(self, sents)
| 28 | |
| 29 | # Build the language model graph |
| 30 | def BuildLMGraph(self, sents): |
| 31 | dy.renew_cg() |
| 32 | # initialize the RNN |
| 33 | init_state = self.builder.initial_state() |
| 34 | # parameters -> expressions |
| 35 | R = dy.parameter(self.R) |
| 36 | bias = dy.parameter(self.bias) |
| 37 | |
| 38 | S = vocab.w2i["<s>"] |
| 39 | # get the cids and masks for each step |
| 40 | tot_chars = 0 |
| 41 | cids = [] |
| 42 | masks = [] |
| 43 | |
| 44 | for i in range(len(sents[0])): |
| 45 | cids.append([(vocab.w2i[sent[i]] if len(sent) > i else S) for sent in sents]) |
| 46 | mask = [(1 if len(sent)>i else 0) for sent in sents] |
| 47 | masks.append(mask) |
| 48 | tot_chars += sum(mask) |
| 49 | |
| 50 | # start the rnn with "<s>" |
| 51 | init_ids = cids[0] |
| 52 | s = init_state.add_input(lookup_batch(self.lookup, init_ids)) |
| 53 | |
| 54 | losses = [] |
| 55 | |
| 56 | # feed char vectors into the RNN and predict the next char |
| 57 | for cid, mask in zip(cids[1:], masks[1:]): |
| 58 | score = dy.affine_transform([bias, R, s.output()]) |
| 59 | loss = dy.pickneglogsoftmax_batch(score, cid) |
| 60 | # mask the loss if at least one sentence is shorter |
| 61 | if mask[-1] != 1: |
| 62 | mask_expr = dy.inputVector(mask) |
| 63 | mask_expr = dy.reshape(mask_expr, (1,), len(sents)) |
| 64 | loss = loss * mask_expr |
| 65 | |
| 66 | losses.append(loss) |
| 67 | # update the state of the RNN |
| 68 | cemb = dy.lookup_batch(self.lookup, cid) |
| 69 | s = s.add_input(cemb) |
| 70 | |
| 71 | return dy.sum_batches(dy.esum(losses)), tot_chars |
| 72 | |
| 73 | |
| 74 | def sample(self, first=1, nchars=0, stop=-1): |
no test coverage detected