HMM报错ValueError: transmat_ rows

Question

使用CategoricalHMM，在model.predict(X)[position-1]时报错ValueError: transmat_和为[1., 0., 0.]

目标是预测韩语元音类别，分词后把音节拆分为音素，只保留元音类别序列

import numpy as np
from hmmlearn import hmm
import jamotools


 # Specify the position of the vowel to predict
test_position = 2

# Define the vowel categories
vowel_categories = {'0': ['ㅏ', 'ㅑ', 'ㅗ', 'ㅛ', 'ㅐ', 'ㅘ', 'ㅚ', 'ㅙ'],
                    '1': ['ㅓ', 'ㅕ', 'ㅜ', 'ㅠ', 'ㅔ', 'ㅝ', 'ㅟ', 'ㅞ'],
                    '2': ['ㅡ', 'ㅣ', 'ㅢ']}

states = ['0', '1', '2']
observations = np.array(['ㅏ', 'ㅑ', 'ㅗ', 'ㅛ', 'ㅐ', 'ㅘ', 'ㅚ', 'ㅙ', 'ㅓ', 'ㅕ', 'ㅜ', 'ㅠ', 'ㅔ', 'ㅝ', 'ㅟ', 'ㅞ', 'ㅡ', 'ㅣ', 'ㅢ'])
emission_probability3 = np.array([[0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.01, 0, 0, 0, 0, 0, 0, 0, 0, 0.09, 0.09, 0.09],
                        [0, 0, 0, 0, 0, 0, 0, 0, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09],
                        [0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0532, 0.0526]])

emission_probability2 = np.array([[0.125, 0.125, 0.125, 0.125, 0.125, 0.125, 0.125, 0.125, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                        [0, 0, 0, 0, 0, 0, 0, 0, 0.125, 0.125, 0.125, 0.125, 0.125, 0.125, 0.125, 0.125, 0, 0, 0],
                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.33, 0.34, 0.33]])


# Function to split a Korean character into its components
def split_korean_character(character):
    components = []
    x = jamotools.split_syllables(character)
    for i in x:
        i = str(i)
        components.append(i)
    return components

# Assign vowel categories to each vowel in the text
def mark_words(text):
    marked_text = []
    text_split = text.split()
    for word in text_split:
        word = word.strip("abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ-'!?.,—%1234567890(《》<>)")
        #print("learning word:",word)
        marked_word = ""
        for char in word:
            components = split_korean_character(char)
            word_categories = ""
            for component in components:
                if component in ['ㅏ', 'ㅑ', 'ㅗ', 'ㅛ', 'ㅐ', 'ㅘ', 'ㅚ', 'ㅙ']:
                    category = '0'
                elif component in ['ㅓ', 'ㅕ', 'ㅜ', 'ㅠ', 'ㅔ', 'ㅝ', 'ㅟ', 'ㅞ']:
                    category = '1'
                elif component in ['ㅡ', 'ㅣ', 'ㅢ']:
                    category = '2'
                else:
                    category = ''
                word_categories += category
            marked_word += word_categories
        if len(marked_word) > test_position + 1:
            marked_text.append(marked_word)
        #print("marked word:",marked_word)
    #print("length of training words:",len(marked_text))
    return marked_text

# Learn from the annotated data
def learn_markov_chain(data):
    model = hmm.CategoricalHMM(n_components=len(states), n_features= 19, init_params='st')
    model.emissionprob_ = np.array(emission_probability2, dtype=float)
    observed = []
    lengths = []
    for sentence in data:
        lengths.append(len(sentence))
        observed.extend([vowel[0] for vowel in sentence])
        #print("sentence:",sentence)
    X = np.array([[int(obs)] for obs in observed])
    lengths = np.array(lengths)
    lengths = lengths.astype(int)
    model.fit(X, lengths)

    return model

# Test the model by hiding the specified syllable at the test position in each word
def predict_vowel_category(model, text, position):
    print("prediction begins")
    sentence = mark_words(text)
    corretto = 0
    fiadata = 0
    corretto_fiadata = 0
    for word_idx, word in enumerate(sentence):
        #print("\nsentence:", word_idx, '\n', 'word:', word, 'len_sen:',len(sentence))
        if position <= len(word):
            hidden_word = word[:position-1]  + word[position:]
            #print("hidden_word:",hidden_word)
            X = np.array([[int(char)] for char in hidden_word])
   
            # calculate correct ratio
            predicted_category = states[model.predict(X)[position-1]]
            real_category = word[position-1]
            print(f"Word '{text.split()[word_idx]}': Predicted category at position {position} is {predicted_category}, Real category is {real_category}")
            if predicted_category == real_category:
                corretto += 1

            # calculate correct ratio of the words vialating vowel harmony
            add = 0
            len_drop = 0
            
            for i in X:
                if i!= 2:
                    add+=i
                    len_drop+=1 
            if add != len_drop and add!= 0:
                fiadata += 1
                if predicted_category == real_category:
                    corretto_fiadata += 1
                print('add:', add, 'len_drop:', len_drop, 'Vowel Harmony Vialation')
    print('ratio:', corretto / len(sentence))
    print('fiadata:', fiadata, 'ratio of vialated:', corretto_fiadata/fiadata)

# Training text
korean2 =  open('korean2.txt','r',encoding='utf-8')
training_text = korean2.read()
korean2.close()

marked_data = mark_words(training_text)
for i in range(200):
    model = learn_markov_chain(marked_data)
    print(f"training time:{i}")

# Test text
with open('korean1.txt', 'r', encoding='utf-8') as korean1:
    test_text = korean1.read()  # Test text
    test_split = test_text.split()
    elements_to_remove = []
    for i in test_split:
        if len(i) < test_position+1:
            elements_to_remove.append(i)
    test_split = [x for x in test_split if x not in elements_to_remove]
    test_text = ' '.join(test_split)

predicted_category = predict_vowel_category(model, test_text, test_position)

"""
Some rows of transmat_ have zero sum because no transition from the state was ever observed.
Some rows of transmat_ have zero sum because no transition from the state was ever observed.
training time:0
prediction begins
X: [[0]
 [0]
 [1]]
Traceback (most recent call last):
  File "c:\Users\Desktop\项目\概率论\hmm2.py", line 140, in <module>
    predicted_category = predict_vowel_category(model, test_text, test_position)
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "c:\Users\Desktop\项目\概率论\hmm2.py", line 97, in predict_vowel_category
    predicted_category = states[model.predict(X)[position-1]]
                                ^^^^^^^^^^^^^^^^
  File "C:\conda\envs\mne\Lib\site-packages\hmmlearn\_emissions.py", line 27, in <lambda>
    return functools.wraps(func)(lambda *args, **kwargs: func(*args, **kwargs))
                                                         ^^^^^^^^^^^^^^^^^^^^^
  File "C:\conda\envs\mne\Lib\site-packages\hmmlearn\base.py", line 375, in predict
    _, state_sequence = self.decode(X, lengths)
                        ^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\conda\envs\mne\Lib\site-packages\hmmlearn\_emissions.py", line 27, in <lambda>
    return functools.wraps(func)(lambda *args, **kwargs: func(*args, **kwargs))
                                                         ^^^^^^^^^^^^^^^^^^^^^
  File "C:\conda\envs\mne\Lib\site-packages\hmmlearn\base.py", line 336, in decode
    self._check()
  File "C:\conda\envs\mne\Lib\site-packages\hmmlearn\hmm.py", line 139, in _check
    super()._check()
  File "C:\conda\envs\mne\Lib\site-packages\hmmlearn\base.py", line 977, in _check
    self._check_sum_1("transmat_")
  File "C:\conda\envs\mne\Lib\site-packages\hmmlearn\base.py", line 951, in _check_sum_1
    raise ValueError(
ValueError: transmat_ rows must sum to 1 (got row sums of [1. 0. 0.])
"""

Answer 1

以下内容由AIGC及阿里嘎多学长共同生成、有用望采纳：

---

从你提供的代码和错误信息来看，主要问题在于 transmat_ 的行和不为 1。这通常是因为在初始化或训练过程中，状态转移矩阵没有正确地设置。以下是一些具体的解决方案：

1. 确保 transmat_ 的行和为 1

在初始化 CategoricalHMM 时，需要确保状态转移矩阵的每一行的和为 1。你可以通过以下方式初始化 transmat_：

import numpy as np

# 假设有 3 个状态
n_states = 3

# 初始化状态转移矩阵，每一行的和为 1
transmat_init = np.ones((n_states, n_states)) / n_states

# 创建模型
model = hmm.CategoricalHMM(n_components=n_states, n_features=len(observations), init_params='st')
model.transmat_ = transmat_init

2. 检查 emission_probability 是否正确初始化

确保 emission_probability 的每一行的和也为 1。你可以通过以下方式检查：

print(np.sum(emission_probability2, axis=1))
print(np.sum(emission_probability3, axis=1))

如果输出不是 [1. 1. 1.]，则需要调整 emission_probability。

3. 确保模型训练数据的完整性

在训练模型时，确保数据包含所有状态的转移。可以通过增加更多的训练数据或调整数据预处理步骤来实现。

4. 处理模型预测时的数据格式问题

确保输入数据的格式与模型训练时的数据格式一致。在预测时，输入数据应该是一个二维数组，其中每一行代表一个观测序列。

5. 优化模型的预测逻辑

重新检查预测函数的逻辑，确保在隐藏音节时，模型能够正确处理并预测下一个音节的类别。

6. 增加模型的健壮性

考虑引入异常检测机制，识别并处理不常见的输入数据。

示例代码调整

以下是调整后的示例代码：

import numpy as np
from hmmlearn import hmm
import jamotools

# Specify the position of the vowel to predict
test_position = 2

# Define the vowel categories
vowel_categories = {'0': ['ㅏ', 'ㅑ', 'ㅗ', 'ㅛ', 'ㅐ', 'ㅘ', 'ㅚ', 'ㅙ'],
                    '1': ['ㅓ', 'ㅕ', 'ㅜ', 'ㅠ', 'ㅔ', 'ㅝ', 'ㅟ', 'ㅞ'],
                    '2': ['ㅡ', 'ㅣ', 'ㅢ']}

states = ['0', '1', '2']
observations = np.array(['ㅏ', 'ㅑ', 'ㅗ', 'ㅛ', 'ㅐ', 'ㅘ', 'ㅚ', 'ㅙ', 'ㅓ', 'ㅕ', 'ㅜ', 'ㅠ', 'ㅔ', 'ㅝ', 'ㅟ', 'ㅞ', 'ㅡ', 'ㅣ', 'ㅢ'])
emission_probability = np.array([[0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.01, 0, 0, 0, 0, 0, 0, 0, 0, 0.09, 0.09, 0.09],
                        [0, 0, 0, 0, 0, 0, 0, 0, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09, 0.09],
                        [0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0526, 0.0532, 0.0526]])

# Function to split a Korean character into its components
def split_korean_character(character):
    components = []
    x = jamotools.split_syllables(character)
    for i in x:
        i = str(i)
        components.append(i)
    return components

# Assign vowel categories to each vowel in the text
def mark_words(text):
    marked_text = []
    text_split = text.split()
    for word in text_split:
        word = word.strip("abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ-'!?.,—%1234567890(《》<>)")
        marked_word = ""
        for char in word:
            components = split_korean_character(char)
            word_categories = ""
            for component in components:
                if component in ['ㅏ', 'ㅑ', 'ㅗ', 'ㅛ', 'ㅐ', 'ㅘ', 'ㅚ', 'ㅙ']:
                    category = '0'
                elif component in ['ㅓ', 'ㅕ', 'ㅜ', 'ㅠ', 'ㅔ', 'ㅝ', 'ㅟ', 'ㅞ']:
                    category = '1'
                elif component in ['ㅡ', 'ㅣ', 'ㅢ']:
                    category = '2'
                else:
                    category = ''
                word_categories += category
            marked_word += word_categories
        if len(marked_word) > test_position + 1:
            marked_text.append(marked_word)
    return marked_text

# Learn from the annotated data
def learn_markov_chain(data):
    model = hmm.CategoricalHMM(n_components=len(states), n_features= 19, init_params='st')
    model.emissionprob_ = emission_probability
    model.transmat_ = np.ones((len(states), len(states))) / len(states)  # 初始化状态转移矩阵
    observed = []
    lengths = []
    for sentence in data:
        lengths.append(len(sentence))
        observed.extend([vowel[0] for vowel in sentence])
    X = np.array([[int(obs)] for obs in observed])
    lengths = np.array(lengths)
    lengths = lengths.astype(int)
    model.fit(X, lengths)

    return model

# Test the model by hiding the specified syllable at the test position in each word
def predict_vowel_category(model, text, position):
    sentence = mark_words(text)
    corretto = 0
    for word_idx, word in enumerate(sentence):
        if position <= len(word):
            hidden_word = word[:position-1]  + word[position:]
            X = np.array([[int(char)] for char in hidden_word])
            predicted_category = states[model.predict(X)[position-1]]
            real_category = word[position-1]
            if predicted_category == real_category:
                corretto += 1
    print('ratio:', corretto / len(sentence))

# Training text
korean2 =  open('korean2.txt','r',encoding='utf-8')
training_text = korean2.read()
korean2.close()

marked_data = mark_words(training_text)
model = learn_markov_chain(marked_data)

# Test text
with open('korean1.txt', 'r', encoding='utf-8') as korean1:
    test_text = korean1.read()  # Test text
    test_split = test_text.split()
    elements_to_remove = []
    for i in test_split:
        if len(i) < test_position+1:
            elements_to_remove.append(i)
    test_split = [x for x in test_split if x not in elements_to_remove]
    test_text = ' '.join(test_split)

predict_vowel_category(model, test_text, test_position)

这段代码首先初始化状态转移矩阵，确保每一行的和为 1，并在训练模型时使用这些初始化值。这样可以避免在预测时出现 ValueError。