python - 内存错误和训练部署分离

Question

我实际上试图将训练和部署部分分开，因为编译程序需要很长时间。

有人建议我使用 pickle dump 和 load，但将训练和部署部分分开。我尝试使用它但没有用。

def main():
    print "Fetching data..."
    train_df = util.get_training_data('../data/training_set_rel3.tsv')
    valid_df = util.get_validation_data('../data/valid_set.tsv')

    print "Standardizing scores..."
    train_df, valid_df = util.append_standardized_column(train_df, valid_df, 'score')

    print "Calculating perplexity feature..."

    train_df, valid_df = Perplexity().fill_perplexity_columns(train_df, valid_df)

    print "Calculating number of sentences feature..."

    train_df, valid_df = fill_sentence_column(train_df, valid_df)

    print "Cleaning for spelling and word count..."
    # cleaned up data for spelling feature
    vectorizer_train_spelling = util.vectorizer_clean_spelling(train_df)
    train_essays_spelling = vectorizer_train_spelling['essay'].values
    vectorizer_valid_spelling = util.vectorizer_clean_spelling(valid_df)
    valid_essays_spelling = vectorizer_valid_spelling['essay'].values

    print "Calculating total words feature..."

    train_df, valid_df = fill_total_words_column(train_df, valid_df, train_essays_spelling, valid_essays_spelling)

    print "Calculating unique words feature..."

    train_df, valid_df = fill_unique_words_column(train_df, valid_df, train_essays_spelling, valid_essays_spelling)

    print "Calculating spelling feature..."
    # spelling feature
    train_df, valid_df = fill_spelling_column(train_df, valid_df, train_essays_spelling, valid_essays_spelling)
print "Calculating pos tags features..."

train_df, valid_df = fill_pos_columns(train_df, valid_df)

print "Cleaning for TFIDF..."
# cleaned up data for tfidf vector feature
vectorizer_train = util.vectorizer_clean(train_df)
train_essays = vectorizer_train['essay'].values
vectorizer_valid = util.vectorizer_clean(valid_df)
valid_essays = vectorizer_valid['essay'].values

print "Calculating TFIDF features with unigram..."
train_df, valid_df = fill_tfidf_column(train_df, valid_df, train_essays, valid_essays, 1)

# print "Calculating TFIDF features with trigram..."
# train_df, valid_df = fill_tfidf_column(train_df, valid_df, train_essays, valid_essays, 3)

print train_df.head()

print valid_df.head()

COLS = ['essay_set', 'spelling_correct', 'std_sentence_count', 'std_unique_words', 'std_total_words',
        'std_unique_words',
        'ADJ', 'ADP', 'ADV', 'CONJ', 'DET', 'NOUN', 'NUM', 'PRT', 'PRON', 'VERB', '.', 'X', 'std_perplexity',
        'std_score']

train_df = train_df[COLS].join(train_df.filter(regex=("tfidf_*")))
valid_df = valid_df[COLS].join(valid_df.filter(regex=("tfidf_*")))

print train_df.shape
print valid_df.shape

max_essay_set = max(train_df['essay_set'])

linreg_scores_df = pd.DataFrame(columns=['essay_set', 'p', 'spearman'])

lasso_scores_df = pd.DataFrame(columns=['essay_set', 'alpha', 'p', 'spearman'])
ridge_scores_df = pd.DataFrame(columns=['essay_set', 'alpha', 'p', 'spearman'])

alphas = [x * 1.0 / 20 for x in range(20, 0, -1)]

for i in range(1, max_essay_set + 1):

    print ""

    train_x = np.asarray((train_df[train_df['essay_set'] == i]).drop(['essay_set', 'std_score'], axis=1))
    train_std_scores = np.asarray((train_df[train_df['essay_set'] == i])['std_score'], dtype="|S6").astype(np.float)

    regr = LinReg(fit_intercept=False, copy_X=False)
    regr.fit(train_x, train_std_scores)

    valid_x = np.asarray((valid_df[valid_df['essay_set'] == i]).drop(['essay_set', 'std_score'], axis=1))
    valid_pred_std_scores = regr.predict(valid_x)

    linreg_spear, p = Spearman(a=(valid_df[valid_df['essay_set'] == i])["std_score"], b=valid_pred_std_scores)
    linreg_scores_df = linreg_scores_df.append({'essay_set': i, 'p': p, 'spearman': linreg_spear},
                                               ignore_index=True)

    print "Linear for Essay Set " + str(i) + ":", linreg_spear

    for a in alphas:
        ridge = linear_model.Ridge(alpha=a)
        ridge.fit(train_x, train_std_scores)
        valid_pred_std_scores_ridge = ridge.predict(valid_x)

        ridge_spear, p = Spearman(a=(valid_df[valid_df['essay_set'] == i])["std_score"],
                                  b=valid_pred_std_scores_ridge)
        ridge_scores_df = ridge_scores_df.append({'essay_set': i, 'alpha': a, 'p': p, 'spearman': ridge_spear},
                                                 ignore_index=True)

        print "Alpha = " + str(a) + " Ridge for Essay Set " + str(i) + ":", ridge_spear

        lasso = linear_model.Lasso(alpha=a)
        lasso.fit(train_x, train_std_scores)
        valid_pred_std_scores_lasso = lasso.predict(valid_x)

        lasso_spear, p = Spearman(a=(valid_df[valid_df['essay_set'] == i])["std_score"],
                                  b=valid_pred_std_scores_lasso)
        lasso_scores_df = lasso_scores_df.append({'essay_set': i, 'alpha': a, 'p': p, 'spearman': lasso_spear},
                                                 ignore_index=True)

        print "Alpha = " + str(a) + "Lasso for Essay Set " + str(i) + ":", lasso_spear

print linreg_scores_df
print ridge_scores_df
print lasso_scores_df

linreg_scores_df.to_pickle('linreg_scores-01.pickle')
ridge_scores_df.to_pickle('ridge_scores-01.pickle')
lasso_scores_df.to_pickle('lasso_scores-01.pickle')


s1 = pickle.dumps(linreg_scores_df)
clf1 = pickle.loads(s)

s2= pickle.dumps(ridge_scores_df)
clf2 = pickle.loads(s)
s3= pickle.dumps(lasso_scores_df)
clf3 = pickle.loads(s)

这不是load和dump的正确使用方式吗。运行代码时出现内存错误，并且每次都开始训练。如何将训练和部署分开？

Answer 1

问题:

• 您的代码每次开始时都会训练数据
• 你使用了错误的酸洗“技术”

解决方案:

• 将数据训练部分放到一个函数中。

  • 开始时，检查 pickled 训练文件是否已经存在:
    • 如果是:加载并使用它
    • 如果不是:调用对训练数据进行数据预处理并对其进行 pickle 的函数

• 您正在 pickling pandas 数据帧，您需要使用正确的 (pandas) 方法进行 pickling/loading - 而不是来自 module pickle 的“原始”pickling 方法

您可以阅读更多关于酸洗的 pandas 函数(当前 0.24.x)here :

original_df = pd.DataFrame({"foo": range(5), "bar": range(5, 
original_df.to_pickle("./dummy.pkl")
unpickled_df = pd.read_pickle("./dummy.pkl")        # read pickle
print(unpickled_df)

Output:

   foo  bar
0    0    5
1    1    6
2    2    7
3    3    8
4    4    9

您正在使用 loads method泡菜本身。它用于腌制以字符串形式提供的内容(不是包含文件名的字符串)。

在更改为 pandas 方法后，它应该可以工作:您将 (df.to_pickle) 处理过的内容放入 jar (file)将其存储在架子上(硬盘)，如果您饿了(用于数据处理)，您可以将其取出并打开(pd.read_pickle) 并使用它。