├── results └── .gitkeep ├── tests └── .gitkeep ├── environment.yml ├── utils ├── logger.py ├── word_encodings.py ├── image_manipulation.py └── convnet14_cifar10_mnist_joint.py ├── layers ├── extracting.py ├── encoding.py ├── reading.py └── writing.py ├── README.md ├── models └── convnet14.py ├── image_association_task_lstm.py ├── data ├── babi_data.py └── image_association_data.py ├── image_association_task.py ├── babi_task_single.py └── LICENSE /results/.gitkeep: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /tests/.gitkeep: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /environment.yml: -------------------------------------------------------------------------------- 1 | name: H-Mem 2 | 3 | channels: 4 | - defaults 5 | 6 | dependencies: 7 | - python=3.7 8 | - tensorflow-gpu=2.1 9 | -------------------------------------------------------------------------------- /utils/logger.py: -------------------------------------------------------------------------------- 1 | """CSV logger""" 2 | 3 | import csv 4 | 5 | from tensorflow.keras.callbacks import Callback 6 | 7 | 8 | class MyCSVLogger(Callback): 9 | 10 | def __init__(self, filename): 11 | self.filename = filename 12 | 13 | def on_test_begin(self, logs=None): 14 | self.csv_file = open(self.filename, "a") 15 | 16 | class CustomDialect(csv.excel): 17 | delimiter = ',' 18 | 19 | self.fieldnames = ['error [%]'] 20 | self.writer = csv.DictWriter(self.csv_file, self.fieldnames, dialect=CustomDialect) 21 | self.writer.writeheader() 22 | 23 | def on_test_batch_begin(self, batch, logs=None): 24 | pass 25 | 26 | def on_test_batch_end(self, batch, logs=None): 27 | logs = {'error [%]': 100.0 - logs['accuracy'] * 100.0} 28 | self.writer.writerow(logs) 29 | self.csv_file.flush() 30 | 31 | def on_test_end(self, logs=None): 32 | self.csv_file.close() 33 | self.writer = None 34 | -------------------------------------------------------------------------------- /utils/word_encodings.py: -------------------------------------------------------------------------------- 1 | """Word encodings.""" 2 | 3 | import numpy as np 4 | 5 | 6 | def position_encoding(sentence_size, embedding_size): 7 | """Position Encoding. 8 | 9 | Encodes the position of words within the sentence (implementation based on 10 | https://arxiv.org/pdf/1503.08895.pdf [1]). 11 | 12 | Arguments: 13 | sentence_size: int, the size of the sentence (number of words). 14 | embedding_size: int, the size of the word embedding. 15 | 16 | Returns: 17 | A encoding matrix represented by a Numpy array with shape `[sentence_size, embedding_size]`. 18 | 19 | """ 20 | encoding = np.ones((embedding_size, sentence_size), dtype=np.float32) 21 | ls = sentence_size + 1 22 | le = embedding_size + 1 23 | for i in range(1, le): 24 | for j in range(1, ls): 25 | encoding[i - 1, j - 1] = (i - (embedding_size + 1) / 2) * (j - (sentence_size + 1) / 2) 26 | encoding = 1 + 4 * encoding / embedding_size / sentence_size 27 | 28 | # Make position encoding of time words identity to avoid modifying them. 29 | encoding[:, -1] = 1.0 30 | 31 | return np.transpose(encoding) 32 | -------------------------------------------------------------------------------- /utils/image_manipulation.py: -------------------------------------------------------------------------------- 1 | """Image manipulations.""" 2 | 3 | import numpy as np 4 | 5 | 6 | def merge(a, b): 7 | """Merge two images to one. 8 | 9 | The images are stacked column wise (left `a` and right `b`). 10 | 11 | Arguments: 12 | a: iterable, The images. 13 | b: iterable, The images. 14 | 15 | Returns: 16 | A Numpy array containing the merged images. 17 | 18 | """ 19 | rows_a, cols_a, channels_a = a.shape 20 | rows_b, cols_b, channels_b = b.shape 21 | 22 | rows = max(rows_a, rows_b) 23 | cols = cols_a + cols_b 24 | channels = max(channels_a, channels_b) 25 | 26 | c = np.zeros(shape=(rows, cols, channels)) 27 | c[:rows_a, :cols_a] = a 28 | c[:rows_b, cols_a:] = b 29 | 30 | return c 31 | 32 | 33 | def pad(x, pad_width): 34 | """Pad images in `x` with zeros. 35 | 36 | Arguments: 37 | x: iterable, The images to pad. 38 | pad_width: sequence, array_like, int, Number of values padded to the edges of each axis. ((before_1, 39 | after_1), … (before_N, after_N)) unique pad widths for each axis. ((before, after),) yields same before 40 | and after pad for each axis. (pad,) or int is a shortcut for before = after = pad width for all axes. 41 | 42 | Returns: 43 | A Numpy array containing the padded images. 44 | 45 | """ 46 | y = [] 47 | for item in x: 48 | y.append(np.pad(item, pad_width=pad_width)) 49 | 50 | y = np.array(y) 51 | 52 | return y 53 | 54 | 55 | def expand_channels(x, num_channels=3): 56 | y = [] 57 | for i, item in enumerate(x): 58 | rows, cols = item.shape 59 | c = np.zeros(shape=(rows, cols, num_channels)) 60 | c[:, :, :] = item[:, :, np.newaxis] 61 | y.append(c) 62 | 63 | y = np.array(y) 64 | 65 | return y 66 | -------------------------------------------------------------------------------- /layers/extracting.py: -------------------------------------------------------------------------------- 1 | """Extracting layer that computes key and value.""" 2 | 3 | import tensorflow as tf 4 | from tensorflow.keras.layers import Dense, Layer 5 | 6 | 7 | class Extracting(Layer): 8 | 9 | def __init__(self, 10 | units, 11 | use_bias, 12 | activation, 13 | kernel_initializer, 14 | kernel_regularizer, 15 | **kwargs): 16 | super().__init__(**kwargs) 17 | 18 | self.units = units 19 | self.use_bias = use_bias 20 | self.activation = activation 21 | self.kernel_initializer = kernel_initializer 22 | self.kernel_regularizer = kernel_regularizer 23 | 24 | self.dense1 = Dense(units=self.units, 25 | use_bias=self.use_bias, 26 | activation=self.activation, 27 | kernel_initializer=self.kernel_initializer, 28 | kernel_regularizer=self.kernel_regularizer) 29 | self.dense2 = Dense(units=self.units, 30 | use_bias=self.use_bias, 31 | activation=self.activation, 32 | kernel_initializer=self.kernel_initializer, 33 | kernel_regularizer=self.kernel_regularizer) 34 | 35 | def build(self, input_shape): 36 | super().build(input_shape) 37 | 38 | def call(self, inputs, mask=None): 39 | if mask is not None: 40 | mask = tf.cast(mask, dtype=self.dtype) 41 | mask = tf.expand_dims(mask, axis=-1) 42 | else: 43 | mask = 1.0 44 | 45 | k = mask * self.dense1(inputs) 46 | v = mask * self.dense2(inputs) 47 | 48 | return tf.concat([k, v], axis=-1) 49 | 50 | def compute_mask(self, inputs, mask=None): 51 | return mask 52 | -------------------------------------------------------------------------------- /utils/convnet14_cifar10_mnist_joint.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | """Model for classifying CIFAR10 and MNIST jointly.""" 3 | 4 | import numpy as np 5 | import tensorflow as tf 6 | 7 | from models.convnet14 import ConvNet14 8 | from utils.image_manipulation import expand_channels, pad 9 | 10 | EPOCHS = 100 11 | BATCH_SIZE = 128 12 | PAD_WIDTH = ((2, 2), (2, 2), (0, 0)) 13 | 14 | # Load the data. 15 | cifar10 = tf.keras.datasets.cifar10 16 | (x_train_cifar10, y_train_cifar10), (x_test_cifar10, y_test_cifar10) = cifar10.load_data() 17 | 18 | mnist = tf.keras.datasets.mnist 19 | (x_train_mnist, y_train_mnist), (x_test_mnist, y_test_mnist) = mnist.load_data() 20 | x_train_mnist = expand_channels(x_train_mnist, num_channels=3) 21 | x_test_mnist = expand_channels(x_test_mnist, num_channels=3) 22 | x_train_mnist = pad(x_train_mnist, PAD_WIDTH) 23 | x_test_mnist = pad(x_test_mnist, PAD_WIDTH) 24 | 25 | # Concatenate MNISt and CIFAR10 26 | x_train = np.concatenate((x_train_cifar10, x_train_mnist)) 27 | y_train = np.concatenate((y_train_cifar10.flatten(), y_train_mnist+10)) 28 | x_test = np.concatenate((x_test_cifar10, x_test_mnist)) 29 | y_test = np.concatenate((y_test_cifar10.flatten(), y_test_mnist+10)) 30 | 31 | idc = np.random.RandomState(seed=42).permutation(x_train.shape[0]) 32 | x_train = x_train[idc] 33 | y_train = y_train[idc] 34 | idc = np.random.RandomState(seed=42).permutation(x_test.shape[0]) 35 | x_test = x_test[idc] 36 | y_test = y_test[idc] 37 | 38 | input_shape = x_train.shape[1:] 39 | 40 | # Normalize pixel values to be between 0 and 1 41 | x_train, x_test = x_train / 255.0, x_test / 255.0 42 | 43 | 44 | def lr_scheduler(epoch): 45 | if epoch < 50: 46 | return 0.001 47 | else: 48 | return 0.001 * tf.math.exp(0.1 * (50 - epoch)) 49 | 50 | 51 | # Build the model. 52 | model = ConvNet14(output_size=20) 53 | 54 | # Compile the model. 55 | model.compile(optimizer=tf.keras.optimizers.Adam(), 56 | loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True), 57 | metrics=['accuracy']) 58 | 59 | # Train and evaluate the model 60 | callback = tf.keras.callbacks.LearningRateScheduler(lr_scheduler) 61 | model.fit(x=x_train, y=y_train, epochs=EPOCHS, batch_size=BATCH_SIZE, validation_split=0.1, 62 | callbacks=[callback]) 63 | 64 | model.evaluate(x=x_test, y=y_test, verbose=2) 65 | 66 | # Save the models weights. 67 | model.save_weights('saved_models/weights/ConvNet14-CIFAR10-MNIST/ConvNet14-CIFAR10-MNIST', save_format='tf') 68 | 69 | model.summary() 70 | -------------------------------------------------------------------------------- /layers/encoding.py: -------------------------------------------------------------------------------- 1 | """Sentence encoding.""" 2 | 3 | import tensorflow as tf 4 | from tensorflow.keras.constraints import Constraint 5 | from tensorflow.keras.layers import Layer 6 | 7 | from utils.word_encodings import position_encoding 8 | 9 | 10 | class Encoding(Layer): 11 | 12 | def __init__(self, 13 | encodings_type, 14 | encodings_constraint='mask_time_word', 15 | **kwargs): 16 | super().__init__(**kwargs) 17 | 18 | self.encodings_type = encodings_type.lower() 19 | self.encodings_constraint = encodings_constraint.lower() 20 | 21 | if self.encodings_type not in ('identity_encoding', 'position_encoding', 'learned_encoding'): 22 | raise ValueError('Could not interpret encodings type:', self.encodings_type) 23 | 24 | if self.encodings_constraint not in ('none', 'mask_time_word'): 25 | raise ValueError('Could not interpret encodings constraint:', self.encodings_type) 26 | 27 | self.constraint = self.MaskTimeWord() if self.encodings_constraint == 'mask_time_word' else None 28 | 29 | def build(self, input_shape): 30 | if self.encodings_type.lower() == 'identity_encoding': 31 | self.encoding = tf.ones((input_shape[-2], input_shape[-1])) 32 | if self.encodings_type.lower() == 'position_encoding': 33 | self.encoding = position_encoding(input_shape[-2], input_shape[-1]) 34 | if self.encodings_type.lower() == 'learned_encoding': 35 | self.encoding = self.add_weight(shape=(input_shape[-2], input_shape[-1]), trainable=True, 36 | initializer=tf.initializers.Ones(), 37 | constraint=self.constraint, 38 | dtype=self.dtype, name='encoding') 39 | 40 | super().build(input_shape) 41 | 42 | def call(self, inputs, mask=None): 43 | mask = tf.cast(mask, dtype=self.dtype) 44 | mask = tf.expand_dims(mask, axis=-1) 45 | 46 | return tf.reduce_sum(mask * self.encoding * inputs, axis=-2) 47 | 48 | def compute_mask(self, inputs, mask=None): 49 | if mask is None: 50 | return None 51 | 52 | return tf.reduce_any(mask, axis=-1) 53 | 54 | def compute_output_shape(self, input_shape): 55 | return (input_shape[0], input_shape[-1]) 56 | 57 | class MaskTimeWord(Constraint): 58 | """Make encoding of time words identity to avoid modifying them.""" 59 | 60 | def __init__(self, 61 | **kwargs): 62 | super().__init__(**kwargs) 63 | 64 | def __call__(self, w): 65 | indices = [[w.shape[0]-1]] 66 | updates = tf.ones((1, w.shape[1])) 67 | new_w = tf.tensor_scatter_nd_update(w, indices, updates) 68 | 69 | return new_w 70 | -------------------------------------------------------------------------------- /layers/reading.py: -------------------------------------------------------------------------------- 1 | """Reading layers that read from memory.""" 2 | 3 | import tensorflow as tf 4 | import tensorflow.keras.backend as K 5 | from tensorflow.keras.layers import Dense, Layer 6 | 7 | 8 | class Reading(Layer): 9 | 10 | def __init__(self, 11 | units, 12 | use_bias, 13 | activation, 14 | kernel_initializer, 15 | kernel_regularizer, 16 | **kwargs): 17 | super().__init__(**kwargs) 18 | 19 | self.units = units 20 | self.use_bias = use_bias 21 | self.activation = activation 22 | self.kernel_initializer = kernel_initializer 23 | self.kernel_regularizer = kernel_regularizer 24 | 25 | self.dense = Dense(units=self.units, 26 | use_bias=self.use_bias, 27 | activation=self.activation, 28 | kernel_initializer=self.kernel_initializer, 29 | kernel_regularizer=self.kernel_regularizer) 30 | 31 | def build(self, input_shape): 32 | super().build(input_shape) 33 | 34 | def call(self, inputs, constants): 35 | memory_matrix = constants[0] 36 | 37 | k = self.dense(inputs) 38 | 39 | v = K.batch_dot(k, memory_matrix) 40 | 41 | return v 42 | 43 | def compute_mask(self, inputs, mask=None): 44 | return mask 45 | 46 | 47 | class ReadingCell(Layer): 48 | 49 | def __init__(self, 50 | units, 51 | use_bias, 52 | activation, 53 | kernel_initializer, 54 | kernel_regularizer, 55 | **kwargs): 56 | super().__init__(**kwargs) 57 | 58 | self.units = units 59 | self.use_bias = use_bias 60 | self.activation = activation 61 | self.kernel_initializer = kernel_initializer 62 | self.kernel_regularizer = kernel_regularizer 63 | 64 | self.dense = Dense(units=self.units, 65 | use_bias=self.use_bias, 66 | activation=self.activation, 67 | kernel_initializer=self.kernel_initializer, 68 | kernel_regularizer=self.kernel_regularizer) 69 | 70 | @property 71 | def state_size(self): 72 | return self.units 73 | 74 | def build(self, input_shape): 75 | super().build(input_shape) 76 | 77 | def call(self, inputs, states, constants): 78 | v = states[0] 79 | memory_matrix = constants[0] 80 | 81 | k = self.dense(tf.concat([inputs, v], axis=1)) 82 | 83 | v = K.batch_dot(k, memory_matrix) 84 | 85 | return v, v 86 | 87 | def compute_mask(self, inputs, mask=None): 88 | return mask 89 | 90 | def get_initial_state(self, inputs=None, batch_size=None, dtype=None): 91 | return tf.zeros((batch_size, self.units)) 92 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/h-mem-harnessing-synaptic-plasticity-with/question-answering-on-babi)](https://paperswithcode.com/sota/question-answering-on-babi?p=h-mem-harnessing-synaptic-plasticity-with) 2 | 3 | # H-Mem: Harnessing synaptic plasticity with Hebbian Memory Networks 4 | This is the code used in the paper "[H-Mem: Harnessing synaptic plasticity with Hebbian Memory 5 | Networks](https://www.biorxiv.org/content/10.1101/2020.07.01.180372v2)" for training H-Mem on a single-shot 6 | image association task and on the bAbI question-answering tasks. 7 | 8 | ![H-Mem schema](https://i.imgur.com/fK3UWaP.png) 9 | 10 | ## Setup 11 | You need [TensorFlow](https://www.tensorflow.org/) to run this code. We tested it on TensorFlow version 2.1. 12 | Additional dependencies are listed in [environment.yml](environment.yml). If you use 13 | [Conda](https://docs.conda.io/en/latest/), run 14 | 15 | ```bash 16 | conda env create --file=environment.yml 17 | ``` 18 | 19 | to install the required packages and their dependencies. 20 | 21 | ## Usage 22 | 23 | ### Single-shot associations with H-Mem 24 | To start training on the single-shot image association task, run 25 | 26 | ```bash 27 | python image_association_task.py 28 | ``` 29 | 30 | Set the command line argument `--delay` to set the between-image delay (in the paper we used delays ranging from 0 to 40). Run the following command 31 | 32 | ```bash 33 | python image_association_task_lstm.py 34 | ``` 35 | 36 | to start training the LSTM model on this task (the default value for the between-image delay is 0; you can change it with the command line argument `--delay`). 37 | 38 | ### Question answering with H-Mem 39 | Run the following command 40 | 41 | ```bash 42 | python babi_task_single.py 43 | ``` 44 | 45 | to start training on bAbI task 1 in the 10k training examples setting. Set the command line argument `--task_id` to train on other tasks. You can try different model configurations by changing various command line arguments. For example, 46 | 47 | ```bash 48 | python babi_task_single.py --task_id=4 --memory_size=20 --epochs=50 --logging=1 49 | ``` 50 | 51 | will train the model with an associative memory of size 20 on task 4 for 50 epochs. The results will be stored in `results/`. 52 | 53 | ### Memory-dependent memorization 54 | In our extended model we have added an 'read-before-write' step. This model will be used if the 55 | command line argument `--read_before_write` is set to `1`. Run the following command 56 | 57 | ```bash 58 | python babi_task_single.py --task_id=16 --epochs=250 --read_before_write=1 59 | ``` 60 | 61 | to start training on bAbI task 16 in the 10k training examples setting (note that we trained the extended 62 | model for 250 epochs---instead of 100 epochs). You should get an accuracy of about 100% on this task. Compare 63 | to the original model, which does not solve task 16, by running the following command 64 | 65 | ```bash 66 | python babi_task_single.py --task_id=16 --epochs=250 67 | ``` 68 | 69 | ## References 70 | * Limbacher, T., & Legenstein, R. (2020). H-Mem: Harnessing synaptic plasticity with Hebbian Memory Networks. Advances in Neural Information Processing Systems, 33. 71 | https://www.biorxiv.org/content/10.1101/2020.07.01.180372v2 72 | -------------------------------------------------------------------------------- /models/convnet14.py: -------------------------------------------------------------------------------- 1 | """Convolutional network with 14 weight layers.""" 2 | 3 | import tensorflow as tf 4 | from tensorflow.keras.layers import BatchNormalization, Conv2D, Dense, Dropout, Flatten, MaxPool2D 5 | from tensorflow.keras.models import Sequential 6 | 7 | BatchNormalization._USE_V2_BEHAVIOR = False 8 | 9 | 10 | class ConvNet14(Sequential): 11 | 12 | def __init__(self, 13 | output_size=10, 14 | include_top=True, 15 | **kwargs): 16 | super().__init__(**kwargs) 17 | 18 | self.output_size = output_size 19 | self.include_top = include_top 20 | 21 | weight_decay = 1e-3 22 | self.kernel_regularizer = tf.keras.regularizers.l2(weight_decay) 23 | 24 | self.add(Conv2D(32, (3, 3), activation='elu', kernel_initializer='he_uniform', 25 | kernel_regularizer=self.kernel_regularizer, padding='same')) 26 | self.add(BatchNormalization()) 27 | self.add(Conv2D(32, (3, 3), activation='elu', kernel_initializer='he_uniform', 28 | kernel_regularizer=self.kernel_regularizer, padding='same')) 29 | self.add(BatchNormalization()) 30 | self.add(MaxPool2D((2, 2))) 31 | self.add(Dropout(0.1)) 32 | 33 | self.add(Conv2D(64, (3, 3), activation='elu', kernel_initializer='he_uniform', 34 | kernel_regularizer=self.kernel_regularizer, padding='same')) 35 | self.add(BatchNormalization()) 36 | self.add(Conv2D(64, (3, 3), activation='elu', kernel_initializer='he_uniform', 37 | kernel_regularizer=self.kernel_regularizer, padding='same')) 38 | self.add(BatchNormalization()) 39 | self.add(MaxPool2D((2, 2))) 40 | self.add(Dropout(0.1)) 41 | 42 | self.add(Conv2D(128, (3, 3), activation='elu', kernel_initializer='he_uniform', 43 | kernel_regularizer=self.kernel_regularizer, padding='same')) 44 | self.add(BatchNormalization()) 45 | self.add(Conv2D(128, (3, 3), activation='elu', kernel_initializer='he_uniform', 46 | kernel_regularizer=self.kernel_regularizer, padding='same')) 47 | self.add(BatchNormalization()) 48 | self.add(MaxPool2D((2, 2))) 49 | self.add(Dropout(0.2)) 50 | 51 | self.add(Conv2D(256, (3, 3), activation='elu', kernel_initializer='he_uniform', 52 | kernel_regularizer=self.kernel_regularizer, padding='same')) 53 | self.add(BatchNormalization()) 54 | self.add(Conv2D(256, (3, 3), activation='elu', kernel_initializer='he_uniform', 55 | kernel_regularizer=self.kernel_regularizer, padding='same')) 56 | self.add(BatchNormalization()) 57 | self.add(MaxPool2D((2, 2))) 58 | self.add(Dropout(0.3)) 59 | 60 | self.add(Conv2D(512, (3, 3), activation='elu', kernel_initializer='he_uniform', 61 | kernel_regularizer=self.kernel_regularizer, padding='same')) 62 | self.add(BatchNormalization()) 63 | self.add(Conv2D(512, (3, 3), activation='elu', kernel_initializer='he_uniform', 64 | kernel_regularizer=self.kernel_regularizer, padding='same')) 65 | self.add(BatchNormalization()) 66 | self.add(MaxPool2D((2, 2))) 67 | self.add(Dropout(0.3)) 68 | if self.include_top: 69 | self.add(Flatten()) 70 | self.add(Dense(512, activation='elu', kernel_initializer='he_uniform')) 71 | self.add(BatchNormalization()) 72 | self.add(Dense(256, activation='elu', kernel_initializer='he_uniform')) 73 | self.add(BatchNormalization()) 74 | self.add(Dense(128, activation='elu', kernel_initializer='he_uniform')) 75 | self.add(BatchNormalization()) 76 | self.add(Dropout(0.4)) 77 | self.add(Dense(self.output_size)) 78 | -------------------------------------------------------------------------------- /layers/writing.py: -------------------------------------------------------------------------------- 1 | """Writing layers that write to memory.""" 2 | 3 | import tensorflow as tf 4 | from tensorflow.keras.layers import Layer 5 | import tensorflow.keras.backend as K 6 | 7 | 8 | class Writing(Layer): 9 | 10 | def __init__(self, 11 | units, 12 | gamma, 13 | learn_gamma=False, 14 | **kwargs): 15 | super().__init__(**kwargs) 16 | 17 | self.units = units 18 | self._gamma = gamma 19 | self.learn_gamma = learn_gamma 20 | 21 | def build(self, input_shape): 22 | self.gamma = self.add_weight(shape=(1,), trainable=self.learn_gamma, 23 | initializer=tf.keras.initializers.Constant(self._gamma), 24 | dtype=self.dtype, name='gamma') 25 | 26 | super().build(input_shape) 27 | 28 | def call(self, inputs, mask=None): 29 | k, v = tf.split(inputs, 2, axis=-1) 30 | 31 | k = tf.expand_dims(k, 2) 32 | v = tf.expand_dims(v, 1) 33 | 34 | hebb = self.gamma * k * v 35 | 36 | memory_matrix = hebb 37 | 38 | return memory_matrix 39 | 40 | def compute_mask(self, inputs, mask=None): 41 | return mask 42 | 43 | 44 | class WritingCell(Layer): 45 | 46 | def __init__(self, 47 | units, 48 | gamma_pos, 49 | gamma_neg, 50 | w_assoc_max, 51 | use_bias=False, 52 | read_before_write=False, 53 | kernel_initializer=None, 54 | kernel_regularizer=None, 55 | learn_gamma_pos=False, 56 | learn_gamma_neg=False, 57 | **kwargs): 58 | super().__init__(**kwargs) 59 | 60 | self.units = units 61 | self.w_max = w_assoc_max 62 | self._gamma_pos = gamma_pos 63 | self._gamma_neg = gamma_neg 64 | self.use_bias = use_bias 65 | self.read_before_write = read_before_write 66 | self.kernel_initializer = kernel_initializer 67 | self.kernel_regularizer = kernel_regularizer 68 | self.learn_gamma_pos = learn_gamma_pos 69 | self.learn_gamma_neg = learn_gamma_neg 70 | 71 | if self.read_before_write: 72 | self.dense = tf.keras.layers.Dense(units=self.units, 73 | use_bias=self.use_bias, 74 | kernel_initializer=self.kernel_initializer, 75 | kernel_regularizer=self.kernel_regularizer) 76 | 77 | self.ln1 = tf.keras.layers.LayerNormalization() 78 | self.ln2 = tf.keras.layers.LayerNormalization() 79 | 80 | @property 81 | def state_size(self): 82 | return tf.TensorShape((self.units, self.units)) 83 | 84 | def build(self, input_shape): 85 | self.gamma_pos = self.add_weight(shape=(1,), trainable=self.learn_gamma_pos, 86 | initializer=tf.keras.initializers.Constant(self._gamma_pos), 87 | dtype=self.dtype, name='gamma_pos') 88 | self.gamma_neg = self.add_weight(shape=(1,), trainable=self.learn_gamma_neg, 89 | initializer=tf.keras.initializers.Constant(self._gamma_neg), 90 | dtype=self.dtype, name='gamma_neg') 91 | 92 | super().build(input_shape) 93 | 94 | def call(self, inputs, states, mask=None): 95 | memory_matrix = states[0] 96 | k, v = tf.split(inputs, 2, axis=-1) 97 | 98 | if self.read_before_write: 99 | k = self.ln1(k) 100 | v_h = K.batch_dot(k, memory_matrix) 101 | 102 | v = self.dense(tf.concat([v, v_h], axis=1)) 103 | v = self.ln2(v) 104 | 105 | k = tf.expand_dims(k, 2) 106 | v = tf.expand_dims(v, 1) 107 | 108 | hebb = self.gamma_pos * (self.w_max - memory_matrix) * k * v - self.gamma_neg * memory_matrix * k**2 109 | 110 | memory_matrix = hebb + memory_matrix 111 | 112 | return memory_matrix, memory_matrix 113 | 114 | def compute_mask(self, inputs, mask=None): 115 | return mask 116 | 117 | def get_initial_state(self, inputs=None, batch_size=None, dtype=None): 118 | return tf.zeros((batch_size, self.units, self.units)) 119 | -------------------------------------------------------------------------------- /image_association_task_lstm.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | """Runs an LSTM on a single-shot image association task.""" 3 | 4 | import argparse 5 | 6 | import numpy as np 7 | import tensorflow as tf 8 | from tensorflow.keras.layers import TimeDistributed 9 | 10 | from data.image_association_data import load_data 11 | from models.convnet14 import ConvNet14 as ConvNet 12 | 13 | strategy = tf.distribute.MirroredStrategy() 14 | 15 | parser = argparse.ArgumentParser() 16 | parser.add_argument('--delay', type=int, default=0) 17 | parser.add_argument('--timesteps', type=int, default=3) 18 | parser.add_argument('--delay_padding', type=str, default='random', help='`zeros` or `random`') 19 | 20 | parser.add_argument('--epochs', type=int, default=100) 21 | parser.add_argument('--batch_size_per_replica', type=int, default=32) 22 | parser.add_argument('--max_grad_norm', type=float, default=10.0) 23 | parser.add_argument('--learning_rate', type=float, default=0.001) 24 | parser.add_argument('--validation_split', type=float, default=0.1) 25 | 26 | parser.add_argument('--retrain_convnet', type=int, default=0) 27 | parser.add_argument('--use_pretrained_convnet', type=int, default=0) 28 | 29 | parser.add_argument('--hidden_size', type=int, default=200) 30 | parser.add_argument('--dense_size', type=int, default=128) 31 | parser.add_argument('--gamma_pos', type=float, default=0.01) 32 | parser.add_argument('--gamma_neg', type=float, default=0.01) 33 | parser.add_argument('--w_assoc_max', type=float, default=1.0) 34 | 35 | parser.add_argument('--verbose', type=int, default=1) 36 | args = parser.parse_args() 37 | 38 | batch_size = args.batch_size_per_replica * strategy.num_replicas_in_sync 39 | 40 | # Load the data. 41 | (x_train, y_train), (x_test, y_test) = load_data(timesteps=args.timesteps, merge=True, 42 | data_dir='data/image_association_task/') 43 | 44 | num_train = y_train.size - int(args.validation_split * y_train.size) 45 | num_val = int(args.validation_split * y_train.size) 46 | num_test = y_test.size 47 | 48 | x_val = [x_train[0][-num_val:], x_train[1][-num_val:]] 49 | y_val = y_train[-num_val:] 50 | 51 | x_train = [x_train[0][:num_train], x_train[1][:num_train]] 52 | y_train = y_train[:num_train] 53 | 54 | timesteps_with_delay = args.timesteps * (args.delay + 1) 55 | 56 | input_shape = (timesteps_with_delay+1, ) + x_train[0].shape[2:] 57 | 58 | 59 | # Create the datasets. 60 | def dataset_generator(x, y, seed): 61 | rng = np.random.RandomState(seed=seed) 62 | for a, b, y in zip(x[0], x[1], y): 63 | size = (timesteps_with_delay, ) + a.shape[1:-1] + (1,) 64 | if args.delay_padding == 'random': 65 | aa = rng.uniform(size=size).repeat(a.shape[-1], axis=3) 66 | elif args.delay_padding == 'zeros': 67 | aa = np.zeros(shape=size).repeat(a.shape[-1], axis=3) 68 | aa[::args.delay+1, :] = a 69 | inputs = np.concatenate([aa, b[np.newaxis, :, :, :]]) 70 | 71 | yield {'inputs': inputs}, y 72 | 73 | 74 | output_types = ({'inputs': 'float32'}, 'uint8') 75 | output_shapes = ({'inputs': [None, None, None, None]}, []) 76 | train_dataset = tf.data.Dataset.from_generator(generator=lambda: dataset_generator(x_train, y_train, 42), 77 | output_types=output_types, 78 | output_shapes=output_shapes) 79 | val_dataset = tf.data.Dataset.from_generator(generator=lambda: dataset_generator(x_val, y_val, 43), 80 | output_types=output_types, 81 | output_shapes=output_shapes) 82 | test_dataset = tf.data.Dataset.from_generator(generator=lambda: dataset_generator(x_test, y_test, 44), 83 | output_types=output_types, 84 | output_shapes=output_shapes) 85 | 86 | train_dataset = train_dataset.cache().repeat(args.epochs * batch_size).shuffle(10000).batch(batch_size) 87 | val_dataset = val_dataset.batch(batch_size) 88 | test_dataset = test_dataset.batch(batch_size) 89 | 90 | # Load pretrained model. 91 | conv_net = ConvNet(include_top=False) 92 | if args.use_pretrained_convnet: 93 | conv_net.load_weights('saved_models/weights/ConvNet14-CIFAR10-MNIST/ConvNet14-CIFAR10-MNIST') 94 | conv_net.trainable = bool(args.retrain_convnet) 95 | 96 | with strategy.scope(): 97 | # Build the model. 98 | inputs = tf.keras.layers.Input(input_shape, name='inputs') 99 | 100 | features = TimeDistributed(conv_net, name='conv')(inputs) 101 | features = TimeDistributed(tf.keras.layers.Flatten(), name='flatten')(features) 102 | features = TimeDistributed(tf.keras.layers.Dense(args.dense_size, 103 | use_bias=False, 104 | activation='relu', 105 | kernel_initializer='he_uniform', 106 | kernel_regularizer=tf.keras.regularizers.l2(1e-3)), 107 | name='dense')(features) 108 | features = TimeDistributed(tf.keras.layers.BatchNormalization(), name='batch_norm')(features) 109 | features = TimeDistributed(tf.keras.layers.Dropout(0.3), name='dropout')(features) 110 | 111 | states = tf.keras.layers.LSTM(args.hidden_size, name='states')(features) 112 | 113 | outputs = tf.keras.layers.Dense(10, kernel_initializer='he_uniform', use_bias=False)(states) 114 | 115 | model = tf.keras.Model(inputs=inputs, outputs=outputs) 116 | 117 | # Compile the model. 118 | optimizer_kwargs = {'clipnorm': args.max_grad_norm} if args.max_grad_norm else {} 119 | model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=args.learning_rate, **optimizer_kwargs), 120 | loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True), 121 | metrics=['accuracy']) 122 | 123 | model.summary() 124 | 125 | 126 | # Train and evaluate. 127 | def lr_scheduler(epoch): 128 | if epoch < 50: 129 | return 0.001 130 | else: 131 | return 0.001 * tf.math.exp(0.1 * (50 - epoch)) 132 | 133 | 134 | callback = tf.keras.callbacks.LearningRateScheduler(lr_scheduler) 135 | model.fit(train_dataset, 136 | epochs=args.epochs, 137 | steps_per_epoch=np.ceil(num_train/batch_size), 138 | validation_data=val_dataset if num_val > 0 else None, 139 | validation_steps=np.ceil(num_val/batch_size), 140 | callbacks=[callback], 141 | verbose=args.verbose) 142 | 143 | model.evaluate(test_dataset, steps=np.ceil(num_test/batch_size), verbose=2) 144 | -------------------------------------------------------------------------------- /data/babi_data.py: -------------------------------------------------------------------------------- 1 | """Utilities for downloading and parsing bAbI task data. 2 | 3 | Modified from https://github.com/domluna/memn2n/blob/master/data_utils.py. 4 | 5 | """ 6 | 7 | import os 8 | import re 9 | import shutil 10 | import urllib.request 11 | 12 | import numpy as np 13 | 14 | tasks = { 15 | 1: 'single_supporting_fact', 16 | 2: 'two_supporting_facts', 17 | 3: 'three_supporting_facts', 18 | 4: 'two_arg_relations', 19 | 5: 'three_arg_relations', 20 | 6: 'yes_no_questions', 21 | 7: 'counting', 22 | 8: 'lists_sets', 23 | 9: 'simple_negation', 24 | 10: 'indefinite_knowledge', 25 | 11: 'basic_coreference', 26 | 12: 'conjunction', 27 | 13: 'compound_coreference', 28 | 14: 'time_reasoning', 29 | 15: 'basic_deduction', 30 | 16: 'basic_induction', 31 | 17: 'positional_reasoning', 32 | 18: 'size_reasoning', 33 | 19: 'path_finding', 34 | 20: 'agents_motivations' 35 | } 36 | 37 | 38 | def download(extract=True): 39 | """Downloads the data set. 40 | 41 | Arguments: 42 | extract: boolean, whether to extract the downloaded archive (default=`True`). 43 | 44 | Returns: 45 | data_dir: string, the data directory. 46 | 47 | """ 48 | url = 'https://s3.amazonaws.com/text-datasets/' 49 | file_name = 'babi_tasks_1-20_v1-2.tar.gz' 50 | data_dir = 'data/' 51 | file_path = data_dir + file_name 52 | 53 | if not os.path.exists(file_path): 54 | print('Downloading ' + url + file_name + '...') 55 | print('-') 56 | with urllib.request.urlopen(url + file_name) as response, open(file_path, 'wb') as out_file: 57 | shutil.copyfileobj(response, out_file) 58 | shutil.unpack_archive(file_path, data_dir) 59 | shutil.move(data_dir + 'tasks_1-20_v1-2', data_dir + 'babi_tasks_1-20_v1-2') 60 | 61 | return data_dir + 'babi_tasks_1-20_v1-2' 62 | 63 | 64 | def load_task(data_dir, task_id, training_set_size='1k', only_supporting=False): 65 | """Loads the nth task. There are 20 tasks in total. 66 | 67 | Arguments: 68 | data_dir: string, the data directory. 69 | task_id: int, the ID of the task (valid values are in `range(1, 21)`). 70 | training_set_size: string, the size of the training set to load (`1k` or `10k`, default=`1k`). 71 | only_supporting: boolean, if `True` only supporting facts are loaded (default=`False`). 72 | 73 | Returns: 74 | A Python tuple containing the training and testing data for the task. 75 | 76 | """ 77 | assert task_id > 0 and task_id < 21 78 | 79 | data_dir = data_dir + '/en/' if training_set_size == '1k' else data_dir + '/en-10k/' 80 | files = os.listdir(data_dir) 81 | files = [os.path.join(data_dir, f) for f in files] 82 | s = 'qa{}_'.format(task_id) 83 | train_file = [f for f in files if s in f and 'train' in f][0] 84 | test_file = [f for f in files if s in f and 'test' in f][0] 85 | train_data = _get_stories(train_file, only_supporting) 86 | test_data = _get_stories(test_file, only_supporting) 87 | 88 | return train_data, test_data 89 | 90 | 91 | def vectorize_data(data, word_idx, max_num_sentences, sentence_size, query_size): 92 | """Vectorize stories, queries and answers. 93 | 94 | If a sentence length < `sentence_size`, the sentence will be padded with `0`s. If a story length < 95 | `max_num_sentences`, the story will be padded with empty sentences. Empty sentences are 1-D arrays of 96 | length `sentence_size` filled with `0`s. The answer array is returned as a one-hot encoding. 97 | 98 | Arguments: 99 | data: iterable, containing stories, queries and answers. 100 | word_idx: dict, mapping words to unique integers. 101 | max_num_sentences: int, the maximum number of sentences to extract. 102 | sentence_size: int, the maximum number of words in a sentence. 103 | query_size: int, the maximum number of words in a query. 104 | 105 | Returns: 106 | A Python tuple containing vectorized stories, queries, and answers. 107 | 108 | """ 109 | S = [] 110 | Q = [] 111 | A = [] 112 | for story, query, answer in data: 113 | if len(story) > max_num_sentences: 114 | continue 115 | 116 | ss = [] 117 | for i, sentence in enumerate(story, 1): 118 | # Pad to sentence_size, i.e., add nil words, and add story. 119 | ls = max(0, sentence_size - len(sentence)) 120 | ss.append([word_idx[w] for w in sentence] + [0] * ls) 121 | 122 | # Make the last word of each sentence the time 'word' which corresponds to vector of lookup table. 123 | for i in range(len(ss)): 124 | ss[i][-1] = len(word_idx) - max_num_sentences - i + len(ss) 125 | 126 | # Pad stories to max_num_sentences (i.e., add empty stories). 127 | ls = max(0, max_num_sentences - len(ss)) 128 | for _ in range(ls): 129 | ss.append([0] * sentence_size) 130 | 131 | # Pad queries to query_size (i.e., add nil words). 132 | lq = max(0, query_size - len(query)) 133 | q = [word_idx[w] for w in query] + [0] * lq 134 | 135 | y = np.zeros(len(word_idx) + 1) # 0 is reserved for nil word. 136 | for a in answer: 137 | y[word_idx[a]] = 1 138 | 139 | S.append(ss) 140 | Q.append(q) 141 | A.append(y) 142 | 143 | return np.array(S), np.array(Q), np.array(A) 144 | 145 | 146 | def _get_stories(f, only_supporting=False): 147 | """Given a file name, read the file, retrieve the stories, and then convert the sentences into a single 148 | story. 149 | 150 | If only_supporting is true, only the sentences that support the answer are kept. 151 | 152 | Arguments: 153 | f: string, the file name. 154 | only_supporting: boolean, if `True` only supporting facts are loaded (default=`False`). 155 | 156 | Returns: 157 | A list of Python tuples containing stories, queries, and answers. 158 | 159 | """ 160 | with open(f) as f: 161 | data = _parse_stories(f.readlines(), only_supporting=only_supporting) 162 | 163 | return data 164 | 165 | 166 | def _parse_stories(lines, only_supporting=False): 167 | """Parse stories provided in the bAbI tasks format. 168 | 169 | If only_supporting is true, only the sentences that support the answer are kept. 170 | 171 | Arguments: 172 | lines: iterable, containing the sentences of a full story (story, query, and answer). 173 | only_supporting: boolean, if `True` only supporting facts are loaded (default=`False`). 174 | 175 | Returns: 176 | A Python list containing the parsed stories. 177 | 178 | """ 179 | data = [] 180 | story = [] 181 | for line in lines: 182 | line = str.lower(line) 183 | nid, line = line.split(' ', 1) 184 | nid = int(nid) 185 | if nid == 1: 186 | story = [] 187 | if '\t' in line: # Question 188 | q, a, supporting = line.split('\t') 189 | q = _tokenize(q) 190 | a = [a] # Answer is one vocab word even ie it's actually multiple words. 191 | substory = None 192 | 193 | # Remove question marks 194 | if q[-1] == '?': 195 | q = q[:-1] 196 | 197 | if only_supporting: 198 | # Only select the related substory. 199 | supporting = map(int, supporting.split()) 200 | substory = [story[i - 1] for i in supporting] 201 | else: 202 | # Provide all the substories. 203 | substory = [x for x in story if x] 204 | 205 | data.append((substory, q, a)) 206 | story.append('') 207 | else: # Regular sentence 208 | sent = _tokenize(line) 209 | # Remove periods 210 | if sent[-1] == '.': 211 | sent = sent[:-1] 212 | story.append(sent) 213 | 214 | return data 215 | 216 | 217 | def _tokenize(sent): 218 | """Return the tokens of a sentence including punctuation. 219 | 220 | Arguments: 221 | sent: iterable, containing the sentence. 222 | 223 | Returns: 224 | A Python list containing the tokens in the sentence. 225 | 226 | Examples: 227 | 228 | ```python 229 | tokenize('Bob dropped the apple. Where is the apple?') 230 | 231 | ['Bob', 'dropped', 'the', 'apple', '.', 'Where', 'is', 'the', 'apple', '?'] 232 | ``` 233 | 234 | """ 235 | return [x.strip() for x in re.split(r'(\W+)+?', sent) if x.strip()] 236 | -------------------------------------------------------------------------------- /image_association_task.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | """Runs H-Mem on a single-shot image association task.""" 3 | 4 | import argparse 5 | import random 6 | 7 | import numpy as np 8 | import tensorflow as tf 9 | from tensorflow.keras.layers import TimeDistributed 10 | from tensorflow.keras import Model 11 | 12 | from data.image_association_data import load_data 13 | from layers.extracting import Extracting 14 | from layers.reading import Reading 15 | from layers.writing import WritingCell 16 | from models.convnet14 import ConvNet14 as ConvNet 17 | 18 | strategy = tf.distribute.MirroredStrategy() 19 | 20 | parser = argparse.ArgumentParser() 21 | parser.add_argument('--delay', type=int, default=0) 22 | parser.add_argument('--timesteps', type=int, default=3) 23 | parser.add_argument('--delay_padding', type=str, default='random', help='`zeros` or `random`') 24 | 25 | parser.add_argument('--epochs', type=int, default=100) 26 | parser.add_argument('--learning_rate', type=float, default=0.001) 27 | parser.add_argument('--batch_size_per_replica', type=int, default=32) 28 | parser.add_argument('--random_state', type=int, default=None) 29 | parser.add_argument('--max_grad_norm', type=float, default=10.0) 30 | parser.add_argument('--validation_split', type=float, default=0.1) 31 | 32 | parser.add_argument('--retrain_convnet', type=int, default=0) 33 | parser.add_argument('--use_pretrained_convnet', type=int, default=0) 34 | 35 | parser.add_argument('--memory_size', type=int, default=200) 36 | parser.add_argument('--dense_size', type=int, default=128) 37 | parser.add_argument('--gamma_pos', type=float, default=0.01) 38 | parser.add_argument('--gamma_neg', type=float, default=0.01) 39 | parser.add_argument('--w_assoc_max', type=float, default=1.0) 40 | 41 | parser.add_argument('--verbose', type=int, default=1) 42 | args = parser.parse_args() 43 | 44 | batch_size = args.batch_size_per_replica * strategy.num_replicas_in_sync 45 | 46 | # Set random seeds. 47 | np.random.seed(args.random_state) 48 | random.seed(args.random_state) 49 | tf.random.set_seed(args.random_state) 50 | 51 | # Load the data. 52 | (x_train, y_train), (x_test, y_test) = load_data(timesteps=args.timesteps, merge=True, 53 | data_dir='data/image_association_task/') 54 | 55 | num_train = y_train.size - int(args.validation_split * y_train.size) 56 | num_val = int(args.validation_split * y_train.size) 57 | num_test = y_test.size 58 | 59 | x_val = [x_train[0][-num_val:], x_train[1][-num_val:]] 60 | y_val = y_train[-num_val:] 61 | 62 | x_train = [x_train[0][:num_train], x_train[1][:num_train]] 63 | y_train = y_train[:num_train] 64 | 65 | timesteps_with_delay = args.timesteps * (args.delay + 1) 66 | 67 | input_a_shape = (timesteps_with_delay, ) + x_train[0].shape[2:] 68 | input_b_shape = x_train[1].shape[1:] 69 | 70 | 71 | # Create the datasets. 72 | def dataset_generator(x, y, seed): 73 | rng = np.random.RandomState(seed=seed) 74 | for a, b, y in zip(x[0], x[1], y): 75 | size = (timesteps_with_delay, ) + a.shape[1:-1] + (1,) 76 | if args.delay_padding == 'random': 77 | aa = rng.uniform(size=size).repeat(a.shape[-1], axis=3) 78 | elif args.delay_padding == 'zeros': 79 | aa = np.zeros(shape=size).repeat(a.shape[-1], axis=3) 80 | aa[::args.delay+1, :] = a 81 | 82 | yield {'input_a': aa, 'input_b': b}, y 83 | 84 | 85 | output_types = ({'input_a': 'float32', 'input_b': 'float32'}, 'uint8') 86 | output_shapes = ({'input_a': [None, None, None, None], 'input_b': [None, None, None]}, []) 87 | train_dataset = tf.data.Dataset.from_generator(generator=lambda: dataset_generator(x_train, y_train, 42), 88 | output_types=output_types, 89 | output_shapes=output_shapes) 90 | val_dataset = tf.data.Dataset.from_generator(generator=lambda: dataset_generator(x_val, y_val, 43), 91 | output_types=output_types, 92 | output_shapes=output_shapes) 93 | test_dataset = tf.data.Dataset.from_generator(generator=lambda: dataset_generator(x_test, y_test, 44), 94 | output_types=output_types, 95 | output_shapes=output_shapes) 96 | 97 | train_dataset = train_dataset.cache().repeat(args.epochs * batch_size).shuffle(10000).batch(batch_size) 98 | val_dataset = val_dataset.batch(batch_size) 99 | test_dataset = test_dataset.batch(batch_size) 100 | 101 | # Load pretrained model. 102 | conv_net = ConvNet(include_top=False, name='conv_b') 103 | if args.use_pretrained_convnet: 104 | conv_net.load_weights('saved_models/weights/ConvNet14-CIFAR10-MNIST/ConvNet14-CIFAR10-MNIST') 105 | conv_net.trainable = bool(args.retrain_convnet) 106 | 107 | with strategy.scope(): 108 | # Build the model. 109 | input_a = tf.keras.layers.Input(input_a_shape, name='input_a') 110 | input_b = tf.keras.layers.Input(input_b_shape, name='input_b') 111 | 112 | features_a = TimeDistributed(conv_net, name='conv_a')(input_a) 113 | features_a = TimeDistributed(tf.keras.layers.Flatten(), name='flatten_a')(features_a) 114 | features_a = TimeDistributed(tf.keras.layers.Dense(args.dense_size, 115 | use_bias=False, 116 | activation='relu', 117 | kernel_initializer='he_uniform', 118 | kernel_regularizer=tf.keras.regularizers.l2(1e-3)), 119 | name='dense_a')(features_a) 120 | features_a = TimeDistributed(tf.keras.layers.BatchNormalization(), name='batch_norm_a')(features_a) 121 | features_a = TimeDistributed(tf.keras.layers.Dropout(0.3), name='dropout_a')(features_a) 122 | 123 | features_b = conv_net(input_b) 124 | features_b = tf.keras.layers.Flatten(name='flatten_b')(features_b) 125 | features_b = tf.keras.layers.Dense(args.dense_size, 126 | use_bias=False, 127 | activation='relu', 128 | kernel_initializer='he_uniform', 129 | kernel_regularizer=tf.keras.regularizers.l2(1e-3), 130 | name='dense_b')(features_b) 131 | features_b = tf.keras.layers.BatchNormalization(name='batch_norm_b')(features_b) 132 | features_b = tf.keras.layers.Dropout(0.3, name='dropout_b')(features_b) 133 | 134 | entities = Extracting(units=args.memory_size, 135 | use_bias=False, 136 | activation='relu', 137 | kernel_initializer='he_uniform', 138 | kernel_regularizer=tf.keras.regularizers.l2(1e-3), 139 | name='entity_extracting')(features_a) 140 | 141 | memory_matrix = tf.keras.layers.RNN(WritingCell(units=args.memory_size, 142 | gamma_pos=args.gamma_pos, 143 | gamma_neg=args.gamma_neg, 144 | w_assoc_max=args.w_assoc_max, 145 | learn_gamma_pos=False, 146 | learn_gamma_neg=False), 147 | name='entity_writing')(entities) 148 | 149 | queried_value = Reading(units=args.memory_size, 150 | use_bias=False, 151 | activation='relu', 152 | kernel_initializer='he_uniform', 153 | kernel_regularizer=tf.keras.regularizers.l2(1e-3), 154 | name='entity_reading')(features_b, constants=[memory_matrix]) 155 | 156 | outputs = tf.keras.layers.Dense(10, 157 | use_bias=False, 158 | kernel_initializer='he_uniform', 159 | name='output')(queried_value) 160 | 161 | model = Model(inputs=[input_a, input_b], outputs=outputs) 162 | 163 | # Compile the model. 164 | optimizer_kwargs = {'clipnorm': args.max_grad_norm} if args.max_grad_norm else {} 165 | model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=args.learning_rate, **optimizer_kwargs), 166 | loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True), 167 | metrics=['accuracy']) 168 | 169 | model.summary() 170 | 171 | 172 | # Train and evaluate. 173 | def lr_scheduler(epoch): 174 | if epoch < 50: 175 | return 0.001 176 | else: 177 | return 0.001 * tf.math.exp(0.1 * (50 - epoch)) 178 | 179 | 180 | callback = tf.keras.callbacks.LearningRateScheduler(lr_scheduler) 181 | model.fit(train_dataset, 182 | epochs=args.epochs, 183 | steps_per_epoch=np.ceil(num_train/batch_size), 184 | validation_data=val_dataset if num_val > 0 else None, 185 | validation_steps=np.ceil(num_val/batch_size), 186 | callbacks=[callback], 187 | verbose=args.verbose) 188 | 189 | model.evaluate(test_dataset, steps=np.ceil(num_test/batch_size), verbose=2) 190 | -------------------------------------------------------------------------------- /data/image_association_data.py: -------------------------------------------------------------------------------- 1 | """Creates the data for the image association tasks.""" 2 | 3 | import os 4 | import pathlib 5 | 6 | import numpy as np 7 | import tensorflow as tf 8 | 9 | from utils import image_manipulation 10 | 11 | 12 | def load_data(timesteps, pad_equal=False, merge=False, data_dir='data', seed=42): 13 | pad_equal = True if merge else pad_equal 14 | 15 | if not os.path.exists(data_dir): 16 | os.makedirs(data_dir) 17 | 18 | suffix = '_{0}{1}.npy'.format(timesteps, '_merged' if merge else '') 19 | 20 | x_train_files = [] 21 | x_test_files = [] 22 | if merge: 23 | for i in ['_a', '_b']: 24 | x_train_files.append(pathlib.Path(data_dir, 'x_train' + i + suffix)) 25 | x_test_files.append(pathlib.Path(data_dir, 'x_test' + i + suffix)) 26 | else: 27 | for i in ['_a', '_b', '_c']: 28 | x_train_files.append(pathlib.Path(data_dir, 'x_train' + i + suffix)) 29 | x_test_files.append(pathlib.Path(data_dir, 'x_test' + i + suffix)) 30 | 31 | y_train_file = pathlib.Path(data_dir, 'y_train' + suffix) 32 | y_test_file = pathlib.Path(data_dir, 'y_test' + suffix) 33 | 34 | cifar10_train, cifar10_test = _get_cifar10_dataset() 35 | mnist_train, mnist_test = _get_mnist_dataset(num_channels=1) 36 | 37 | if not all([f.is_file() for f in x_train_files]) or not y_train_file.is_file(): 38 | x_train, y_train = _combine_data(cifar10_train, mnist_train, pad_equal) 39 | x_train, y_train = _create_dataset(x_train, y_train, timesteps, merge, seed) 40 | 41 | for i in range(len(x_train_files)): 42 | np.save(x_train_files[i], x_train[i]) 43 | np.save(y_train_file, y_train) 44 | 45 | x_train = [] 46 | for i in range(len(x_train_files)): 47 | x_train.append(np.load(x_train_files[i], mmap_mode=None)[:12500]) 48 | y_train = np.load(y_train_file, mmap_mode=None)[:12500] 49 | 50 | if not all([f.is_file() for f in x_test_files]) or not y_test_file.is_file(): 51 | x_test, y_test = _combine_data(cifar10_test, mnist_test, pad_equal) 52 | x_test, y_test = _create_dataset(x_test, y_test, timesteps, merge, seed) 53 | 54 | for i in range(len(x_test_files)): 55 | np.save(x_test_files[i], x_test[i]) 56 | np.save(y_test_file, y_test) 57 | 58 | x_test = [] 59 | for i in range(len(x_test_files)): 60 | x_test.append(np.load(x_test_files[i], mmap_mode=None)[:2230]) 61 | y_test = np.load(y_test_file, mmap_mode=None)[:2230] 62 | 63 | return (x_train, y_train), (x_test, y_test) 64 | 65 | 66 | def _create_dataset(features, labels, timesteps, merge, seed): 67 | features_a, features_b = features 68 | 69 | num_classes = np.unique(labels).size 70 | shape_a = features_a.shape[1:] 71 | shape_b = features_b.shape[1:] 72 | 73 | features_a = np.reshape(features_a, (-1, num_classes) + shape_a, order='F') 74 | features_b = np.reshape(features_b, (-1, num_classes) + shape_b, order='F') 75 | labels = np.reshape(labels, (-1, num_classes), order='F') 76 | 77 | a1, a2 = np.split(features_a, (timesteps * len(features_a) // (timesteps+1), )) 78 | b1, b2 = np.split(features_b, (timesteps * len(features_b) // (timesteps+1), )) 79 | y1, y2 = np.split(labels, (timesteps * len(labels) // (timesteps+1), )) 80 | 81 | a1 = np.reshape(a1, (-1, timesteps) + shape_a) 82 | a2 = np.reshape(a2, (-1, ) + shape_a) 83 | b1 = np.reshape(b1, (-1, timesteps) + shape_b) 84 | y1 = np.reshape(y1, (-1, timesteps)) 85 | y2 = np.reshape(y2, (-1, )) 86 | 87 | a1 = a1[:len(a1) // num_classes * num_classes] 88 | b1 = b1[:len(a1) // num_classes * num_classes] 89 | y1 = y1[:len(a1) // num_classes * num_classes] 90 | 91 | cols = [] 92 | for i in range(y1.shape[1]): 93 | cols.append(np.unique(y1[:, i])) 94 | unique_cols = np.unique(cols, axis=0) 95 | 96 | x_a = a1 97 | x_b = b1 98 | 99 | # Targets are always the first `len(a1) // len(unique_cols)` elements in each timestep. We shuffle it 100 | # afterwards. 101 | y = -1 * np.ones(y1.shape[0], dtype=y1.dtype) 102 | x_c = np.zeros((a1.shape[0], ) + a2.shape[1:]) 103 | target_mask = np.zeros(y1.shape, dtype=y1.dtype) 104 | for j, col in enumerate(unique_cols): 105 | for i in col: 106 | idc = np.nonzero(y1[:, j] == i)[0][:a1.shape[0] // num_classes] 107 | y[idc + j * a1.shape[0] // len(unique_cols)] = i 108 | 109 | target_mask[idc + j * a1.shape[0] // len(unique_cols), j] = 1 110 | 111 | idc2 = np.nonzero(y2 == i)[0][:a1.shape[0] // num_classes] 112 | x_c[idc + j * a1.shape[0] // len(unique_cols)] = a2[idc2] 113 | 114 | rows = np.split(np.indices(y1.shape)[0], 115 | np.arange(num_classes // timesteps, len(y1), num_classes // timesteps)) 116 | cols = np.split(np.indices(y1.shape)[1], 117 | np.arange(num_classes // timesteps, len(y1), num_classes // timesteps)) 118 | 119 | def shuffle_cols(cols, seed): 120 | return np.array([np.random.RandomState(seed=seed).permutation(c) for c in cols]) 121 | 122 | def shuffle_rows(rows, seed): 123 | y = rows.shape[1] 124 | tmp = rows 125 | for i in range(y): 126 | tmp[:, i] = np.random.RandomState(seed=seed+i).permutation(rows[:, i]) 127 | 128 | return tmp 129 | 130 | row_list_a, row_list_b = [], [] 131 | col_list_a, col_list_b = [], [] 132 | for i, (r, c) in enumerate(zip(rows, cols)): 133 | row_list_a.append(shuffle_rows(r, seed=seed*(i+1))) 134 | col_list_a.append(shuffle_cols(c, seed=seed*(i+2))) 135 | 136 | row_list_b.append(shuffle_rows(r, seed=seed*(i+3))) 137 | col_list_b.append(shuffle_cols(c, seed=seed*(i+4))) 138 | 139 | rows_a = np.concatenate(row_list_a) 140 | cols_a = np.concatenate(col_list_a) 141 | 142 | rows_b = np.concatenate(row_list_b) 143 | cols_b = np.concatenate(col_list_b) 144 | 145 | x_a = x_a[rows_a, cols_a] 146 | y_a = y1[rows_a, cols_a] 147 | if timesteps > 1: 148 | x_b = x_b[rows_b, cols_b] 149 | y_b = y1[rows_b, cols_b] 150 | else: 151 | y_b = y1 152 | x_c = np.stack([x_c] * timesteps, axis=1) 153 | x_c = x_c[rows_a, cols_a] 154 | 155 | target_mask = target_mask[rows_a, cols_a] 156 | x_c = x_c[np.nonzero(target_mask == 1)] 157 | y = y_b[np.nonzero(target_mask == 1)] 158 | 159 | idc = np.random.RandomState(seed=seed+10).permutation(y.shape[0]) 160 | x_a = x_a[idc] 161 | x_b = x_b[idc] 162 | x_c = x_c[idc] 163 | y_a = y_a[idc] 164 | y_b = y_b[idc] 165 | y = y[idc] 166 | 167 | if merge: 168 | x_ab = [] 169 | for a, b in zip(x_a.reshape((-1,) + x_a.shape[2:]), x_b.reshape((-1,) + x_b.shape[2:])): 170 | x_ab.append(image_manipulation.merge(a, b)) 171 | 172 | x_ab = np.array(x_ab).reshape((-1, timesteps) + x_ab[0].shape) 173 | 174 | pad_width = ((0, 0), (0, x_c.shape[2]), (0, 0)) 175 | x_c = image_manipulation.pad(x_c, pad_width=pad_width) 176 | 177 | return (x_ab, x_c), y 178 | else: 179 | return (x_a, x_b, x_c), y 180 | 181 | 182 | def _combine_data(a, b, pad): 183 | features_a, labels_a = a 184 | features_b, labels_b = b 185 | 186 | labels_a = labels_a.flatten() 187 | labels_b = labels_b.flatten() 188 | 189 | num_classes_a = np.unique(labels_a).size 190 | num_classes_b = np.unique(labels_b).size 191 | min_num_examples_a = min(np.unique(labels_a, return_counts=True)[1]) 192 | min_num_examples_b = min(np.unique(labels_b, return_counts=True)[1]) 193 | 194 | assert num_classes_a == num_classes_b 195 | 196 | pad_width_dim1_a = pad_width_dim1_b = pad_width_dim2_a = pad_width_dim2_b = 0 197 | if pad: 198 | if features_a.shape[1] > features_b.shape[1]: 199 | pad_width_dim1_b = (features_a.shape[1] - features_b.shape[1]) // 2 200 | if features_a.shape[1] < features_b.shape[1]: 201 | pad_width_dim1_a = (features_b.shape[1] - features_a.shape[1]) // 2 202 | if features_a.shape[2] > features_b.shape[2]: 203 | pad_width_dim2_b = (features_a.shape[2] - features_b.shape[2]) // 2 204 | if features_a.shape[2] < features_b.shape[2]: 205 | pad_width_dim2_a = (features_b.shape[2] - features_a.shape[2]) // 2 206 | 207 | x_a = [] 208 | x_b = [] 209 | y = [] 210 | for i in range(num_classes_a): 211 | idc_a = np.where(labels_a == i)[0] 212 | idc_b = np.where(labels_b == i)[0] 213 | num = min(idc_a.size, idc_b.size, min_num_examples_a, min_num_examples_b) 214 | x_a.append(features_a[idc_a[:num]]) 215 | x_b.append(features_b[idc_b[:num]]) 216 | y.append(labels_a[idc_a[:num]]) 217 | 218 | x_a = np.concatenate(x_a) 219 | x_b = np.concatenate(x_b) 220 | y = np.concatenate(y) 221 | 222 | pad_width_a = ((pad_width_dim1_a, pad_width_dim1_a), (pad_width_dim2_a, pad_width_dim2_a), (0, 0)) 223 | pad_width_b = ((pad_width_dim1_b, pad_width_dim1_b), (pad_width_dim2_b, pad_width_dim2_b), (0, 0)) 224 | x_a = image_manipulation.pad(x_a, pad_width=pad_width_a) 225 | x_b = image_manipulation.pad(x_b, pad_width=pad_width_b) 226 | 227 | return (x_a, x_b), y 228 | 229 | 230 | def _get_mnist_dataset(num_channels=1, pad_width=((0, 0), (0, 0), (0, 0))): 231 | mnist = tf.keras.datasets.mnist 232 | (x_train, y_train), (x_test, y_test) = mnist.load_data() 233 | x_train, x_test = x_train / 255.0, x_test / 255.0 234 | if num_channels == 1: 235 | x_train = x_train[..., tf.newaxis] 236 | x_test = x_test[..., tf.newaxis] 237 | else: 238 | x_train = image_manipulation.expand_channels(x_train, num_channels=num_channels) 239 | x_test = image_manipulation.expand_channels(x_test, num_channels=num_channels) 240 | 241 | x_train = image_manipulation.pad(x_train, pad_width) 242 | x_test = image_manipulation.pad(x_test, pad_width) 243 | 244 | return (x_train, y_train), (x_test, y_test) 245 | 246 | 247 | def _get_cifar10_dataset(): 248 | cifar10 = tf.keras.datasets.cifar10 249 | (x_train, y_train), (x_test, y_test) = cifar10.load_data() 250 | x_train, x_test = x_train / 255.0, x_test / 255.0 251 | 252 | return (x_train, y_train.flatten()), (x_test, y_test.flatten()) 253 | -------------------------------------------------------------------------------- /babi_task_single.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | """Runs H-Mem on a single bAbI task.""" 3 | 4 | import argparse 5 | import os 6 | import random 7 | from functools import reduce 8 | from itertools import chain 9 | 10 | import numpy as np 11 | import tensorflow as tf 12 | from data.babi_data import download, load_task, tasks, vectorize_data 13 | from layers.encoding import Encoding 14 | from layers.extracting import Extracting 15 | from layers.reading import ReadingCell 16 | from layers.writing import WritingCell 17 | from tensorflow.keras import Model 18 | from tensorflow.keras.layers import TimeDistributed 19 | from utils.logger import MyCSVLogger 20 | 21 | strategy = tf.distribute.MirroredStrategy() 22 | 23 | parser = argparse.ArgumentParser() 24 | parser.add_argument('--task_id', type=int, default=1) 25 | parser.add_argument('--max_num_sentences', type=int, default=-1) 26 | parser.add_argument('--training_set_size', type=str, default='10k', help='`1k` or `10k`') 27 | 28 | parser.add_argument('--epochs', type=int, default=100) 29 | parser.add_argument('--learning_rate', type=float, default=0.003) 30 | parser.add_argument('--batch_size_per_replica', type=int, default=128) 31 | parser.add_argument('--random_state', type=int, default=None) 32 | parser.add_argument('--max_grad_norm', type=float, default=20.0) 33 | parser.add_argument('--validation_split', type=float, default=0.1) 34 | 35 | parser.add_argument('--hops', type=int, default=3) 36 | parser.add_argument('--memory_size', type=int, default=100) 37 | parser.add_argument('--embeddings_size', type=int, default=80) 38 | parser.add_argument('--read_before_write', type=int, default=0) 39 | parser.add_argument('--gamma_pos', type=float, default=0.01) 40 | parser.add_argument('--gamma_neg', type=float, default=0.01) 41 | parser.add_argument('--w_assoc_max', type=float, default=1.0) 42 | parser.add_argument('--encodings_type', type=str, default='learned_encoding', 43 | help='`identity_encoding`, `position_encoding` or `learned_encoding`') 44 | 45 | parser.add_argument('--verbose', type=int, default=1) 46 | parser.add_argument('--logging', type=int, default=0) 47 | args = parser.parse_args() 48 | 49 | batch_size = args.batch_size_per_replica * strategy.num_replicas_in_sync 50 | 51 | # Set random seeds. 52 | np.random.seed(args.random_state) 53 | random.seed(args.random_state) 54 | tf.random.set_seed(args.random_state) 55 | 56 | if args.logging: 57 | logdir = 'results/' 58 | 59 | if not os.path.exists(logdir): 60 | os.makedirs(logdir) 61 | 62 | # Download bAbI data set. 63 | data_dir = download() 64 | 65 | if args.verbose: 66 | print('Extracting stories for the challenge: {0}, {1}'.format(args.task_id, tasks[args.task_id])) 67 | 68 | # Load the data. 69 | train, test = load_task(data_dir, args.task_id, args.training_set_size) 70 | data = train + test 71 | 72 | vocab = sorted(reduce(lambda x, y: x | y, (set(list(chain.from_iterable(s)) + q + a) for s, q, a in data))) 73 | word_idx = dict((c, i + 1) for i, c in enumerate(vocab)) 74 | 75 | max_story_size = max(map(len, (s for s, _, _ in data))) 76 | 77 | max_num_sentences = max_story_size if args.max_num_sentences == -1 else min(args.max_num_sentences, 78 | max_story_size) 79 | 80 | out_size = len(word_idx) + 1 # +1 for nil word. 81 | 82 | # Add time words/indexes 83 | for i in range(max_num_sentences): 84 | word_idx['time{}'.format(i+1)] = 'time{}'.format(i+1) 85 | 86 | vocab_size = len(word_idx) + 1 # +1 for nil word. 87 | mean_story_size = int(np.mean([len(s) for s, _, _ in data])) 88 | max_sentence_size = max(map(len, chain.from_iterable(s for s, _, _ in data))) + 1 # +1 for time word. 89 | max_query_size = max(map(len, (q for _, q, _ in data))) 90 | 91 | if args.verbose: 92 | print('-') 93 | print('Vocab size:', vocab_size, 'unique words (including "nil" word and "time" words)') 94 | print('Story max length:', max_story_size, 'sentences') 95 | print('Story mean length:', mean_story_size, 'sentences') 96 | print('Story max length:', max_sentence_size, 'words (including "time" word)') 97 | print('Query max length:', max_query_size, 'words') 98 | print('-') 99 | print('Here\'s what a "story" tuple looks like (story, query, answer):') 100 | print(data[0]) 101 | print('-') 102 | print('Vectorizing the stories...') 103 | 104 | # Vectorize the data. 105 | max_words = max(max_sentence_size, max_query_size) 106 | trainS, trainQ, trainA = vectorize_data(train, word_idx, max_num_sentences, max_words, max_words) 107 | testS, testQ, testA = vectorize_data(test, word_idx, max_num_sentences, max_words, max_words) 108 | 109 | trainQ = np.repeat(np.expand_dims(trainQ, axis=1), args.hops, axis=1) 110 | testQ = np.repeat(np.expand_dims(testQ, axis=1), args.hops, axis=1) 111 | 112 | story_shape = trainS.shape[1:] 113 | query_shape = trainQ.shape[1:] 114 | 115 | x_train = [trainS, trainQ] 116 | y_train = np.argmax(trainA, axis=1) 117 | 118 | x_test = [testS, testQ] 119 | y_test = np.argmax(testA, axis=1) 120 | 121 | if args.verbose: 122 | print('-') 123 | print('Stories: integer tensor of shape (samples, max_length, max_words): {0}'.format(trainS.shape)) 124 | print('Here\'s what a vectorized story looks like (sentence, word):') 125 | print(trainS[0]) 126 | print('-') 127 | print('Queries: integer tensor of shape (samples, length): {0}'.format(trainQ.shape)) 128 | print('Here\'s what a vectorized query looks like:') 129 | print(trainQ[0]) 130 | print('-') 131 | print('Answers: binary tensor of shape (samples, vocab_size): {0}'.format(trainA.shape)) 132 | print('Here\'s what a vectorized answer looks like:') 133 | print(trainA[0]) 134 | print('-') 135 | print('Training...') 136 | 137 | with strategy.scope(): 138 | # Build the model. 139 | story_input = tf.keras.layers.Input(story_shape, name='story_input') 140 | query_input = tf.keras.layers.Input(query_shape, name='query_input') 141 | 142 | embedding = tf.keras.layers.Embedding(input_dim=vocab_size, 143 | output_dim=args.embeddings_size, 144 | embeddings_initializer='he_uniform', 145 | embeddings_regularizer=None, 146 | mask_zero=True, 147 | name='embedding') 148 | story_embedded = TimeDistributed(embedding, name='story_embedding')(story_input) 149 | query_embedded = TimeDistributed(embedding, name='query_embedding')(query_input) 150 | 151 | encoding = Encoding(args.encodings_type, name='encoding') 152 | story_encoded = TimeDistributed(encoding, name='story_encoding')(story_embedded) 153 | query_encoded = TimeDistributed(encoding, name='query_encoding')(query_embedded) 154 | 155 | story_encoded = tf.keras.layers.BatchNormalization(name='batch_norm_story')(story_encoded) 156 | query_encoded = tf.keras.layers.BatchNormalization(name='batch_norm_query')(query_encoded) 157 | 158 | entities = Extracting(units=args.memory_size, 159 | use_bias=False, 160 | activation='relu', 161 | kernel_initializer='he_uniform', 162 | kernel_regularizer=tf.keras.regularizers.l2(1e-3), 163 | name='entity_extracting')(story_encoded) 164 | 165 | memory_matrix = tf.keras.layers.RNN(WritingCell(units=args.memory_size, 166 | read_before_write=args.read_before_write, 167 | use_bias=False, 168 | gamma_pos=args.gamma_pos, 169 | gamma_neg=args.gamma_neg, 170 | w_assoc_max=args.w_assoc_max, 171 | kernel_initializer='he_uniform', 172 | kernel_regularizer=tf.keras.regularizers.l2(1e-3)), 173 | name='entity_writing')(entities) 174 | 175 | queried_value = tf.keras.layers.RNN(ReadingCell(units=args.memory_size, 176 | use_bias=False, 177 | activation='relu', 178 | kernel_initializer='he_uniform', 179 | kernel_regularizer=tf.keras.regularizers.l2(1e-3)), 180 | name='entity_reading')(query_encoded, constants=[memory_matrix]) 181 | 182 | outputs = tf.keras.layers.Dense(vocab_size, 183 | use_bias=False, 184 | kernel_initializer='he_uniform', 185 | name='output')(queried_value) 186 | 187 | model = Model(inputs=[story_input, query_input], outputs=outputs) 188 | 189 | # Compile the model. 190 | optimizer_kwargs = {'clipnorm': args.max_grad_norm} if args.max_grad_norm else {} 191 | model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=args.learning_rate, **optimizer_kwargs), 192 | loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True), 193 | metrics=['accuracy']) 194 | 195 | model.summary() 196 | 197 | 198 | # Train and evaluate. 199 | def lr_scheduler(epoch): 200 | if args.read_before_write: 201 | if epoch < 150: 202 | return args.learning_rate 203 | else: 204 | return args.learning_rate * tf.math.exp(0.01 * (150 - epoch)) 205 | else: 206 | return args.learning_rate * 0.85**tf.math.floor(epoch / 20) 207 | 208 | 209 | callbacks = [] 210 | callbacks.append(tf.keras.callbacks.LearningRateScheduler(lr_scheduler, verbose=0)) 211 | if args.logging: 212 | callbacks.append(tf.keras.callbacks.CSVLogger(os.path.join(logdir, '{0}_{1}_{2}_{3}-{4}.log'.format( 213 | args.task_id, args.training_set_size, args.encodings_type, args.hops, args.random_state)))) 214 | 215 | model.fit(x=x_train, y=y_train, epochs=args.epochs, validation_split=args.validation_split, 216 | batch_size=batch_size, callbacks=callbacks, verbose=args.verbose) 217 | 218 | callbacks = [] 219 | if args.logging: 220 | callbacks.append(MyCSVLogger(os.path.join(logdir, '{0}_{1}_{2}_{3}-{4}.log'.format( 221 | args.task_id, args.training_set_size, args.encodings_type, args.hops, args.random_state)))) 222 | 223 | model.evaluate(x=x_test, y=y_test, callbacks=callbacks, verbose=2) 224 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | --------------------------------------------------------------------------------