Related
I'm using the HuggingFace Transformers BERT model, and I want to compute a summary vector (a.k.a. embedding) over the tokens in a sentence, using either the mean or max function. The complication is that some tokens are [PAD], so I want to ignore the vectors for those tokens when computing the average or max.
Here's an example. I initially instantiate a BertTokenizer and a BertModel:
import torch
import transformers
from transformers import AutoTokenizer, AutoModel
transformer_name = 'bert-base-uncased'
tokenizer = AutoTokenizer.from_pretrained(transformer_name, use_fast=True)
model = AutoModel.from_pretrained(transformer_name)
I then input some sentences into the tokenizer and get out input_ids and attention_mask. Notably, an attention_mask value of 0 means that the token was a [PAD] that I can ignore.
sentences = ['Deep learning is difficult yet very rewarding.',
'Deep learning is not easy.',
'But is rewarding if done right.']
tokenizer_result = tokenizer(sentences, max_length=32, padding=True, return_attention_mask=True, return_tensors='pt')
input_ids = tokenizer_result.input_ids
attention_mask = tokenizer_result.attention_mask
print(input_ids.shape) # torch.Size([3, 11])
print(input_ids)
# tensor([[ 101, 2784, 4083, 2003, 3697, 2664, 2200, 10377, 2075, 1012, 102],
# [ 101, 2784, 4083, 2003, 2025, 3733, 1012, 102, 0, 0, 0],
# [ 101, 2021, 2003, 10377, 2075, 2065, 2589, 2157, 1012, 102, 0]])
print(attention_mask.shape) # torch.Size([3, 11])
print(attention_mask)
# tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
# [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
# [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0]])
Now, I call the BERT model to get the 768-D token embeddings (the top-layer hidden states).
model_result = model(input_ids, attention_mask=attention_mask, return_dict=True)
token_embeddings = model_result.last_hidden_state
print(token_embeddings.shape) # torch.Size([3, 11, 768])
So at this point, I have:
token embeddings in a [3, 11, 768] matrix: 3 sentences, 11 tokens, 768-D vector for each token.
attention mask in a [3, 11] matrix: 3 sentences, 11 tokens. A 1 value indicates non-[PAD].
How do I compute the mean / max over the vectors for the valid, non-[PAD] tokens?
I tried using the attention mask as a mask and then called torch.max(), but I don't get the right dimensions:
masked_token_embeddings = token_embeddings[attention_mask==1]
print(masked_token_embeddings.shape) # torch.Size([29, 768] <-- WRONG. SHOULD BE [3, 11, 768]
pooled = torch.max(masked_token_embeddings, 1)
print(pooled.values.shape) # torch.Size([29]) <-- WRONG. SHOULD BE [3, 768]
What I really want is a tensor of shape [3, 768]. That is, a 768-D vector for each of the 3 sentences.
For max, you can multiply with attention_mask:
pooled = torch.max((token_embeddings * attention_mask.unsqueeze(-1)), axis=1)
For mean, you can sum along the axis and divide by attention_mask along that axis:
mean_pooled = token_embeddings.sum(axis=1) / attention_mask.sum(axis=-1).unsqueeze(-1)
In addition to #Quang, you can have a look at sentence_transformers Pooling layer.
For max pooling, they do this:
input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
token_embeddings[input_mask_expanded == 0] = -1e9 # Set padding tokens to large negative value
pooled = torch.max(token_embeddings, 1)[0]
And for mean pooling they do the following:
input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
sum_embeddings = torch.sum(token_embeddings * input_mask_expanded, 1)
sum_mask = input_mask_expanded.sum(1)
sum_mask = torch.clamp(sum_mask, min=1e-9)
pooled = sum_embeddings / sum_mask
The max pooling presented in the accepted answer will suffer when the max is negative, and the implementation from sentence transformers changes token_embeddings, which throw an error when you want to use the embedding for back propagation:
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation:
If you're interested on anything back-prop related, you can do something like this:
input_mask_expanded = torch.where(attention_mask==0, -1e-9, 0.).unsqueeze(-1).expand(token_embeddings.size()).float()
pooled = torch.max(token_embeddings-input_mask_expanded, 1)[0] # Set padding tokens to large negative value
It's the same idea of making all masked tokens to be very small, but it doesn't change the token_embeddings while at it.
Alex is right.
Look on hidden states for strings that go into tokenizer. For different strings, padding will have different embeddings.
So, in order to properly pool embeddings, you need to ignore those padding vectors.
Let's say you want to get embeddings out of the last 4 layers of BERT (as it yields the best classification results):
#iterate over the last 4 layers and get embeddings for
#strings without having embeddings from PAD tokens
m = []
for i in range(len(hidden_states[0])):
m.append([hidden_states[j+9][i,:,:][tokens["attention_mask"][i] !=0] for j in range(4)])
#average over all tokens embeddings
means = []
for i in range(len(hidden_states[0])):
means.append(torch.stack(m[i]).mean(dim=1))
#stack embeddings for all strings
pooled = torch.stack(means).reshape(-1,1,3072)
I have a dataset (31 features including the class). This dataset is about to be used for a classification problem. I thought to check the correlation between the features using Pearson correlation exists in pandas. When I set the Pearson's threshold > 0.5, I get the following:
import pandas as pd
data = pd.read_csv("../dataset.csv")
cor = data.corr(method='pearson')
cor_target = abs(cor['Class'])
result = cor_target[cor_target > 0.5]
print(result)
The result is:
Class 1.0
Name: Class, dtype: float64
It turns out that all 30 features are not correlated at all. What does this mean? Is it always a good indicator that features are independent?
Thank you.
Your assumptions are somewhat wrong.
Take for an example:
import pandas as pd
data = pd.DataFrame({'a': [1, 2, 3, 4, 5], 'b': [1, 2, 3, 4, 5], 'Class' : [0, 1, 1, 0, 1]})
cor = data.corr(method='pearson')
print(cor)
cor_target = abs(cor['Class'])
print(cor_target)
result = cor_target[cor_target > 0.5]
print(result)
a b Class
a 1.000000 1.000000 0.288675
b 1.000000 1.000000 0.288675
Class 0.288675 0.288675 1.000000
a 0.288675
b 0.288675
Class 1.000000
Name: Class, dtype: float64
Class 1.0
Name: Class, dtype: float64
Feature set a and b are exactly the same, they have 1.0 correlation, but you'll still get only 1.
Remove the class labels, and only observe the correlation between the intermediate features.
Observe the correlation matrix and select the ones with low correlation.
import pandas as pd
data = pd.DataFrame({'a': [1, 2, 3, 4, 5], 'b': [1, 2, 3, 4, 5], 'Class' : [0, 1, 1, 0, 1]})
cor = data[['a', 'b']].corr(method='pearson')
print(cor)
cor_target = abs(cor)
a b
a 1.0 1.0
b 1.0 1.0
If you want to use labels, try scikit-learn's feature importance, https://scikit-learn.org/stable/modules/feature_selection.html
How can I build a meta-classifier in scikit-learn out of N binary classifiers which will return 1 if any of the classifiers returns 1?
Currently I've tried VotingClassifier, but it lacks the logic that I need, both with voting equal to hard and soft. Pipeline seems to be oriented towards sequential computation
I can write the logic by myself, but I am wondering if there is anything built-in?
The built-in options are only soft and hard voting. As you mentioned, we can create a custom function to this meta-classifier, which uses OR logic based on the source code here. This custom meta classifier can fit into the pipeline as well.
from sklearn.utils.validation import check_is_fitted
class CustomMetaClassifier(VotingClassifier):
def predict(self, X):
""" Predict class labels for X.
Parameters
----------
X : {array-like, sparse matrix}, shape = [n_samples, n_features]
The input samples.
Returns
----------
maj : array-like, shape = [n_samples]
Predicted class labels.
"""
check_is_fitted(self, 'estimators_')
maj = np.max(eclf1._predict(X), 1)
maj = self.le_.inverse_transform(maj)
return maj
>>> import numpy as np
>>> from sklearn.linear_model import LogisticRegression
>>> from sklearn.naive_bayes import GaussianNB
>>> from sklearn.ensemble import RandomForestClassifier, VotingClassifier
>>> clf1 = LogisticRegression(solver='lbfgs', multi_class='multinomial',
... random_state=1)
>>> clf2 = RandomForestClassifier(n_estimators=50, random_state=1)
>>> clf3 = GaussianNB()
>>> X = np.array([[-1, -1], [-2, -1], [-3, -2], [1, 1], [2, 1], [3, 2]])
>>> y = np.array([1, 1, 1, 2, 2, 2])
>>> eclf1 = CustomMetaClassifier(estimators=[
... ('lr', clf1), ('rf', clf2), ('gnb', clf3)])
>>> eclf1 = eclf1.fit(X, y)
>>> eclf1.predict(X)
array([1, 1, 1, 2, 2, 2])
I'm trying to calculate probabilities for a multi-class dataset using scikit learn. However, for some reason, I'm getting a the same probabilities for every example. Any idea what's happening? Does this have to do with my model, my use of the library, or something else? Appreciate any help!
svm_model = svm.SVC(probability=True, kernel='rbf',C=1, decision_function_shape='ovr', gamma=0.001,verbose=100)
svm_model.fit(train_X,train_y)
preds= svm_model.predict_proba(test_X)
train_X looks like this
array([[2350, 5550, 2750.0, ..., 23478, 1, 3],
[2500, 5500, 3095.5, ..., 23674, 0, 3],
[3300, 6900, 3600.0, ..., 6529, 0, 3],
...,
[2150, 6175, 2500.0, ..., 11209, 0, 3],
[2095, 5395, 2595.4, ..., 10070, 0, 3],
[1650, 2850, 2000.0, ..., 25463, 1, 3]], dtype=object)
train_y looks like this
0 1
1 2
10 2
100 2
1000 2
10000 2
10001 2
10002 2
10003 2
10004 2
10005 2
10006 2
10007 2
10008 1
10009 1
1001 2
10010 2
test_X looks like this
array([[2190, 3937, 2200.5, ..., 24891, 1, 5],
[2695, 7000, 2850.0, ..., 5491, 1, 4],
[2950, 12000, 4039.5, ..., 22367, 0, 4],
...,
[2850, 5200, 3000.0, ..., 15576, 1, 1],
[3200, 16000, 4100.0, ..., 1320, 0, 3],
[2100, 3750, 2400.0, ..., 6022, 0, 1]], dtype=object)
My results look like
array([[ 0.07819139, 0.22727628, 0.69453233],
[ 0.07819139, 0.22727628, 0.69453233],
[ 0.07819139, 0.22727628, 0.69453233],
...,
[ 0.07819139, 0.22727628, 0.69453233],
[ 0.07819139, 0.22727628, 0.69453233],
[ 0.07819139, 0.22727628, 0.69453233]])
Start with preprocessing!.
It's very important to standardize your data to zero-mean and unit-variance.
The scikit-learn docs say this:
Support Vector Machine algorithms are not scale invariant, so it is highly recommended to scale your data. For example, scale each attribute on the input vector X to [0,1] or [-1,+1], or standardize it to have mean 0 and variance 1. Note that the same scaling must be applied to the test vector to obtain meaningful results. See section Preprocessing data for more details on scaling and normalization
sklearns Section on Preprocessing
sklearns StandardScaler.
The next step after this is parameter-tuning (C, gamma and co.). This is usually done by GridSearch. But i usually expect people to try a simple LinearSVM first before trying the Kernel-SVM (less hyper-parameters, less computation-time, better generalization for non-optimal parameter-chosings).
I am trying to set up a cost-sensitive binary classification learning in TensorFlow, which would put different penalties on false positives and false negatives. Does anyone know how to create a loss function from a set of penalty weights $(w_1, w_2, w_3, w_4)$ for (true positive, false positive, false negative, true negative).
I went over the standard cost functions offered, but can't figure out how to combine them to get something similar to the above.
Following #Cauchyzhou's answer, if you have the logits, and the sparse labels as well as a cost_matrix whose shape is [L, L], where L is the number of unique labels, you can simply use the function below to calculate the loss
def sparse_cost_sensitive_loss (logits, labels, cost_matrix):
batch_cost_matrix = tf.nn.embedding_lookup(cost_matrix, labels)
eps = 1e-6
probability = tf.clip_by_value(tf.nn.softmax(logits), eps, 1-eps)
cost_values = tf.log(1-probability)*batch_cost_matrix
loss = tf.reduce_mean(-tf.reduce_sum(cost_values, axis=1))
return loss
I am not aware of anyone who has built a cost sensitive neural network classifier but Alejandro Correa Bahnsen has published academic papers for cost sensitive logistic regression and cost sensitive decision trees and a very well documented python cost sensitive classification library named CostCla. CostCla is pretty easy to use if you are familiar with scikit-learn.
You should be able to use the Bayes minimum risk model in the library to minimize the cost of your neural network since it fits a cost model to output prediction probabilities of any classifier.
Note that CostCla is intended to work with potentially different costs for each sample. You give it a cost matrix for your training and test samples. However, you can just make all the rows in the cost matrix the same if that applies to your problem.
Here are a couple of additional academic papers on the subject:
The Foundations of Cost-Sensitive Learning
Optimal ROC Curve for a Combination of Classifiers
cost_matrix:
[[0,1,100],
[1,0,1],
[1,20,0]]
label:
[1,2]
y*:
[[0,1,0],
[0,0,1]]
y(prediction):
[[0.2,0.3,0.5],
[0.1,0.2,0.7]]
label,cost_matrix-->cost_embedding:
[[1,0,1],
[1,20,0]]
It obvious 0.3 in [0.2,0.3,0.5] refers to right lable probility of [0,1,0], so it should not contibute to loss.
0.7 in [0.1,0.2,0.7] is the same. In other words, the pos with value 1 in y* not contibute to loss.
So I have (1-y*):
[[1,0,1],
[1,1,0]]
Then the entropy is target*log(predict) + (1-target) * log(1-predict),and value 0 in y*,should use (1-target)*log(1-predict), so I use (1-predict) said (1-y)
1-y:
[[0.8,*0.7*,0.5],
[0.9,0.8,*0.3*]]
(italic num is useless)
the custom loss is
[[1,0,1], [1,20,0]] * log([[0.8,0.7,0.5],[0.9,0.8,0.3]]) *
[[1,0,1],[1,1,0]]
and you can see the (1-y*) can be drop here
so the loss is -tf.reduce_mean(cost_embedding*log(1-y))
,to make it applicable , should be:
-tf.reduce_mean(cost_embedding*log(tf.clip((1-y),1e-10)))
the demo is below
import tensorflow as tf
import numpy as np
hidden_units = 50
num_class = 3
class Model():
def __init__(self,name_scope,is_custom):
self.name_scope = name_scope
self.is_custom = is_custom
self.input_x = tf.placeholder(tf.float32,[None,hidden_units])
self.input_y = tf.placeholder(tf.int32,[None])
self.instantiate_weights()
self.logits = self.inference()
self.predictions = tf.argmax(self.logits,axis=1)
self.losses,self.train_op = self.opitmizer()
def instantiate_weights(self):
with tf.variable_scope(self.name_scope + 'FC'):
self.W = tf.get_variable('W',[hidden_units,num_class])
self.b = tf.get_variable('b',[num_class])
self.cost_matrix = tf.constant(
np.array([[0,1,100],[1,0,100],[20,5,0]]),
dtype = tf.float32
)
def inference(self):
return tf.matmul(self.input_x,self.W) + self.b
def opitmizer(self):
if not self.is_custom:
loss = tf.nn.sparse_softmax_cross_entropy_with_logits\
(labels=self.input_y,logits=self.logits)
else:
batch_cost_matrix = tf.nn.embedding_lookup(
self.cost_matrix,self.input_y
)
loss = - tf.log(1 - tf.nn.softmax(self.logits))\
* batch_cost_matrix
train_op = tf.train.AdamOptimizer().minimize(loss)
return loss,train_op
import random
batch_size = 128
norm_model = Model('norm',False)
custom_model = Model('cost',True)
split_point = int(0.9 * dataset_size)
train_set = datasets[:split_point]
test_set = datasets[split_point:]
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(100):
batch_index = random.sample(range(split_point),batch_size)
train_batch = train_set[batch_index]
train_labels = lables[batch_index]
_,eval_predict,eval_loss = sess.run([norm_model.train_op,
norm_model.predictions,norm_model.losses],
feed_dict={
norm_model.input_x:train_batch,
norm_model.input_y:train_labels
})
_,eval_predict1,eval_loss1 = sess.run([custom_model.train_op,
custom_model.predictions,custom_model.losses],
feed_dict={
custom_model.input_x:train_batch,
custom_model.input_y:train_labels
})
# print '默认',eval_predict,'\n自定义',eval_predict1
print np.sum(((eval_predict == train_labels)==True).astype(np.int)),\
np.sum(((eval_predict1 == train_labels)==True).astype(np.int))
if i%10 == 0:
print '默认测试',sess.run(norm_model.predictions,
feed_dict={
norm_model.input_x:test_set,
norm_model.input_y:lables[split_point:]
})
print '自定义测试',sess.run(custom_model.predictions,
feed_dict={
custom_model.input_x:test_set,
custom_model.input_y:lables[split_point:]
})
Here is other solution where you can use any tensorflow loss and make it cost sensitive using kwarg weights ... note that unlike most cases here you need to use cost as '1' instead of '0' when you want to keep loss as it is ...
Some advantages of this approach are:
it extends tf.losses.Loss and satisfies the call api
reduction kwarg of the original loss remains functional and the behaviour is propagated to CostSensitiveLoss
you can also pass your own extra weights to new loss instances. Note that internally generated weights are used by wrapped self.loss
import numpy as np
from keras.api._v2 import keras as tk
import tensorflow as tf
from keras.utils import losses_utils
import typing as t
class CostSensitiveLoss(tk.losses.Loss):
def __init__(
self,
cost_matrix: t.List, loss: tk.losses.Loss,
):
super().__init__(reduction=loss.reduction, name=loss.name)
self.loss = loss
self.cost_matrix = cost_matrix
self._cost_matrix = tf.constant(cost_matrix, dtype=tf.float32)
#classmethod
def from_config(cls, config):
config['loss'] = tk.losses.deserialize(config['loss'])
return cls(**config)
def get_config(self):
return {
'cost_matrix': self.cost_matrix,
'loss': tk.losses.serialize(self.loss),
'reduction': self.reduction, 'name': self.name
}
def call(self, y_true, y_pred):
# if y_true is one hot encoded then get integer indices
if y_true.ndim == 1:
y_true_index = y_true
elif y_true.ndim == 2:
y_true_index = tf.argmax(y_true, axis=1)
else:
raise Exception(f"`y_true.ndim` {y_true.ndim} not supported")
# get cost for batch
cost_for_batch = tf.nn.embedding_lookup(self._cost_matrix, y_true_index)
cost_for_batch *= y_pred
cost_for_batch = tf.reduce_sum(cost_for_batch, axis=1)
# get loss
return self.loss(y_true, y_pred, cost_for_batch)
if __name__ == '__main__':
# for debug purpose I have kept 'none' you can
# safely use other options like 'sum', 'auto'
_loss = tk.losses.MeanAbsoluteError(reduction='none')
# some cost matrices the first cost matrix is the case when you are
# not using cost sensitive weights
_cs_loss_1 = CostSensitiveLoss(
cost_matrix=[[1, 1, 1], [1, 1, 1], [1, 1, 1], ],
loss=_loss
)
_cs_loss_2 = CostSensitiveLoss(
cost_matrix=[[1, 2, 2], [4, 1, 4], [8, 8, 1], ],
loss=_loss
)
_cs_loss_3 = CostSensitiveLoss(
cost_matrix=[[1, 4, 8], [2, 1, 8], [2, 4, 1], ],
loss=_loss
)
_y_true = np.asarray(
[
[1, 0, 0],
[0, 1, 0],
[0, 0, 1],
[1, 0, 0],
[0, 1, 0],
[0, 0, 1],
[1, 0, 0],
[0, 1, 0],
[0, 0, 1],
]
)
_y_pred = np.asarray(
[
[0.8, 0.1, 0.1],
[0.1, 0.8, 0.1],
[0.1, 0.1, 0.8],
[0.1, 0.8, 0.1],
[0.1, 0.1, 0.8],
[0.8, 0.1, 0.1],
[0.1, 0.1, 0.8],
[0.8, 0.1, 0.1],
[0.1, 0.8, 0.1],
]
)
print("loss ........................")
print(_loss(_y_true, _y_pred).numpy())
print("cs_loss_1 ...................")
print(_cs_loss_1(_y_true, _y_pred).numpy())
print("cs_loss_2 ...................")
print(_cs_loss_2(_y_true, _y_pred).numpy())
print("cs_loss_3 ...................")
print(_cs_loss_3(_y_true, _y_pred).numpy())