Hierarchical Attention Network - model.fit generates error 'ValueError: Input dimension mis-match'

ぐ巨炮叔叔 提交于 2020-05-14 21:28:13

问题


For background, I am referring to the Hierarchical Attention Network used for sentiment classification.

For code: my full code is posted below, but it is just simple revision of the original code posted by the author on the link above. And I explain my changes below. For training data: here For word embeddings: this is the Glove embedding here Key config: Keras 2.0.9, Scikit-Learn 0.19.1, Theano 0.9.0

The original code posted in the link above takes a 3D shape input, i.e., (review, sentence, word). And the attention mechanism is applied to sentence, and also words. So it had two attention components, as you can see in the fourth code block on the webpage.

I would like to change it to one that takes only a 2D shape input. I do this by

  1. changing the input shape and input embedding matrix (see inline comment in my code below)
  2. changing the model building part by removing the sentence attention component, keeping only the word attention component (see inline comment in my code below)

However, the code generates an error when 'model.fit' is called. I post the complete code and error below.

CODE:

import numpy as np
import pandas as pd
import re

from bs4 import BeautifulSoup

import os

from keras.preprocessing.text import Tokenizer, text_to_word_sequence
from keras.utils import plot_model
from keras.utils.np_utils import to_categorical

from keras.layers import Dense, Input
from keras.layers import Embedding, GRU, Bidirectional, TimeDistributed
from keras.models import Model

from keras import backend as K
from keras.engine.topology import Layer
from keras import initializers

MAX_SENT_LENGTH = 100
MAX_NB_WORDS = 20000
EMBEDDING_DIM = 100
VALIDATION_SPLIT = 0.2


def clean_str(string):
    """
    Tokenization/string cleaning for dataset
    Every dataset is lower cased except
    """
    string = re.sub(r"\\", "", string)
    string = re.sub(r"\'", "", string)
    string = re.sub(r"\"", "", string)
    return string.strip().lower()


#replace this to your own file path
data_train = pd.read_csv('/home/zz/Work/wop/data/sentiment/labeledTrainData_small.tsv', sep='\t')
print(data_train.shape)

labels = []
texts = []

for idx in range(data_train.review.shape[0]):
    text = BeautifulSoup(data_train.review[idx])
    text = clean_str(text.get_text().encode('ascii', 'ignore').decode('ascii'))
    texts.append(text)
    labels.append(data_train.sentiment[idx])

tokenizer = Tokenizer(nb_words=MAX_NB_WORDS)
tokenizer.fit_on_texts(texts)

##################################
# Change 1. The input shape is now 2D (sentence, words) instead of 3D
##################################
data = np.zeros((len(texts), MAX_SENT_LENGTH), dtype='int32')
for i, content in enumerate(texts):
    wordTokens = text_to_word_sequence(content)
    k = 0
    for _, word in enumerate(wordTokens):
        if k < MAX_SENT_LENGTH and tokenizer.word_index[word] < MAX_NB_WORDS:
            data[i, k] = tokenizer.word_index[word]
            k = k + 1
##################################

word_index = tokenizer.word_index
print('Total %s unique tokens.' % len(word_index))

labels = to_categorical(np.asarray(labels))
print('Shape of data tensor:', data.shape)
print('Shape of label tensor:', labels.shape)

indices = np.arange(data.shape[0])
np.random.shuffle(indices)
data = data[indices]
labels = labels[indices]
nb_validation_samples = int(VALIDATION_SPLIT * data.shape[0])

x_train = data[:-nb_validation_samples]
y_train = labels[:-nb_validation_samples]
x_val = data[-nb_validation_samples:]
y_val = labels[-nb_validation_samples:]

print('Number of positive and negative reviews in traing and validation set')
print(y_train.sum(axis=0))
print(y_val.sum(axis=0))

#replace with your own embedding file path
GLOVE_DIR = "/home/zz/Work/data/glove.6B"
embeddings_index = {}
f = open(os.path.join(GLOVE_DIR, 'glove.6B.100d.txt'))
for line in f:
    values = line.split()
    word = values[0]
    coefs = np.asarray(values[1:], dtype='float32')
    embeddings_index[word] = coefs
f.close()

print('Total %s word vectors.' % len(embeddings_index))

embedding_matrix = np.random.random((len(word_index) + 1, EMBEDDING_DIM))
for word, i in word_index.items():
    embedding_vector = embeddings_index.get(word)
    if embedding_vector is not None:
        # words not found in embedding index will be all-zeros.
        embedding_matrix[i] = embedding_vector

# building Hierachical Attention network
embedding_matrix = np.random.random((len(word_index) + 1, EMBEDDING_DIM))
for word, i in word_index.items():
    embedding_vector = embeddings_index.get(word)
    if embedding_vector is not None:
        # words not found in embedding index will be all-zeros.
        embedding_matrix[i] = embedding_vector

embedding_layer = Embedding(len(word_index) + 1,
                            EMBEDDING_DIM,
                            weights=[embedding_matrix],
                            input_length=MAX_SENT_LENGTH,
                            trainable=True,
                            mask_zero=True)


class AttLayer(Layer):
def __init__(self, attention_dim,**kwargs):
    self.init = initializers.get('normal')
    self.supports_masking = True
    self.attention_dim = attention_dim
    super(AttLayer, self).__init__(**kwargs)

def build(self, input_shape):
    assert len(input_shape) == 3
    self.W = K.variable(self.init((input_shape[-1], self.attention_dim)))
    self.b = K.variable(self.init((self.attention_dim,)))
    self.u = K.variable(self.init((self.attention_dim, 1)))
    self.trainable_weights = [self.W, self.b, self.u]
    super(AttLayer, self).build(input_shape)

def compute_mask(self, inputs, mask=None):
    return None

def call(self, x, mask=None):
    # size of x :[batch_size, sel_len, attention_dim]
    # size of u :[batch_size, attention_dim]
    # uit = tanh(xW+b)
    uit = K.tile(K.expand_dims(self.W, axis=0), (K.shape(x)[0], 1, 1))
    uit = tf.matmul(x, uit)
    uit = K.tanh(K.bias_add(uit, self.b))
    ait = K.dot(uit, self.u)
    ait = K.squeeze(ait, -1)

    ait = K.exp(ait)

    if mask is not None:
        # Cast the mask to floatX to avoid float64 upcasting in theano
        ait *= K.cast(mask, K.floatx())
    ait /= K.cast(K.sum(ait, axis=1, keepdims=True) + K.epsilon(), K.floatx())
    ait = K.expand_dims(ait)
    weighted_input = x * ait
    output = K.sum(weighted_input, axis=1)

    return output

def compute_output_shape(self, input_shape):
    return (input_shape[0], input_shape[-1])

#################################################
# Change 2. The model contains only one attention block now
#################################################
sentence_input = Input(shape=(MAX_SENT_LENGTH,), dtype='int32')
embedded_sequences = embedding_layer(sentence_input)
l_lstm = Bidirectional(GRU(100, return_sequences=True))(embedded_sequences)
l_dense = TimeDistributed(Dense(200))(l_lstm)
l_att = AttLayer(100)(l_dense)
############################################

preds = Dense(2, activation='softmax')(l_att)
model = Model(sentence_input, preds)

#### clone the model #### Line X
model_copy = clone_model(model)

plot_model(model, to_file="model.png")
model.summary()
model.compile(loss='categorical_crossentropy',
              optimizer='adam',
              metrics=['accuracy'])

print("model fitting - Hierachical attention network")
model.fit(x_train, y_train, validation_data=(x_val, y_val),
          nb_epoch=10, batch_size=50,verbose=2)

ERROR: the last line of the code generates an error trace of:

Epoch 1/10
Traceback (most recent call last):
  File "/home/zz/Programs/anaconda3/lib/python3.6/site-packages/theano/compile/function_module.py", line 884, in __call__
    self.fn() if output_subset is None else\
ValueError: Input dimension mis-match. (input[0].shape[1] = 50, input[1].shape[1] = 100)

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/home/zz/Work/wop/code/python/src/3rdparty/han/textClassfierHATT2D.py", line 187, in <module>
    nb_epoch=10, batch_size=50,verbose=2)
  File "/home/zz/Programs/anaconda3/lib/python3.6/site-packages/keras/engine/training.py", line 1631, in fit
    validation_steps=validation_steps)
  File "/home/zz/Programs/anaconda3/lib/python3.6/site-packages/keras/engine/training.py", line 1213, in _fit_loop
    outs = f(ins_batch)
  File "/home/zz/Programs/anaconda3/lib/python3.6/site-packages/keras/backend/theano_backend.py", line 1223, in __call__
    return self.function(*inputs)
  File "/home/zz/Programs/anaconda3/lib/python3.6/site-packages/theano/compile/function_module.py", line 898, in __call__
    storage_map=getattr(self.fn, 'storage_map', None))
  File "/home/zz/Programs/anaconda3/lib/python3.6/site-packages/theano/gof/link.py", line 325, in raise_with_op
    reraise(exc_type, exc_value, exc_trace)
  File "/home/zz/Programs/anaconda3/lib/python3.6/site-packages/six.py", line 692, in reraise
    raise value.with_traceback(tb)
  File "/home/zz/Programs/anaconda3/lib/python3.6/site-packages/theano/compile/function_module.py", line 884, in __call__
    self.fn() if output_subset is None else\
ValueError: Input dimension mis-match. (input[0].shape[1] = 50, input[1].shape[1] = 100)
Apply node that caused the error: Elemwise{mul,no_inplace}(InplaceDimShuffle{x,0}.0, Elemwise{Cast{float32}}.0)
Toposort index: 459
Inputs types: [TensorType(float32, row), TensorType(float32, matrix)]
Inputs shapes: [(1, 50), (50, 100)]
Inputs strides: [(200, 4), (400, 4)]
Inputs values: ['not shown', 'not shown']
Outputs clients: [[Sum{axis=[1], acc_dtype=float64}(Elemwise{mul,no_inplace}.0)]]

HINT: Re-running with most Theano optimization disabled could give you a back-trace of when this node was created. This can be done with by setting the Theano flag 'optimizer=fast_compile'. If that does not work, Theano optimizations can be disabled with 'optimizer=None'.
HINT: Use the Theano flag 'exception_verbosity=high' for a debugprint and storage map footprint of this apply node.

I would really appreciate some advice on this, many thanks in advance!


回答1:


In the referenced tutorial, it choose to use theano instead of tensorflow because the behavior of dot in tensorflow is different from that in numpy. But I am not familiar with theano so it's hard for me to make it work correctly by using theano backend. I rather use a series of operations to mimic the behavior of dot in numpy. Following I've changed the K.dot to a series of operations.

import tensorflow as tf
import numpy as np
import pandas as pd
import re
from bs4 import BeautifulSoup
import os
from keras.preprocessing.text import Tokenizer, text_to_word_sequence
from keras.utils import plot_model
from keras.utils.np_utils import to_categorical
from keras.layers import Dense, Input
from keras.layers import Embedding, GRU, Bidirectional, TimeDistributed, Lambda
from keras.models import Model
from keras import backend as K
from keras.engine.topology import Layer
from keras import initializers

class AttLayer(Layer):
    def __init__(self, attention_dim, **kwargs):
        self.init = initializers.get('normal')
        self.supports_masking = True
        self.attention_dim = attention_dim
        super(AttLayer, self).__init__(**kwargs)

    def build(self, input_shape):
        assert len(input_shape) == 3
        self.W = K.variable(self.init((input_shape[-1], self.attention_dim)))
        self.b = K.variable(self.init((self.attention_dim,)))
        self.u = K.variable(self.init((self.attention_dim, 1)))
        self.trainable_weights = [self.W, self.b, self.u]
        super(AttLayer, self).build(input_shape)

    def compute_mask(self, inputs, mask=None):
        return None

    def call(self, x, mask=None):
        # size of x :[batch_size, sel_len, attention_dim]
        # size of u :[batch_size, attention_dim]
        # uit = tanh(xW+b)
        uit = K.tile(K.expand_dims(self.W, axis=0), (K.shape(x)[0], 1, 1))
        uit = tf.matmul(x, uit)
        uit = K.tanh(K.bias_add(uit, self.b))
        ait = K.dot(uit, self.u)
        ait = K.squeeze(ait, -1)

        ait = K.exp(ait)

        if mask is not None:
            # Cast the mask to floatX to avoid float64 upcasting in theano
            ait *= K.cast(mask, K.floatx())
        ait /= K.cast(K.sum(ait, axis=1, keepdims=True) + K.epsilon(), K.floatx())
        ait = K.expand_dims(ait)
        weighted_input = x * ait
        output = K.sum(weighted_input, axis=1)

        return output

    def compute_output_shape(self, input_shape):
        return (input_shape[0], input_shape[-1])

    # https://github.com/keras-team/keras/issues/5401
    # solve the problem of keras.models.clone_model
    # and model.save_weights, model.load_weights
    def get_config(self):
        config = {'attention_dim': self.attention_dim}
        base_config = super(AttLayer, self).get_config()
        return dict(list(base_config.items()) + list(config.items()))

Also compute_mask now returns None because there is no sel_len axis in AttLayer's output.

Following is a script validating that the two operations are equivalent:

B = 8
S = 100
E = 200
A = 50
X = np.random.randn(B, S, E)
W = np.random.randn(E, A)
np_result = np.dot(X, W) #shape correct

X_ph = tf.placeholder(tf.float64)
W_ph = tf.placeholder(tf.float64)

tf_dot = tf.matmul(X_ph, 
                   tf.tile(
                           tf.expand_dims(W_ph, axis=0),
                           (K.shape(X_ph)[0], 1, 1)))

with tf.Session() as sess:
    tf_result = sess.run(tf_dot,
                         feed_dict = {X_ph:X, W_ph:W})

print(np.allclose(np_result, tf_result)) #True

Training history(I set batch_size as 8):

Train on 20000 samples, validate on 5000 samples
Epoch 1/10
20000/20000 [==============================] - 1247s 62ms/step - loss: 0.4203 - acc: 0.8044 - val_loss: 0.3520 - val_acc: 0.8468
Epoch 2/10
20000/20000 [==============================] - 985s 49ms/step - loss: 0.2344 - acc: 0.9070 - val_loss: 0.3411 - val_acc: 0.8586
Epoch 3/10
20000/20000 [==============================] - 996s 50ms/step - loss: 0.0982 - acc: 0.9628 - val_loss: 0.4474 - val_acc: 0.8512
Epoch 4/10
20000/20000 [==============================] - 966s 48ms/step - loss: 0.0285 - acc: 0.9904 - val_loss: 0.7837 - val_acc: 0.8408
Epoch 5/10
20000/20000 [==============================] - 912s 46ms/step - loss: 0.0179 - acc: 0.9936 - val_loss: 1.0177 - val_acc: 0.8440
Epoch 6/10
20000/20000 [==============================] - 910s 45ms/step - loss: 0.0105 - acc: 0.9963 - val_loss: 1.0635 - val_acc: 0.8418
Epoch 7/10
20000/20000 [==============================] - 909s 45ms/step - loss: 0.0101 - acc: 0.9964 - val_loss: 1.0966 - val_acc: 0.8372
Epoch 8/10
20000/20000 [==============================] - 909s 45ms/step - loss: 0.0057 - acc: 0.9981 - val_loss: 1.2678 - val_acc: 0.8392
Epoch 9/10
20000/20000 [==============================] - 910s 46ms/step - loss: 0.0077 - acc: 0.9974 - val_loss: 1.2166 - val_acc: 0.8258
Epoch 10/10
20000/20000 [==============================] - 910s 46ms/step - loss: 0.0056 - acc: 0.9985 - val_loss: 1.4640 - val_acc: 0.8204


来源:https://stackoverflow.com/questions/54967216/hierarchical-attention-network-model-fit-generates-error-valueerror-input-di

易学教程内所有资源均来自网络或用户发布的内容,如有违反法律规定的内容欢迎反馈
该文章没有解决你所遇到的问题?点击提问,说说你的问题,让更多的人一起探讨吧!