fix error

chapter_natural-language-processing/nmt.md

@@ -102,6 +102,7 @@ for i in range(len(input_seqs)):
     Y[i] = nd.array(output_vocab.to_indices(output_seqs[i]), ctx=ctx)
 
 dataset = gluon.data.ArrayDataset(X, Y)
+
 ```
 
 ### 编码器、含注意力机制的解码器和解码器初始状态
@@ -166,7 +167,7 @@ class Decoder(Block):
         single_layer_state = [state[0][-1].expand_dims(0)]
         encoder_outputs = encoder_outputs.reshape((self.max_seq_len, 1,
                                                    self.encoder_hidden_dim))
-        # hidden尺寸: [(1, 1, decoder_hidden_dim)]
+        # single_layer_state尺寸: [(1, 1, decoder_hidden_dim)]
         # hidden_broadcast尺寸: (max_seq_len, 1, decoder_hidden_dim)
         hidden_broadcast = nd.broadcast_axis(single_layer_state[0], axis=0,
                                              size=self.max_seq_len)
@@ -243,7 +244,7 @@ def translate(encoder, decoder, decoder_init_state, fr_ens, ctx, max_seq_len):
         encoder_outputs, encoder_state = encoder(inputs.expand_dims(0),
                                                  encoder_state)
         encoder_outputs = encoder_outputs.flatten()
-        # 编码器的第一个输入为BOS字符。
+        # 解码器的第一个输入为BOS字符。
         decoder_input = nd.array([output_vocab.token_to_idx[BOS]], ctx=ctx)
         decoder_state = decoder_init_state(encoder_state[0])
         output_tokens = []
@@ -295,7 +296,7 @@ def train(encoder, decoder, decoder_init_state, max_seq_len, ctx, eval_fr_ens):
 
                 # encoder_outputs尺寸: (max_seq_len, encoder_hidden_dim)
                 encoder_outputs = encoder_outputs.flatten()
-                # 编码器的第一个输入为BOS字符。
+                # 解码器的第一个输入为BOS字符。
                 decoder_input = nd.array([output_vocab.token_to_idx[BOS]],
                                          ctx=ctx)
                 decoder_state = decoder_init_state(encoder_state[0])
@@ -320,10 +321,14 @@ def train(encoder, decoder, decoder_init_state, max_seq_len, ctx, eval_fr_ens):
             h, remainder = divmod((cur_time - prev_time).seconds, 3600)
             m, s = divmod(remainder, 60)
             time_str = 'Time %02d:%02d:%02d' % (h, m, s)
-            print_loss_avg = total_loss / epoch_period / len(data_iter)
+            if epoch == 1:
+                print_loss_avg = total_loss / len(data_iter)
+            else:
+                print_loss_avg = total_loss / epoch_period / len(data_iter)
             loss_str = 'Epoch %d, Loss %f, ' % (epoch, print_loss_avg)
             print(loss_str + time_str)
-            total_loss = 0.0
+            if epoch != 1:
+                total_loss = 0.0
             prev_time = cur_time
 
             translate(encoder, decoder, decoder_init_state, eval_fr_ens, ctx,