,28 @@ class TestParallelExecutorBase(unittest.TestCase): memory_opt=True, iter=10, batch_size=None, - allow_op_delay=False): + allow_op_delay=False, + feed_dict={}): main = fluid.Program() startup = fluid.Program() with fluid.program_guard(main, startup): - loss = method() + loss = method(use_feed=len(feed_dict) > 0) adam = fluid.optimizer.Adam() adam.minimize(loss) if memory_opt: fluid.memory_optimize(main) - exe = fluid.ParallelExecutor( - loss_name=loss.name, - use_cuda=True, - allow_op_delay=allow_op_delay) + exe = fluid.ParallelExecutor(loss_name=loss.name, use_cuda=True) if batch_size is not None: batch_size *= fluid.core.get_cuda_device_count() begin = time.time() - first_loss, = exe.run([loss.name]) + first_loss, = exe.run([loss.name], feed_dict=feed_dict) first_loss = numpy.array(first_loss) for i in xrange(iter): - exe.run([]) + exe.run([], feed_dict=feed_dict) - last_loss, = exe.run([loss.name]) + last_loss, = exe.run([loss.name], feed_dict=feed_dict) end = time.time() if batch_size is not None: @@ -242,9 +251,19 @@ class TestMNIST(TestParallelExecutorBase): self.check_network_convergence(simple_fc_net) self.check_network_convergence(simple_fc_net, allow_op_delay=True) + img = numpy.zeros(shape=[32, 784], dtype='float32') + label = numpy.ones(shape=[32, 1], dtype='int64') + self.check_network_convergence( + simple_fc_net, feed_dict={"image": img, + "label": label}) + def test_batchnorm_fc(self): self.check_network_convergence(fc_with_batchnorm) - self.check_network_convergence(fc_with_batchnorm, allow_op_delay=True) + img = numpy.zeros(shape=[32, 784], dtype='float32') + label = numpy.ones(shape=[32, 1], dtype='int64') + self.check_network_convergence( + fc_with_batchnorm, feed_dict={"image": img, + "label": label}) class TestResnet(TestParallelExecutorBase): @@ -400,7 +419,8 @@ def prepare_batch_input(insts, src_pad_idx, trg_pad_idx, n_head): import transformer_model -def transformer(): +def transformer(use_feed): + assert not use_feed, "transfomer doesn't support feed yet" return transformer_model.transformer( ModelHyperParams.src_vocab_size + 1, ModelHyperParams.trg_vocab_size + 1, ModelHyperParams.max_length + 1,