debug/format protobuf to human-readable codes

[Superjomn]

called with

from paddle.v2.fluid import debuger print debuger.pprint_program_codes(framework.default_main_program().desc)

it is easy to customize the debug information for a specific operator, just add a new handler into the op_repr_handlers.

The code sample adds a handler for fill_constant, and it looks like aval = 1. [shape=[1]]

// block-0 parent--1 // variables Tensor tmp_19 (tensor(type=float32, shape=[1L])) Tensor tmp_17 (tensor(type=float32, shape=[1L])) Tensor tmp_15 (tensor(type=float32, shape=[1L])) Tensor tmp_14 (tensor(type=float32, shape=[1L])) Tensor tmp_6 (tensor(type=float32, shape=[1L])) Tensor tmp_5 (tensor(type=float32, shape=[1L])) Tensor tmp_3 (tensor(type=float32, shape=[1L])) Tensor tmp_2 (tensor(type=float32, shape=[1L])) Tensor tmp_0 (tensor(type=float32, shape=[1L])) Tensor moment_8 (tensor(type=float32, shape=[32L, 32L])) Tensor moment_7 (tensor(type=float32, shape=[1L, 224L])) Tensor tmp_9 (tensor(type=float32, shape=[1L])) Tensor moment_6 (tensor(type=float32, shape=[30000L, 16L])) Tensor tmp_13 (tensor(type=float32, shape=[1L])) Tensor moment_4 (tensor(type=float32, shape=[32L, 30000L])) Tensor tmp_12 (tensor(type=float32, shape=[1L])) Tensor moment_2 (tensor(type=float32, shape=[32L, 128L])) Tensor tmp_11 (tensor(type=float32, shape=[1L])) Tensor tmp_18 (tensor(type=float32, shape=[1L])) Tensor moment_5 (tensor(type=float32, shape=[16L, 128L])) Tensor tmp_7 (tensor(type=float32, shape=[1L])) Tensor moment_3 (tensor(type=float32, shape=[128L])) Tensor lstm_0.tmp_3 (tensor(type=float32, shape=[-1L, 32L])) Tensor tmp_4 (tensor(type=float32, shape=[1L])) Tensor dynamic_rnn_0.tmp_0 (tensor(type=bool, shape=[1L])) Tensor lstm_0.tmp_2 (tensor(type=float32, shape=[-1L, 128L])) LoDTensor lstm_0.tmp_1 (level=1, tensor(type=float32, shape=[-1L, 32L])) LoDTensor embedding_1.tmp_0 (level=1, tensor(type=float32, shape=[-1L, 16L])) LoDTensor target_language_next_word (level=1, tensor(type=int64, shape=[-1L, 1L])) Tensor tmp_8 (tensor(type=float32, shape=[1L])) LoDTensor embedding_0.tmp_0 (level=1, tensor(type=float32, shape=[-1L, 16L])) Tensor tmp_1 (tensor(type=float32, shape=[1L])) Tensor vemb (tensor(type=float32, shape=[30000L, 16L])) LoDTensor fc_0.tmp_1 (level=1, tensor(type=float32, shape=[-1L, 128L])) LoDTensor target_language_word (level=1, tensor(type=int64, shape=[-1L, 1L])) Tensor moment_1 (tensor(type=float32, shape=[30000L])) Tensor fc_0.w_0 (tensor(type=float32, shape=[16L, 128L])) Tensor sequence_pool_0.tmp_1 (tensor(type=float32, shape=[])) Tensor moment_0 (tensor(type=float32, shape=[32L])) Tensor cross_entropy_0.tmp_0 (tensor(type=float32, shape=[-1L, 1L])) Tensor lstm_0.b_0 (tensor(type=float32, shape=[1L, 224L])) LoDTensor src_word_id (level=1, tensor(type=int64, shape=[-1L, 1L])) Tensor lod_rank_table_0 (tensor(type=bool, shape=[])) Tensor fc_2.b_0 (tensor(type=float32, shape=[30000L])) Tensor fc_0.b_0 (tensor(type=float32, shape=[128L])) Tensor tmp_10 (tensor(type=float32, shape=[1L])) Tensor fc_1.b_0 (tensor(type=float32, shape=[32L])) LoDTensor fc_0.tmp_0 (level=1, tensor(type=float32, shape=[-1L, 128L])) Tensor _generated_var_0 (tensor(type=bool, shape=[])) LoDTensor fc_0.tmp_2 (level=1, tensor(type=float32, shape=[-1L, 128L])) Tensor fill_constant_1.tmp_0 (tensor(type=int64, shape=[1L])) Tensor lstm_0.w_0 (tensor(type=float32, shape=[32L, 128L])) Tensor tmp_16 (tensor(type=float32, shape=[1L])) Tensor dynamic_rnn_mem_array_0 (tensor(type=bool, shape=[])) Tensor mean_0.tmp_0 (tensor(type=float32, shape=[1L])) Tensor dynamic_rnn_max_seq_len_0 (tensor(type=int64, shape=[1L])) Tensor dynamic_rnn_input_array_0 (tensor(type=bool, shape=[])) Tensor fill_constant_0.tmp_0 (tensor(type=int64, shape=[1L])) Tensor array_to_lod_tensor_0.tmp_0 (tensor(type=float32, shape=[-1L, 30000L])) Tensor fc_1.w_0 (tensor(type=float32, shape=[16L, 32L])) Tensor fc_1.w_1 (tensor(type=float32, shape=[32L, 32L])) Tensor fc_2.w_0 (tensor(type=float32, shape=[32L, 30000L])) Tensor moment_9 (tensor(type=float32, shape=[16L, 32L])) Tensor dynamic_rnn_0_output_array_fc_2.tmp_2_0 (tensor(type=bool, shape=[])) Tensor learning_rate_0 (tensor(type=float32, shape=[1L])) Tensor sequence_pool_0.tmp_0 (tensor(type=float32, shape=[-1L, 32L])) LoDTensor lstm_0.tmp_0 (level=1, tensor(type=float32, shape=[-1L, 32L])) // operators embedding_0.tmp_0 = lookup_table(Ids=src_word_id, W=vemb) [{padding_idx=-1,is_sparse=True}] fc_0.tmp_0 = mul(X=embedding_0.tmp_0, Y=fc_0.w_0) [{y_num_col_dims=1,x_num_col_dims=1}] fc_0.tmp_1 = elementwise_add(X=fc_0.tmp_0, Y=fc_0.b_0) [{axis=1}] fc_0.tmp_2 = tanh(X=fc_0.tmp_1) [{}] lstm_0.tmp_3, lstm_0.tmp_2, lstm_0.tmp_1, lstm_0.tmp_0 = lstm(Bias=lstm_0.b_0, C0=[], H0=[], Input=fc_0.tmp_2, Weight=lstm_0.w_0) [{candidate_activation=tanh,use_peepholes=True,is_reverse=False,gate_activation=s igmoid,cell_activation=tanh}] sequence_pool_0.tmp_1, sequence_pool_0.tmp_0 = sequence_pool(X=lstm_0.tmp_0) [{pooltype=LAST}] embedding_1.tmp_0 = lookup_table(Ids=target_language_word, W=vemb) [{padding_idx=-1,is_sparse=True}] fill_constant_0.tmp_0 = 0.0 [shape=[1]] fill_constant_1.tmp_0 = 0.0 [shape=[1]] lod_rank_table_0 = lod_rank_table(X=embedding_1.tmp_0) [{level=0}] dynamic_rnn_max_seq_len_0 = max_sequence_len(RankTable=lod_rank_table_0) [{}] dynamic_rnn_0.tmp_0 = less_than(X=fill_constant_1.tmp_0, Y=dynamic_rnn_max_seq_len_0) [{axis=-1}] dynamic_rnn_input_array_0 = lod_tensor_to_array(RankTable=lod_rank_table_0, X=embedding_1.tmp_0) [{}] dynamic_rnn_mem_array_0 = write_to_array(I=fill_constant_0.tmp_0, X=sequence_pool_0.tmp_0) [{}] [u'fill_constant_1.tmp_0', u'dynamic_rnn_0_output_array_fc_2.tmp_2_0', u'dynamic_rnn_mem_array_0', u'dynamic_rnn_0.tmp_0'], _generated_var_0 = while(Condition=dynamic_rnn_0.tmp_0, X=[u'fc_1.w_1', u'fc_1.w_0', u' fill_constant_1.tmp_0', u'dynamic_rnn_input_array_0', u'fc_2.w_0', u'dynamic_rnn_mem_array_0', u'dynamic_rnn_max_seq_len_0', u'fc_1.b_0', u'lod_rank_table_0', u'fc_2.b_0']) [{sub_block=1}] array_to_lod_tensor_0.tmp_0 = array_to_lod_tensor(RankTable=lod_rank_table_0, X=dynamic_rnn_0_output_array_fc_2.tmp_2_0) [{}] cross_entropy_0.tmp_0 = cross_entropy(Label=target_language_next_word, X=array_to_lod_tensor_0.tmp_0) [{soft_label=False}] mean_0.tmp_0 = mean(X=cross_entropy_0.tmp_0) [{}] tmp_0 = 1.0 [shape=[1]] tmp_1 = elementwise_mul(X=learning_rate_0, Y=tmp_0) [{axis=-1}] tmp_2 = 1.0 [shape=[1]] tmp_3 = elementwise_mul(X=learning_rate_0, Y=tmp_2) [{axis=-1}] tmp_4 = 1.0 [shape=[1]] tmp_5 = elementwise_mul(X=learning_rate_0, Y=tmp_4) [{axis=-1}] tmp_6 = 1.0 [shape=[1]] tmp_7 = elementwise_mul(X=learning_rate_0, Y=tmp_6) [{axis=-1}] tmp_8 = 1.0 [shape=[1]] tmp_9 = elementwise_mul(X=learning_rate_0, Y=tmp_8) [{axis=-1}] tmp_10 = 1.0 [shape=[1]] tmp_11 = elementwise_mul(X=learning_rate_0, Y=tmp_10) [{axis=-1}] tmp_12 = 1.0 [shape=[1]] tmp_13 = elementwise_mul(X=learning_rate_0, Y=tmp_12) [{axis=-1}] tmp_14 = 1.0 [shape=[1]] tmp_15 = elementwise_mul(X=learning_rate_0, Y=tmp_14) [{axis=-1}] tmp_16 = 1.0 [shape=[1]] tmp_17 = elementwise_mul(X=learning_rate_0, Y=tmp_16) [{axis=-1}] tmp_18 = 1.0 [shape=[1]] tmp_19 = elementwise_mul(X=learning_rate_0, Y=tmp_18) [{axis=-1}]

[jacquesqiao]

what is the meaning of repr?

[Superjomn]

represent, similar to the repr method in python.

[jacquesqiao]

can these repr methods be used independently?

[Superjomn]

yes, just input proto and get a str.

[jacquesqiao]

Great work!