Fix the AttentionWrapper(contrib_seq2seq) to run on version higher th…

…an 1.2
j-min · Dec 13, 2017 · da0defe · da0defe
1 parent e5afbb9
commit da0defe
Showing 1 changed file with 27 additions and 17 deletions.
diff --git a/RNN_seq2seq/contrib_seq2seq/04_AttentionWrapper.ipynb b/RNN_seq2seq/contrib_seq2seq/04_AttentionWrapper.ipynb
@@ -556,19 +556,29 @@
     "                num_units=self.attn_size,\n",
     "                memory=self.enc_outputs,\n",
     "                memory_sequence_length=self.enc_sequence_length,\n",
-    "                normalize=False,\n",
+    "             #   normalize=False,\n",
     "                name='LuongAttention')\n",
     "\n",
-    "            dec_cell = tf.contrib.seq2seq.DynamicAttentionWrapper(\n",
+    "            \n",
+    "            dec_cell = tf.contrib.seq2seq.AttentionWrapper(\n",
     "                cell=dec_cell,\n",
     "                attention_mechanism=attn_mech,\n",
-    "                attention_size=self.attn_size,\n",
-    "                # attention_history=False (in ver 1.2)\n",
+    "                attention_layer_size=self.attn_size,\n",
+    "              #  attention_history=False, # (in ver 1.2)\n",
     "                name='Attention_Wrapper')\n",
     "            \n",
-    "            initial_state = tf.contrib.seq2seq.DynamicAttentionWrapperState(\n",
-    "                cell_state=self.enc_last_state,\n",
-    "                attention=_zero_state_tensors(self.attn_size, batch_size, tf.float32))\n",
+    "#             outputs = tf.contrib.rnn.OutputProjectionWrapper(\n",
+    "#                 dec_cell, batch_size, reuse=False\n",
+    "#             )\n",
+    "            \n",
+    "            initial_state=dec_cell.zero_state(dtype=tf.float32, batch_size=batch_size)\n",
+    "\n",
+    "#             initial_state = tf.contrib.seq2seq.AttentionWrapperState(\n",
+    "#                 cell_state=self.enc_last_state,\n",
+    "#                 attention=_zero_state_tensors(self.attn_size, batch_size, tf.float32),\n",
+    "#                  time=0, alignments=(), \n",
+    "#                 alignment_history=()\n",
+    "#             )\n",
     "\n",
     "            # output projection (replacing `OutputProjectionWrapper`)\n",
     "            output_layer = Dense(dec_vocab_size+2, name='output_projection')\n",
@@ -593,7 +603,7 @@
     "                    initial_state=initial_state,\n",
     "                    output_layer=output_layer) \n",
     "\n",
-    "                train_dec_outputs, train_dec_last_state = tf.contrib.seq2seq.dynamic_decode(\n",
+    "                train_dec_outputs, train_dec_last_state, _ = tf.contrib.seq2seq.dynamic_decode(\n",
     "                    training_decoder,\n",
     "                    output_time_major=False,\n",
     "                    impute_finished=True,\n",
@@ -642,7 +652,7 @@
     "                    initial_state=initial_state,\n",
     "                    output_layer=output_layer)\n",
     "                \n",
-    "                infer_dec_outputs, infer_dec_last_state = tf.contrib.seq2seq.dynamic_decode(\n",
+    "                infer_dec_outputs, infer_dec_last_state, _ = tf.contrib.seq2seq.dynamic_decode(\n",
     "                    inference_decoder,\n",
     "                    output_time_major=False,\n",
     "                    impute_finished=True,\n",
@@ -659,7 +669,7 @@
     "        self.training_op = self.optimizer(self.learning_rate, name='training_op').minimize(self.batch_loss)\n",
     "        \n",
     "    def save(self, sess, var_list=None, save_path=None):\n",
-    "        print(f'Saving model at {save_path}')\n",
+    "        print('Saving model at {save_path}')\n",
     "        if hasattr(self, 'training_variables'):\n",
     "            var_list = self.training_variables\n",
     "        saver = tf.train.Saver(var_list)\n",
@@ -741,10 +751,10 @@
     "                print('Epoch', epoch)\n",
     "                for input_batch, target_batch, batch_preds in zip(input_batches, target_batches, all_preds):\n",
     "                    for input_sent, target_sent, pred in zip(input_batch, target_batch, batch_preds):\n",
-    "                        print(f'\\tInput: {input_sent}')\n",
-    "                        print(f'\\tPrediction:', idx2sent(pred, reverse_vocab=dec_reverse_vocab))\n",
-    "                        print(f'\\tTarget:, {target_sent}')\n",
-    "                print(f'\\tepoch loss: {epoch_loss:.2f}\\n')\n",
+    "                        print('\\tInput: {input_sent}')\n",
+    "                        print('\\tPrediction:', idx2sent(pred, reverse_vocab=dec_reverse_vocab))\n",
+    "                        print('\\tTarget:, {target_sent}')\n",
+    "                print('\\tepoch loss: {epoch_loss:.2f}\\n')\n",
     "                \n",
     "        if save_path:\n",
     "            self.save(sess, save_path=save_path)\n",
@@ -1103,7 +1113,7 @@
   "language_info": {
    "codemirror_mode": {
     "name": "ipython",
-    "version": 3
+    "version": 3.0
    },
    "file_extension": ".py",
    "mimetype": "text/x-python",
@@ -1114,5 +1124,5 @@
   }
  },
  "nbformat": 4,
- "nbformat_minor": 2
-}
+ "nbformat_minor": 0
+}