Python dynet.log方法代码示例

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import log [as 别名]
def forward(self, observations):

        def log_sum_exp(scores):
            npval = scores.npvalue()
            argmax_score = np.argmax(npval)
            max_score_expr = dy.pick(scores, argmax_score)
            max_score_expr_broadcast = dy.concatenate([max_score_expr] * self.tagset_size)
            return max_score_expr + dy.log(dy.sum_dim(dy.transpose(dy.exp(scores - max_score_expr_broadcast)),[1]))

        init_alphas = [-1e10] * self.tagset_size
        init_alphas[t2i[START_TAG]] = 0
        for_expr = dy.inputVector(init_alphas)
        for obs in observations:
            alphas_t = []
            for next_tag in range(self.tagset_size):
                obs_broadcast = dy.concatenate([dy.pick(obs, next_tag)] * self.tagset_size)
                next_tag_expr = for_expr + self.transitions[next_tag] + obs_broadcast
                alphas_t.append(log_sum_exp(next_tag_expr))
            for_expr = dy.concatenate(alphas_t)
        terminal_expr = for_expr + self.transitions[t2i["<STOP>"]]
        alpha = log_sum_exp(terminal_expr)
        return alpha 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import log [as 别名]
def init_logger():
    if not os.path.exists(root_dir):
        os.mkdir(root_dir)
    log_formatter = logging.Formatter("%(message)s")
    logger = logging.getLogger()
    file_handler = logging.FileHandler("{0}/info.log".format(root_dir), mode='w')
    file_handler.setFormatter(log_formatter)
    logger.addHandler(file_handler)
    console_handler = logging.StreamHandler()
    console_handler.setFormatter(log_formatter)
    logger.addHandler(console_handler)
    logger.setLevel(logging.INFO)
    return logger


# ===-----------------------------------------------------------------------===
# Set up logging
# ===-----------------------------------------------------------------------=== 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import log [as 别名]
def forward(self, observations):

        def log_sum_exp(scores):
            npval = scores.npvalue()
            argmax_score = np.argmax(npval)
            max_score_expr = dy.pick(scores, argmax_score)
            max_score_expr_broadcast = dy.concatenate([max_score_expr] * self.dim_output)
            return max_score_expr + dy.log(dy.sum_elems(dy.transpose(dy.exp(scores - max_score_expr_broadcast))))

        init_alphas = [-1e10] * self.dim_output
        init_alphas[self.sp_s] = 0
        for_expr = dy.inputVector(init_alphas)
        for obs in observations:
            alphas_t = []
            for next_tag in range(self.dim_output):
                obs_broadcast = dy.concatenate([dy.pick(obs, next_tag)] * self.dim_output)
                next_tag_expr = for_expr + self.trans[next_tag] + obs_broadcast
                alphas_t.append(log_sum_exp(next_tag_expr))
            for_expr = dy.concatenate(alphas_t)
        terminal_expr = for_expr + self.trans[self.sp_e]
        alpha = log_sum_exp(terminal_expr)
        return alpha 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import log [as 别名]
def get_POStagging_loss(self, X, Y):
        losses = []
        probs = self.encode_pt(X, train=True)
        for prob, y in zip(probs, Y):
            losses.append(-dy.log(dy.pick(prob, y)))
        loss = dy.esum(losses)
        return loss 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import log [as 别名]
def _pick_sample(self, probs, temperature=1.0):
        probs = probs / np.sum(probs)
        scaled_prediction = np.log(probs) / temperature
        scaled_prediction = (scaled_prediction -
                             np.logaddexp.reduce(scaled_prediction))
        scaled_prediction = np.exp(scaled_prediction)
        # print np.sum(probs)
        # probs = probs / np.sum(probs)
        return np.random.choice(np.arange(256), p=scaled_prediction) 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import log [as 别名]
def compute_loss(gold_seq,
                 scores,
                 index_to_token_maps,
                 gold_tok_to_id,
                 noise=0.00000001):
    """ Computes the loss of a gold sequence given scores.

    Inputs:
        gold_seq (list of str): A sequence of gold tokens.
        scores (list of dy.Expression): Expressions representing the scores of
            potential output tokens for each token in gold_seq.
        index_to_tok_maps (list of dict str->list of int): Maps from index in the
            sequence to a dictionary mapping from a string to a set of integers.
        gold_tok_to_id (lambda (str, str)->list of int): Maps from the gold token
            and some lookup function to the indices in the probability distribution
            where the gold token occurs.
        noise (float, optional): The amount of noise to add to the loss.

    Returns:
        dy.Expression representing the sum of losses over the sequence.
    """
    assert len(gold_seq) == len(scores)
    assert len(index_to_token_maps) == len(scores)

    losses = []
    for i, gold_tok in enumerate(gold_seq):
        score = scores[i]
        token_map = index_to_token_maps[i]

        gold_indices = gold_tok_to_id(gold_tok, token_map)
        assert len(gold_indices) > 0
        if len(gold_indices) == 1:
            losses.append(dy.pickneglogsoftmax(score, gold_indices[0]))
        else:
            prob_of_tok = dy.zeroes(1)
            probdist = dy.softmax(score)
            for index in gold_indices:
                prob_of_tok += probdist[index]
            prob_of_tok += noise
            losses.append(-dy.log(prob_of_tok))

    return dy.esum(losses) 

# 需要导入模块: import dynet [as 别名]
# 或者: from dynet import log [as 别名]
def test(epoch, trees_test, policy, policy_save_test, best_test, best_test_idx):
    if epoch % 10 == 0:
        pass
    metric_total = [0] * 6
    micro_total = [0.] * 3
    wrong_at_total = [0.] * 10
    policy.disable_dropout()
    for i_episode in tqdm(range(len(trees_test))):
        dy.renew_cg()
        policy.re_init()
        # prob_l = []
        T = trees_test[i_episode]
        T_rollout = copy_tree(T, min(args.n_rollout_test, (len(T.terms) - 1) * 2))  # a list of T's copy
        policy.init_history(args.n_rollout_test)

        for i in range(len(T.terms) - 1):
            if i == 0:
                prob, pairs = policy.selection_by_tree(T, mode='test')
                prob = dy.log(prob).npvalue()
                indices = np.argsort(prob)[-args.n_rollout_test:]
                prob_total = prob[indices]
                selected_pairs = [pairs[idx] for idx in indices]
                pair_from_tree_idx = [0] * len(T_rollout)
            else:
                prob_per = []
                pairs_per = []
                for T_idx in range(len(T_rollout)):
                    prob, pairs = policy.selection_by_tree(T_rollout[T_idx], mode='test', idx=T_idx)
                    prob = dy.log(prob) + prob_total[T_idx]
                    prob_per.append(prob.npvalue())
                    pairs_per.append(pairs)
                prob_total, T_rollout, selected_pairs, pair_from_tree_idx = find_top_k(T_rollout, prob_per, pairs_per,
                                                                                       args.n_rollout_test)
            for tree_idx, (tree_i, pair_i, from_idx) in enumerate(zip(T_rollout, selected_pairs, pair_from_tree_idx)):
                pair_i = tuple(pair_i)
                tree_i.update(pair_i)
                policy.update_history(pair_i, from_idx=from_idx, to_idx=tree_idx)
        # best candidate
        metric_total, micro_total, wrong_at_total, wrong_total = T.evaluate(metric_total, micro_total,
                                                                            wrong_at_total, reward_type='print')
        # if args.debug:
        #     for tmp_T in T_rollout:
        #         tmp_total = [0] * 6
        #         print tmp_T.evaluate(data=tmp_total, reward_type='print')
        # T.re_init()
    for m_idx in range(5):
        metric_total[m_idx] = round(metric_total[m_idx] / len(trees_test), 3)
    for w_idx in range(len(wrong_at_total)):
        wrong_at_total[w_idx] = round(wrong_at_total[w_idx] / len(trees_test), 3)
    best_test, best_test_idx = update_best(metric_total, best_test, best_test_idx, epoch)
    if epoch % 1 == 0:
        print '[test]epoch {}:{} {} {} {}'.format(epoch, metric_total, micro_total, get_micro_f1(micro_total),
                                                  wrong_at_total),
        print 'best_test', best_test, best_test_idx

    return policy_save_test, best_test, best_test_idx