diff --git a/doc/v2/design/mkl/mkldnn.md b/doc/v2/design/mkl/mkldnn.md index 1bd2e7bc34ee79eb753b3520d97e5e7beca89b0b..bd5bcf6f67168c21cebb046a629b948d1661e75c 100644 --- a/doc/v2/design/mkl/mkldnn.md +++ b/doc/v2/design/mkl/mkldnn.md @@ -5,7 +5,7 @@ 充分展现英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。
-
+
Figure 1. PaddlePaddle on IA
@@ -42,16 +42,43 @@ Figure 1. PaddlePaddle on IA MKL，MKLML以及MKL-DNN三者关系如下表： -| Name | Open Source | License | Descriptions | -| :---------- | :--------------- | :---------- | :------------ | -| MKL | No | Proprietary | Accelerate math processing routines | -| MKLML | No | Proprietary | Small package of MKL, especially for Machine Learning | -| MKL-DNN | Yes | Apache 2.0 | Accelerate primitives processing routines especially for Deep Neural Networks | + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
NameOpen SourceLicenseDescriptions
MKLNoProprietaryAccelerate math processing routines
MKLMLNoProprietarySmall package of MKL, especially for Machine Learning
MKL-DNNYesApache 2.0Accelerate primitives processing routines especially for Deep Neural Networks
MKLML可以与MKL-DNN共同使用，以此达到最好的性能。
-
+
Figure 2. PaddlePaddle with MKL Engines
@@ -103,7 +130,7 @@ MKL-DNN的库目前只有动态库`libmkldnn.so`。 所以我们定义了一个`MKLDNNMatrix`用于管理MKL-DNN数据的不同格式以及相互之间的转换。
-
+
Figure 3. MKLDNNMatrix
@@ -113,7 +140,7 @@ Figure 3. MKLDNNMatrix 子类只需要使用定义好的接口，实现具体的函数功能即可。
-
+
Figure 4. MKLDNNLayer
@@ -150,7 +177,7 @@ Figure 4. MKLDNNLayer 所以整体上，在实现每个子类的时候就不需要关心分支的事情了。
-
+
Figure 5. Merge Gradients
diff --git a/doc/v2/images/FullyConnected.jpg b/doc/v2/images/FullyConnected.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b2241f401434e527f95ee4e0e541a3f2ff78fd1e Binary files /dev/null and b/doc/v2/images/FullyConnected.jpg differ diff --git a/doc/v2/images/add_security_group.png b/doc/v2/images/add_security_group.png new file mode 100644 index 0000000000000000000000000000000000000000..bd34f46c9b0ada7027fd53e553e7d033255d25fc Binary files /dev/null and b/doc/v2/images/add_security_group.png differ diff --git a/doc/v2/images/bi_lstm.jpg b/doc/v2/images/bi_lstm.jpg new file mode 100644 index 0000000000000000000000000000000000000000..adec1606d64d6e35ffe7e62abfa9a09309b05c84 Binary files /dev/null and b/doc/v2/images/bi_lstm.jpg differ diff --git a/doc/v2/images/checkpointing.png b/doc/v2/images/checkpointing.png new file mode 100644 index 0000000000000000000000000000000000000000..c221e8474f90f37e31416cbb19c9452207a0d14c Binary files /dev/null and b/doc/v2/images/checkpointing.png differ diff --git a/doc/v2/images/create_efs.png b/doc/v2/images/create_efs.png new file mode 100644 index 0000000000000000000000000000000000000000..e5f1526033d1daf401700989af1d25919bcb7675 Binary files /dev/null and b/doc/v2/images/create_efs.png differ diff --git a/doc/v2/images/csr.png b/doc/v2/images/csr.png new file mode 100644 index 0000000000000000000000000000000000000000..3dc10b8de4f6d3f517624956b1694b689405a031 Binary files /dev/null and b/doc/v2/images/csr.png differ diff --git a/doc/v2/images/data_dispatch.png b/doc/v2/images/data_dispatch.png new file mode 100644 index 0000000000000000000000000000000000000000..5bdcc24d6a6d193cb014f8c38b362451fded5e54 Binary files /dev/null and b/doc/v2/images/data_dispatch.png differ diff --git a/doc/v2/images/dataset.graffle b/doc/v2/images/dataset.graffle new file mode 100644 index 0000000000000000000000000000000000000000..c10a423ed16a23229a9ee33d11bfc82bb59646c8 Binary files /dev/null and b/doc/v2/images/dataset.graffle differ diff --git a/doc/v2/images/dataset.png b/doc/v2/images/dataset.png new file mode 100644 index 0000000000000000000000000000000000000000..2fb7f1cce3b6dd21489392557826e95a9f207c34 Binary files /dev/null and b/doc/v2/images/dataset.png differ diff --git a/doc/v2/images/doc_en.png b/doc/v2/images/doc_en.png new file mode 100644 index 0000000000000000000000000000000000000000..ed6b9178fba91a3bdf45ae797a9924f84146fbc8 Binary files /dev/null and b/doc/v2/images/doc_en.png differ diff --git a/doc/v2/images/efs_mount.png b/doc/v2/images/efs_mount.png new file mode 100644 index 0000000000000000000000000000000000000000..0f9e3cab98445707e5e9baa18ddabe15cdf04576 Binary files /dev/null and b/doc/v2/images/efs_mount.png differ diff --git a/doc/v2/images/encoder-decoder-attention-model.png b/doc/v2/images/encoder-decoder-attention-model.png new file mode 100644 index 0000000000000000000000000000000000000000..79f911d4ba12ac0c0d1a936c9df639c302786914 Binary files /dev/null and b/doc/v2/images/encoder-decoder-attention-model.png differ diff --git a/doc/v2/images/engine.png b/doc/v2/images/engine.png new file mode 100644 index 0000000000000000000000000000000000000000..1f5f65c2cc765a514a3ba9e7b7f468e1dc4b0c3b Binary files /dev/null and b/doc/v2/images/engine.png differ diff --git a/doc/v2/images/file_storage.graffle b/doc/v2/images/file_storage.graffle new file mode 100644 index 0000000000000000000000000000000000000000..50a17e70fa255495337c529a3bf12a5c0024a5be Binary files /dev/null and b/doc/v2/images/file_storage.graffle differ diff --git a/doc/v2/images/file_storage.png b/doc/v2/images/file_storage.png new file mode 100644 index 0000000000000000000000000000000000000000..fccb4e3e7e738224c7f1584326bd5f351ce799aa Binary files /dev/null and b/doc/v2/images/file_storage.png differ diff --git a/doc/v2/images/glossary_rnn.dot b/doc/v2/images/glossary_rnn.dot new file mode 100644 index 0000000000000000000000000000000000000000..2cd0fb1820c44b0e8e0b869f9d39fcad27efa758 --- /dev/null +++ b/doc/v2/images/glossary_rnn.dot @@ -0,0 +1,42 @@ +digraph G{ + subgraph cluster_timestep0 { + label="recurrent timestep i-1" + bgcolor=lightgray + node [style=filled,color=white] + fc0_0 [label="fc 0"] + fc0_1 [label="fc 1"] + fc0_2 [label="fc 2"] + + fc0_0 -> fc0_1 + fc0_1 -> fc0_2 + } + + subgraph cluster_timestep1 { + label="recurrent timestep i" + node [style=filled]; + fc1_0 [label="fc 0"] + fc1_1 [label="fc 1"] + fc1_2 [label="fc 2"] + color=blue + + fc1_0 -> fc1_1 + fc1_1 -> fc1_2 + } + + subgraph cluster_timestep2 { + label="recurrent timestep i+1" + bgcolor=lightgray + node [style=filled,color=white] + fc2_0 [label="fc 0"] + fc2_1 [label="fc 1"] + fc2_2 [label="fc 2"] + + fc2_0 -> fc2_1 + fc2_1 -> fc2_2 + } + + + fc0_1 -> fc1_1 [style="dotted" constraint=false] + fc1_1 -> fc2_1 [style="dotted" constraint=false] + +} \ No newline at end of file diff --git a/doc/v2/images/glossary_rnn_with_memory.dot b/doc/v2/images/glossary_rnn_with_memory.dot new file mode 100644 index 0000000000000000000000000000000000000000..0f101ec2d8f15aec76c57f328046b6b55cf0c7eb --- /dev/null +++ b/doc/v2/images/glossary_rnn_with_memory.dot @@ -0,0 +1,48 @@ +digraph G{ + subgraph cluster_timestep0 { + label="recurrent timestep i-1" + bgcolor=lightgray + node [style=filled,color=white] + fc0_0 [label="fc 0"] + fc0_1 [label="fc 1"] + fc0_2 [label="fc 2"] + m0 [label="memory"] + fc0_0 -> fc0_1 + fc0_1 -> fc0_2 + fc0_1 -> m0 + m0 -> fc0_1 + } + + subgraph cluster_timestep1 { + label="recurrent timestep i" + node [style=filled]; + fc1_0 [label="fc 0"] + fc1_1 [label="fc 1"] + fc1_2 [label="fc 2"] + m1 [label="memory"] + color=blue + fc1_0 -> fc1_1 + fc1_1 -> fc1_2 + fc1_1 -> m1 + m1 -> fc1_1 + } + + subgraph cluster_timestep2 { + label="recurrent timestep i+1" + bgcolor=lightgray + node [style=filled,color=white] + fc2_0 [label="fc 0"] + fc2_1 [label="fc 1"] + fc2_2 [label="fc 2"] + m2 [label="memory"] + fc2_0 -> fc2_1 + fc2_1 -> fc2_2 + fc2_1 -> m2 + m2 -> fc2_1 + } + + + m0 -> m1 [style="dotted" constraint=false] + m1 -> m2 [style="dotted" constraint=false] + +} \ No newline at end of file diff --git a/doc/v2/images/gradients.png b/doc/v2/images/gradients.png new file mode 100644 index 0000000000000000000000000000000000000000..f031bcf8e4cec14e63075b8b9d2c7bbd9f1b1a3c Binary files /dev/null and b/doc/v2/images/gradients.png differ diff --git a/doc/v2/images/init_lock.graffle b/doc/v2/images/init_lock.graffle new file mode 100644 index 0000000000000000000000000000000000000000..fa9149f21b1311eed48ef72ec55e556559d0fc94 Binary files /dev/null and b/doc/v2/images/init_lock.graffle differ diff --git a/doc/v2/images/init_lock.png b/doc/v2/images/init_lock.png new file mode 100644 index 0000000000000000000000000000000000000000..92404ee6d6c0f9a7727952bae3c869ba338ecd7f Binary files /dev/null and b/doc/v2/images/init_lock.png differ diff --git a/doc/v2/images/k8s-paddle-arch.png b/doc/v2/images/k8s-paddle-arch.png new file mode 100644 index 0000000000000000000000000000000000000000..b3800c4fe81302d35e49f7dbacb9221c4dfa5cde Binary files /dev/null and b/doc/v2/images/k8s-paddle-arch.png differ diff --git a/doc/v2/images/layers.png b/doc/v2/images/layers.png new file mode 100644 index 0000000000000000000000000000000000000000..306f79b7a844610915eb8944128f57d2b7a3065a Binary files /dev/null and b/doc/v2/images/layers.png differ diff --git a/doc/v2/images/managed_policy.png b/doc/v2/images/managed_policy.png new file mode 100644 index 0000000000000000000000000000000000000000..c7ecda555b81d7750e9292a9ab72d2f517f76a2a Binary files /dev/null and b/doc/v2/images/managed_policy.png differ diff --git a/doc/v2/images/matrix.png b/doc/v2/images/matrix.png new file mode 100644 index 0000000000000000000000000000000000000000..c33ce9cf0335e47cc8c1253304d0fe179186e6f2 Binary files /dev/null and b/doc/v2/images/matrix.png differ diff --git a/doc/v2/images/nvvp1.png b/doc/v2/images/nvvp1.png new file mode 100644 index 0000000000000000000000000000000000000000..1af23ac3c52929b2b0645d2f9fa4d4c6db1f6e77 Binary files /dev/null and b/doc/v2/images/nvvp1.png differ diff --git a/doc/v2/images/nvvp2.png b/doc/v2/images/nvvp2.png new file mode 100644 index 0000000000000000000000000000000000000000..177c9db708da6863d1075f3e615f5962dbe18b29 Binary files /dev/null and b/doc/v2/images/nvvp2.png differ diff --git a/doc/v2/images/nvvp3.png b/doc/v2/images/nvvp3.png new file mode 100644 index 0000000000000000000000000000000000000000..d8f393667d6569b6f1e61ffccac43fae5888b6db Binary files /dev/null and b/doc/v2/images/nvvp3.png differ diff --git a/doc/v2/images/nvvp4.png b/doc/v2/images/nvvp4.png new file mode 100644 index 0000000000000000000000000000000000000000..51f2f3e183295de6cf8ddaf2b3b8a0862aa35f01 Binary files /dev/null and b/doc/v2/images/nvvp4.png differ diff --git a/doc/v2/images/overview.png b/doc/v2/images/overview.png new file mode 100644 index 0000000000000000000000000000000000000000..8fb7bbb9dd654bf363d701d0c8cd4a557043d188 Binary files /dev/null and b/doc/v2/images/overview.png differ diff --git a/doc/v2/images/paddle-cloud-in-data-center.png b/doc/v2/images/paddle-cloud-in-data-center.png new file mode 100644 index 0000000000000000000000000000000000000000..da5d1a77562480ad1d886f5f21dbd84001d3d508 Binary files /dev/null and b/doc/v2/images/paddle-cloud-in-data-center.png differ diff --git a/doc/v2/images/paddle-etcd.graffle b/doc/v2/images/paddle-etcd.graffle new file mode 100644 index 0000000000000000000000000000000000000000..f973dc9b9dbf72e9bc31e2d32822916cd281f8d9 Binary files /dev/null and b/doc/v2/images/paddle-etcd.graffle differ diff --git a/doc/v2/images/paddle-etcd.png b/doc/v2/images/paddle-etcd.png new file mode 100644 index 0000000000000000000000000000000000000000..57981ceb4b94f0f7d6dfa63f3d28c0402bf9cc31 Binary files /dev/null and b/doc/v2/images/paddle-etcd.png differ diff --git a/doc/v2/images/paddle-model-sharding.graffle b/doc/v2/images/paddle-model-sharding.graffle new file mode 100644 index 0000000000000000000000000000000000000000..fba30f0ca2b47f0d202a432821d95e55aac37ec8 Binary files /dev/null and b/doc/v2/images/paddle-model-sharding.graffle differ diff --git a/doc/v2/images/paddle-model-sharding.png b/doc/v2/images/paddle-model-sharding.png new file mode 100644 index 0000000000000000000000000000000000000000..8c3f6724ef46c6527e63a4cd8cb0b50fe0167124 Binary files /dev/null and b/doc/v2/images/paddle-model-sharding.png differ diff --git a/doc/v2/images/paddle-ps-0.png b/doc/v2/images/paddle-ps-0.png new file mode 100644 index 0000000000000000000000000000000000000000..47ef32806f182cab003da77f1556823b3f6d1721 Binary files /dev/null and b/doc/v2/images/paddle-ps-0.png differ diff --git a/doc/v2/images/paddle-ps-1.png b/doc/v2/images/paddle-ps-1.png new file mode 100644 index 0000000000000000000000000000000000000000..f3125db73096c52bac6e7c60e1675552857c0774 Binary files /dev/null and b/doc/v2/images/paddle-ps-1.png differ diff --git a/doc/v2/images/paddle-ps.graffle b/doc/v2/images/paddle-ps.graffle new file mode 100644 index 0000000000000000000000000000000000000000..0e536ffdd91cd696008b4c01bad3cb53edebdc16 Binary files /dev/null and b/doc/v2/images/paddle-ps.graffle differ diff --git a/doc/v2/images/paddle-task-queues.graffle b/doc/v2/images/paddle-task-queues.graffle new file mode 100644 index 0000000000000000000000000000000000000000..4263ed8bfd2ef0e55058828bf23f2fac3595e5fd Binary files /dev/null and b/doc/v2/images/paddle-task-queues.graffle differ diff --git a/doc/v2/images/paddle-task-queues.png b/doc/v2/images/paddle-task-queues.png new file mode 100644 index 0000000000000000000000000000000000000000..5f980266795776752cebd0c346b85c4a75a47780 Binary files /dev/null and b/doc/v2/images/paddle-task-queues.png differ diff --git a/doc/v2/images/paddle-task-states.graffle b/doc/v2/images/paddle-task-states.graffle new file mode 100644 index 0000000000000000000000000000000000000000..cf1a0b9246d9386a949d2dbb8c32fe84f72eea83 Binary files /dev/null and b/doc/v2/images/paddle-task-states.graffle differ diff --git a/doc/v2/images/paddle-task-states.png b/doc/v2/images/paddle-task-states.png new file mode 100644 index 0000000000000000000000000000000000000000..4ae43cb66c071aee9eb90d875e2373b29af9c3e0 Binary files /dev/null and b/doc/v2/images/paddle-task-states.png differ diff --git a/doc/v2/images/ps_cn.png b/doc/v2/images/ps_cn.png new file mode 100644 index 0000000000000000000000000000000000000000..f9525739cc8bc6506adde642aafa0a85ae3ebebc Binary files /dev/null and b/doc/v2/images/ps_cn.png differ diff --git a/doc/v2/images/ps_en.png b/doc/v2/images/ps_en.png new file mode 100644 index 0000000000000000000000000000000000000000..6537d3d56589ca9f19a77a50a970e4b5275e6ce0 Binary files /dev/null and b/doc/v2/images/ps_en.png differ diff --git a/doc/v2/images/pserver_and_trainer.png b/doc/v2/images/pserver_and_trainer.png new file mode 100644 index 0000000000000000000000000000000000000000..f41fe48920590333ad332bb51eb18e03dc251541 Binary files /dev/null and b/doc/v2/images/pserver_and_trainer.png differ diff --git a/doc/v2/images/pserver_init.graffle b/doc/v2/images/pserver_init.graffle new file mode 100644 index 0000000000000000000000000000000000000000..5f3f1f52be8aa7f9049a8fcd6b7c93c8560c1676 Binary files /dev/null and b/doc/v2/images/pserver_init.graffle differ diff --git a/doc/v2/images/pserver_init.png b/doc/v2/images/pserver_init.png new file mode 100644 index 0000000000000000000000000000000000000000..dfe491ff98dd7db1c336093c80964a260df2cd90 Binary files /dev/null and b/doc/v2/images/pserver_init.png differ diff --git a/doc/v2/images/route53_create_recordset.png b/doc/v2/images/route53_create_recordset.png new file mode 100644 index 0000000000000000000000000000000000000000..34e476c7beac30fcdde13fccc4cc8d08b4be3d35 Binary files /dev/null and b/doc/v2/images/route53_create_recordset.png differ diff --git a/doc/v2/images/route53_create_zone.png b/doc/v2/images/route53_create_zone.png new file mode 100644 index 0000000000000000000000000000000000000000..25b7ddb831c5cba97f4b2edddd27da3234d621af Binary files /dev/null and b/doc/v2/images/route53_create_zone.png differ diff --git a/doc/v2/images/sequence_data.png b/doc/v2/images/sequence_data.png new file mode 100644 index 0000000000000000000000000000000000000000..6e47a46b8955dfe977e85898fe3c9f33ed28de7e Binary files /dev/null and b/doc/v2/images/sequence_data.png differ diff --git a/doc/v2/images/simple_full_hierarchical_recurrent.dot b/doc/v2/images/simple_full_hierarchical_recurrent.dot new file mode 100644 index 0000000000000000000000000000000000000000..ff278a0323bb2c3ef07bf6f016a3a8df05783581 --- /dev/null +++ b/doc/v2/images/simple_full_hierarchical_recurrent.dot @@ -0,0 +1,30 @@ +digraph G { + rankdir=LR; + + subgraph cluster_t0 { + a [label="4"] + b [label="5"] + c [label="2"] + } + + subgraph cluster_t1 { + d [label="0"] + e [label="9"] + } + + subgraph cluster_t2 { + f [label="8"] + g [label="1"] + h [label="4"] + } + + a -> b; + b -> c; + c -> d [constraint=false]; + + d -> e; + e -> f [constraint=false]; + + f -> g; + g -> h; +} \ No newline at end of file diff --git a/doc/v2/images/simple_full_recurrent.dot b/doc/v2/images/simple_full_recurrent.dot new file mode 100644 index 0000000000000000000000000000000000000000..cee281fbac993afbd0cc3416570f95965cdf0a59 --- /dev/null +++ b/doc/v2/images/simple_full_recurrent.dot @@ -0,0 +1,19 @@ +digraph G { + rankdir=LR; + a [label="4"] + b [label="5"] + c [label="2"] + d [label="0"] + e [label="9"] + f [label="8"] + g [label="1"] + h [label="4"] + + a -> b; + b -> c; + c -> d; + d -> e; + e -> f; + f -> g; + g -> h; +} \ No newline at end of file diff --git a/doc/v2/images/submit-job.graffle b/doc/v2/images/submit-job.graffle new file mode 100644 index 0000000000000000000000000000000000000000..677cdfb6d9a32168bf71729eb841fa1ca0dd31d6 Binary files /dev/null and b/doc/v2/images/submit-job.graffle differ diff --git a/doc/v2/images/submit-job.png b/doc/v2/images/submit-job.png new file mode 100644 index 0000000000000000000000000000000000000000..3046a460a7ba708079e88a560debaa215a694680 Binary files /dev/null and b/doc/v2/images/submit-job.png differ diff --git a/doc/v2/images/trainer.graffle b/doc/v2/images/trainer.graffle new file mode 100644 index 0000000000000000000000000000000000000000..43415ed8cf61a5acfa34f8e56b9577f338dbf254 Binary files /dev/null and b/doc/v2/images/trainer.graffle differ diff --git a/doc/v2/images/trainer.png b/doc/v2/images/trainer.png new file mode 100644 index 0000000000000000000000000000000000000000..6537d3d56589ca9f19a77a50a970e4b5275e6ce0 Binary files /dev/null and b/doc/v2/images/trainer.png differ diff --git a/doc/v2/images/trainer_cn.png b/doc/v2/images/trainer_cn.png new file mode 100644 index 0000000000000000000000000000000000000000..f9525739cc8bc6506adde642aafa0a85ae3ebebc Binary files /dev/null and b/doc/v2/images/trainer_cn.png differ diff --git a/doc/v2/images/worker_security_group.png b/doc/v2/images/worker_security_group.png new file mode 100644 index 0000000000000000000000000000000000000000..57eb0265a34ad4223b69600d2a3dd355482e0bf5 Binary files /dev/null and b/doc/v2/images/worker_security_group.png differ diff --git a/doc/v2/images/workflow_of_CAPI.png b/doc/v2/images/workflow_of_CAPI.png new file mode 100644 index 0000000000000000000000000000000000000000..a4399ade048b3fe10d2d9c714bc34333ca068edb Binary files /dev/null and b/doc/v2/images/workflow_of_CAPI.png differ diff --git a/paddle/fluid/framework/details/fetch_op_handle.cc b/paddle/fluid/framework/details/fetch_op_handle.cc index 1e8ca20b51d43554cf1898b41b31c27b90e6c642..a3cae8c64cdff8594c8971b0458c443f54375f11 100644 --- a/paddle/fluid/framework/details/fetch_op_handle.cc +++ b/paddle/fluid/framework/details/fetch_op_handle.cc @@ -49,7 +49,9 @@ void FetchOpHandle::RunImpl() { platform::DeviceContextPool::Instance().Get(platform::CPUPlace()); for (auto *input : inputs_) { auto *var = static_cast(input); - var->generated_op_->Wait(cpu_ctx); + if (var->generated_op_) { + var->generated_op_->Wait(cpu_ctx); + } } tensors_.resize(inputs_.size()); auto *var_handle = static_cast(inputs_[0]); diff --git a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc index b055bb48f608c9fd9cc671d175cb463d25dc489b..16aa5d067ab7a222af8fbb6ca8ec18222ecd799b 100644 --- a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc +++ b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc @@ -36,7 +36,9 @@ void NCCLAllReduceOpHandle::RunImpl() { // Wait input done for (auto *in : inputs_) { auto &p = static_cast(in)->place_; - in->generated_op_->Wait(dev_ctxes_[p]); + if (in->generated_op_) { + in->generated_op_->Wait(dev_ctxes_[p]); + } } auto &var_name = static_cast(this->inputs_[0])->name_; diff --git a/paddle/fluid/framework/details/send_op_handle.cc b/paddle/fluid/framework/details/send_op_handle.cc index 0763f92171e7813ec0ee8ca4f3aa42b76205130a..bd97c5260dbba935e422793e0aa6aac8b6875627 100644 --- a/paddle/fluid/framework/details/send_op_handle.cc +++ b/paddle/fluid/framework/details/send_op_handle.cc @@ -32,7 +32,9 @@ void SendOpHandle::RunImpl() { if (in->DebugString() == "dummy") { // HACK continue; } - in->generated_op_->Wait(dev_ctxes_[p]); + if (in->generated_op_) { + in->generated_op_->Wait(dev_ctxes_[p]); + } } auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get(); // FIXME(wuyi): can not use RunAndRecordEvent here, for it will cause dead diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt index 50f635a41a99b2ae292d13afde5637a3bf4e6f8c..db953e5cf35094fbfe36d84da5a9b54dff3d7287 100644 --- a/paddle/fluid/inference/CMakeLists.txt +++ b/paddle/fluid/inference/CMakeLists.txt @@ -21,6 +21,8 @@ endif() if(WITH_TESTING) add_subdirectory(tests/book) + # analysis test depends the models that generate by python/paddle/fluid/tests/book + add_subdirectory(analysis) endif() if (TENSORRT_FOUND) diff --git a/paddle/fluid/inference/analysis/CMakeLists.txt b/paddle/fluid/inference/analysis/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..de7becae4d25d48111fea8d2123bc85aef70230a --- /dev/null +++ b/paddle/fluid/inference/analysis/CMakeLists.txt @@ -0,0 +1 @@ +cc_library(dot SRCS dot.cc) diff --git a/paddle/fluid/inference/analysis/dot.cc b/paddle/fluid/inference/analysis/dot.cc new file mode 100644 index 0000000000000000000000000000000000000000..d5471ffcb594a6915e9e65c0fee5adc5f5bdf40c --- /dev/null +++ b/paddle/fluid/inference/analysis/dot.cc @@ -0,0 +1,23 @@ +// 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. + +#include "paddle/fluid/inference/analysis/dot.h" + +namespace paddle { +namespace inference { +namespace analysis { +size_t Dot::counter = 0; +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/dot.h b/paddle/fluid/inference/analysis/dot.h new file mode 100644 index 0000000000000000000000000000000000000000..4bf1840fdda8508b52d7274a338c5b1c95baf354 --- /dev/null +++ b/paddle/fluid/inference/analysis/dot.h @@ -0,0 +1,155 @@ +// 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. + +/* + * This file implements some helper classes and methods for DOT programming + * support. It will give a visualization of the graph and that helps to debug + * the logics of each Pass. + */ +#pragma once + +#include +#include +#include +#include +#include + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * A Dot template that helps to build a DOT graph definition. + */ +class Dot { + public: + static size_t counter; + + struct Attr { + std::string key; + std::string value; + + Attr(const std::string& key, const std::string& value) + : key(key), value(value) {} + + std::string repr() const { + std::stringstream ss; + ss << key << "=" << '"' << value << '"'; + return ss.str(); + } + }; + + struct Node { + std::string name; + std::vector attrs; + + Node(const std::string& name, const std::vector& attrs) + : name(name), + attrs(attrs), + id_("node_" + std::to_string(Dot::counter++)) {} + + std::string id() const { return id_; } + + std::string repr() const { + std::stringstream ss; + CHECK(!name.empty()); + ss << id_; + for (size_t i = 0; i < attrs.size(); i++) { + if (i == 0) { + ss << "[label=" << '"' << name << '"' << " "; + } + ss << attrs[i].repr(); + ss << ((i < attrs.size() - 1) ? " " : "]"); + } + return ss.str(); + } + + private: + std::string id_; + }; + + struct Edge { + std::string source; + std::string target; + std::vector attrs; + + Edge(const std::string& source, const std::string& target, + const std::vector& attrs) + : source(source), target(target), attrs(attrs) {} + + std::string repr() const { + std::stringstream ss; + CHECK(!source.empty()); + CHECK(!target.empty()); + ss << source << "->" << target; + for (size_t i = 0; i < attrs.size(); i++) { + if (i == 0) { + ss << "["; + } + ss << attrs[i].repr(); + ss << ((i < attrs.size() - 1) ? " " : "]"); + } + return ss.str(); + } + }; + + Dot() = default; + + explicit Dot(const std::vector& attrs) : attrs_(attrs) {} + + void AddNode(const std::string& name, const std::vector& attrs) { + CHECK(!nodes_.count(name)) << "duplicate Node '" << name << "'"; + nodes_.emplace(name, Node{name, attrs}); + } + + void AddEdge(const std::string& source, const std::string& target, + const std::vector& attrs) { + CHECK(!source.empty()); + CHECK(!target.empty()); + auto sid = nodes_.at(source).id(); + auto tid = nodes_.at(target).id(); + edges_.emplace_back(sid, tid, attrs); + } + + // Compile to DOT language codes. + std::string Build() const { + std::stringstream ss; + const std::string indent = " "; + ss << "digraph G {" << '\n'; + + // Add graph attrs + for (const auto& attr : attrs_) { + ss << indent << attr.repr() << '\n'; + } + // add nodes + for (auto& item : nodes_) { + ss << indent << item.second.repr() << '\n'; + } + // add edges + for (auto& edge : edges_) { + ss << indent << edge.repr() << '\n'; + } + ss << "} // end G"; + return ss.str(); + } + + private: + std::unordered_map nodes_; + std::vector edges_; + std::vector attrs_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/engine.h b/paddle/fluid/inference/engine.h index 6b0ac92fa908427a89a6a5fa74dacc3b24abd1c3..c0e5f3b201bc74dc357d6d9dfa9749316b61df45 100644 --- a/paddle/fluid/inference/engine.h +++ b/paddle/fluid/inference/engine.h @@ -14,11 +14,15 @@ limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/framework.pb.h" namespace paddle { namespace inference { +struct Buffer; +enum class DeviceType { UNK = -1, CPU, GPU }; + /* * EngineBase is the base class of all inference engines. An inference engine * takes a paddle program as input, and outputs the result in fluid Tensor @@ -45,8 +49,20 @@ class EngineBase { // Execute the engine, that will run the inference network. virtual void Execute(int batch_size) = 0; + // Return the IO buffer that allocated in engine. One can read/write directly + // on the buffer. If the buffer's buffer is nullptr, one can also allocate + // memory and maintain it outside the engine. + virtual Buffer& buffer(const std::string& name) = 0; + virtual ~EngineBase() {} }; // class EngineBase +struct Buffer { + void* buffer{nullptr}; // buffer should be allocated only once. + int max_size; // buffer allocated space. + int size; // data size. + DeviceType device{DeviceType::UNK}; // tells which device this buffer is on. +}; + } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/CMakeLists.txt b/paddle/fluid/inference/tensorrt/CMakeLists.txt index 288789d6e484100820c937e6081701f1e9245706..41632dabcb9dae2db101b6eed2ab2689a5f4b27b 100644 --- a/paddle/fluid/inference/tensorrt/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/CMakeLists.txt @@ -1,4 +1,5 @@ +nv_library(tensorrt_engine SRCS engine.cc) nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader) -nv_test(test_tensorrt_engine SRCS test_engine.cc engine.cc DEPS dynload_cuda) -set(ENGINE_FILE ${CMAKE_CURRENT_SOURCE_DIR}/engine.cc) +nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine) + add_subdirectory(convert) diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt index bf494d921a1ad57269eade6c4e60d0745db7944e..5178c54c08400125d190078dac6c52d021f8488b 100644 --- a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -1,4 +1,4 @@ nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES}) -nv_test(test_trt_activation_op SRCS test_activation_op.cc io_converter.cc ${ENGINE_FILE} activation_op.cc - DEPS ${FLUID_CORE_MODULES} activation_op) +nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc + DEPS ${FLUID_CORE_MODULES} activation_op tensorrt_engine) nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor) diff --git a/paddle/fluid/inference/tensorrt/engine.cc b/paddle/fluid/inference/tensorrt/engine.cc index 0a69ab9bdde653e37e0f917364d2c766d6c75682..1c296e33a610493b889359c43629003fd76b893c 100644 --- a/paddle/fluid/inference/tensorrt/engine.cc +++ b/paddle/fluid/inference/tensorrt/engine.cc @@ -30,16 +30,24 @@ void TensorRTEngine::Build(const DescType& paddle_model) { } void TensorRTEngine::Execute(int batch_size) { - infer_context_->enqueue(batch_size, buffers_.data(), *stream_, nullptr); + std::vector buffers; + for (auto& buf : buffers_) { + PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated"); + PADDLE_ENFORCE_GT(buf.max_size, 0); + PADDLE_ENFORCE(buf.device == DeviceType::GPU); + buffers.push_back(buf.buffer); + } + infer_context_->enqueue(batch_size, buffers.data(), *stream_, nullptr); cudaStreamSynchronize(*stream_); } TensorRTEngine::~TensorRTEngine() { // clean buffer - for (auto& buffer : buffers_) { - if (buffer != nullptr) { - PADDLE_ENFORCE_EQ(0, cudaFree(buffer)); - buffer = nullptr; + for (auto& buf : buffers_) { + if (buf.buffer != nullptr) { + PADDLE_ENFORCE_EQ(0, cudaFree(buf.buffer)); + buf.buffer = nullptr; + buf.max_size = 0; } } } @@ -59,7 +67,7 @@ void TensorRTEngine::FreezeNetwork() { infer_context_.reset(infer_engine_->createExecutionContext()); // allocate GPU buffers. - buffers_.resize(buffer_sizes_.size(), nullptr); + buffers_.resize(buffer_sizes_.size()); for (auto& item : buffer_sizes_) { if (item.second == 0) { auto slot_offset = infer_engine_->getBindingIndex(item.first.c_str()); @@ -67,7 +75,11 @@ void TensorRTEngine::FreezeNetwork() { infer_engine_->getBindingDataType(slot_offset))] * AccumDims(infer_engine_->getBindingDimensions(slot_offset)); } - PADDLE_ENFORCE_EQ(0, cudaMalloc(&buffer(item.first), item.second)); + auto& buf = buffer(item.first); + CHECK(buf.buffer == nullptr); // buffer should be allocated only once. + PADDLE_ENFORCE_EQ(0, cudaMalloc(&buf.buffer, item.second)); + buf.size = buf.max_size = item.second; + buf.device = DeviceType::GPU; } } @@ -113,7 +125,7 @@ void TensorRTEngine::DeclareOutput(const std::string& name) { } void* TensorRTEngine::GetOutputInGPU(const std::string& name) { - return buffer(name); + return buffer(name).buffer; } void TensorRTEngine::GetOutputInCPU(const std::string& name, void* dst, @@ -123,11 +135,13 @@ void TensorRTEngine::GetOutputInCPU(const std::string& name, void* dst, PADDLE_ENFORCE(it != buffer_sizes_.end()); PADDLE_ENFORCE_GT(it->second, 0); PADDLE_ENFORCE_GE(max_size, it->second); - PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(dst, buffer(name), it->second, + auto& buf = buffer(name); + PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated before"); + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(dst, buf.buffer, it->second, cudaMemcpyDeviceToHost, *stream_)); } -void*& TensorRTEngine::buffer(const std::string& name) { +Buffer& TensorRTEngine::buffer(const std::string& name) { PADDLE_ENFORCE(infer_engine_ != nullptr, "call FreezeNetwork first."); auto it = buffer_sizes_.find(name); PADDLE_ENFORCE(it != buffer_sizes_.end()); @@ -137,9 +151,12 @@ void*& TensorRTEngine::buffer(const std::string& name) { void TensorRTEngine::SetInputFromCPU(const std::string& name, void* data, size_t size) { - void* buf = buffer(name); - PADDLE_ENFORCE_EQ( - 0, cudaMemcpyAsync(buf, data, size, cudaMemcpyHostToDevice, *stream_)); + auto& buf = buffer(name); + PADDLE_ENFORCE_NOT_NULL(buf.buffer); + PADDLE_ENFORCE_LE(size, buf.max_size, "buffer is too small"); + PADDLE_ENFORCE(buf.device == DeviceType::GPU); + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(buf.buffer, data, size, + cudaMemcpyHostToDevice, *stream_)); } void TensorRTEngine::SetITensor(const std::string& name, diff --git a/paddle/fluid/inference/tensorrt/engine.h b/paddle/fluid/inference/tensorrt/engine.h index ec919b943d3281dd675b15e2f14adb7b3487f46f..b8298c6059e8644327194a1fcf7a7438cc9a7286 100644 --- a/paddle/fluid/inference/tensorrt/engine.h +++ b/paddle/fluid/inference/tensorrt/engine.h @@ -87,7 +87,9 @@ class TensorRTEngine : public EngineBase { // these memory directly for acceleration, for example, output the converted // data directly to the buffer to save data copy overhead. // NOTE this should be used after calling `FreezeNetwork`. - void*& buffer(const std::string& name); + Buffer& buffer(const std::string& name) override; + + cudaStream_t* stream() { return stream_; } // Fill an input from CPU memory with name and size. void SetInputFromCPU(const std::string& name, void* data, size_t size); @@ -116,7 +118,7 @@ class TensorRTEngine : public EngineBase { cudaStream_t* stream_; nvinfer1::ILogger& logger_; - std::vector buffers_; + std::vector buffers_; // max data size for the buffers. std::unordered_map buffer_sizes_; std::unordered_map diff --git a/paddle/fluid/operators/load_op.cc b/paddle/fluid/operators/load_op.cc index 6ffe0bec5e38432676ecadfa1abbbe70a1425bb1..c6bd2bf3dfca77dc078eb04b1d90c7d90883203f 100644 --- a/paddle/fluid/operators/load_op.cc +++ b/paddle/fluid/operators/load_op.cc @@ -46,19 +46,6 @@ class LoadOp : public framework::OperatorBase { auto *tensor = out_var->GetMutable(); DeserializeFromStream(fin, tensor, *dev_ctx); - - if (platform::is_gpu_place(place)) { - // copy CPU to GPU - framework::LoDTensor cpu_tensor; - cpu_tensor.ShareDataWith(*tensor); - cpu_tensor.set_lod(tensor->lod()); - - // reset tensor - out_var->Clear(); - tensor = out_var->GetMutable(); - tensor->set_lod(cpu_tensor.lod()); - TensorCopy(cpu_tensor, place, *dev_ctx, tensor); - } } }; diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc index cfddd8e8711f8005e0eff7ef7a2980f535b2f851..50bc0aba6aa0f056dc0b2d49f6b3b745433e0756 100644 --- a/paddle/fluid/platform/profiler.cc +++ b/paddle/fluid/platform/profiler.cc @@ -463,7 +463,7 @@ void SetProfileListener() { std::mt19937 rng; rng.seed(std::random_device()()); std::uniform_int_distribution dist6( - 1, std::numeric_limits::max()); + 1, std::numeric_limits::max()); profiler_lister_id = dist6(rng); } int64_t ListenerId() { return profiler_lister_id; } diff --git a/python/paddle/dataset/wmt16.py b/python/paddle/dataset/wmt16.py index ad23338a96df6856c7e15cb5e3bb713021a55bf0..540d43b692e0f65460f558dd74a52715ff4db68d 100644 --- a/python/paddle/dataset/wmt16.py +++ b/python/paddle/dataset/wmt16.py @@ -96,7 +96,7 @@ def __get_dict_size(src_dict_size, trg_dict_size, src_lang): src_dict_size = min(src_dict_size, (TOTAL_EN_WORDS if src_lang == "en" else TOTAL_DE_WORDS)) trg_dict_size = min(trg_dict_size, (TOTAL_DE_WORDS if src_lang == "en" else - TOTAL_ENG_WORDS)) + TOTAL_EN_WORDS)) return src_dict_size, trg_dict_size diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor.py b/python/paddle/fluid/tests/unittests/test_parallel_executor.py index 9056f5e66fceb42397c9a923d802320dd772725b..4eb25a6e00b7564ac17db568ec78c1c84933af43 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -import numpy +import numpy as np import unittest import paddle.fluid as fluid @@ -243,7 +243,7 @@ class TestParallelExecutorBase(unittest.TestCase): begin = time.time() first_loss, = run_executor( exe=exe, feed=feed_dict, fetch_list=[loss.name]) - first_loss = numpy.array(first_loss) + first_loss = np.array(first_loss) for i in xrange(iter): run_executor(exe=exe, feed=feed_dict, fetch_list=[]) @@ -256,7 +256,7 @@ class TestParallelExecutorBase(unittest.TestCase): print "%.4f Instance per second" % ( (batch_size * iter + 2) / (end - begin)) - last_loss = numpy.array(last_loss) + last_loss = np.array(last_loss) print first_loss, last_loss # self.assertGreater(first_loss[0], last_loss[0]) @@ -284,8 +284,8 @@ 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') + img = np.zeros(shape=[32, 784], dtype='float32') + label = np.ones(shape=[32, 1], dtype='int64') self.check_network_convergence( simple_fc_net, feed_dict={"image": img, "label": label}) @@ -294,8 +294,8 @@ class TestMNIST(TestParallelExecutorBase): self.check_simple_fc_convergence() def check_simple_fc_parallel_accuracy(self): - img = numpy.zeros(shape=[32, 784], dtype='float32') - label = numpy.ones(shape=[32, 1], dtype='int64') + img = np.zeros(shape=[32, 784], dtype='float32') + label = np.ones(shape=[32, 1], dtype='int64') single_first_loss, single_last_loss = self.check_network_convergence( method=simple_fc_net, seed=1000, @@ -319,8 +319,8 @@ class TestMNIST(TestParallelExecutorBase): def check_batchnorm_fc_convergence(self): self.check_network_convergence(fc_with_batchnorm) - img = numpy.zeros(shape=[32, 784], dtype='float32') - label = numpy.ones(shape=[32, 1], dtype='int64') + img = np.zeros(shape=[32, 784], dtype='float32') + label = np.ones(shape=[32, 1], dtype='int64') self.check_network_convergence( fc_with_batchnorm, feed_dict={"image": img, "label": label}) @@ -404,9 +404,6 @@ class ModelHyperParams(object): dropout = 0.1 -import numpy as np - - def prepare_batch_input(insts, src_pad_idx, trg_pad_idx, n_head): """ Pad the instances to the max sequence length in batch, and generate the @@ -533,9 +530,8 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase): opt.minimize(loss) batch_size = 32 - image = numpy.random.normal(size=(batch_size, - 784)).astype('float32') - label = numpy.random.randint(0, 10, (batch_size, 1), dtype="int64") + image = np.random.normal(size=(batch_size, 784)).astype('float32') + label = np.random.randint(0, 10, (batch_size, 1), dtype="int64") place = fluid.CUDAPlace(0) exe = fluid.Executor(place) @@ -552,12 +548,12 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase): for i in xrange(5): test_loss, = test_exe.run([loss.name], feed=feed_dict) - test_loss = numpy.array(test_loss) + test_loss = np.array(test_loss) train_loss, = train_exe.run([loss.name], feed=feed_dict) - train_loss = numpy.array(train_loss) + train_loss = np.array(train_loss) self.assertTrue( - numpy.allclose( + np.allclose( train_loss, test_loss, atol=1e-8), "Train loss: " + str(train_loss) + "\n Test loss:" + str(test_loss)) @@ -712,7 +708,7 @@ class TestCRFModel(unittest.TestCase): data = train_data() for i in xrange(10): cur_batch = next(data) - print map(numpy.array, + print map(np.array, pe.run(feed=feeder.feed(cur_batch), fetch_list=[avg_cost.name]))[0] @@ -721,3 +717,84 @@ class TestCRFModel(unittest.TestCase): def test_update_dense_parameter(self): self.check_network_convergence(is_sparse=False) + + +# test fetch all the variables of global_block + +import paddle.dataset.flowers as flowers +import math + + +def Lenet(data, class_dim): + conv1 = fluid.layers.conv2d(data, 32, 5, 1, act=None) + bn1 = fluid.layers.batch_norm(conv1, act='relu') + pool1 = fluid.layers.pool2d(bn1, 2, 'max', 2) + conv2 = fluid.layers.conv2d(pool1, 50, 5, 1, act=None) + bn2 = fluid.layers.batch_norm(conv2, act='relu') + pool2 = fluid.layers.pool2d(bn2, 2, 'max', 2) + + fc1 = fluid.layers.fc(pool2, size=500, act='relu') + fc2 = fluid.layers.fc(fc1, size=class_dim, act='softmax') + + return fc2 + + +class TestFetchOp(unittest.TestCase): + def parallel_exe(self, train_inputs, seed): + main = fluid.Program() + startup = fluid.Program() + startup.random_seed = seed + with fluid.program_guard(main, startup): + data = fluid.layers.data( + name='image', shape=[3, 224, 224], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + out = Lenet(data, class_dim=102) + loss = fluid.layers.cross_entropy(input=out, label=label) + loss = fluid.layers.mean(loss) + + opt = fluid.optimizer.Momentum( + learning_rate=0.1, + momentum=0.9, + regularization=fluid.regularizer.L2Decay(1e-4)) + + opt.minimize(loss) + + # TODO(zcd): I found that onece the memory optimizer is open, + # parallel_exe doesn't fetch some variable, such as conv2d_0.b_0@GRAD, + # conv2d_1.b_0@GRAD. Those variables should not be pruned. + # fluid.memory_optimize(main) + + place = fluid.CUDAPlace(0) + exe = fluid.Executor(place) + exe.run(startup) + + feeder = fluid.DataFeeder(place=place, feed_list=[data, label]) + pe = fluid.ParallelExecutor( + use_cuda=True, loss_name=loss.name, main_program=main) + + fetch_list = [] + all_vars = main.global_block().vars + for k, v in all_vars.iteritems(): + if 'tmp' not in k and k[0] is not '_' or v.persistable: + fetch_list.append(k) + + for data in train_inputs: + ret = pe.run(fetch_list, feed=feeder.feed(data)) + for i in range(len(fetch_list)): + assert not math.isnan(np.sum(ret[i])) and \ + not math.isinf(np.sum(ret[i])) + + def test_update_sparse_parameter(self): + tst_reader = paddle.batch(flowers.test(use_xmap=False), batch_size=16) + tst_reader_iter = tst_reader() + + iters = 3 + train_inputs = [] + for i in range(iters): + train_inputs.append(tst_reader_iter.next()) + + self.parallel_exe(train_inputs, seed=1) + + +if __name__ == '__main__': + unittest.main()