Python utils.xprint函数代码示例

def build_lstm9(embeddings, shape, settings):
    """2 layer LSTM
    """
    model = Sequential()
    model.add(
        Embedding(
            embeddings.shape[0],
            embeddings.shape[1],
            input_length=shape['max_length'],
            trainable=False,
            weights=[embeddings],
            mask_zero=False
        )
    )
    model.add(TimeDistributed(Dense(shape['n_hidden'], use_bias=False), name='td9a'))
    model.add(Bidirectional(LSTM(shape['n_hidden'], return_sequences=True,
                                 recurrent_dropout=settings['dropout'],
                                 dropout=settings['dropout']), name='bidi9a'))
    # model.add(GlobalMaxPool1D())
    # model.add(BatchNormalization())
    # model.add(Dropout(settings['dropout'] / 2.0))

    # model.add(TimeDistributed(Dense(shape['n_hidden'], use_bias=False), name='td9b'))
    model.add(Bidirectional(LSTM(shape['n_hidden'], return_sequences=True,
                                 recurrent_dropout=settings['dropout'],
                                 dropout=settings['dropout']), name='bidi9b'))
    model.add(GlobalMaxPool1D(name='mp9'))
    model.add(BatchNormalization(name='bn9'))
    model.add(Dropout(settings['dropout'] / 2.0, name='drop9b'))

    model.add(Dense(shape['n_class'], activation='sigmoid', name='den9b'))
    xprint('build_lstm9: embeddings=%s shape=%s' % (dim(embeddings), shape))
    return model

def build_lstm4(embeddings, shape, settings):
    model = Sequential()
    model.add(
        Embedding(
            embeddings.shape[0],
            embeddings.shape[1],
            input_length=shape['max_length'],
            trainable=False,
            weights=[embeddings],
            mask_zero=False,
            name='eembed'
        )
    )
    model.add(TimeDistributed(Dense(shape['n_hidden'], use_bias=False, name='td4')))
    model.add(Bidirectional(LSTM(shape['n_hidden'], return_sequences=True,
                                 recurrent_dropout=settings['dropout'],
                                 dropout=settings['dropout'])))
    model.add(Flatten(name='flaaten'))
    model.add(BatchNormalization())
    n_dense = int(math.ceil(math.sqrt(shape['n_hidden'] * shape['n_class'])))
    model.add(Dense(n_dense, activation='relu'))
    # model.add(BatchNormalization())
    # x = Dropout(dropout)(x)
    model.add(Dense(shape['n_class'], activation='sigmoid'))
    xprint('build_lstm4: embeddings=%s shape=%s' % (dim(embeddings), shape))
    return model

    def fit(self, train, test_size=0.1):
        model_dir = get_model_dir(self.model_name, 0)
        # RocAucEvaluation saves the trainable part of the model
        model_path = os.path.join(model_dir, 'model')
        os.makedirs(model_dir, exist_ok=True)
        xprint('ClfCharLstm.fit: model_dir=%s' % model_dir)

        y_train = train[LABEL_COLS].values
        X_train = df_to_sentences(train)
        X_val, y_val = None, None
        if test_size > 0.0:
            X_train, X_val, y_train, y_val = train_test_split(X_train, y_train, test_size=test_size)

        lstm_shape = {'n_hidden': self.n_hidden,
                      'max_length': self.max_length,
                      'n_class': len(LABEL_COLS)}
        lstm_settings = {'dropout': self.dropout,
                         'lr': self.learn_rate}
        lstm, self.best_epochs = do_train(X_train, y_train, X_val, y_val, lstm_shape, lstm_settings, {},
                        epochs=self.epochs, batch_size=self.batch_size, frozen=self.frozen,
                        lstm_type=self.lstm_type, model_path=model_path)

        with open(os.path.join(model_dir, 'config.json'), 'wt') as f:
            f.write(lstm.to_json())

        print('****: best_epochs=%s - %s' % (self.best_epochs, self.description))

def describe(y):
    """Return table of values
        min, mean, max
    """
    MEASURES = ['min', 'mean', 'max']
    stats = np.zeros((3, len(LABEL_COLS)), dtype=np.float64)
    xprint('stats=%s' % dim(stats))
    for j, col in enumerate(LABEL_COLS):
        stats[0, j] = y[:, j].min()
        stats[1, j] = y[:, j].mean()
        stats[2, j] = y[:, j].max()

    def draw(name, vals, sep='|'):
        vals = ['%12s' % v for v in ([name] + vals)]
        xprint((' %s ' % sep).join(vals))

    def draw_bar():
        bar = '-' * 12
        draw(bar, [bar] * len(LABEL_COLS), sep='+')

    draw_bar()
    draw('', LABEL_COLS)
    draw_bar()
    for i, measure in enumerate(MEASURES):
        draw(measure, ['%10.4f' % z for z in stats[i, :]])
    draw_bar()

 def evaluate(self, get_clf):
     auc = np.zeros((self.n, len(LABEL_COLS)), dtype=np.float64)
     for i in range(self.n):
         ok, auc[i, :] = self._evaluate(get_clf, i)
         if not ok:
             return ok, auc
         show_auc(auc[:i + 1, :])
     xprint('program=%s train=%s' % (sys.argv[0], dim(self.train)))
     return True, auc

def do_train(train_texts, train_labels, dev_texts, dev_labels,
    lstm_shape, lstm_settings, lstm_optimizer, batch_size=100, epochs=5, by_sentence=True,
    frozen=False, lstm_type=1, model_path=None):
    """Train a Keras model on the sentences in `train_texts`
        All the sentences in a text have the text's label
    """

    print('do_train: train_texts=%s dev_texts=%s' % (dim(train_texts), dim(dev_texts)))

    embeddings, char_index, _ = get_char_embeddings()

    n_train_sents = count_sentences(char_index, train_texts, batch_size, 'train')
    X_train, y_train = make_char_sentences(char_index, lstm_shape['max_length'], batch_size,
        train_texts, train_labels, 'train', n_train_sents)
    validation_data = None
    if dev_texts is not None:
        n_dev_sents = count_sentences(char_index, dev_texts, batch_size, 'dev')
        X_val, y_val = make_char_sentences(char_index, lstm_shape['max_length'], batch_size,
            dev_texts, dev_labels, 'dev', n_dev_sents)
        validation_data = (X_val, y_val)
    sentence_cache.flush()

    model = build_lstm[lstm_type](embeddings, lstm_shape, lstm_settings)
    compile_lstm(model, lstm_settings['lr'])

    callback_list = None
    if validation_data is not None:
        ra_val = RocAucEvaluation(validation_data=validation_data, interval=1, frozen=frozen,
            model_path=model_path)
        early = EarlyStopping(monitor='val_auc', mode='max', patience=1, verbose=1)
        callback_list = [ra_val, early]

    model.fit(X_train, y_train, batch_size=batch_size, epochs=epochs,
              validation_data=validation_data, callbacks=callback_list, verbose=1)
    best_epoch_frozen = ra_val.best_epoch
    ra_val.best_epoch = -1

    best_epoch_unfrozen = -1
    if not frozen:
        xprint("Unfreezing")
        for layer in model.layers:
            layer.trainable = True
        compile_lstm(model, lstm_settings['lr'] / 10)
        if validation_data is not None:
            # Reload the best model so far
            lstm_weights = [embeddings] + ra_val.top_weights
            model.set_weights(lstm_weights)
            # Reset early stopping
            early = EarlyStopping(monitor='val_auc', mode='max', patience=1, verbose=1)
            callback_list = [ra_val, early]
        model.fit(X_train, y_train, batch_size=batch_size, epochs=epochs,
              validation_data=validation_data, callbacks=callback_list, verbose=1)
        best_epoch_unfrozen = ra_val.best_epoch

    return model, (best_epoch_frozen, best_epoch_unfrozen)

 def load(cls, path, char_index, max_length, frozen):
     xprint('SentimentAnalyser.load: path=%s max_length=%d' % (path, max_length))
     with open(os.path.join(path, 'config.json'), 'rt') as f:
         model = model_from_json(f.read())
     with open(os.path.join(path, 'model'), 'rb') as f:
         lstm_weights = pickle.load(f)
     if frozen:
         embeddings, char_index, index_char = get_char_embeddings()
         lstm_weights = [embeddings] + lstm_weights
     model.set_weights(lstm_weights)
     return cls(char_index, model, max_length=max_length)

def split_data(df, indexes, frac):
    show_values('df', df)

    n = int(len(df) * frac)
    train = df.loc[indexes[:n]]
    test = df.loc[indexes[n:]]

    show_values('train', train)
    show_values('test', test)
    xprint('split_data: %.2f of %d: train=%d test=%d' % (frac, len(df), len(train), len(test)))
    return train, test

def make_submission_reductions(get_clf, submission_name, predict_methods):
    seed_random()
    os.makedirs(SUBMISSION_DIR, exist_ok=True)

    train, test, subm = load_data()
    clf = get_clf()
    clf.fit(train, test_size=0.0)
    reductions = clf.predict_reductions(test, predict_methods)

    ok = True
    for method in predict_methods:
        submission_path = join(SUBMISSION_DIR, '%s.%s.%s.csv' % (
            submission_name, get_n_samples_str(), method))
        if os.path.exists(submission_path):
            xprint('make_submission_reductions: submission_path=%s already exists' % submission_path)
            ok = False
            break
        xprint('make_submission_reduction: method=%s' % method)
        pred = reductions[method]
        describe(pred)

        # Create the submission file.
        submid = pd.DataFrame({'id': subm['id']})
        submission = pd.concat([submid, pd.DataFrame(pred, columns=LABEL_COLS)], axis=1)
        submission.to_csv(submission_path, index=False)
        xprint('make_submission_reductions: Saved in %s' % submission_path)

        xprint('program=%s train=%s test=%s submission=%s' % (sys.argv[0], dim(train), dim(test),
            dim(submission)))

    if clf is not None:
        del clf
    return ok

def process_summary(path, n_rank):
    print('=' * 100)
    print('path=%s' % path)
    completed_tests = load_json(path)
    xprint('run_summary_path=%s' % path)
    best = {}
    try:
        best = display_results(completed_tests, do_max, n_rank)
        # display_results(completed_tests, True)
    except Exception as e:

        print('Bad summary: %s' % e)
    print('&' * 100)
    return best

def show_auc(auc):
    n = auc.shape[0]
    mean_auc = auc.mean(axis=0)
    auc_mean = auc.mean(axis=1)

    xprint('-' * 110, 'n=%d' % n)
    for i in range(n):
        xprint('%5d: auc=%.3f %s' % (i, auc[i, :].mean(), label_score(auc[i, :])))
    xprint('%5s: auc=%.3f %s' % ('Mean', mean_auc.mean(), label_score(mean_auc)))
    xprint('-' * 110)
    xprint('auc=%.3f +- %.3f (%.0f%%) range=%.3f (%.0f%%)' % (
         auc_mean.mean(), auc_mean.std(),
         100.0 * auc_mean.std() / auc_mean.mean(),
         auc_mean.max() - auc_mean.min(),
         100.0 * (auc_mean.max() - auc_mean.min()) / auc_mean.mean()
    ))

    def fit(self, train, test_size=0.1):
        print('ClfSpacy.fit', '-' * 80)
        (model1_path, config1_path), (model2_path, config2_path), epoch_path = self._get_paths(True)
        if not self.force_fit:
            if self.frozen:
                if (os.path.exists(model1_path) and os.path.exists(config1_path) and
                    SaveAllEpochs.epoch_dict(epoch_path)['epoch1'] == self.epochs):
                    xprint('model1_path already exists. re-using')
                    return
            else:
                if (os.path.exists(model2_path) and os.path.exists(config2_path) and
                    SaveAllEpochs.epoch_dict(epoch_path)['epoch2'] == self.epochs2):
                    xprint('model2_path already exists. re-using')
                    return
        do_fit1 = (not (os.path.exists(model1_path) and os.path.exists(config1_path)) or
                   SaveAllEpochs.epoch_dict(epoch_path)['epoch1'] < self.epochs)
        do_fit2 = (not self.frozen and (not (os.path.exists(model2_path) and
                                             os.path.exists(config2_path)) or
                   SaveAllEpochs.epoch_dict(epoch_path)['epoch2'] < self.epochs2))

        y_train = train[LABEL_COLS].values
        X_train = df_to_sentences(train)
        X_val, y_val = None, None
        if test_size > 0.0:
            X_train, X_val, y_train, y_val = train_test_split(X_train, y_train, test_size=test_size)

        lstm_shape = {'n_hidden': self.n_hidden,
                      'max_length': self.max_length,
                      'n_class': len(LABEL_COLS)}
        lstm_settings = {'dropout': self.dropout,
                         'lr': self.learn_rate}
        lstm, self.best_epochs = do_train(X_train, y_train, X_val, y_val, lstm_shape, lstm_settings,
            {}, batch_size=self.batch_size, lstm_type=self.lstm_type,
            do_fit1=do_fit1, epochs1=self.epochs, model1_path=model1_path, config1_path=config1_path,
            do_fit2=do_fit2, epochs2=self.epochs2, model2_path=model2_path, config2_path=config2_path,
            epoch_path=epoch_path)

        assert do_fit1
        if do_fit1:
            assert os.path.exists(model1_path), model1_path
            assert os.path.exists(config1_path), config1_path
        if do_fit2:
            assert os.path.exists(model2_path), model2_path
            assert os.path.exists(config2_path), config2_path

        print('****: best_epochs=%s - %s Add 1 to these' % (self.best_epochs, self.description))
        del lstm

def show_scores(scores, force=False):
    global scores_t0, scores_len

    if not force:
        if not scores or len(scores) == scores_len:
            return
        if time.clock() < scores_t0 + 60.0:
            return
    scores_t0 = time.clock()
    scores_len = len(scores)

    scores.sort(key=lambda x: (-x[0], x[2]))
    xprint('!' * 80)
    with open('all.results3.txt', 'wt') as f:
        for i, (score, col_scores, params, desc) in enumerate(scores):
            if i < 10:
                xprint('%4d: auc=%.3f %s %s %s' % (i, score, col_scores, params, desc))
            print('%4d: auc=%.3f %s %s %s' % (i, score, col_scores, params, desc), file=f)

 def _get_paths(self, create_dir):
     model_dir = get_model_dir(self.model_name, 0)
     if create_dir:
         os.makedirs(model_dir, exist_ok=True)
     # RocAucEvaluation saves the trainable part of the model
     model1_path = os.path.join(model_dir, 'model')
     config1_path = os.path.join(model_dir, 'config.json')
     model2_path = os.path.join(model_dir, 'model2')
     config2_path = os.path.join(model_dir, 'config2.json')
     epoch_path = os.path.join(model_dir, 'epochs.json')
     if not self._shown_paths:
         xprint('model1_path=%s exists=%s' % (model1_path, os.path.exists(model1_path)))
         xprint('config1_path=%s exists=%s' % (config1_path, os.path.exists(config1_path)))
         xprint('model2_path=%s exists=%s' % (model2_path, os.path.exists(model2_path)))
         xprint('config2_path=%s exists=%s' % (config1_path, os.path.exists(config2_path)))
         xprint('epoch_path=%s exists=%s' % (epoch_path, os.path.exists(epoch_path)))
         self._shown_paths = True
     return (model1_path, config1_path), (model2_path, config2_path), epoch_path

 def evaluate_reductions(self, get_clf, predict_methods):
     predict_methods_all = predict_methods + ['BEST']
     auc_reductions = {method: np.zeros((self.n, len(LABEL_COLS)), dtype=np.float64)
                       for method in predict_methods_all}
     best_methods = []
     for i in range(self.n):
         ok, reductions, best = self._evaluate_reductions(get_clf, i, predict_methods)
         best_methods.append(best)
         if not ok:
             return ok, {}, best_methods
         for method in predict_methods_all:
             auc = auc_reductions[method]
             auci = reductions[method]
             auc[i, :] = auci
             print('evaluate_reductions: method=%s' % method)
             show_auc(auc[:i + 1, :])
     xprint('program=%s train=%s' % (sys.argv[0], dim(self.train)))
     return True, auc_reductions, best_methods