Repository: keon/pointer-networks
Branch: master
Commit: bca17beee594
Files: 9
Total size: 11.9 KB

Directory structure:
gitextract_f0mounp3/

├── .gitignore
├── LICENSE
├── PointerLSTM.py
├── README.md
├── model_weight_100.hdf5
├── requirement.txt
├── run.py
├── sort_data.py
└── tsp_data.py

================================================
FILE CONTENTS
================================================

================================================
FILE: .gitignore
================================================
# Byte-compiled / optimized / DLL files
__pycache__/
*.py[cod]
*$py.class

# C extensions
*.so

# Distribution / packaging
.Python
env/
build/
develop-eggs/
dist/
downloads/
eggs/
.eggs/
lib/
lib64/
parts/
sdist/
var/
*.egg-info/
.installed.cfg
*.egg

# PyInstaller
#  Usually these files are written by a python script from a template
#  before PyInstaller builds the exe, so as to inject date/other infos into it.
*.manifest
*.spec

# Installer logs
pip-log.txt
pip-delete-this-directory.txt

# Unit test / coverage reports
htmlcov/
.tox/
.coverage
.coverage.*
.cache
nosetests.xml
coverage.xml
*,cover
.hypothesis/

# Translations
*.mo
*.pot

# Django stuff:
*.log
local_settings.py

# Flask stuff:
instance/
.webassets-cache

# Scrapy stuff:
.scrapy

# Sphinx documentation
docs/_build/

# PyBuilder
target/

# IPython Notebook
.ipynb_checkpoints

# pyenv
.python-version

# celery beat schedule file
celerybeat-schedule

# dotenv
.env

# virtualenv
venv/
ENV/

# Spyder project settings
.spyderproject

# Rope project settings
.ropeproject


================================================
FILE: LICENSE
================================================
MIT License

Copyright (c) 2017 Keon Kim

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.


================================================
FILE: PointerLSTM.py
================================================
import keras.backend as K
from keras.activations import tanh, softmax
from keras.engine import InputSpec
from keras.layers import LSTM
import keras


class Attention(keras.layers.Layer):
    """
        Attention layer
    """

    def __init__(self, hidden_dimensions, name='attention'):
        super(Attention, self).__init__(name=name, trainable=True)
        self.W1 = keras.layers.Dense(hidden_dimensions, use_bias=False)
        self.W2 = keras.layers.Dense(hidden_dimensions, use_bias=False)
        self.V = keras.layers.Dense(1, use_bias=False)

    def call(self, encoder_outputs, dec_output, mask=None):

        w1_e = self.W1(encoder_outputs)
        w2_d = self.W2(dec_output)
        tanh_output = tanh(w1_e + w2_d)
        v_dot_tanh = self.V(tanh_output)
        if mask is not None:
            v_dot_tanh += (mask * -1e9)
        attention_weights = softmax(v_dot_tanh, axis=1)
        att_shape = K.shape(attention_weights)
        return K.reshape(attention_weights, (att_shape[0], att_shape[1]))


class Decoder(keras.layers.Layer):
    """
        Decoder class for PointerLayer
    """

    def __init__(self, hidden_dimensions):
        super(Decoder, self).__init__()
        self.lstm = keras.layers.LSTM(
            hidden_dimensions, return_sequences=False, return_state=True)

    def call(self, x, hidden_states):
        dec_output, state_h, state_c = self.lstm(
            x, initial_state=hidden_states)
        return dec_output, [state_h, state_c]

    def get_initial_state(self, inputs):
        return self.lstm.get_initial_state(inputs)

    def process_inputs(self, x_input, initial_states, constants):
        return self.lstm._process_inputs(x_input, initial_states, constants)


class PointerLSTM(keras.layers.Layer):
    """
        PointerLSTM
    """

    def __init__(self, hidden_dimensions, name='pointer', **kwargs):
        super(PointerLSTM, self).__init__(
            hidden_dimensions, name=name, **kwargs)
        self.hidden_dimensions = hidden_dimensions
        self.attention = Attention(hidden_dimensions)
        self.decoder = Decoder(hidden_dimensions)

    def build(self, input_shape):
        super(PointerLSTM, self).build(input_shape)
        self.input_spec = [InputSpec(shape=input_shape)]

    def call(self, x, training=None, mask=None, states=None):
        """
        :param Tensor x: Should be the output of the decoder
        :param Tensor states: last state of the decoder
        :param Tensor mask: The mask to apply
        :return: Pointers probabilities
        """

        input_shape = self.input_spec[0].shape
        en_seq = x
        x_input = x[:, input_shape[1] - 1, :]
        x_input = K.repeat(x_input, input_shape[1])
        if states:
            initial_states = states
        else:
            initial_states = self.decoder.get_initial_state(x_input)

        constants = []
        preprocessed_input, _, constants = self.decoder.process_inputs(
            x_input, initial_states, constants)
        constants.append(en_seq)
        last_output, outputs, states = K.rnn(self.step, preprocessed_input,
                                             initial_states,
                                             go_backwards=self.decoder.lstm.go_backwards,
                                             constants=constants,
                                             input_length=input_shape[1])

        return outputs

    def step(self, x_input, states):
        x_input = K.expand_dims(x_input,1)
        input_shape = self.input_spec[0].shape
        en_seq = states[-1]
        _, [h, c] = self.decoder(x_input, states[:-1])
        dec_seq = K.repeat(h, input_shape[1])
        probs = self.attention(dec_seq, en_seq)
        return probs, [h, c]

    def get_output_shape_for(self, input_shape):
        # output shape is not affected by the attention component
        return (input_shape[0], input_shape[1], input_shape[1])

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


================================================
FILE: README.md
================================================
## Code upgrade of [Pointer Networks](http://arxiv.org/pdf/1511.06391v4.pdf) to run on keras>=2.4.3 and tensorflow>=2.2.0 
The original author code is at https://github.com/keon/pointer-networks.git.


================================================
FILE: requirement.txt
================================================
keras==2.4.3
tensorflow==2.2.0

================================================
FILE: run.py
================================================
from keras.models import Model
from keras.layers import LSTM, Input
from keras.callbacks import LearningRateScheduler
from keras.utils.np_utils import to_categorical
from pointer import PointerLSTM
import pickle
import tsp_data as tsp
import numpy as np
import keras


def scheduler(epoch):
    if epoch < nb_epochs/4:
        return learning_rate
    elif epoch < nb_epochs/2:
        return learning_rate*0.5
    return learning_rate*0.1

print("preparing dataset...")
t = tsp.Tsp()
X, Y = t.next_batch(10000)
x_test, y_test = t.next_batch(1000)

YY = []
for y in Y:
    YY.append(to_categorical(y))
YY = np.asarray(YY)

hidden_size = 128
seq_len = 10
nb_epochs = 10000
learning_rate = 0.1

print("building model...")
main_input = Input(shape=(seq_len, 2), name='main_input')

encoder,state_h, state_c = LSTM(hidden_size,return_sequences = True, name="encoder",return_state=True)(main_input)
decoder = PointerLSTM(hidden_size, name="decoder")(encoder,states=[state_h, state_c])

model = Model(main_input, decoder)
print(model.summary())
model.compile(optimizer='adam',
              loss='categorical_crossentropy',
              metrics=['accuracy'])

model.fit(X, YY, epochs=nb_epochs, batch_size=64,)
print(model.predict(x_test))
print('evaluate : ',model.evaluate(x_test,to_categorical(y_test)))
print("------")
print(to_categorical(y_test))
model.save_weights('model_weight_100.hdf5')


================================================
FILE: sort_data.py
================================================
from __future__ import absolute_import, division, print_function
import numpy as np


class DataGenerator(object):
    def next_batch(self, batch_size, N, train_mode=True):
        """Return the next `batch_size` examples from this data set."""

        # A sequence of random numbers from [0, 1]
        encoder_batch = []

        # Sorted sequence that we feed to encoder
        # In inference we feed an unordered sequence again
        decoder_batch = []

        # Ordered sequence where one hot vector encodes
        # position in the input array
        target_batch = []
        for _ in range(batch_size):
            encoder_batch.append(np.zeros([N, 1]))
        for _ in range(batch_size):
            decoder_batch.append(np.zeros([N, 1]))
            target_batch.append(np.zeros([N, N]))

        encoder_batch = np.asarray(encoder_batch)
        decoder_batch = np.asarray(decoder_batch)
        target_batch = np.asarray(target_batch)

        for b in range(batch_size):
            shuffle = np.random.permutation(N)
            sequence = np.sort(np.random.random(N))
            shuffled_sequence = sequence[shuffle]

            for i in range(N):
                encoder_batch[b][i] = shuffled_sequence[i]
                if train_mode:
                    decoder_batch[b][i] = sequence[i]
                else:
                    decoder_batch[b][i] = shuffled_sequence[i]
                target_batch[b, i][shuffle[i]] = 1.0

            # Points to the stop symbol
            # target_batch[b, N][0] = 1.0

        return encoder_batch, decoder_batch, target_batch


if __name__ == "__main__":
    seq_len = 3
    batch_size = 3
    dataset = DataGenerator()
    enc_input, dec_input, targets = dataset.next_batch(batch_size, seq_len)
    print("batch_size", batch_size, "seq_len", seq_len)
    print("-------------encoder input-------------")
    print(enc_input.shape)
    print(enc_input)
    print("-------------decoder input-------------")
    print(dec_input.shape)
    print(dec_input)
    print("-------------   targets   -------------")
    print(targets.shape)
    print(targets)


================================================
FILE: tsp_data.py
================================================
import math
import numpy as np
import random
import itertools


class Tsp:
    def next_batch(self, batch_size=1):
        X, Y = [], []
        for b in range(batch_size):
            print("preparing dataset... %s/%s" % (b, batch_size))
            points = self.generate_data()
            solved = self.solve_tsp_dynamic(points)
            X.append(points), Y.append(solved)
        return np.asarray(X), np.asarray(Y)

    def length(self, x, y):
        return (math.sqrt((x[0]-y[0])**2 + (x[1]-y[1])**2))

    def solve_tsp_dynamic(self, points):
        # calc all lengths
        all_distances = [[self.length(x, y) for y in points] for x in points]
        # initial value - just distance from 0 to
        # every other point + keep the track of edges
        A = {(frozenset([0, idx+1]), idx+1): (dist, [0, idx+1])
             for idx, dist in enumerate(all_distances[0][1:])}
        cnt = len(points)
        for m in range(2, cnt):
            B = {}
            for S in [frozenset(C) | {0}
                      for C in itertools.combinations(range(1, cnt), m)]:
                for j in S - {0}:
                    B[(S, j)] = min([(A[(S-{j}, k)][0] + all_distances[k][j],
                                      A[(S-{j}, k)][1] + [j])
                                     for k in S if k != 0 and k != j])
            A = B
        res = min([(A[d][0] + all_distances[0][d[1]], A[d][1])
                   for d in iter(A)])
        return res[1]

    def generate_data(self, N=10):
        radius = 1
        rangeX = (0, 10)
        rangeY = (0, 10)
        qty = N

        deltas = set()
        for x in range(-radius, radius+1):
            for y in range(-radius, radius+1):
                if x*x + y*y <= radius*radius:
                    deltas.add((x, y))

        randPoints = []
        excluded = set()
        i = 0
        while i < qty:
            x = random.randrange(*rangeX)
            y = random.randrange(*rangeY)
            if (x, y) in excluded:
                continue
            randPoints.append((x, y))
            i += 1
            excluded.update((x+dx, y+dy) for (dx, dy) in deltas)
        return randPoints

if __name__ == "__main__":
    p = Tsp()
    X, Y = p.next_batch(1)
    print(X)
    print(Y)