Python dynet.scalarInput方法代码示例

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import scalarInput [as 别名]
def _compute_guided_attention(self, att_vect, decoder_step, input_size, output_size):
        if output_size <= 1 or input_size <= 1:
            return dy.scalarInput(0)

        target_probs = []

        t1 = float(decoder_step) / output_size

        for encoder_step in range(input_size):
            target_probs.append(1.0 - np.exp(-((float(encoder_step) / input_size - t1) ** 2) / 0.08))

        # print target_probs
        target_probs = dy.inputVector(target_probs)
        # print (target_probs.npvalue().shape, att_vect.npvalue().shape)

        return dy.transpose(target_probs) * att_vect 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import scalarInput [as 别名]
def _make_input(self, seq):
        x_list = [self.phone_lookup[self.encodings.char2int['START']]]
        for pi in seq:
            if pi.char not in self.encodings.char2int:
                print("Unknown input: '" + pi.char + "'")
            else:
                char_emb = self.phone_lookup[self.encodings.char2int[pi.char]]
                context = []
                for feature in pi.context:
                    if feature in self.encodings.context2int:
                        context.append(self.feature_lookup[self.encodings.context2int[feature]])
                if len(context) == 0:
                    x_list.append(char_emb)
                else:
                    x_list.append(char_emb + dy.esum(context) * dy.scalarInput(1.0 / len(context)))
        x_list.append(self.phone_lookup[self.encodings.char2int['STOP']])
        return x_list 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import scalarInput [as 别名]
def score_sentence(self, observations, tags):
        if len(tags) == 0:
            tags = [t2i[NONE_TAG]] * len(observations)
        assert len(observations) == len(tags)
        score_seq = [0]
        score = dy.scalarInput(0)
        tags = [t2i[START_TAG]] + tags
        for i, obs in enumerate(observations):
            score = score + dy.pick(self.transitions[tags[i + 1]], tags[i]) + dy.pick(obs, tags[i + 1])
            score_seq.append(score.value())
        score = score + dy.pick(self.transitions[t2i[END_TAG]], tags[-1])
        return score 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import scalarInput [as 别名]
def viterbi_loss(self, sentence, gold_tags, use_margins=True):
        observations = self.build_tagging_graph(sentence)
        viterbi_tags, viterbi_score = self.viterbi_decoding(observations, gold_tags, use_margins)
        if viterbi_tags != gold_tags:
            gold_score = self.score_sentence(observations, gold_tags)
            return (viterbi_score - gold_score), viterbi_tags
        else:
            return dy.scalarInput(0), viterbi_tags 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import scalarInput [as 别名]
def score_sentence(self, observations, tags):
        if not len(observations) == len(tags):
            raise AssertionError("len(observations) != len(tags)")

        score_seq = [0]
        score = dy.scalarInput(0)
        tags = [self.sp_s] + tags
        for i, obs in enumerate(observations):
            score = score + dy.pick(self.trans[tags[i+1]], tags[i]) + dy.pick(obs, tags[i+1])
            score_seq.append(score.value())
        score = score + dy.pick(self.trans[self.sp_e], tags[-1])
        return score 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import scalarInput [as 别名]
def learn(self, characters, target_mgc, guided_att=True):
        num_mgc = target_mgc.shape[0]
        # print num_mgc
        dy.renew_cg()

        for pi in characters:
            if pi.char not in self.encodings.char2int:
                print("Unknown input: '" + pi.char + "' - skipping file")
                return 0

        style_probs = self.compute_gold_style_probs(target_mgc)

        output_mgc, output_stop, output_attention = self._predict(characters, target_mgc, style_probs=style_probs)
        losses = []
        index = 0
        for mgc, real_mgc in zip(output_mgc, target_mgc):
            t_mgc = dy.inputVector(real_mgc)
            # losses.append(self._compute_binary_divergence(mgc, t_mgc) )
            losses.append(dy.l1_distance(mgc, t_mgc))

            if index % 3 == 0:
                # attention loss
                if guided_att:
                    att = output_attention[index // 3]
                    losses.append(self._compute_guided_attention(att, index // 3, len(characters) + 2, num_mgc // 3))
                # EOS loss
                stop = output_stop[index // 3]
                if index >= num_mgc - 6:
                    losses.append(dy.l1_distance(stop, dy.scalarInput(-0.8)))
                else:
                    losses.append(dy.l1_distance(stop, dy.scalarInput(0.8)))
            index += 1
        loss = dy.esum(losses)
        loss_val = loss.value() / num_mgc
        loss.backward()
        self.trainer.update()
        return loss_val 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import scalarInput [as 别名]
def __call__(self, x, y):
        W = dy.parameter(self.W)
        w_x = dy.parameter(self.w_x)
        w_y = dy.parameter(self.w_y)
        b = dy.parameter(self.b)

        out = dy.transpose(x) * W * y
        out += dy.dot_product(w_x, x)
        out += dy.dot_product(w_y, y)
        out = dy.concatenate([dy.scalarInput(0)] * (self.n_out - 1) + [out])
        out += b

        return out 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import scalarInput [as 别名]
def __call__(self, a, b, c):
        enc = [dy.rectify(self.a_mlp(a)),  # HOTFIX rectify here?
               dy.rectify(self.b_mlp(b)),
               dy.rectify(self.c_mlp(c))]
        enc = [dy.concatenate([dy.scalarInput(1), x]) for x in enc]
        return self.multilinear(*enc) 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import scalarInput [as 别名]
def synthesize(self, mgc, batch_size, sample=True, temperature=1.0):
        synth = []
        total_audio_len = mgc.shape[0] * len(self.upsample_w_s)
        num_batches = total_audio_len / batch_size
        if total_audio_len % batch_size != 0:
            num_batches + 1
        last_rnn_state = None
        last_sample = 127
        w_index = 0
        last_proc = 0
        for iBatch in range(num_batches):
            dy.renew_cg()
            # bias=dy.inputVector([0]*self.RNN_SIZE)
            # gain=dy.inputVector([1.0]*self.RNN_SIZE)
            start = batch_size * iBatch
            stop = batch_size * (iBatch + 1)
            if stop >= total_audio_len:
                stop = total_audio_len - 1
            upsampled = self._upsample(mgc, start, stop)
            rnn = self.rnn.initial_state()
            if last_rnn_state is not None:
                rnn_state = [dy.inputVector(s) for s in last_rnn_state]
                rnn = rnn.set_s(rnn_state)

            out_list = []
            for index in range(stop - start):
                w_index += 1
                curr_proc = w_index * 100 / total_audio_len
                if curr_proc % 5 == 0 and curr_proc != last_proc:
                    last_proc = curr_proc
                    sys.stdout.write(' ' + str(curr_proc))
                    sys.stdout.flush()

                if self.OUTPUT_EMB_SIZE != 1:
                    rnn_input = dy.concatenate([self.output_lookup[last_sample], upsampled[index]])
                else:
                    rnn_input = dy.concatenate([dy.scalarInput(float(last_sample) / 127.0 - 1.0), upsampled[index]])
                rnn = rnn.add_input(rnn_input)
                rnn_output = rnn.output()  # dy.layer_norm(rnn.output(), gain, bias)
                hidden = rnn_output
                for w, b in zip(self.mlp_w, self.mlp_b):
                    hidden = dy.tanh(w.expr(update=True) * hidden + b.expr(update=True))
                softmax_output = dy.softmax(
                    self.softmax_w.expr(update=True) * hidden + self.softmax_b.expr(update=True))
                out_list.append(softmax_output)

                if sample:
                    last_sample = self._pick_sample(softmax_output.npvalue(),
                                                    temperature=temperature)  # np.argmax(softmax_output.npvalue())
                else:
                    last_sample = np.argmax(softmax_output.npvalue())
                # last_sample = np.argmax(softmax_output.npvalue())
                synth.append(last_sample)

            rnn_state = rnn.s()
            last_rnn_state = [s.value() for s in rnn_state]

        return synth 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import scalarInput [as 别名]
def greedy_search(self, char_seq, truth = None, mu =0.):
        init_state = self.params['lstm'].initial_state().add_input(self.param_exprs['<bos>'])
        init_y = dy.tanh(self.param_exprs['pW'] * init_state.output() + self.param_exprs['pb'])
        init_score = dy.scalarInput(0.)
        init_sentence = Sentence(score=init_score.scalar_value(),score_expr=init_score,LSTMState =init_state, y= init_y , prevState = None, wlen=None, golden=True)
        
        if truth is not None:
            cembs = [ dy.dropout(dy.lookup(self.params['embed'],char),self.options['dropout_rate']) for char in char_seq ]
        else:
            cembs = [dy.lookup(self.params['embed'],char) for char in char_seq ]

        start_agenda = init_sentence
        agenda = [start_agenda]

        for idx, _ in enumerate(char_seq,1): # from left to right, character by character
            now = None
            for wlen in xrange(1,min(idx,self.options['max_word_len'])+1): # generate word candidate vectors
                # join segmentation sent + word
                word = self.word_repr(char_seq[idx-wlen:idx], cembs[idx-wlen:idx])
                sent = agenda[idx-wlen]

                if truth is not None:
                    word = dy.dropout(word,self.options['dropout_rate'])
                
                word_score = dy.dot_product(word,self.param_exprs['U'])

                if truth is not None:
                    golden =  sent.golden and truth[idx-1]==wlen
                    margin = dy.scalarInput(mu*wlen if truth[idx-1]!=wlen else 0.)
                    score = margin + sent.score_expr + dy.dot_product(sent.y, word) + word_score
                else:
                    golden = False
                    score = sent.score_expr + dy.dot_product(sent.y, word) + word_score


                good = (now is None or now.score < score.scalar_value())
                if golden or good:
                    new_state = sent.LSTMState.add_input(word)
                    new_y = dy.tanh(self.param_exprs['pW'] * new_state.output() + self.param_exprs['pb'])
                    new_sent = Sentence(score=score.scalar_value(),score_expr=score,LSTMState=new_state,y=new_y, prevState=sent, wlen=wlen, golden=golden)
                    if good:
                        now = new_sent
                    if golden:
                        golden_sent = new_sent

            agenda.append(now)
            if truth is not None and truth[idx-1]>0 and (not now.golden):
                return (now.score_expr - golden_sent.score_expr)

        if truth is not None:
            return (now.score_expr - golden_sent.score_expr)

        return agenda