Merge branch 'master' into master

Author: neubig

01-intro/bow.py

@@ -22,22 +22,21 @@ def read_dataset(filename):
 ntags = len(t2i)
 
 # Start DyNet and define trainer
-model = dy.Model()
+model = dy.ParameterCollection()
 trainer = dy.AdamTrainer(model)
 
 # Define the model
-W_sm = model.add_lookup_parameters((nwords, ntags)) # Word weights
-b_sm = model.add_parameters((ntags))                # Softmax bias
+W = model.add_lookup_parameters((nwords, ntags)) # Word weights
+b = model.add_parameters((ntags))                # Softmax bias
 
 # A function to calculate scores for one value
 def calc_scores(words):
   # Create a computation graph, and add parameters
   dy.renew_cg()
-  b_sm_exp = dy.parameter(b_sm)
   # Take the sum of all the embedding vectors for each word
-  score = dy.esum([dy.lookup(W_sm, x) for x in words])
+  score = dy.esum([dy.lookup(W, x) for x in words])
   # Add the bias vector and return
-  return score + b_sm_exp
+  return score + b
 
 for ITER in range(100):
   # Perform training

01-intro/cbow.py

@@ -22,7 +22,7 @@ def read_dataset(filename):
 ntags = len(t2i)
 
 # Start DyNet and define trainer
-model = dy.Model()
+model = dy.ParameterCollection()
 trainer = dy.AdamTrainer(model)
 
 # Define the model
@@ -35,9 +35,7 @@ def read_dataset(filename):
 def calc_scores(words):
   dy.renew_cg()
   cbow = dy.esum([dy.lookup(W_emb, x) for x in words])
-  W_sm_exp = dy.parameter(W_sm)
-  b_sm_exp = dy.parameter(b_sm)
-  return W_sm_exp * cbow + b_sm_exp
+  return W_sm * cbow + b_sm
 
 for ITER in range(100):
   # Perform training

01-intro/deep-cbow.py

@@ -22,7 +22,7 @@ def read_dataset(filename):
 ntags = len(t2i)
 
 # Start DyNet and define trainer
-model = dy.Model()
+model = dy.ParameterCollection()
 trainer = dy.AdamTrainer(model)
 
 # Define the model
@@ -40,8 +40,8 @@ def calc_scores(words):
   dy.renew_cg()
   h = dy.esum([dy.lookup(W_emb, x) for x in words])
   for W_h_i, b_h_i in zip(W_h, b_h):
-    h = dy.tanh( dy.parameter(W_h_i) * h + dy.parameter(b_h_i) )
-  return dy.parameter(W_sm) * h + dy.parameter(b_sm)
+    h = dy.tanh( W_h_i * h + b_h_i )
+  return W_sm * h + b_sm
 
 for ITER in range(100):
   # Perform training

02-lm/loglin-lm.py

@@ -29,7 +29,7 @@ def read_dataset(filename):
 nwords = len(w2i)
 
 # Start DyNet and define trainer
-model = dy.Model()
+model = dy.ParameterCollection()
 trainer = dy.SimpleSGDTrainer(model, learning_rate=0.1)
 
 # Define the model
@@ -39,7 +39,7 @@ def read_dataset(filename):
 # A function to calculate scores for one value
 def calc_score_of_history(words):
   # Create a list of things to sum up with only the bias vector at first
-  score_vecs = [dy.parameter(b_sm)]
+  score_vecs = [b_sm]
   for word_id, lookup_param in zip(words, W_sm): 
     score_vecs.append(lookup_param[word_id])
   return dy.esum(score_vecs)

02-lm/nn-lm-batch.py

@@ -30,15 +30,15 @@ def read_dataset(filename):
 nwords = len(w2i)
 
 # Start DyNet and define trainer
-model = dy.Model()
+model = dy.ParameterCollection()
 trainer = dy.AdamTrainer(model, alpha=0.001)
 
 # Define the model
 W_emb = model.add_lookup_parameters((nwords, EMB_SIZE)) # Word weights at each position
-W_h_p = model.add_parameters((HID_SIZE, EMB_SIZE * N))    # Weights of the softmax
-b_h_p = model.add_parameters((HID_SIZE))                  # Weights of the softmax
-W_sm_p = model.add_parameters((nwords, HID_SIZE))         # Weights of the softmax
-b_sm_p = model.add_parameters((nwords))                   # Softmax bias
+W_h = model.add_parameters((HID_SIZE, EMB_SIZE * N))    # Weights of the softmax
+b_h = model.add_parameters((HID_SIZE))                  # Weights of the softmax
+W_sm = model.add_parameters((nwords, HID_SIZE))         # Weights of the softmax
+b_sm = model.add_parameters((nwords))                   # Softmax bias
 
 # A function to calculate scores for one value
 def calc_score_of_histories(words, dropout=0.0):
@@ -47,15 +47,11 @@ def calc_score_of_histories(words, dropout=0.0):
   # Lookup the embeddings and concatenate them
   emb = dy.concatenate([dy.lookup_batch(W_emb, x) for x in words])
   # Create the hidden layer
-  W_h = dy.parameter(W_h_p)
-  b_h = dy.parameter(b_h_p)
   h = dy.tanh(dy.affine_transform([b_h, W_h, emb]))
   # Perform dropout
   if dropout != 0.0:
     h = dy.dropout(h, dropout)
   # Calculate the score and return
-  W_sm = dy.parameter(W_sm_p)
-  b_sm = dy.parameter(b_sm_p)
   return dy.affine_transform([b_sm, W_sm, h])
 
 # Calculate the loss value for the entire sentence

02-lm/nn-lm-optim.py

@@ -30,32 +30,28 @@ def read_dataset(filename):
 nwords = len(w2i)
 
 # Start DyNet and define trainer
-model = dy.Model()
+model = dy.ParameterCollection()
 
 # CHANGE 1: Use Adam instead of Simple SGD
 trainer = dy.AdamTrainer(model, alpha=0.001)
 
 # Define the model
 W_emb = model.add_lookup_parameters((nwords, EMB_SIZE)) # Word weights at each position
-W_h_p = model.add_parameters((HID_SIZE, EMB_SIZE * N))    # Weights of the softmax
-b_h_p = model.add_parameters((HID_SIZE))                  # Weights of the softmax
-W_sm_p = model.add_parameters((nwords, HID_SIZE))         # Weights of the softmax
-b_sm_p = model.add_parameters((nwords))                   # Softmax bias
+W_h = model.add_parameters((HID_SIZE, EMB_SIZE * N))    # Weights of the softmax
+b_h = model.add_parameters((HID_SIZE))                  # Weights of the softmax
+W_sm = model.add_parameters((nwords, HID_SIZE))         # Weights of the softmax
+b_sm = model.add_parameters((nwords))                   # Softmax bias
 
 # A function to calculate scores for one value
 def calc_score_of_history(words, dropout=0.0):
   # Lookup the embeddings and concatenate them
   emb = dy.concatenate([W_emb[x] for x in words])
   # Create the hidden layer
-  W_h = dy.parameter(W_h_p)
-  b_h = dy.parameter(b_h_p)
   h = dy.tanh(dy.affine_transform([b_h, W_h, emb]))
   # CHANGE 2: perform dropout
   if dropout != 0.0:
     h = dy.dropout(h, dropout)
   # Calculate the score and return
-  W_sm = dy.parameter(W_sm_p)
-  b_sm = dy.parameter(b_sm_p)
   return dy.affine_transform([b_sm, W_sm, h])
 
 # Calculate the loss value for the entire sentence

02-lm/nn-lm.py

@@ -30,27 +30,23 @@ def read_dataset(filename):
 nwords = len(w2i)
 
 # Start DyNet and define trainer
-model = dy.Model()
+model = dy.ParameterCollection()
 trainer = dy.SimpleSGDTrainer(model, learning_rate=0.1)
 
 # Define the model
 W_emb = model.add_lookup_parameters((nwords, EMB_SIZE)) # Word weights at each position
-W_h_p = model.add_parameters((HID_SIZE, EMB_SIZE * N))    # Weights of the softmax
-b_h_p = model.add_parameters((HID_SIZE))                  # Weights of the softmax
-W_sm_p = model.add_parameters((nwords, HID_SIZE))         # Weights of the softmax
-b_sm_p = model.add_parameters((nwords))                   # Softmax bias
+W_h = model.add_parameters((HID_SIZE, EMB_SIZE * N))    # Weights of the softmax
+b_h = model.add_parameters((HID_SIZE))                  # Weights of the softmax
+W_sm = model.add_parameters((nwords, HID_SIZE))         # Weights of the softmax
+b_sm = model.add_parameters((nwords))                   # Softmax bias
 
 # A function to calculate scores for one value
 def calc_score_of_history(words):
   # Lookup the embeddings and concatenate them
   emb = dy.concatenate([W_emb[x] for x in words])
   # Create the hidden layer
-  W_h = dy.parameter(W_h_p)
-  b_h = dy.parameter(b_h_p)
   h = dy.tanh(dy.affine_transform([b_h, W_h, emb]))
   # Calculate the score and return
-  W_sm = dy.parameter(W_sm_p)
-  b_sm = dy.parameter(b_sm_p)
   return dy.affine_transform([b_sm, W_sm, h])
 
 # Calculate the loss value for the entire sentence