gnmt 模型参数分析

# tensorflow/nmt的模型参数分析,配置为enc:2,dec:2,uni,bahdanau
# src_emb:620, trg_emb:650, src_hid:1000, trg_hid:1200
# Trainable variables
  embeddings/encoder/embedding_encoder:0, (30005, 620), /device:GPU:0
  embeddings/decoder/embedding_decoder:0, (30005, 650), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_0/gru_cell/reset_gate/matrix_0:0, (620, 1000), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_0/gru_cell/reset_gate/matrix_1:0, (1000, 1000), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_0/gru_cell/update_gate/matrix_0:0, (620, 1000), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_0/gru_cell/update_gate/matrix_1:0, (1000, 1000), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_0/gru_cell/candidate/matrix_0:0, (620, 1000), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_0/gru_cell/candidate/matrix_1:0, (1000, 1000), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_0/gru_cell/candidate/bias:0, (1000,), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_1/gru_cell/reset_gate/matrix_0:0, (1000, 1000), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_1/gru_cell/reset_gate/matrix_1:0, (1000, 1000), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_1/gru_cell/update_gate/matrix_0:0, (1000, 1000), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_1/gru_cell/update_gate/matrix_1:0, (1000, 1000), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_1/gru_cell/candidate/matrix_0:0, (1000, 1000), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_1/gru_cell/candidate/matrix_1:0, (1000, 1000), /device:GPU:0
  dynamic_seq2seq/encoder/rnn/multi_rnn_cell/cell_1/gru_cell/candidate/bias:0, (1000,), /device:GPU:0
  dynamic_seq2seq/decoder/memory_layer/kernel:0, (1000, 1200), 
# memory_layer 负责将注意力机制获得的context的维度转换为目标端hidden states 大小
  dynamic_seq2seq/decoder/attention/gru_cell/reset_gate/matrix_0:0, (1850, 1200), /device:GPU:0
# decoder端的GRU的输入大小为1850=1200+650,即将hidden states 和y_{t-1}直接拼接起来
  dynamic_seq2seq/decoder/attention/gru_cell/reset_gate/matrix_1:0, (1200, 1200), /device:GPU:0
  dynamic_seq2seq/decoder/attention/gru_cell/update_gate/matrix_0:0, (1850, 1200), /device:GPU:0
  dynamic_seq2seq/decoder/attention/gru_cell/update_gate/matrix_1:0, (1200, 1200), /device:GPU:0
  dynamic_seq2seq/decoder/attention/gru_cell/candidate/matrix_0:0, (1850, 1200), /device:GPU:0
  dynamic_seq2seq/decoder/attention/gru_cell/candidate/matrix_1:0, (1200, 1200), /device:GPU:0
  dynamic_seq2seq/decoder/attention/gru_cell/candidate/bias:0, (1200,), /device:GPU:0
  dynamic_seq2seq/decoder/attention/bahdanau_attention/query_layer/kernel:0, (1200, 1200), /device:GPU:0
  dynamic_seq2seq/decoder/attention/bahdanau_attention/attention_v:0, (1200,), /device:GPU:0
  dynamic_seq2seq/decoder/attention/attention_layer/kernel:0, (2200, 1200), /device:GPU:0
# 对应output阶段,将h_t和ctx_t拼接起来(2200=1000+120),作为输出层
  dynamic_seq2seq/decoder/output_projection/kernel:0, (1200, 30005),
# softmax阶段 

 

©️2020 CSDN 皮肤主题: 大白 设计师: CSDN官方博客 返回首页
实付0元
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、C币套餐、付费专栏及课程。

余额充值