From 9be0db7523abdfa59c19115585f1ee96d73d08c6 Mon Sep 17 00:00:00 2001
From: Alex Auvolat <alex.auvolat@ens.fr>
Date: Wed, 10 Jun 2015 15:22:18 -0400
Subject: Changes

---
 lstm.py | 75 +++++++++++++++++++++++++++++++++++++++++------------------------
 1 file changed, 48 insertions(+), 27 deletions(-)

(limited to 'lstm.py')

diff --git a/lstm.py b/lstm.py
index 32cdb9b..e294793 100644
--- a/lstm.py
+++ b/lstm.py
@@ -13,36 +13,48 @@ from blocks.graph import ComputationGraph, apply_noise, apply_dropout
 
 # An epoch will be composed of 'num_seqs' sequences of len 'seq_len'
 # divided in chunks of lengh 'seq_div_size'
-num_seqs = 10
+num_seqs = 20
 seq_len = 2000
 seq_div_size = 100
 
 io_dim = 256
 
-hidden_dims = [512, 512]
+hidden_dims = [512, 512, 512]
 activation_function = Tanh()
 
-all_hidden_for_output = False
+i2h_all = True             # input to all hidden layers or only first layer
+h2o_all = True             # all hiden layers to output or only last layer
 
 w_noise_std = 0.01
 i_dropout = 0.5
 
-step_rule = 'adadelta'
+step_rule = 'momentum'
+learning_rate = 0.1
+momentum = 0.9
 
 
-param_desc = '%s-%sHO-n%s-d%s-%dx%d(%d)-%s' % (
+param_desc = '%s-%sIH,%sHO-n%s-d%s-%dx%d(%d)-%s' % (
                  repr(hidden_dims),
-                 'all' if all_hidden_for_output else 'last',
+                 'all' if i2h_all else 'first',
+                 'all' if h2o_all else 'last',
                  repr(w_noise_std),
                  repr(i_dropout),
                  num_seqs, seq_len, seq_div_size,
                  step_rule
                 ) 
 
+save_freq = 5
+
+# parameters for sample generation
+sample_len = 60
+sample_temperature = 0.3
+
 if step_rule == 'rmsprop':
     step_rule = RMSProp()
 elif step_rule == 'adadelta':
     step_rule = AdaDelta()
+elif step_rule == 'momentum':
+    step_rule = Momentum(learning_rate=learning_rate, momentum=momentum)
 else:
     assert(False)
 
@@ -52,7 +64,9 @@ class Model():
 
         in_onehot = tensor.eq(tensor.arange(io_dim, dtype='int16').reshape((1, 1, io_dim)),
                               inp[:, :, None])
+        in_onehot.name = 'in_onehot'
 
+        # Construct hidden states
         dims = [io_dim] + hidden_dims
         states = [in_onehot.dimshuffle(1, 0, 2)]
         bricks = []
@@ -65,38 +79,44 @@ class Model():
 
             linear = Linear(input_dim=dims[i-1], output_dim=4*dims[i],
                             name="lstm_in_%d"%i)
+            bricks.append(linear)
+            inter = linear.apply(states[-1])
+
+            if i2h_all and i > 1:
+                linear2 = Linear(input_dim=dims[0], output_dim=4*dims[i],
+                                 name="lstm_in0_%d"%i)
+                bricks.append(linear2)
+                inter = inter + linear2.apply(states[0])
+                inter.name = 'inter_bis_%d'%i
+
             lstm = LSTM(dim=dims[i], activation=activation_function,
                         name="lstm_rec_%d"%i)
+            bricks.append(lstm)
 
-            new_states, new_cells = lstm.apply(linear.apply(states[-1]),
+            new_states, new_cells = lstm.apply(inter,
                                                states=init_state,
                                                cells=init_cell)
             updates.append((init_state, new_states[-1, :, :]))
             updates.append((init_cell, new_cells[-1, :, :]))
 
             states.append(new_states)
-            bricks = bricks + [linear, lstm]
 
-        states = [s.dimshuffle(1, 0, 2).reshape((inp.shape[0] * inp.shape[1], dim))
-                        for dim, s in zip(dims, states)]
+        states = [s.dimshuffle(1, 0, 2) for s in states]
 
-        if all_hidden_for_output:
-            top_linear = MLP(dims=[sum(hidden_dims), io_dim],
-                             activations=[Softmax()],
-                             name="pred_mlp")
+        # Construct output from hidden states
+        out = None
+        layers = zip(dims, states)[1:]
+        if not h2o_all:
+            layers = [layers[-1]]
+        for i, (dim, state) in enumerate(layers):
+            top_linear = Linear(input_dim=dim, output_dim=io_dim,
+                                name='top_linear_%d'%i)
             bricks.append(top_linear)
+            out_i = top_linear.apply(state)
+            out = out_i if out is None else out + out_i
+            out.name = 'out_part_%d'%i
 
-            out = top_linear.apply(tensor.concatenate(states[1:], axis=1))
-        else:
-            top_linear = MLP(dims=[hidden_dims[-1], io_dim],
-                             activations=[None],
-                             name="pred_mlp")
-            bricks.append(top_linear)
-
-            out = top_linear.apply(states[-1])
-
-        out = out.reshape((inp.shape[0], inp.shape[1], io_dim))
-
+        # Do prediction and calculate cost
         pred = out.argmax(axis=2)
 
         cost = Softmax().categorical_cross_entropy(inp[:, 1:].flatten(),
@@ -104,13 +124,13 @@ class Model():
                                                                            io_dim)))
         error_rate = tensor.neq(inp[:, 1:].flatten(), pred[:, :-1].flatten()).mean()
 
-        # Initialize
+        # Initialize all bricks
         for brick in bricks:
             brick.weights_init = IsotropicGaussian(0.1)
             brick.biases_init = Constant(0.)
             brick.initialize()
 
-        # apply noise
+        # Apply noise and dropout
         cg = ComputationGraph([cost, error_rate])
         if w_noise_std > 0:
             noise_vars = VariableFilter(roles=[WEIGHT])(cg)
@@ -123,6 +143,7 @@ class Model():
         self.error_rate = error_rate
         self.cost_reg = cost_reg
         self.error_rate_reg = error_rate_reg
+        self.out = out
         self.pred = pred
 
         self.updates = updates
-- 
cgit v1.2.3