[layers] Add LSTM

bob/learn/tensorflow/layers/LSTM.py

+#!/usr/bin/env python
+
+import tensorflow as tf
+
+from tensorflow.python.layers import base
+
+def lstm(inputs, n_hidden, name=None):
+    """
+    """
+    return LSTM(n_hidden=n_hidden, name=name)(inputs)
+
+
+class LSTM(base.Layer):
+    """
+    """
+
+    def __init__(self,
+                 n_hidden,
+                 name=None,
+                 **kwargs):
+        super(LSTM, self).__init__(name=name,
+                                   **kwargs)
+        self.n_hidden = n_hidden
+        self.lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(self.n_hidden,
+                                                      forget_bias=1.0)
+
+    def call(self, inputs, training=False):
+        """
+        """
+
+        outputs, states = tf.nn.static_rnn(self.lstm_cell,
+                                           inputs,
+                                           dtype=tf.float32)
+
+        return outputs[-1]

bob/learn/tensorflow/layers/__init__.py

 from .Layer import Layer
 from .Conv1D import Conv1D
 #from .Maxout import maxout
+from .LSTM import lstm
 
 
 # gets sphinx autodoc done right - don't remove it
@@ -19,7 +20,7 @@ def __appropriate__(*args):
 
 __appropriate__(
     Layer,
-    Conv1D
+    Conv1D,
+    LSTM
     )
 __all__ = [_ for _ in dir() if not _.startswith('_')]
-

bob/learn/tensorflow/test/test_lstm.py

@@ -7,6 +7,8 @@ from bob.learn.tensorflow.loss import BaseLoss
 from bob.learn.tensorflow.trainers import Trainer, constant
 from bob.learn.tensorflow.utils import load_real_mnist, load_mnist
 from bob.learn.tensorflow.utils.session import Session
+from bob.learn.tensorflow.layers import lstm
+
 import tensorflow as tf
 import shutil
 
@@ -68,7 +70,7 @@ def test_dnn_trainer():
     # Preparing the architecture
     input_pl = train_data_shuffler("data", from_queue=False)
 
-    version = "lstm"
+    version = "bob"
 
     # Original code using MLP
     if version == "mlp":
@@ -88,6 +90,13 @@ def test_dnn_trainer():
         graph = outputs[-1]
         graph = slim.fully_connected(graph, n_classes, activation_fn=None)
 
+    elif version == "bob":
+        slim = tf.contrib.slim
+        graph = input_pl[:, n_input:]
+        graph = tf.reshape(graph, (-1, n_steps, n_input))
+        graph = tf.unstack(graph, n_steps, 1)
+        graph = lstm(graph, n_hidden)
+        graph = slim.fully_connected(graph, n_classes, activation_fn=None)
 
     # Loss for the softmax
     loss = BaseLoss(tf.nn.sparse_softmax_cross_entropy_with_logits, tf.reduce_mean)