def _create_output_layer(self):
output_layer = []
for i in range(0, self.number_of_outputs):
perceptron = Perceptron(
layer=len(self.layers),
activation_function=self.activation_function)
output_layer.append(perceptron)
self._connect_neuron(perceptron)
return output_layer
def train_perceptron():
network = Perceptron()
input_count = len(dataset[0].inputs)
print('----------------------------')
print('Generating layers')
for _ in range(input_count):
network.s_layer.add_neuron(None, lambda value: value)
print('S-layer generated')
a_neurons_count = 2 ** input_count - 1
for position in range(a_neurons_count):
neuron = ANeuron(None, lambda value: int(value >= 0))
# инициализация весов нейронов А слоя
neuron.input_weights = [
random.choice([-1, 0, 1]) for i in range(input_count)
]
neuron.calculate_bias()
network.a_layer.neurons.append(neuron)
print('A-layer generated')
for _ in range(NUMBER_COUNT):
network.r_layer.add_neuron(a_neurons_count, lambda: 0, lambda value: 1 if value >=0 else -1, 0.01, 0)
print('R-layer generated')
network.train(dataset)
network.optimize(dataset)
return network
def test_predict(self):
p = Perceptron()
p.weights = [1, 2]
self.assertEqual(p.predict([1, 2]), 1)
p.weights = [-1, -2]
self.assertEqual(p.predict([1, 2]), 0)
from perceptron.perceptron import Perceptron
import numpy as np
X = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
y = np.array([[0], [1], [1], [1]])
print("[INFO] training perceptron...")
p = Perceptron(X.shape[1], alpha=0.1)
p.fit(X, y, epochs=20)
print("[INFO[ testing perceptron...")
for (x, target) in zip(X, y):
pred = p.predict(x)
print("[INFO] data = {}. ground truth={}, pred = {}".format(
x, target[0], pred))
def __init__(self, input_num):
'''初始化线性单元，设置输入参数的个数'''
Perceptron.__init__(self, input_num, f)
from matplotlib import pyplot as plt from sklearn.model_selection import train_test_split from perceptron.perceptron import Perceptron from utils.abalone_data import get_abalone x, y = get_abalone() print(x.shape) print(y.shape) x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.20) model = Perceptron(input_shape=[1, 9]) model.train(x_train, y_train, 0.0005, 100) plt.plot(model.historical_error) plt.show()
def setUp(self):
self.perceptron = Perceptron()
class Parser():
"""A transition-based dependency parser.
This parser implements the arc-standard algorithm for dependency
parsing. When being presented with an input sentence, it first
tags the sentence for parts of speech, and then uses a multi-class
perceptron classifier to predict a sequence of *moves*
(transitions) that construct a dependency tree for the input
sentence. Moves are encoded as integers as follows:
SHIFT = 0, LEFT-ARC = 1, RIGHT-ARC = 2, SWAP = 3
At any given point in the predicted sequence, the state of the
parser can be specified by: a buffer containing the words in the
input sentence that the parser has not yet started to process; a
stack holding the indices of those words that are currently being
processed; and a partial dependency tree, represented as a list of
indices such that `tree[i]` gives the index of the head (parent
node) of the word at position `i`, or 0 in case the corresponding
word has not yet been assigned a head.
Attributes:
tagger: A part-of-speech tagger.
classifier: A multi-class perceptron classifier used to
predict the next move of the parser.
"""
def __init__(self, tagger):
"""Initializes a new parser."""
self.tagger = tagger
self.classifier = Perceptron()
def initial_config(self, words):
"""Initializes the config for the parser
Args:
words: the words of a sentence
Returns:
a initial parser config
"""
config = {}
config['score'] = 0
config['pred_tree'] = [0] * len(words)
config['stack'] = []
config['buffer'] = list(range(len(words)))
config['next_move'] = 0
config['is_gold'] = True
return config
def predict(self, feat, candidates):
"""Calls the predict function of the classifier and applies the softmax
function on the scores
Args:
feat: a feature vector
candidates: the possible moves
Returns:
the possible moves with their respective scores
"""
_, scores = self.classifier.predict(feat, candidates)
if scores:
# apply softmax on the scores
scores_lst = [(k, v) for k, v in scores.items()]
softmax_scores = softmax(list(zip(*scores_lst))[1])
scores = dict(list(zip(list(zip(*scores_lst))[0], softmax_scores)))
return scores
def update_and_reset_config(self, config, feat, gold_move):
"""This functions is called when the gold_tree falls of the beam. It
updates the classifier and resets the parser config such that only the
gold configuration is in the beam.
Args:
config: the parser gold config
feat: a featire vector
gold_move: the correct move
Returns:
the new config
"""
config['next_move'] = gold_move
self.classifier.update(feat, gold_move)
return [config]
def parse(self, words, gold_tree=None, beam_size=10):
"""Parses a sentence and also updates the classifier
if a gold tree was passed to the function
Args:
words: The input sentence, a list of words.
gold_tree: if a gold_tree is passed, the classifier is trained
beam_size: the width of the beam, when using beam search
Returns:
A pair consisting of the predicted tags and the predicted
dependency tree for the input sentence.
"""
if gold_tree:
word_order = self.get_word_order(gold_tree)
tags = self.tagger.tag(words)
possible_configs = [self.initial_config(words)]
while any(config['next_move'] != None for config in possible_configs):
old_possible_configs = possible_configs
possible_configs = []
for config in old_possible_configs:
config = self.move(config)
candidates = self.valid_moves(config)
if candidates:
feat = self.features(words, tags, config)
scores = self.predict(feat, candidates)
if gold_tree:
gold_move = self.gold_move(config, gold_tree, \
word_order)
if config['is_gold'] and gold_move not in scores:
possible_configs = self.update_and_reset_config( \
config, feat, gold_move)
break
# add new configs for the possible moves
for curr_move, curr_score in scores.items():
# create a copy of the config and append it to the list
new_config = deepcopy(config)
if curr_score > 0:
new_config['score'] += log(curr_score)
else:
new_config['score'] += float("-inf")
new_config['next_move'] = curr_move
if gold_tree and gold_move != curr_move:
new_config['is_gold'] = False
possible_configs.append(new_config)
else:
config['next_move'] = None
possible_configs.append(config)
# delete the configs with the lowest scores
while len(possible_configs) > beam_size:
worst_conf_ind, worst_conf = \
min(enumerate(possible_configs),
key = lambda t: t[1]['score'])
if gold_tree and worst_conf['is_gold'] == True:
feat = self.features(words, tags, worst_conf)
possible_configs = self.update_and_reset_config( \
worst_conf, feat, worst_conf['next_move'])
else:
del possible_configs[worst_conf_ind]
# return best tree
best_config = max(possible_configs, key=lambda t: t['score'])
return tags, best_config['pred_tree']
def valid_moves(self, config):
"""Returns the valid moves for the specified parser
configuration.
Args:
config: the current parser configuration
Returns:
The list of valid moves for the specified parser
configuration.
"""
moves = []
if len(config['buffer']) > 0:
moves.append(0)
if len(config['stack']) > 2:
moves.append(1)
if len(config['stack']) > 1:
moves.append(2)
if len(config['stack']
) > 2 and config['stack'][-1] > config['stack'][-2]:
moves.append(3)
return moves
def move(self, config):
"""Executes a single move.
Args:
config: the current parser configuration
Returns:
The new parser configuration
"""
if config['next_move'] == 0:
config['stack'].append(config['buffer'].pop(0))
elif config['next_move'] == 1:
config['pred_tree'][config['stack'][-2]] = config['stack'][-1]
del config['stack'][-2]
elif config['next_move'] == 2:
config['pred_tree'][config['stack'][-1]] = config['stack'][-2]
del config['stack'][-1]
elif config['next_move'] == 3:
config['buffer'].insert(0, config['stack'].pop(-2))
return config
def is_descendant(self, tree, ancestor, descendant):
"""Returns true if a certain node is a descendant of another node or
ancestor == descendant
Args:
tree: the dependency tree
ancestor: the ancestor node
descendant: the descendant node
Returns:
True or False
"""
if ancestor == descendant:
return True
if descendant:
return self.is_descendant(tree, ancestor, tree[descendant])
else:
return False
def get_word_order(self, gold_tree):
"""Returns the word order such that the tree would be projective
Args:
gold_tree: the pependency tree of a sentence
Returns:
list of word indices
"""
words = list(range(len(gold_tree)))
tree = gold_tree.copy()
word_order = [words.pop(0)]
del tree[0]
while words:
node = word_order[-1]
# children and their children
if node in tree:
for i in range(len(words)):
if self.is_descendant(gold_tree, node, words[i]):
word_order.append(words.pop(i))
del tree[i]
break
# siblings and their children
elif gold_tree[node] in tree:
for i in range(len(words)):
if self.is_descendant(gold_tree, gold_tree[node],
words[i]):
word_order.append(words.pop(i))
del tree[i]
break
# parent
elif gold_tree[node] in words:
ind = words.index(gold_tree[node])
word_order.append(words.pop(ind))
del tree[ind]
else:
while node:
node = gold_tree[node]
# relatives
if node in tree:
for i in range(len(words)):
if self.is_descendant(gold_tree, node, words[i]):
word_order.append(words.pop(i))
del tree[i]
break
break
# ancestors
if node in words:
ind = words.index(node)
word_order.append(words.pop(ind))
del tree[ind]
return word_order
def train(self, data, beam_size=10, n_epochs=1, trunc_data=None):
"""Trains the parser on training data.
Args:
data: Training data, a list of sentences with gold trees.
beam_size: the width of the beam, when using beam search
n_epochs: for how many epochs the parser should be trained
trunc_data: if it should stop after processing only a port of the
data (only used during development)
"""
print("Training syntactic parser:")
for e in range(n_epochs):
print("Epoch:", e + 1, "/", n_epochs)
train_sentences_tags_trees = zip( get_sentences(data), \
get_tags(data), \
get_trees(data) )
for i, (words, gold_tags, gold_tree) in \
enumerate(train_sentences_tags_trees):
self.parse(words, gold_tree, beam_size=beam_size)
print("\rUpdated with sentence #{}".format(i), end="")
if trunc_data and i >= trunc_data:
break
print("")
self.finalize()
def gold_move(self, config, gold_tree, word_order):
"""Returns the gold-standard move for the specified parser
configuration.
The gold-standard move is the first possible move from the
following list: LEFT-ARC, RIGHT-ARC, SHIFT, SWAP.
Args:
buffer: the current configuration of the parser
gold_tree: The gold-standard dependency tree.
word_order: the projective word order
Returns:
The gold-standard move for the specified parser
configuration, or `None` if no move is possible.
"""
buffer = config['buffer']
stack = config['stack']
pred_tree = config['pred_tree']
left_arc_possible = False
if len(stack) > 2 and stack[-1] == gold_tree[stack[-2]]:
left_arc_possible = True
for j in range(len(pred_tree)):
if gold_tree[j] == stack[-2]:
if pred_tree[j] == 0:
left_arc_possible = False
right_arc_possible = False
if len(stack) > 1 and stack[-2] == gold_tree[stack[-1]]:
right_arc_possible = True
for j in range(len(pred_tree)):
if gold_tree[j] == stack[-1]:
if pred_tree[j] == 0:
right_arc_possible = False
swap_possible = False
if len(stack) > 2 and \
word_order.index(stack[-1]) < word_order.index(stack[-2]):
swap_possible = True
if left_arc_possible:
return 1
elif right_arc_possible:
return 2
elif swap_possible:
return 3
elif len(buffer) > 0:
return 0
else:
return None
def features(self, words, tags, config):
"""Extracts features for the specified parser configuration.
Args:
words: The input sentence, a list of words.
tags: The list of tags for the input sentence.
config: the current configuration of the parser
Returns:
A feature vector for the specified configuration.
"""
buffer = config['buffer']
stack = config['stack']
pred_tree = config['pred_tree']
feat = []
# Single word features
b1_w = words[buffer[0]] if buffer else "<empty>"
b1_t = tags[buffer[0]] if buffer else "<empty>"
b1_wt = b1_w + " " + b1_t
b2_w = words[buffer[1]] if len(buffer) > 1 else "<empty>"
b2_t = tags[buffer[1]] if len(buffer) > 1 else "<empty>"
b2_wt = b2_w + " " + b2_t
b3_w = words[buffer[2]] if len(buffer) > 2 else "<empty>"
b3_t = tags[buffer[2]] if len(buffer) > 2 else "<empty>"
b3_wt = b3_w + " " + b3_t
s1_w = words[stack[-1]] if stack else "<empty>"
s1_t = tags[stack[-1]] if stack else "<empty>"
s1_wt = s1_w + " " + s1_t
s2_w = words[stack[-2]] if len(stack) > 1 else "<empty>"
s2_t = tags[stack[-2]] if len(stack) > 1 else "<empty>"
s2_wt = s2_w + " " + s2_t
'''
for i in pred_tree:
if stack and pred_tree[stack[-1]] == i:
feat.append("tag" + str(i) + str(tags[i]))
'''
# Triple word features
def is_parent(parent, child):
if child == 0:
return False
if parent == child:
return True
return is_parent(parent, pred_tree[child])
# Child that is the most on the left
def lc1(parent):
for i in range(0, len(words)):
if is_parent(parent, i):
return i
return -1
# Child that is the most on the right
def rc1(parent):
for i in range(0, len(words), -1):
if is_parent(parent, i):
return i
return -1
lc1_s1 = lc1(stack[-1]) if stack else -1
rc1_s1 = rc1(stack[-1]) if stack else -1
lc1_s2 = lc1(stack[-2]) if len(stack) > 1 else -1
rc1_s2 = rc1(stack[-2]) if len(stack) > 1 else -1
s2_t_s1_t_b1_t = s2_t + " " + s1_t + " " + b1_t
if lc1_s1 >= 0:
s2_t_s1_t_lc1_s1_t = s2_t + " " + s1_t + " " + tags[lc1_s1]
else:
s2_t_s1_t_lc1_s1_t = "<empty>"
if rc1_s1 >= 0:
s2_t_s1_t_rc1_s1_t = s2_t + " " + s1_t + " " + tags[rc1_s1]
else:
s2_t_s1_t_rc1_s1_t = "<empty>"
if lc1_s2 >= 0:
s2_t_s1_t_lc1_s2_t = s2_t + " " + s1_t + " " + tags[rc1_s2]
else:
s2_t_s1_t_lc1_s2_t = "<empty>"
if rc1_s2 >= 0:
s2_t_s1_t_rc1_s2_t = s2_t + " " + s1_t + " " + tags[rc1_s2]
else:
s2_t_s1_t_rc1_s2_t = "<empty>"
if lc1_s2 >= 0:
s2_t_s1_w_rc1_s2_t = s2_t + " " + s1_w + " " + tags[rc1_s2]
else:
s2_t_s1_w_rc1_s2_t = "<empty>"
if lc1_s1 >= 0:
s2_t_s1_w_lc1_s1_t = s2_t + " " + s1_w + " " + tags[lc1_s1]
else:
s2_t_s1_w_lc1_s1_t = "<empty>"
feat.append("b1_w:" + b1_w)
feat.append("b1_t:" + b1_t)
feat.append("b1_wt:" + b1_wt)
feat.append("b2_w:" + b2_w)
feat.append("b2_t:" + b2_t)
feat.append("b2_wt:" + b2_wt)
feat.append("b3_w:" + b3_w)
feat.append("b3_t:" + b3_t)
feat.append("b3_wt:" + b3_wt)
feat.append("s1_w:" + s1_w)
feat.append("s1_t:" + s1_t)
feat.append("s1_wt:" + s1_wt)
feat.append("s2_w:" + s2_w)
feat.append("s2_t:" + s2_t)
feat.append("s2_wt:" + s2_wt)
feat.append("s1_wt_s2_wt:" + s1_wt + " " + s2_wt)
feat.append("s1_wt_s2_w:" + s1_wt + " " + s2_w)
feat.append("s1_wt_s2_t:" + s1_wt + " " + s2_t)
feat.append("s1_w_s2_wt:" + s1_w + " " + s2_wt)
feat.append("s1_t_s2_wt:" + s1_t + " " + s2_wt)
feat.append("s1_w_s2_w:" + s1_w + " " + s2_w)
feat.append("s1_t_s2_t:" + s1_t + " " + s2_t)
feat.append("s1_t_b1_t:" + s1_t + " " + b1_t)
feat.append("s2_t_s1_t_b1_t:" + s2_t_s1_t_b1_t)
feat.append("s2_t_s1_t_lc1_s1_t:" + s2_t_s1_t_lc1_s1_t)
feat.append("s2_t_s1_t_rc1_s1_t:" + s2_t_s1_t_rc1_s1_t)
feat.append("s2_t_s1_t_lc1_s2_t:" + s2_t_s1_t_lc1_s2_t)
feat.append("s2_t_s1_t_rc1_s2_t:" + s2_t_s1_t_rc1_s2_t)
feat.append("s2_t_s1_w_rc1_s2_t:" + s2_t_s1_w_rc1_s2_t)
feat.append("s2_t_s1_w_lc1_s1_t:" + s2_t_s1_w_lc1_s1_t)
return feat
def finalize(self):
"""Averages the weight vectors."""
self.classifier.finalize()
def __init__(self, tagger):
"""Initializes a new parser."""
self.tagger = tagger
self.classifier = Perceptron()
#And Implementation
#Third attribute is bias (always one)
from perceptron.perceptron import Perceptron
from perceptron.tester import Tester
# The weights are set manually.
perAnd = Perceptron([0.5,0.5,-0.6])
# All states are tested.
print(perAnd.binaryOutput([-1,-1,1]))
print(perAnd.binaryOutput([-1,1,1]))
print(perAnd.binaryOutput([1,-1,1]))
print(perAnd.binaryOutput([1,1,1]))
# Tester shows the success rate of the classification. (In this easy case it is 1.0 (100%).
tester = Tester()
tester.setDataset([[-1,-1,1,-1],[-1,1,1,-1],[1,-1,1,-1],[1,1,1,1]])
print("Success rate:",tester.testPerceptron(perAnd))
#And Implementation by training
#Third attribute is bias (always one)
from perceptron.perceptron import Perceptron
from perceptron.tester import Tester
# Initial weights are way off.
perAnd = Perceptron([0.3, -0.2, 0.6])
# The perceptron is not working before training
print("Before Training:")
print(perAnd.binaryOutput([-1, -1, 1]))
print(perAnd.binaryOutput([-1, 1, 1]))
print(perAnd.binaryOutput([1, -1, 1]))
print(perAnd.binaryOutput([1, 1, 1]))
# As the states are limited, we use them several times to train the perceptron.
trainingDataset = [[-1, -1, 1, -1], [-1, 1, 1, -1], [1, -1, 1,
-1], [1, 1, 1, 1],
[-1, -1, 1, -1], [-1, 1, 1, -1], [1, -1, 1,
-1], [1, 1, 1, 1],
[-1, -1, 1, -1], [-1, 1, 1, -1], [1, -1, 1,
-1], [1, 1, 1, 1],
[-1, -1, 1, -1], [-1, 1, 1, -1], [1, -1, 1,
-1], [1, 1, 1, 1],
[-1, -1, 1, -1], [-1, 1, 1, -1], [1, -1, 1,
-1], [1, 1, 1, 1],
[-1, -1, 1, -1], [-1, 1, 1, -1], [1, -1, 1, -1],
[1, 1, 1, 1]]
perAnd.train(trainingDataset)
from perceptron.perceptron import Perceptron
from linear_discriminant_analysis.fisher_lda import FisherLDA
import pandas as pd
import numpy as np
# Pereceptron Test
df = pd.read_csv('datasets/dataset_1.csv', names=['X1', 'X2', 'y'])
df['y'] = df['y'].replace(0, -1)
print(df.head())
p = Perceptron(0.01, 20, 2, 0, '3', './images/')
X = np.array(df[['X1', 'X2']])
y = df['y']
p.fit(X, y)
# Fishers LDA Test
disc = FisherLDA(dataset=1)
disc.visualize()
class Tagger():
"""A part-of-speech tagger based on a multi-class perceptron
classifier.
This tagger implements a simple, left-to-right tagging algorithm
where the prediction of the tag for the next word in the sentence
can be based on the surrounding words and the previously
predicted tags. The exact features that this prediction is based
on can be controlled with the `features()` method, which should
return a feature vector that can be used as an input to the
multi-class perceptron.
Attributes:
classifier: A multi-class perceptron classifier.
"""
def __init__(self):
"""Initialises a new tagger."""
self.classifier = Perceptron()
def features(self, words, i, pred_tags):
"""Extracts features for the specified tagger configuration.
Args:
words: The input sentence, a list of words.
i: The index of the word that is currently being tagged.
pred_tags: The list of previously predicted tags.
Returns:
A feature vector for the specified configuration.
"""
features = []
for n in range(4):
features.append("w_0=" + words[i])
if words[i][0] == words[i][0].upper():
features.append("capital_word")
if words[i] == words[i].lower():
features.append("lowercase")
if i > 0:
features.append("t_-1=" + pred_tags[i - 1])
features.append("suff1_-1=" + words[i - 1][-1])
features.append("suff2_-1=" + words[i - 1][-2:])
features.append("suff3_-1=" + words[i - 1][-3:])
features.append("pre2_-1=" + words[i - 1][:2])
if i + 1 < len(words):
features.append("w_1" + words[i + 1])
features.append("suff1_1=" + words[i + 1][-1])
features.append("suff2_1=" + words[i + 1][-2:])
features.append("suff3_1=" + words[i + 1][-3:])
features.append("pre1_1=" + words[i + 1][0])
features.append("pre2_1=" + words[i + 1][:2])
features.append("pre3_1=" + words[i + 1][:3])
if i + 2 < len(words):
features.append("w_2" + words[i + 2])
if i + 3 < len(words):
features.append("w_3" + words[i + 3])
features.append("suff1_0=" + words[i][-1])
features.append("suff2_0=" + words[i][-2:])
features.append("suff3_0=" + words[i][-3:])
features.append("pre1_0=" + words[i][0])
features.append("pre2_0=" + words[i][:2])
features.append("pre3_0=" + words[i][:3])
return features
def tag(self, words):
"""Tags a sentence with part-of-speech tags.
Args:
words: The input sentence, a list of words.
Returns:
The list of predicted tags for the input sentence.
"""
pred_tags = []
for i in range(len(words)):
feat = self.features(words, i, pred_tags)
tag, _ = self.classifier.predict(feat)
pred_tags.append(tag)
return pred_tags
def update(self, words, gold_tags):
"""Updates the tagger with a single training instance.
Args:
words: The list of words in the input sentence.
gold_tags: The list of gold-standard tags for the input
sentence.
Returns:
The list of predicted tags for the input sentence.
"""
pred_tags = []
for i in range(len(words)):
feat = self.features(words, i, pred_tags)
pred_tags.append(self.classifier.update(feat, gold_tags[i]))
return pred_tags
def train(self, data, n_epochs=1, trunc_data=None):
"""Train a new tagger on training data.
Args:
data: Training data, a list of tagged sentences.
"""
print("Training POS tagger")
for e in range(n_epochs):
print("Epoch:", e + 1, "/", n_epochs)
train_sentences_tags = zip(get_sentences(data), get_tags(data))
for i, (words, tags) in enumerate(train_sentences_tags):
print("\rUpdated with sentence #{}".format(i), end="")
self.update(words, tags)
if trunc_data and i >= trunc_data:
break
print("")
self.finalize()
def finalize(self):
"""Finalizes the classifier by averaging its weight vectors."""
self.classifier.finalize()
def __init__(self):
"""Initialises a new tagger."""
self.classifier = Perceptron()