Switch to torch rnn, using cudnn

Author: adamlerer

word_language_model/data.py

@@ -13,7 +13,7 @@ def __init__(self):
     def addword(self, word):
         if word not in self.word2idx:
             self.idx2word.append(word)
-            self.word2idx[word] = len(self.idx2word)
+            self.word2idx[word] = len(self.idx2word) - 1
 
         return self.word2idx[word]
 

word_language_model/main.py

@@ -10,20 +10,19 @@
 import torch.nn as nn
 from torch.autograd import Variable
 
-from rnn_modules import *
 import data
 
 parser = argparse.ArgumentParser(description='PyTorch PTB Language Model')
 
 # Data parameters
 parser.add_argument('-data'      , type=str, default='./data/penn', help='Location of the data corpus'              )
 # Model parameters.
-parser.add_argument('-model'     , type=str, default='LSTM'       , help='Type of recurrent net. RNN, LSTM, or GRU.')
+parser.add_argument('-model'     , type=str, default='LSTM'       , help='Type of recurrent net. RNN_TANH, RNN_RELU, LSTM, or GRU.')
 parser.add_argument('-emsize'    , type=int, default=200          , help='Size of word embeddings'                  )
 parser.add_argument('-nhid'      , type=int, default=200          , help='Number of hidden units per layer.'        )
 parser.add_argument('-nlayers'   , type=int, default=2            , help='Number of layers.'                        )
 # Optimization parameters.
-parser.add_argument('-lr'        , type=float, default=1          , help='Initial learning rate.'                   )
+parser.add_argument('-lr'        , type=float, default=20          , help='Initial learning rate.'                   )
 parser.add_argument('-clip'      , type=float, default=0.5        , help='Gradient clipping.'                       )
 parser.add_argument('-maxepoch'  , type=int,   default=6          , help='Upper epoch limit.'                       )
 parser.add_argument('-batchsize' , type=int,   default=20         , help='Batch size.'                              )
@@ -49,18 +48,19 @@
 
 corpus = data.Corpus(args.data)
 
-def batchify(data, bsz, bptt):
-    nbatch = int(math.floor(data.size(0) / bsz / bptt))
-    data = data.narrow(0, 0, nbatch * bptt * bsz)
+def batchify(data, bsz):
+    nbatch = int(math.floor(data.size(0) / bsz))
+    data = data.narrow(0, 0, nbatch * bsz)
     data = data.view(bsz, -1).t().contiguous()
     if args.cuda:
         data = data.cuda()
     return data
 
-train = batchify(corpus.train, args.batchsize, args.bptt)
-valid = batchify(corpus.valid, 10, 1)
-test  = batchify(corpus.test,  10, 1)
-
+eval_bsz = 10
+train = batchify(corpus.train, args.batchsize)
+valid = batchify(corpus.valid, eval_bsz)
+test  = batchify(corpus.test,  eval_bsz)
+#train = train[:123*args.bptt]
 bptt  = args.bptt
 bsz   = args.batchsize
 
@@ -73,22 +73,10 @@ class RNNModel(nn.Container):
     and a decoder. Runs one RNN step at a time.
     """
 
-    @staticmethod
-    def name2module(name):
-        if name == 'RNN':
-            return RNN
-        elif name == 'LSTM':
-            return LSTM
-        elif name == 'GRU':
-            return GRU
-        else:
-            error("Unknown RNN module: " + name)
-
     def __init__(self, rnnType, ntoken, ninp, nhid, nlayers):
-        rnnModule = RNNModel.name2module(rnnType)
         super(RNNModel, self).__init__(
             encoder = nn.sparse.Embedding(ntoken, ninp),
-            rnn = StackedRNN(rnnModule, ninp, nhid, nlayers),
+            rnn = nn.rnn.RNNBase(rnnType, ninp, nhid, nlayers, bias=False),
             decoder = nn.Linear(nhid, ntoken),
         )
 
@@ -99,37 +87,48 @@ def __init__(self, rnnType, ntoken, ninp, nhid, nlayers):
         self.decoder.bias.data.fill_(0)
         self.decoder.weight.data.uniform_(-initrange, initrange)
 
-    def forward(self, hidden, input):
+    def forward(self, input, hidden):
         emb = self.encoder(input)
-        hidden, output = self.rnn(hidden, emb)
-        decoded = self.decoder(output)
-        return hidden, decoded
+        output, hidden = self.rnn(emb, hidden)
+        decoded = self.decoder(output.view(output.size(0)*output.size(1), output.size(2)))
+        return decoded.view(output.size(0), output.size(1), decoded.size(1)), hidden
 
-    def initHidden(self, bsz):
-        return self.rnn.initHidden(bsz)
-
-model = RNNModel(args.model, corpus.dic.ntokens(), args.emsize, args.nhid, args.nlayers)
+ntokens = corpus.dic.ntokens()
+model = RNNModel(args.model, ntokens, args.emsize, args.nhid, args.nlayers)
 if args.cuda:
     model.cuda()
 
 criterion = nn.CrossEntropyLoss()
 
+def initHidden(model, bsz):
+    weight = next(model.parameters()).data
+    if args.model == 'LSTM':
+        return (Variable(weight.new(args.nlayers, bsz, args.nhid).zero_()),
+                Variable(weight.new(args.nlayers, bsz, args.nhid).zero_()))
+    else:
+        return Variable(weight.new(args.nlayers, bsz, args.nhid).zero_())
+
+
 ########################################
 # TRAINING
 ########################################
 
 lr   = args.lr
 clip = args.clip
-reportinterval = args.reportint * args.batchsize
+reportinterval = args.reportint
+
 
 # Perform the forward pass only.
-def evaluate(model, data, criterion):
+def evaluate(model, data, criterion, bsz):
     loss = 0
-    hidden = model.initHidden(data.size(1))
+    hidden = initHidden(model, bsz)
     # Loop over validation data.
-    for i in range(0, data.size(0) - 1):
-        hidden, output = model(hidden, Variable(data[i], requires_grad=False))
-        loss += criterion(output, Variable(data[i+1], requires_grad=False)).data
+    for i in range(0, data.size(0) - 1, bptt):
+        seq_len = min(bptt, data.size(0) - 1 - i)
+        output, hidden = model(Variable(data[i:i+seq_len], requires_grad=False), hidden)
+        targets = data[i+1:i+seq_len+1].view(-1)
+        loss += bptt * criterion(output.view(seq_len*bsz, -1), Variable(targets, requires_grad=False)).data
+        hidden = repackageHidden(hidden)
 
     return loss[0] / data.size(0)
 
@@ -157,46 +156,61 @@ def repackageHidden(h):
     total_loss = 0
     epoch_start_time = time.time()
     # Start with an initial hidden state.
-    hidden = model.initHidden(bsz)
+    hidden = initHidden(model, bsz)
+
     # Loop over the training data.
     loss = 0
     i = 0
     model.zero_grad()
 
     total_loss = 0
     start_time = epoch_start_time = time.time()
-    while i < train.size(0) - 1:
-        hidden, output = model(hidden, Variable(train[i], requires_grad=False))
-        loss += criterion(output, Variable(train[i+1], requires_grad=False))
-        i += 1
+    ntokens = corpus.dic.ntokens()
+    for batch, i in enumerate(range(0, train.size(0) - 1, bptt)):
+        seq_len = min(bptt, train.size(0) - 1 - i)
+        output, hidden = model(Variable(train[i:i+seq_len], requires_grad=False), hidden)
+        targets = train[i+1:i+seq_len+1].view(-1)
+        loss = criterion(output.view(-1, ntokens), Variable(targets, requires_grad=False))
+
+        # FIXME: this is the result of a double bug
+        # bug #1: you can't have dangling nodes in the graph to call backward
+        # bug #2: hidden.sum() doesn't work, gives me an error in backward, which I can't reproduce in a simple way
+        #  File "/data/users/alerer/pytorch/pytorch/torch/autograd/variable.py", line 82, in backward
+        #    self._execution_engine.run_backward(self, gradient, retain_variables)
+        #  File "/data/users/alerer/pytorch/pytorch/torch/autograd/functions/reduce.py", line 27, in backward
+        #    return grad_output.new(*self.input_size).fill_(grad_output[0])
+        #ValueError: fill_ recieved an invalid combination of argument types - got (torch.cuda.FloatTensor), but expected (float value)
+        if args.model == 'LSTM':
+            loss += 0*hidden[0].sum(0).sum(1).sum(2)
+            loss += 0*hidden[1].sum(0).sum(1).sum(2)
+        else:
+            loss += 0*hidden.sum(0).sum(1).sum(2)
 
-        if i % bptt == 0:
-            loss.backward()
+        loss.backward()
 
-            clipped_lr = lr * clipGradient(model, args.clip)
+        clipped_lr = lr * clipGradient(model, args.clip)
 
-            for p in model.parameters():
-                p.data.sub_(p.grad.mul(clipped_lr))
+        for p in model.parameters():
+            p.data.sub_(p.grad.mul(clipped_lr))
 
-            hidden = repackageHidden(hidden)
-            model.zero_grad()
-            total_loss += loss.data
-            loss = 0
+        hidden = repackageHidden(hidden)
+        model.zero_grad()
+        total_loss += loss.data
+        loss = 0
 
-        if i % reportinterval == 0:
+        if batch % reportinterval == 0 and batch > 0:
             cur_loss = total_loss[0] / reportinterval
             elapsed = time.time() - start_time
             print(
                     ('| epoch {:3d} | {:5d}/{:5d} batches | lr {:02.6f} | ms/batch {:5.2f} | '
                     + 'train loss {:5.2f} | train ppl {:8.2f}').format(
-                epoch, i // bptt), train.size(0) // bptt, lr,
-                elapsed * 1000 / reportinterval * bptt,
+                epoch, batch, train.size(0) // bptt, lr, elapsed * 1000 / reportinterval,
                 cur_loss, math.exp(cur_loss)
             ))
             total_loss = 0
             start_time = time.time()
 
-    val_loss = evaluate(model, valid, criterion)
+    val_loss = evaluate(model, valid, criterion, eval_bsz)
 
     print(
         '| end of epoch {:3d} | time: {:5.2f}s | valid loss {:5.2f} | valid ppl {:8.2f}'.format(
@@ -210,7 +224,7 @@ def repackageHidden(h):
     prev_loss = val_loss
 
 # Run on test data.
-test_loss = evaluate(model, test, criterion)
+test_loss = evaluate(model, test, criterion, eval_bsz)
 print(
     '| End of training | test loss {:5.2f} | test ppl {:8.2f}'.format(
     test_loss, math.exp(test_loss)