Review operators required by books.

[qingqing01]

The rest operators needed to be completed at first:

• Top priority:

Op	Author	State	Deadline
batch_norm	@zchen0211	Done
deconv	@zchen0211	done PR
seq_conv	@chengduoZH	Done PR
crf_op	@lcy-seso	Done PR
crf_decoding_op	@lcy-seso	Done
lstm_op	@qingqing01	Done
gru_step	@guoshengCS	Done PR
gru_op	@guoshengCS	Done
first/last_seq_op	@luotao1	Done PR
seq_max_op	@luotao1	Done PR
expand_as_op	@wanghaoshuang	Done PR
how to do regularization	@abhinavarora	Done tasks
beam search	@Superjom
rnn_group_op	@Superjom

• Secondary priority:

Op	Author	State	Deadline
lookup-table by sparse row
hsigmoid_op	@Yancey1989
nce_op	@lcy-seso
chunk_evaluator	@guoshengCS

---

Recognize Digits And Image Classification

• fc_op, conv2d_op, pool2d_op, relu_op, softmax_op, cross_entropy_op, accuracy_op
• momentum_op, add_op/sum_op
• batch_norm_op

Generative Adversarial Net

• fc_op, conv2d_op, pool2d_op, relu_op, softmax_op, cross_entropy_op, accuracy_op
• momentum_op, add_op/sum_op
• batch_norm_op
• deconv_op (working)
• dependency engine / anytime block creation
• Multi optimizers in executors

Word2Vec

• fc_op, concat_op, softmax_op, cross_entropy_op, accuracy_op, sigmoid_op, adagrad_op
• lookup_table_op (use sparse update)

Personalized Recommendation

• lookup_table_op (used dense table by defalut in the book)
• fc_op, tanh_op, cos_sim_op, square_error_op (l2_distance_op), adam_op
• sequence_conv
• sparse matrix as input (maybe can use embeding as inpu)

Understand Sentiment

• lookup_table_op (used dense table by defalut in the book)
• fc_op, softmax_op, cross_entropy_op, accuracy_op, relu_op, adam_op
• lstm_op
• sequence_conv
• sequence_max

Semantic Role Labeling

• lookup_table_op (used dense table by defalut in the book)
• add_op (mixed_op), fc_op,relu_op, sigmoid_op, momentum_op
• lstm_op
• crf_op
• crf_decoding_op

Machine Translation

• lookup_table_op (used dense table by defalut in the book)
• fc_op, concat_op ,softmax_op, cross_entropy_op, accuracy_op, tanh_op, identity_op, adam_op
• gru_op
• gru_step
• first_seq_op, last_seq_op
• rnn_group_op
• beam search
• simple_attention
  • add_op, fc_op, scaling_op (element-wise mul op)
  • seq_pool_op, seq_softmax_op
  • expand_as_op (seq_expand)

[jacquesqiao]

some deadline

1. PythonAPI: [10/15]
   • Complete Python API (paddle-us)
   • Use Python API to write and train Mnist(paddle-us)
   • Sync with paddle-bj about the high level Python API
2. All Operators should be ready(paddle-bj) [10/20]:
3. sparse_update related Operators: [10/17]
   • framework should be ready(SelectedRows/kernel). write some demo Sparse Operator(paddle-us)
   • sync(paddle-bj and paddle-us)
4. some complex operator that need consider：
   crf, rnn, beam_search
5. regularization
   Can use python to compose operator to implement now. so priority is not high.

[zchen0211]

batch_norm, I will further discuss with @Canpio