├── .gitignore ├── RNN ├── data │ └── linux_input.txt ├── optimizer.py ├── linux.py ├── warpeace.py └── lab.py ├── README.md ├── FNN ├── batch_norm.py ├── mnist.py ├── optimizer.py ├── lab.py ├── svhn.py └── cifar10.py └── license /.gitignore: -------------------------------------------------------------------------------- 1 | *~ 2 | *.sw[op] 3 | -------------------------------------------------------------------------------- /RNN/data/linux_input.txt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/cafe/Loss-aware-Binarization/master/RNN/data/linux_input.txt -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Loss-Aware-Binarization 2 | Implementation of ICLR 2017 paper "Loss aware Binarization of Deep Networks", tested with GTX TITAN X, python 2.7, theano 0.9.0 and lasagne 0.2.dev1. 3 | 4 | This repository is divided in two subrepositories: 5 | 6 | - FNN: enables the reproduction of the FNN results(on MNIST, CIFAR-10, SVHN)reported in the article 7 | 8 | - RNN: enables the reproduction of the RNN results(on War and Peace, Linux Kernel) reported in the article 9 | 10 | Requirements 11 | This software is implemented on top of the implementation of [BinaryConnect](https://github.com/MatthieuCourbariaux/BinaryConnect) and has all the same requirements. 12 | 13 | 14 | Example training command on *War and Peace* dataset: 15 | - training LAB 16 | ```sh 17 | python warpeace.py --method="LAB" --lr_start=0.002 --w="w" --len=100 18 | ``` 19 | - training LAB2 20 | ```sh 21 | python warpeace.py --method="LAB" --lr_start=0.002 --w="wa" --len=100 22 | ``` 23 | -------------------------------------------------------------------------------- /FNN/batch_norm.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | 3 | """ 4 | Preliminary implementation of batch normalization for Lasagne. 5 | Does not include a way to properly compute the normalization factors over the 6 | full training set for testing, but can be used as a drop-in for training and 7 | validation. 8 | 9 | Author: Jan Schlüter 10 | """ 11 | 12 | import numpy as np 13 | import lasagne 14 | import theano 15 | import theano.tensor as T 16 | 17 | class BatchNormLayer(lasagne.layers.Layer): 18 | 19 | def __init__(self, incoming, axes=None, epsilon=0.01, alpha=0.5, 20 | nonlinearity=None, **kwargs): 21 | """ 22 | Instantiates a layer performing batch normalization of its inputs, 23 | following Ioffe et al. (http://arxiv.org/abs/1502.03167). 24 | 25 | @param incoming: `Layer` instance or expected input shape 26 | @param axes: int or tuple of int denoting the axes to normalize over; 27 | defaults to all axes except for the second if omitted (this will 28 | do the correct thing for dense layers and convolutional layers) 29 | @param epsilon: small constant added to the standard deviation before 30 | dividing by it, to avoid numeric problems 31 | @param alpha: coefficient for the exponential moving average of 32 | batch-wise means and standard deviations computed during training; 33 | the larger, the more it will depend on the last batches seen 34 | @param nonlinearity: nonlinearity to apply to the output (optional) 35 | """ 36 | super(BatchNormLayer, self).__init__(incoming, **kwargs) 37 | if axes is None: 38 | # default: normalize over all but the second axis 39 | axes = (0,) + tuple(range(2, len(self.input_shape))) 40 | elif isinstance(axes, int): 41 | axes = (axes,) 42 | self.axes = axes 43 | self.epsilon = epsilon 44 | self.alpha = alpha 45 | if nonlinearity is None: 46 | nonlinearity = lasagne.nonlinearities.identity 47 | self.nonlinearity = nonlinearity 48 | shape = list(self.input_shape) 49 | broadcast = [False] * len(shape) 50 | for axis in self.axes: 51 | shape[axis] = 1 52 | broadcast[axis] = True 53 | if any(size is None for size in shape): 54 | raise ValueError("BatchNormLayer needs specified input sizes for " 55 | "all dimensions/axes not normalized over.") 56 | dtype = theano.config.floatX 57 | self.mean = self.add_param(lasagne.init.Constant(0), shape, 'mean', 58 | trainable=False, regularizable=False) 59 | self.std = self.add_param(lasagne.init.Constant(1), shape, 'std', 60 | trainable=False, regularizable=False) 61 | self.beta = self.add_param(lasagne.init.Constant(0), shape, 'beta', 62 | trainable=True, regularizable=True) 63 | self.gamma = self.add_param(lasagne.init.Constant(1), shape, 'gamma', 64 | trainable=True, regularizable=False) 65 | 66 | def get_output_for(self, input, deterministic=False, **kwargs): 67 | if deterministic: 68 | # use stored mean and std 69 | mean = self.mean 70 | std = self.std 71 | else: 72 | # use this batch's mean and std 73 | mean = input.mean(self.axes, keepdims=True) 74 | std = input.std(self.axes, keepdims=True) 75 | # and update the stored mean and std: 76 | # we create (memory-aliased) clones of the stored mean and std 77 | running_mean = theano.clone(self.mean, share_inputs=False) 78 | running_std = theano.clone(self.std, share_inputs=False) 79 | # set a default update for them 80 | running_mean.default_update = ((1 - self.alpha) * running_mean + 81 | self.alpha * mean) 82 | running_std.default_update = ((1 - self.alpha) * running_std + 83 | self.alpha * std) 84 | # and include them in the graph so their default updates will be 85 | # applied (although the expressions will be optimized away later) 86 | mean += 0 * running_mean 87 | std += 0 * running_std 88 | std += self.epsilon 89 | mean = T.addbroadcast(mean, *self.axes) 90 | std = T.addbroadcast(std, *self.axes) 91 | beta = T.addbroadcast(self.beta, *self.axes) 92 | gamma = T.addbroadcast(self.gamma, *self.axes) 93 | normalized = (input - mean) * (gamma / std) + beta 94 | return self.nonlinearity(normalized) 95 | 96 | def batch_norm(layer): 97 | """ 98 | Convenience function to apply batch normalization to a given layer's output. 99 | Will steal the layer's nonlinearity if there is one (effectively introducing 100 | the normalization right before the nonlinearity), and will remove the 101 | layer's bias if there is one (because it would be redundant). 102 | 103 | @param layer: The `Layer` instance to apply the normalization to; note that 104 | it will be irreversibly modified as specified above 105 | @return: A `BatchNormLayer` instance stacked on the given `layer` 106 | """ 107 | nonlinearity = getattr(layer, 'nonlinearity', None) 108 | if nonlinearity is not None: 109 | layer.nonlinearity = lasagne.nonlinearities.identity 110 | if hasattr(layer, 'b'): 111 | del layer.params[layer.b] 112 | layer.b = None 113 | return BatchNormLayer(layer, nonlinearity=nonlinearity) 114 | -------------------------------------------------------------------------------- /FNN/mnist.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | 3 | import sys 4 | import os 5 | import time 6 | import ipdb 7 | import numpy as np 8 | np.random.seed(1234) 9 | from argparse import ArgumentParser 10 | 11 | import theano 12 | import theano.tensor as T 13 | 14 | import lasagne 15 | 16 | import cPickle as pickle 17 | import gzip 18 | 19 | import batch_norm 20 | import lab 21 | import optimizer 22 | 23 | from pylearn2.datasets.mnist import MNIST 24 | from pylearn2.utils import serial 25 | 26 | from collections import OrderedDict 27 | 28 | 29 | def main(method,LR_start,Binarize_weight_only): 30 | 31 | # BN parameters 32 | name = "mnist" 33 | print("dataset = "+str(name)) 34 | 35 | print("Binarize_weight_only="+str(Binarize_weight_only)) 36 | 37 | print("Method = "+str(method)) 38 | 39 | # alpha is the exponential moving average factor 40 | alpha = .1 41 | print("alpha = "+str(alpha)) 42 | epsilon = 1e-4 43 | print("epsilon = "+str(epsilon)) 44 | 45 | batch_size = 100 46 | print("batch_size = "+str(batch_size)) 47 | 48 | num_epochs = 50 49 | print("num_epochs = "+str(num_epochs)) 50 | 51 | # network structure 52 | num_units = 2048 53 | print("num_units = "+str(num_units)) 54 | n_hidden_layers = 3 55 | print("n_hidden_layers = "+str(n_hidden_layers)) 56 | 57 | print("LR_start = "+str(LR_start)) 58 | LR_decay = 0.1 59 | print("LR_decay="+str(LR_decay)) 60 | 61 | if Binarize_weight_only =="w": 62 | activation = lasagne.nonlinearities.rectify 63 | else: 64 | activation = lab.binary_tanh_unit 65 | print("activation = "+ str(activation)) 66 | 67 | 68 | print('Loading MNIST dataset...') 69 | 70 | train_set = MNIST(which_set= 'train', start=0, stop = 50000, center = True) 71 | valid_set = MNIST(which_set= 'train', start=50000, stop = 60000, center = True) 72 | test_set = MNIST(which_set= 'test', center = True) 73 | 74 | # bc01 format 75 | train_set.X = train_set.X.reshape(-1, 1, 28, 28) 76 | valid_set.X = valid_set.X.reshape(-1, 1, 28, 28) 77 | test_set.X = test_set.X.reshape(-1, 1, 28, 28) 78 | 79 | # flatten targets 80 | train_set.y = np.hstack(train_set.y) 81 | valid_set.y = np.hstack(valid_set.y) 82 | test_set.y = np.hstack(test_set.y) 83 | 84 | # Onehot the targets 85 | train_set.y = np.float32(np.eye(10)[train_set.y]) 86 | valid_set.y = np.float32(np.eye(10)[valid_set.y]) 87 | test_set.y = np.float32(np.eye(10)[test_set.y]) 88 | 89 | # for hinge loss 90 | train_set.y = 2* train_set.y - 1. 91 | valid_set.y = 2* valid_set.y - 1. 92 | test_set.y = 2* test_set.y - 1. 93 | 94 | print('Building the MLP...') 95 | 96 | # Prepare Theano variables for inputs and targets 97 | input = T.tensor4('inputs') 98 | target = T.matrix('targets') 99 | LR = T.scalar('LR', dtype=theano.config.floatX) 100 | 101 | mlp = lasagne.layers.InputLayer( 102 | shape=(None, 1, 28, 28), 103 | input_var=input) 104 | 105 | for k in range(n_hidden_layers): 106 | mlp = lab.DenseLayer( 107 | mlp, 108 | nonlinearity=lasagne.nonlinearities.identity, 109 | num_units=num_units, 110 | method = method) 111 | mlp = batch_norm.BatchNormLayer( 112 | mlp, 113 | epsilon=epsilon, 114 | alpha=alpha) 115 | mlp = lasagne.layers.NonlinearityLayer( 116 | mlp, 117 | nonlinearity = activation) 118 | 119 | mlp = lab.DenseLayer( 120 | mlp, 121 | nonlinearity=lasagne.nonlinearities.identity, 122 | num_units=10, 123 | method = method) 124 | 125 | mlp = batch_norm.BatchNormLayer( 126 | mlp, 127 | epsilon=epsilon, 128 | alpha=alpha) 129 | 130 | train_output = lasagne.layers.get_output(mlp, deterministic=False) 131 | 132 | # squared hinge loss 133 | loss = T.mean(T.sqr(T.maximum(0.,1.-target*train_output))) 134 | 135 | if method!="FPN": 136 | 137 | # W updates 138 | W = lasagne.layers.get_all_params(mlp, binary=True) 139 | W_grads = lab.compute_grads(loss,mlp) 140 | updates = optimizer.adam(loss_or_grads=W_grads, params=W, learning_rate=LR) 141 | updates = lab.clipping_scaling(updates,mlp) 142 | 143 | # other parameters updates 144 | params = lasagne.layers.get_all_params(mlp, trainable=True, binary=False) 145 | updates = OrderedDict(updates.items() + optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR).items()) 146 | 147 | ## update 2nd moment, can get from the adam optimizer also 148 | updates3 = OrderedDict() 149 | acc_tag = lasagne.layers.get_all_params(mlp, acc=True) 150 | idx = 0 151 | beta2 = 0.999 152 | for acc_tag_temp in acc_tag: 153 | updates3[acc_tag_temp]= acc_tag_temp*beta2 + W_grads[idx]*W_grads[idx]*(1-beta2) 154 | idx = idx+1 155 | 156 | updates = OrderedDict(updates.items() + updates3.items()) 157 | 158 | else: 159 | params = lasagne.layers.get_all_params(mlp, trainable=True) 160 | updates = optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR) 161 | 162 | test_output = lasagne.layers.get_output(mlp, deterministic=True) 163 | test_loss = T.mean(T.sqr(T.maximum(0.,1.-target*test_output))) 164 | test_err = T.mean(T.neq(T.argmax(test_output, axis=1), T.argmax(target, axis=1)),dtype=theano.config.floatX) 165 | 166 | # Compile a function performing a training step on a mini-batch (by giving the updates dictionary) 167 | # and returning the corresponding training loss: 168 | train_fn = theano.function([input, target, LR], loss, updates=updates) 169 | val_fn = theano.function([input, target], [test_loss, test_err]) 170 | 171 | print('Training...') 172 | 173 | lab.train( 174 | name, method, 175 | train_fn,val_fn, 176 | batch_size, 177 | LR_start,LR_decay, 178 | num_epochs, 179 | train_set.X,train_set.y, 180 | valid_set.X,valid_set.y, 181 | test_set.X,test_set.y) 182 | 183 | 184 | if __name__ == "__main__": 185 | parser = ArgumentParser() 186 | parser.add_argument("--method", type=str, dest="method", 187 | default="LAB", help="Method used") 188 | parser.add_argument("--lr_start", type=float, dest="LR_start", 189 | default=0.01, help="Learning rate") 190 | parser.add_argument("--w", type=str, dest="Binarize_weight_only", 191 | default="w", help="true:only binzrize w, false: binarize w and a") 192 | args = parser.parse_args() 193 | 194 | main(**vars(args)) -------------------------------------------------------------------------------- /FNN/optimizer.py: -------------------------------------------------------------------------------- 1 | 2 | from collections import OrderedDict 3 | 4 | import numpy as np 5 | 6 | import theano 7 | import theano.tensor as T 8 | # from . import utils 9 | 10 | 11 | def get_or_compute_grads(loss_or_grads, params): 12 | if any(not isinstance(p, theano.compile.SharedVariable) for p in params): 13 | raise ValueError("params must contain shared variables only. If it " 14 | "contains arbitrary parameter expressions, then " 15 | "lasagne.utils.collect_shared_vars() may help you.") 16 | if isinstance(loss_or_grads, list): 17 | if not len(loss_or_grads) == len(params): 18 | raise ValueError("Got %d gradient expressions for %d parameters" % 19 | (len(loss_or_grads), len(params))) 20 | return loss_or_grads 21 | else: 22 | return theano.grad(loss_or_grads, params) 23 | 24 | 25 | def sgd(loss_or_grads, params, learning_rate): 26 | grads = get_or_compute_grads(loss_or_grads, params) 27 | updates = OrderedDict() 28 | 29 | for param, grad in zip(params, grads): 30 | updates[param] = param - learning_rate * grad 31 | 32 | return updates 33 | 34 | 35 | def apply_momentum(updates, params=None, momentum=0.9): 36 | if params is None: 37 | params = updates.keys() 38 | updates = OrderedDict(updates) 39 | 40 | for param in params: 41 | value = param.get_value(borrow=True) 42 | velocity = theano.shared(np.zeros(value.shape, dtype=value.dtype), 43 | broadcastable=param.broadcastable) 44 | x = momentum * velocity + updates[param] 45 | updates[velocity] = x - param 46 | updates[param] = x 47 | 48 | return updates 49 | 50 | 51 | def momentum(loss_or_grads, params, learning_rate, momentum=0.9): 52 | updates = sgd(loss_or_grads, params, learning_rate) 53 | return apply_momentum(updates, momentum=momentum) 54 | 55 | 56 | def apply_nesterov_momentum(updates, params=None, momentum=0.9): 57 | if params is None: 58 | params = updates.keys() 59 | updates = OrderedDict(updates) 60 | 61 | for param in params: 62 | value = param.get_value(borrow=True) 63 | velocity = theano.shared(np.zeros(value.shape, dtype=value.dtype), 64 | broadcastable=param.broadcastable) 65 | x = momentum * velocity + updates[param] - param 66 | updates[velocity] = x 67 | updates[param] = momentum * x + updates[param] 68 | 69 | return updates 70 | 71 | 72 | def nesterov_momentum(loss_or_grads, params, learning_rate, momentum=0.9): 73 | updates = sgd(loss_or_grads, params, learning_rate) 74 | return apply_nesterov_momentum(updates, momentum=momentum) 75 | 76 | 77 | def adagrad(loss_or_grads, params, learning_rate=1.0, epsilon=1e-6): 78 | grads = get_or_compute_grads(loss_or_grads, params) 79 | updates = OrderedDict() 80 | 81 | for param, grad in zip(params, grads): 82 | value = param.get_value(borrow=True) 83 | accu = theano.shared(np.zeros(value.shape, dtype=value.dtype), 84 | broadcastable=param.broadcastable) 85 | accu_new = accu + grad ** 2 86 | updates[accu] = accu_new 87 | updates[param] = param - (learning_rate * grad / 88 | T.sqrt(accu_new + epsilon)) 89 | 90 | return updates 91 | 92 | 93 | def rmsprop(loss_or_grads, params, learning_rate=1.0, rho=0.9, epsilon=1e-6): 94 | grads = get_or_compute_grads(loss_or_grads, params) 95 | updates = OrderedDict() 96 | 97 | # Using theano constant to prevent upcasting of float32 98 | one = T.constant(1) 99 | 100 | for param, grad in zip(params, grads): 101 | value = param.get_value(borrow=True) 102 | accu = theano.shared(np.zeros(value.shape, dtype=value.dtype), 103 | broadcastable=param.broadcastable) 104 | accu_new = rho * accu + (one - rho) * grad ** 2 105 | updates[accu] = accu_new 106 | updates[param] = param - (learning_rate * grad / 107 | T.sqrt(accu_new + epsilon)) 108 | 109 | return updates 110 | 111 | 112 | def adadelta(loss_or_grads, params, learning_rate=1.0, rho=0.95, epsilon=1e-6): 113 | grads = get_or_compute_grads(loss_or_grads, params) 114 | updates = OrderedDict() 115 | 116 | # Using theano constant to prevent upcasting of float32 117 | one = T.constant(1) 118 | 119 | for param, grad in zip(params, grads): 120 | value = param.get_value(borrow=True) 121 | # accu: accumulate gradient magnitudes 122 | accu = theano.shared(np.zeros(value.shape, dtype=value.dtype), 123 | broadcastable=param.broadcastable) 124 | # delta_accu: accumulate update magnitudes (recursively!) 125 | delta_accu = theano.shared(np.zeros(value.shape, dtype=value.dtype), 126 | broadcastable=param.broadcastable) 127 | 128 | # update accu (as in rmsprop) 129 | accu_new = rho * accu + (one - rho) * grad ** 2 130 | updates[accu] = accu_new 131 | 132 | # compute parameter update, using the 'old' delta_accu 133 | update = (grad * T.sqrt(delta_accu + epsilon) / 134 | T.sqrt(accu_new + epsilon)) 135 | updates[param] = param - learning_rate * update 136 | 137 | # update delta_accu (as accu, but accumulating updates) 138 | delta_accu_new = rho * delta_accu + (one - rho) * update ** 2 139 | updates[delta_accu] = delta_accu_new 140 | 141 | return updates 142 | 143 | 144 | def adam(loss_or_grads, params, learning_rate=0.001, beta1=0.9, 145 | beta2=0.999, epsilon=1e-8): 146 | # if isinstance(loss_or_grads, list): 147 | # if not len(loss_or_grads) == len(params): 148 | # raise ValueError("Got %d gradient expressions for %d parameters" % 149 | # (len(loss_or_grads), len(params))) 150 | # all_grads = loss_or_grads 151 | # else: 152 | # all_grads = theano.grad(loss_or_grads, params) 153 | all_grads = get_or_compute_grads(loss_or_grads, params) 154 | # t_prev = theano.shared(lasagne.utils.floatX(0.)) 155 | t_prev = theano.shared(np.float32(0.)) 156 | updates = OrderedDict() 157 | 158 | t = t_prev + 1 159 | a_t = learning_rate*T.sqrt(1-beta2**t)/(1-beta1**t) 160 | 161 | for param, g_t in zip(params, all_grads): 162 | value = param.get_value(borrow=True) 163 | m_prev = theano.shared(np.zeros(value.shape, dtype=value.dtype), 164 | broadcastable=param.broadcastable) 165 | v_prev = theano.shared(np.zeros(value.shape, dtype=value.dtype), 166 | broadcastable=param.broadcastable) 167 | 168 | m_t = beta1*m_prev + (1-beta1)*g_t 169 | v_t = beta2*v_prev + (1-beta2)*g_t**2 170 | step = a_t*m_t/(T.sqrt(v_t) + epsilon) 171 | 172 | updates[m_prev] = m_t 173 | updates[v_prev] = v_t 174 | updates[param] = param - step 175 | 176 | updates[t_prev] = t 177 | return updates -------------------------------------------------------------------------------- /RNN/optimizer.py: -------------------------------------------------------------------------------- 1 | 2 | from collections import OrderedDict 3 | 4 | import numpy as np 5 | 6 | import theano 7 | import theano.tensor as T 8 | # from . import utils 9 | 10 | 11 | def get_or_compute_grads(loss_or_grads, params): 12 | if any(not isinstance(p, theano.compile.SharedVariable) for p in params): 13 | raise ValueError("params must contain shared variables only. If it " 14 | "contains arbitrary parameter expressions, then " 15 | "lasagne.utils.collect_shared_vars() may help you.") 16 | if isinstance(loss_or_grads, list): 17 | if not len(loss_or_grads) == len(params): 18 | raise ValueError("Got %d gradient expressions for %d parameters" % 19 | (len(loss_or_grads), len(params))) 20 | return loss_or_grads 21 | else: 22 | return theano.grad(loss_or_grads, params) 23 | 24 | 25 | def sgd(loss_or_grads, params, learning_rate): 26 | grads = get_or_compute_grads(loss_or_grads, params) 27 | updates = OrderedDict() 28 | 29 | for param, grad in zip(params, grads): 30 | updates[param] = param - learning_rate * grad 31 | 32 | return updates 33 | 34 | 35 | def apply_momentum(updates, params=None, momentum=0.9): 36 | if params is None: 37 | params = updates.keys() 38 | updates = OrderedDict(updates) 39 | 40 | for param in params: 41 | value = param.get_value(borrow=True) 42 | velocity = theano.shared(np.zeros(value.shape, dtype=value.dtype), 43 | broadcastable=param.broadcastable) 44 | x = momentum * velocity + updates[param] 45 | updates[velocity] = x - param 46 | updates[param] = x 47 | 48 | return updates 49 | 50 | 51 | def momentum(loss_or_grads, params, learning_rate, momentum=0.9): 52 | updates = sgd(loss_or_grads, params, learning_rate) 53 | return apply_momentum(updates, momentum=momentum) 54 | 55 | 56 | def apply_nesterov_momentum(updates, params=None, momentum=0.9): 57 | if params is None: 58 | params = updates.keys() 59 | updates = OrderedDict(updates) 60 | 61 | for param in params: 62 | value = param.get_value(borrow=True) 63 | velocity = theano.shared(np.zeros(value.shape, dtype=value.dtype), 64 | broadcastable=param.broadcastable) 65 | x = momentum * velocity + updates[param] - param 66 | updates[velocity] = x 67 | updates[param] = momentum * x + updates[param] 68 | 69 | return updates 70 | 71 | 72 | def nesterov_momentum(loss_or_grads, params, learning_rate, momentum=0.9): 73 | updates = sgd(loss_or_grads, params, learning_rate) 74 | return apply_nesterov_momentum(updates, momentum=momentum) 75 | 76 | 77 | def adagrad(loss_or_grads, params, learning_rate=1.0, epsilon=1e-6): 78 | grads = get_or_compute_grads(loss_or_grads, params) 79 | updates = OrderedDict() 80 | 81 | for param, grad in zip(params, grads): 82 | value = param.get_value(borrow=True) 83 | accu = theano.shared(np.zeros(value.shape, dtype=value.dtype), 84 | broadcastable=param.broadcastable) 85 | accu_new = accu + grad ** 2 86 | updates[accu] = accu_new 87 | updates[param] = param - (learning_rate * grad / 88 | T.sqrt(accu_new + epsilon)) 89 | 90 | return updates 91 | 92 | 93 | def rmsprop(loss_or_grads, params, learning_rate=1.0, rho=0.9, epsilon=1e-6): 94 | grads = get_or_compute_grads(loss_or_grads, params) 95 | updates = OrderedDict() 96 | 97 | # Using theano constant to prevent upcasting of float32 98 | one = T.constant(1) 99 | 100 | for param, grad in zip(params, grads): 101 | value = param.get_value(borrow=True) 102 | accu = theano.shared(np.zeros(value.shape, dtype=value.dtype), 103 | broadcastable=param.broadcastable) 104 | accu_new = rho * accu + (one - rho) * grad ** 2 105 | updates[accu] = accu_new 106 | updates[param] = param - (learning_rate * grad / 107 | T.sqrt(accu_new + epsilon)) 108 | 109 | return updates 110 | 111 | 112 | def adadelta(loss_or_grads, params, learning_rate=1.0, rho=0.95, epsilon=1e-6): 113 | grads = get_or_compute_grads(loss_or_grads, params) 114 | updates = OrderedDict() 115 | 116 | # Using theano constant to prevent upcasting of float32 117 | one = T.constant(1) 118 | 119 | for param, grad in zip(params, grads): 120 | value = param.get_value(borrow=True) 121 | # accu: accumulate gradient magnitudes 122 | accu = theano.shared(np.zeros(value.shape, dtype=value.dtype), 123 | broadcastable=param.broadcastable) 124 | # delta_accu: accumulate update magnitudes (recursively!) 125 | delta_accu = theano.shared(np.zeros(value.shape, dtype=value.dtype), 126 | broadcastable=param.broadcastable) 127 | 128 | # update accu (as in rmsprop) 129 | accu_new = rho * accu + (one - rho) * grad ** 2 130 | updates[accu] = accu_new 131 | 132 | # compute parameter update, using the 'old' delta_accu 133 | update = (grad * T.sqrt(delta_accu + epsilon) / 134 | T.sqrt(accu_new + epsilon)) 135 | updates[param] = param - learning_rate * update 136 | 137 | # update delta_accu (as accu, but accumulating updates) 138 | delta_accu_new = rho * delta_accu + (one - rho) * update ** 2 139 | updates[delta_accu] = delta_accu_new 140 | 141 | return updates 142 | 143 | 144 | def adam(loss_or_grads, params, learning_rate=0.001, beta1=0.9, 145 | beta2=0.999, epsilon=1e-8): 146 | # if isinstance(loss_or_grads, list): 147 | # if not len(loss_or_grads) == len(params): 148 | # raise ValueError("Got %d gradient expressions for %d parameters" % 149 | # (len(loss_or_grads), len(params))) 150 | # all_grads = loss_or_grads 151 | # else: 152 | # all_grads = theano.grad(loss_or_grads, params) 153 | all_grads = get_or_compute_grads(loss_or_grads, params) 154 | # t_prev = theano.shared(lasagne.utils.floatX(0.)) 155 | t_prev = theano.shared(np.float32(0.)) 156 | updates = OrderedDict() 157 | 158 | t = t_prev + 1 159 | a_t = learning_rate*T.sqrt(1-beta2**t)/(1-beta1**t) 160 | 161 | for param, g_t in zip(params, all_grads): 162 | value = param.get_value(borrow=True) 163 | m_prev = theano.shared(np.zeros(value.shape, dtype=value.dtype), 164 | broadcastable=param.broadcastable) 165 | v_prev = theano.shared(np.zeros(value.shape, dtype=value.dtype), 166 | broadcastable=param.broadcastable) 167 | 168 | m_t = beta1*m_prev + (1-beta1)*g_t 169 | v_t = beta2*v_prev + (1-beta2)*g_t**2 170 | step = a_t*m_t/(T.sqrt(v_t) + epsilon) 171 | 172 | updates[m_prev] = m_t 173 | updates[v_prev] = v_t 174 | updates[param] = param - step 175 | 176 | updates[t_prev] = t 177 | return updates -------------------------------------------------------------------------------- /RNN/linux.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | 3 | import sys 4 | import os 5 | import time 6 | 7 | import numpy as np 8 | np.random.seed(1234) 9 | 10 | 11 | import theano 12 | import theano.tensor as T 13 | 14 | import lasagne 15 | 16 | import cPickle as pickle 17 | import gzip 18 | 19 | import lab 20 | 21 | from pylearn2.datasets.mnist import MNIST 22 | from pylearn2.utils import serial 23 | 24 | from collections import OrderedDict 25 | import ipdb 26 | from argparse import ArgumentParser 27 | import optimizer 28 | 29 | 30 | def main(method,LR_start,Binarize_weight_only, SEQ_LENGTH): 31 | 32 | lasagne.random.set_rng(np.random.RandomState(1)) 33 | 34 | name = "linux" 35 | print("dataset = "+str(name)) 36 | 37 | print("Binarize_weight_only="+str(Binarize_weight_only)) 38 | 39 | print("Method = "+str(method)) 40 | 41 | # Sequence Length 42 | SEQ_LENGTH = SEQ_LENGTH 43 | # SEQ_LENGTH = 100 #can have diffvalues 50, 100, 200 44 | print("SEQ_LENGTH = "+str(SEQ_LENGTH)) 45 | 46 | # Number of units in the two hidden (LSTM) layers 47 | N_HIDDEN = 512 48 | print("N_HIDDEN = "+str(N_HIDDEN)) 49 | 50 | # All gradients above this will be clipped 51 | GRAD_CLIP=5. #### this clip the gradients at every time step, while T.clip clips the sum of gradients as a whole 52 | print("GRAD_CLIP ="+str(GRAD_CLIP)) 53 | 54 | # Number of epochs to train the net 55 | num_epochs = 200 56 | print("num_epochs = "+str(num_epochs)) 57 | 58 | # Batch Size 59 | batch_size = 100 60 | print("batch_size = "+str(batch_size)) 61 | 62 | print("LR_start = "+str(LR_start)) 63 | LR_decay = 0.98 64 | print("LR_decay="+str(LR_decay)) 65 | 66 | if Binarize_weight_only =="w": 67 | activation = lasagne.nonlinearities.tanh 68 | else: 69 | activation = lab.binary_tanh_unit 70 | print("activation = "+ str(activation)) 71 | 72 | name = name+"_"+Binarize_weight_only 73 | 74 | ## load data, change data file dir 75 | with open('data/linux_input.txt', 'r') as f: 76 | in_text = f.read() 77 | 78 | generation_phrase = "Copyright (C) 1992, 1998-2004 Linus Torvalds, Ingo Molnar\n *\n * This file contains the interrupt probing code and driver APIs.\n */\n\n#include" 79 | #This snippet loads the text file and creates dictionaries to 80 | #encode characters into a vector-space representation and vice-versa. 81 | chars = list(set(in_text)) 82 | data_size, vocab_size = len(in_text), len(chars) 83 | char_to_ix = { ch:i for i,ch in enumerate(chars) } 84 | ix_to_char = { i:ch for i,ch in enumerate(chars) } 85 | 86 | num_splits = [0.9, 0.05, 0.05] 87 | num_splits_all = np.floor(data_size/batch_size/SEQ_LENGTH) 88 | num_train = np.floor(num_splits_all*num_splits[0]) 89 | num_val = np.floor(num_splits_all*num_splits[1]) 90 | num_test = num_splits_all - num_train - num_val 91 | 92 | train_X = in_text[0:(num_train*batch_size*SEQ_LENGTH+1).astype('int32')] 93 | val_X = in_text[(num_train*batch_size*SEQ_LENGTH).astype('int32'):((num_train+num_val)*batch_size*SEQ_LENGTH+1).astype('int32')] 94 | test_X = in_text[((num_train+num_val)*batch_size*SEQ_LENGTH).astype('int32'):(num_splits_all*batch_size*SEQ_LENGTH+1).astype('int32')] 95 | 96 | 97 | ## build model 98 | print('Building the model...') 99 | 100 | # input = T.tensor3('inputs') 101 | target = T.imatrix('target') 102 | LR = T.scalar('LR', dtype=theano.config.floatX) 103 | 104 | # (batch size, SEQ_LENGTH, num_features) 105 | l_in = lasagne.layers.InputLayer(shape=(None, None, vocab_size)) 106 | l_forward_2 = lab.LSTMLayer( 107 | l_in, 108 | num_units=N_HIDDEN, 109 | grad_clipping=GRAD_CLIP, 110 | peepholes=False, 111 | nonlinearity=activation, ### change this activation can change the hidden layer to binary 112 | method=method) ### batch_size*SEQ_LENGTH*N_HIDDEN 113 | 114 | l_shp = lasagne.layers.ReshapeLayer(l_forward_2, (-1, N_HIDDEN)) ## (batch_size*SEQ_LENGTH, N_HIDDEN) 115 | l_out = lasagne.layers.DenseLayer(l_shp, num_units=vocab_size, W = lasagne.init.Normal(), nonlinearity=lasagne.nonlinearities.softmax) 116 | batchsize, seqlen, _ = l_in.input_var.shape 117 | l_shp1 = lasagne.layers.ReshapeLayer(l_out, (batchsize, seqlen, vocab_size)) 118 | l_out1 = lasagne.layers.SliceLayer(l_shp1, -1, 1) 119 | 120 | train_output = lasagne.layers.get_output(l_out, deterministic=False) 121 | loss = T.nnet.categorical_crossentropy(train_output,target.flatten()).mean() 122 | 123 | 124 | if method!= "FPN": 125 | # W updates 126 | W = lasagne.layers.get_all_params(l_out, binary=True) 127 | W_grads = lab.compute_grads(loss,l_out) 128 | updates = optimizer.adam(loss_or_grads=W_grads, params=W, learning_rate=LR, epsilon = 1e-8) ### can choose different methods to update 129 | updates = lab.clipping_scaling(updates,l_out) 130 | 131 | # other parameters updates 132 | params = lasagne.layers.get_all_params(l_out, trainable=True, binary=False) 133 | updates = OrderedDict(updates.items() + optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR, epsilon = 1e-8).items()) 134 | 135 | ## update 2 momentum 136 | updates3 = OrderedDict() 137 | acc_tag = lasagne.layers.get_all_params(l_out, acc=True) 138 | idx = 0 139 | beta2 = 0.999 140 | for acc_tag_temp in acc_tag: 141 | # updates3[acc_tag_temp]=updates.keys()[idx] 142 | updates3[acc_tag_temp]= acc_tag_temp*beta2 + W_grads[idx]*W_grads[idx]*(1-beta2) 143 | idx = idx+1 144 | 145 | updates = OrderedDict(updates.items() + updates3.items()) 146 | 147 | else: 148 | params_other = lasagne.layers.get_all_params(l_out, trainable=True) 149 | 150 | W_grads = [theano.grad(loss, wrt=l_forward_2.W_in_to_ingate), theano.grad(loss, wrt=l_forward_2.W_hid_to_ingate), 151 | theano.grad(loss, wrt=l_forward_2.W_in_to_fotgetgate),theano.grad(loss, wrt=l_forward_2.W_hid_to_forgetgate), 152 | theano.grad(loss, wrt=l_forward_2.W_in_to_cell),theano.grad(loss, wrt=l_forward_2.W_hid_to_cell), 153 | theano.grad(loss, wrt=l_forward_2.W_in_to_outgate),theano.grad(loss, wrt=l_forward_2.W_hid_to_outgate)] 154 | 155 | updates = optimizer.adam(loss_or_grads=loss, params=params_other, learning_rate=LR) 156 | 157 | test_output = lasagne.layers.get_output(l_out, deterministic=True) 158 | test_loss = T.nnet.categorical_crossentropy(test_output,target.flatten()).mean() 159 | 160 | 161 | 162 | train_fn = theano.function([l_in.input_var, target, LR], [loss, W_grads[5]], updates=updates, allow_input_downcast=True) 163 | val_fn = theano.function([l_in.input_var, target], test_loss, allow_input_downcast=True) 164 | probs = theano.function([l_in.input_var],lasagne.layers.get_output(l_out1), allow_input_downcast=True) 165 | 166 | 167 | print('Training...') 168 | 169 | lab.train( 170 | name, method, 171 | train_fn,val_fn, 172 | batch_size, 173 | SEQ_LENGTH, 174 | N_HIDDEN, 175 | LR_start,LR_decay, 176 | num_epochs, 177 | train_X, 178 | val_X, 179 | test_X) 180 | 181 | 182 | if __name__ == "__main__": 183 | parser = ArgumentParser() 184 | parser.add_argument("--method", type=str, dest="method", 185 | default="LAB", help="Method used") 186 | parser.add_argument("--lr_start", type=float, dest="LR_start", 187 | default=2e-3, help="Learning rate") 188 | parser.add_argument("--w", type=str, dest="Binarize_weight_only", 189 | default="w", help="true:only binzrize w, false: binarize w and a") 190 | parser.add_argument("--len", type=int, dest="SEQ_LENGTH", 191 | default=100, help="unrolled timesteps for LSTM") 192 | args = parser.parse_args() 193 | 194 | main(**vars(args)) 195 | -------------------------------------------------------------------------------- /RNN/warpeace.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | 3 | import sys 4 | import os 5 | import time 6 | 7 | import numpy as np 8 | np.random.seed(1234) 9 | 10 | 11 | import theano 12 | import theano.tensor as T 13 | 14 | import lasagne 15 | 16 | import cPickle as pickle 17 | import gzip 18 | 19 | import lab 20 | import optimizer 21 | 22 | # from pylearn2.datasets.mnist import MNIST 23 | from pylearn2.utils import serial 24 | 25 | from collections import OrderedDict 26 | import ipdb 27 | from argparse import ArgumentParser 28 | 29 | 30 | def main(method,LR_start,Binarize_weight_only, SEQ_LENGTH): 31 | 32 | lasagne.random.set_rng(np.random.RandomState(1)) 33 | 34 | name = "warpeace" 35 | print("dataset = "+str(name)) 36 | 37 | print("Binarize_weight_only="+str(Binarize_weight_only)) 38 | 39 | print("Method = "+str(method)) 40 | 41 | SEQ_LENGTH = SEQ_LENGTH 42 | # Sequence Length 43 | # SEQ_LENGTH = 50 #can have diffvalues 50, 100, 200 44 | print("SEQ_LENGTH = "+str(SEQ_LENGTH)) 45 | 46 | # Number of units in the two hidden (LSTM) layers 47 | N_HIDDEN = 512 48 | print("N_HIDDEN = "+str(N_HIDDEN)) 49 | 50 | # All gradients above this will be clipped 51 | GRAD_CLIP=5. #### this clip the gradients at every time step, while T.clip clip the sum of gradients as a whole 52 | print("GRAD_CLIP ="+str(GRAD_CLIP)) 53 | 54 | # Number of epochs to train the net 55 | num_epochs = 200 56 | print("num_epochs = "+str(num_epochs)) 57 | 58 | # Batch Size 59 | batch_size = 100 60 | print("batch_size = "+str(batch_size)) 61 | 62 | print("LR_start = "+str(LR_start)) 63 | LR_decay = 0.98 64 | print("LR_decay="+str(LR_decay)) 65 | 66 | if Binarize_weight_only == "w": 67 | activation = lasagne.nonlinearities.tanh 68 | else: 69 | activation = lab.binary_tanh_unit 70 | print("activation = "+ str(activation)) 71 | 72 | name = name+"_"+Binarize_weight_only 73 | 74 | ## load data, change the data file dir 75 | with open('data/warpeace_input.txt', 'r') as f: 76 | in_text = f.read() 77 | # generation_phrase = "Anna Pavlovna's drawing room was gradually filling. The highest Petersburg society was assembled there: people differing widely in" 78 | generation_phrase = "With\r\nthese words she greeted Prince Vasili Kuragin, a man of high rank and\r\nimportance, who was the first to arrive at her reception. Anna Pavlovna\r\nhad had a cough for some days. She was, as sh" 79 | #This snippet loads the text file and creates dictionaries to 80 | #encode characters into a vector-space representation and vice-versa. 81 | chars = list(set(in_text)) 82 | data_size, vocab_size = len(in_text), len(chars) 83 | char_to_ix = { ch:i for i,ch in enumerate(chars) } 84 | ix_to_char = { i:ch for i,ch in enumerate(chars) } 85 | 86 | num_splits = [0.8, 0.1, 0.1] 87 | num_splits_all = np.floor(data_size/batch_size/SEQ_LENGTH) 88 | num_train = np.floor(num_splits_all*num_splits[0]) 89 | num_val = np.floor(num_splits_all*num_splits[1]) 90 | num_test = num_splits_all - num_train - num_val 91 | 92 | train_X = in_text[0:(num_train*batch_size*SEQ_LENGTH+1).astype('int32')] 93 | val_X = in_text[(num_train*batch_size*SEQ_LENGTH).astype('int32'):((num_train+num_val)*batch_size*SEQ_LENGTH+1).astype('int32')] 94 | test_X = in_text[((num_train+num_val)*batch_size*SEQ_LENGTH).astype('int32'):(num_splits_all*batch_size*SEQ_LENGTH+1).astype('int32')] 95 | 96 | ## build model 97 | print('Building the model...') 98 | # input = T.tensor3('inputs') 99 | target = T.imatrix('target') 100 | LR = T.scalar('LR', dtype=theano.config.floatX) 101 | 102 | # (batch size, SEQ_LENGTH, num_features) 103 | l_in = lasagne.layers.InputLayer(shape=(None, None, vocab_size)) 104 | l_forward_2 = lab.LSTMLayer( 105 | l_in, 106 | num_units=N_HIDDEN, 107 | grad_clipping=GRAD_CLIP, 108 | peepholes=False, 109 | nonlinearity=activation, 110 | method=method) ### batch_size*SEQ_LENGTH*N_HIDDEN 111 | 112 | l_shp = lasagne.layers.ReshapeLayer(l_forward_2, (-1, N_HIDDEN)) ## (batch_size*SEQ_LENGTH, N_HIDDEN) 113 | l_out = lasagne.layers.DenseLayer(l_shp, num_units=vocab_size, W = lasagne.init.Normal(), nonlinearity=lasagne.nonlinearities.softmax) 114 | batchsize, seqlen, _ = l_in.input_var.shape 115 | l_shp1 = lasagne.layers.ReshapeLayer(l_out, (batchsize, seqlen, vocab_size)) 116 | l_out1 = lasagne.layers.SliceLayer(l_shp1, -1, 1) 117 | 118 | train_output = lasagne.layers.get_output(l_out, deterministic=False) 119 | loss = T.nnet.categorical_crossentropy(train_output,target.flatten()).mean() 120 | 121 | 122 | if method!= "FPN": 123 | # W updates 124 | W = lasagne.layers.get_all_params(l_out, binary=True) 125 | W_grads = lab.compute_grads(loss,l_out) 126 | updates = optimizer.adam(loss_or_grads=W_grads, params=W, learning_rate=LR, epsilon = 1e-8) ### can choose different methods to update 127 | updates = lab.clipping_scaling(updates,l_out) 128 | 129 | # other parameters updates 130 | params = lasagne.layers.get_all_params(l_out, trainable=True, binary=False) 131 | updates = OrderedDict(updates.items() + optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR, epsilon = 1e-8).items()) 132 | 133 | ## update 2nd momentum 134 | updates3 = OrderedDict() 135 | acc_tag = lasagne.layers.get_all_params(l_out, acc=True) 136 | idx = 0 137 | beta2 = 0.999 138 | for acc_tag_temp in acc_tag: 139 | updates3[acc_tag_temp]= acc_tag_temp*beta2 + W_grads[idx]*W_grads[idx]*(1-beta2) 140 | idx = idx+1 141 | 142 | updates = OrderedDict(updates.items() + updates3.items()) 143 | 144 | else: 145 | params_other = lasagne.layers.get_all_params(l_out, trainable=True) 146 | 147 | W_grads = [theano.grad(loss, wrt=l_forward_2.W_in_to_ingate), theano.grad(loss, wrt=l_forward_2.W_hid_to_ingate), 148 | theano.grad(loss, wrt=l_forward_2.W_in_to_forgetgate),theano.grad(loss, wrt=l_forward_2.W_hid_to_forgetgate), 149 | theano.grad(loss, wrt=l_forward_2.W_in_to_cell),theano.grad(loss, wrt=l_forward_2.W_hid_to_cell), 150 | theano.grad(loss, wrt=l_forward_2.W_in_to_outgate),theano.grad(loss, wrt=l_forward_2.W_hid_to_outgate)] 151 | 152 | updates = optimizer.adam(loss_or_grads=loss, params=params_other, learning_rate=LR) 153 | 154 | test_output = lasagne.layers.get_output(l_out, deterministic=True) 155 | test_loss = T.nnet.categorical_crossentropy(test_output,target.flatten()).mean() 156 | 157 | train_fn = theano.function([l_in.input_var, target, LR], [loss, W_grads[5]], updates=updates, allow_input_downcast=True) 158 | val_fn = theano.function([l_in.input_var, target], test_loss, allow_input_downcast=True) 159 | probs = theano.function([l_in.input_var],lasagne.layers.get_output(l_out1), allow_input_downcast=True) 160 | 161 | 162 | print('Training...') 163 | 164 | lab.train( 165 | name, method, 166 | train_fn,val_fn, 167 | batch_size, 168 | SEQ_LENGTH, 169 | N_HIDDEN, 170 | LR_start,LR_decay, 171 | num_epochs, 172 | train_X, 173 | val_X, 174 | test_X) 175 | 176 | 177 | if __name__ == "__main__": 178 | parser = ArgumentParser() 179 | parser.add_argument("--method", type=str, dest="method", 180 | default="LAB", help="Method used") 181 | parser.add_argument("--lr_start", type=float, dest="LR_start", 182 | default=2e-3, help="Learning rate") 183 | parser.add_argument("--w", type=str, dest="Binarize_weight_only", 184 | default="w", help="w:only binzrize w, wa: binarize w and a") 185 | parser.add_argument("--len", type=int, dest="SEQ_LENGTH", 186 | default=100, help="unrolled timesteps for LSTM") 187 | args = parser.parse_args() 188 | 189 | main(**vars(args)) -------------------------------------------------------------------------------- /FNN/lab.py: -------------------------------------------------------------------------------- 1 | import time 2 | 3 | from collections import OrderedDict 4 | 5 | import numpy as np 6 | np.random.seed(1234) 7 | 8 | import ipdb 9 | 10 | import theano 11 | import theano.tensor as T 12 | 13 | import lasagne 14 | from theano.scalar.basic import UnaryScalarOp, same_out_nocomplex 15 | from theano.tensor.elemwise import Elemwise 16 | from theano.sandbox.rng_mrg import MRG_RandomStreams as RandomStreams 17 | 18 | class Round3(UnaryScalarOp): 19 | 20 | def c_code(self, node, name, (x,), (z,), sub): 21 | return "%(z)s = round(%(x)s);" % locals() 22 | 23 | def grad(self, inputs, gout): 24 | (gz,) = gout 25 | return gz, 26 | 27 | round3_scalar = Round3(same_out_nocomplex, name='round3') 28 | round3 = Elemwise(round3_scalar) 29 | 30 | def hard_sigmoid(x): 31 | return T.clip((x+1.)/2.,0,1) 32 | 33 | def binary_tanh_unit(x): 34 | return 2.*round3(hard_sigmoid(x))-1. 35 | 36 | def binary_sigmoid_unit(x): 37 | return round3(hard_sigmoid(x)) 38 | 39 | 40 | # The binarization function 41 | def binarization(W,Wacc,method): 42 | 43 | if method == "FPN": 44 | Wb = W 45 | 46 | elif method == "LAB": 47 | L = (T.sqrt(Wacc) + 1e-8) 48 | Wb = hard_sigmoid(W) 49 | Wb = round3(Wb) 50 | Wb = T.cast(T.switch(Wb,1.,-1.), theano.config.floatX) 51 | 52 | alpha = (T.abs_(L*W).sum()/L.sum()).astype('float32') 53 | Wb = alpha*Wb 54 | 55 | return Wb 56 | 57 | 58 | # This class extends the Lasagne DenseLayer to support LAB 59 | class DenseLayer(lasagne.layers.DenseLayer): 60 | 61 | def __init__(self, incoming, num_units, method, **kwargs): 62 | 63 | self.method = method 64 | num_inputs = int(np.prod(incoming.output_shape[1:])) 65 | g_init = np.float32(np.sqrt(1.5/ (num_inputs + num_units))) 66 | if self.method !="FPN": 67 | super(DenseLayer, self).__init__(incoming, num_units, W=lasagne.init.Uniform((-g_init,g_init)), **kwargs) 68 | # add the binary tag to weights 69 | self.params[self.W]=set(['binary']) 70 | else: 71 | super(DenseLayer, self).__init__(incoming, num_units, **kwargs) 72 | # add the acc tag to 2nd momentum 73 | self.acc_W = theano.shared(np.zeros((self.W.get_value(borrow=True)).shape, dtype='float32')) 74 | self.params[self.acc_W]=set(['acc']) 75 | 76 | def get_output_for(self, input, deterministic=False, **kwargs): 77 | 78 | self.Wb = binarization(self.W, self.acc_W, self.method) 79 | Wr = self.W 80 | self.W = self.Wb 81 | 82 | rvalue = super(DenseLayer, self).get_output_for(input, **kwargs) 83 | 84 | self.W = Wr 85 | 86 | return rvalue 87 | 88 | # This class extends the Lasagne Conv2DLayer to support LAB 89 | class Conv2DLayer(lasagne.layers.Conv2DLayer): 90 | 91 | def __init__(self, incoming, num_filters, filter_size, method, **kwargs): 92 | 93 | self.method = method 94 | 95 | num_inputs = int(np.prod(filter_size)*incoming.output_shape[1]) 96 | num_units = int(np.prod(filter_size)*num_filters) # theoretically, I should divide num_units by the pool_shape 97 | g_init = np.float32(np.sqrt(1.5/ (num_inputs + num_units))) 98 | 99 | if self.method!="FPN": 100 | super(Conv2DLayer, self).__init__(incoming, num_filters, filter_size, W=lasagne.init.Uniform((-g_init,g_init)), **kwargs) 101 | # add the binary tag to weights 102 | self.params[self.W]=set(['binary']) 103 | else: 104 | super(Conv2DLayer, self).__init__(incoming, num_filters, filter_size, **kwargs) 105 | 106 | self.acc_W = theano.shared(np.zeros((self.W.get_value(borrow=True)).shape, dtype='float32')) 107 | self.params[self.acc_W]=set(['acc']) 108 | 109 | 110 | def get_output_for(self, input, deterministic=False, **kwargs): 111 | 112 | self.Wb = binarization(self.W, self.acc_W, self.method) 113 | Wr = self.W 114 | self.W = self.Wb 115 | rvalue = super(Conv2DLayer, self).get_output_for(input, **kwargs) 116 | self.W = Wr 117 | 118 | return rvalue 119 | 120 | def compute_grads(loss,network): 121 | 122 | layers = lasagne.layers.get_all_layers(network) 123 | grads = [] 124 | 125 | for layer in layers: 126 | params = layer.get_params(binary=True) 127 | if params: 128 | grads.append(theano.grad(loss, wrt=layer.Wb)) 129 | 130 | return grads 131 | 132 | # This functions clips the weights after the parameter update 133 | def clipping_scaling(updates,network): 134 | 135 | layers = lasagne.layers.get_all_layers(network) 136 | updates = OrderedDict(updates) 137 | 138 | for layer in layers: 139 | params = layer.get_params(binary=True) 140 | for param in params: 141 | updates[param] = T.clip(updates[param],-1.,1.) 142 | return updates 143 | 144 | # Given a dataset and a model, this function trains the model on the dataset for several epochs 145 | # (There is no default train function in Lasagne yet) 146 | def train(name,method,train_fn,val_fn, 147 | batch_size, 148 | LR_start,LR_decay, 149 | num_epochs, 150 | X_train,y_train, 151 | X_val,y_val, 152 | X_test,y_test): 153 | 154 | # This function trains the model a full epoch (on the whole dataset) 155 | def train_epoch(X,y,LR): 156 | 157 | loss = 0 158 | batches = len(X)/batch_size 159 | # move shuffle here to save memory 160 | shuffled_range = range(len(X)) 161 | np.random.shuffle(shuffled_range) 162 | 163 | for i in range(batches): 164 | tmp_ind = shuffled_range[i*batch_size:(i+1)*batch_size] 165 | newloss = train_fn(X[tmp_ind],y[tmp_ind],LR) 166 | loss +=newloss 167 | loss/=batches 168 | 169 | return loss 170 | 171 | # This function tests the model a full epoch (on the whole dataset) 172 | def val_epoch(X,y): 173 | 174 | err = 0 175 | loss = 0 176 | batches = len(X)/batch_size 177 | 178 | for i in range(batches): 179 | new_loss, new_err = val_fn(X[i*batch_size:(i+1)*batch_size], y[i*batch_size:(i+1)*batch_size]) 180 | err += new_err 181 | loss += new_loss 182 | 183 | err = err / batches * 100 184 | loss /= batches 185 | 186 | return err, loss 187 | 188 | 189 | best_val_err = 100 190 | best_epoch = 1 191 | LR = LR_start 192 | # We iterate over epochs: 193 | for epoch in range(1, num_epochs+1): 194 | 195 | start_time = time.time() 196 | train_loss = train_epoch(X_train,y_train,LR) 197 | 198 | val_err, val_loss = val_epoch(X_val,y_val) 199 | 200 | # test if validation error went down 201 | if val_err <= best_val_err: 202 | 203 | best_val_err = val_err 204 | best_epoch = epoch+1 205 | 206 | test_err, test_loss = val_epoch(X_test,y_test) 207 | 208 | epoch_duration = time.time() - start_time 209 | 210 | # Then we print the results for this epoch: 211 | print("Epoch "+str(epoch)+" of "+str(num_epochs)+" took "+str(epoch_duration)+"s") 212 | print(" LR: "+str(LR)) 213 | print(" training loss: "+str(train_loss)) 214 | print(" validation loss: "+str(val_loss)) 215 | print(" validation error rate: "+str(val_err)+"%") 216 | print(" best epoch: "+str(best_epoch)) 217 | print(" best validation error rate: "+str(best_val_err)+"%") 218 | print(" test loss: "+str(test_loss)) 219 | print(" test error rate: "+str(test_err)+"%") 220 | 221 | 222 | with open("{0}_lr{1}_{2}.txt".format(name, LR_start, method), "a") as myfile: 223 | myfile.write("{0} {1:.5f} {2:.5f} {3:.5f} {4:.5f} {5:.5f} {6:.5f} {7:.5f}\n".format(epoch, 224 | train_loss, val_loss, test_loss, val_err, test_err, epoch_duration, LR)) 225 | 226 | 227 | ## Learning rate update scheme 228 | if name=="mnist": 229 | if epoch == 15 or epoch==25: 230 | LR*=LR_decay 231 | elif name=="cifar": 232 | if epoch % 15 ==0: 233 | LR*=LR_decay 234 | elif name=="svhn": 235 | if epoch == 15 or epoch==25: 236 | LR *=LR_decay -------------------------------------------------------------------------------- /FNN/svhn.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | 3 | import sys 4 | import os 5 | import time 6 | import ipdb 7 | 8 | import numpy as np 9 | np.random.seed(1234) 10 | 11 | import theano 12 | import theano.tensor as T 13 | 14 | import lasagne 15 | 16 | import cPickle as pickle 17 | import gzip 18 | 19 | import batch_norm 20 | import lab 21 | import optimizer 22 | 23 | from pylearn2.datasets.svhn import SVHN 24 | from pylearn2.utils import serial 25 | 26 | from pylearn2.datasets.zca_dataset import ZCA_Dataset 27 | from pylearn2.utils import serial 28 | 29 | from collections import OrderedDict 30 | from argparse import ArgumentParser 31 | 32 | def main(method,LR_start,Binarize_weight_only): 33 | 34 | name = "svhn" 35 | print("dataset = "+str(name)) 36 | 37 | print("Binarize_weight_only="+str(Binarize_weight_only)) 38 | 39 | print("Method = "+str(method)) 40 | 41 | # alpha is the exponential moving average factor 42 | alpha = .1 43 | print("alpha = "+str(alpha)) 44 | epsilon = 1e-4 45 | print("epsilon = "+str(epsilon)) 46 | 47 | # Training parameters 48 | batch_size = 50 49 | print("batch_size = "+str(batch_size)) 50 | 51 | num_epochs = 50 52 | print("num_epochs = "+str(num_epochs)) 53 | 54 | print("LR_start = "+str(LR_start)) 55 | LR_decay = 0.1 56 | print("LR_decay="+str(LR_decay)) 57 | # BTW, LR decay might good for the BN moving average... 58 | 59 | if Binarize_weight_only =="w": 60 | activation = lasagne.nonlinearities.rectify 61 | else: 62 | activation = lab.binary_tanh_unit 63 | print("activation = "+ str(activation)) 64 | 65 | ## number of filters in the first convolutional layer 66 | K = 64 67 | print("K="+str(K)) 68 | 69 | print('Building the CNN...') 70 | 71 | # Prepare Theano variables for inputs and targets 72 | input = T.tensor4('inputs') 73 | target = T.matrix('targets') 74 | LR = T.scalar('LR', dtype=theano.config.floatX) 75 | 76 | l_in = lasagne.layers.InputLayer( 77 | shape=(None, 3, 32, 32), 78 | input_var=input) 79 | 80 | # 128C3-128C3-P2 81 | l_cnn1 = lab.Conv2DLayer( 82 | l_in, 83 | num_filters=K, 84 | filter_size=(3, 3), 85 | pad=1, 86 | nonlinearity=lasagne.nonlinearities.identity, 87 | method = method) 88 | 89 | l_bn1 = batch_norm.BatchNormLayer( 90 | l_cnn1, 91 | epsilon=epsilon, 92 | alpha=alpha) 93 | 94 | l_nl1 = lasagne.layers.NonlinearityLayer( 95 | l_bn1, 96 | nonlinearity = activation) 97 | 98 | l_cnn2 = lab.Conv2DLayer( 99 | l_nl1, 100 | num_filters=K, 101 | filter_size=(3, 3), 102 | pad=1, 103 | nonlinearity=lasagne.nonlinearities.identity, 104 | method = method) 105 | 106 | l_mp1 = lasagne.layers.MaxPool2DLayer(l_cnn2, pool_size=(2, 2)) 107 | 108 | l_bn2 = batch_norm.BatchNormLayer( 109 | l_mp1, 110 | epsilon=epsilon, 111 | alpha=alpha) 112 | 113 | l_nl2 = lasagne.layers.NonlinearityLayer( 114 | l_bn2, 115 | nonlinearity = activation) 116 | # 256C3-256C3-P2 117 | l_cnn3 = lab.Conv2DLayer( 118 | l_nl2, 119 | num_filters=2*K, 120 | filter_size=(3, 3), 121 | pad=1, 122 | nonlinearity=lasagne.nonlinearities.identity, 123 | method = method) 124 | 125 | l_bn3 = batch_norm.BatchNormLayer( 126 | l_cnn3, 127 | epsilon=epsilon, 128 | alpha=alpha) 129 | 130 | l_nl3 = lasagne.layers.NonlinearityLayer( 131 | l_bn3, 132 | nonlinearity = activation) 133 | 134 | l_cnn4 = lab.Conv2DLayer( 135 | l_nl3, 136 | num_filters=2*K, 137 | filter_size=(3, 3), 138 | pad=1, 139 | nonlinearity=lasagne.nonlinearities.identity, 140 | method = method) 141 | 142 | l_mp2 = lasagne.layers.MaxPool2DLayer(l_cnn4, pool_size=(2, 2)) 143 | 144 | l_bn4 = batch_norm.BatchNormLayer( 145 | l_mp2, 146 | epsilon=epsilon, 147 | alpha=alpha) 148 | 149 | l_nl4 = lasagne.layers.NonlinearityLayer( 150 | l_bn4, 151 | nonlinearity = activation) 152 | 153 | # 512C3-512C3-P2 154 | l_cnn5 = lab.Conv2DLayer( 155 | l_nl4, 156 | num_filters=4*K, 157 | filter_size=(3, 3), 158 | pad=1, 159 | nonlinearity=lasagne.nonlinearities.identity, 160 | method = method) 161 | 162 | l_bn5 = batch_norm.BatchNormLayer( 163 | l_cnn5, 164 | epsilon=epsilon, 165 | alpha=alpha) 166 | 167 | l_nl5 = lasagne.layers.NonlinearityLayer( 168 | l_bn5, 169 | nonlinearity = activation) 170 | 171 | l_cnn6 = lab.Conv2DLayer( 172 | l_nl5, 173 | num_filters=4*K, 174 | filter_size=(3, 3), 175 | pad=1, 176 | nonlinearity=lasagne.nonlinearities.identity, 177 | method = method) 178 | 179 | l_mp3 = lasagne.layers.MaxPool2DLayer(l_cnn6, pool_size=(2, 2)) 180 | 181 | l_bn6 = batch_norm.BatchNormLayer( 182 | l_mp3, 183 | epsilon=epsilon, 184 | alpha=alpha) 185 | 186 | l_nl6 = lasagne.layers.NonlinearityLayer( 187 | l_bn6, 188 | nonlinearity = activation) 189 | 190 | # print(cnn.output_shape) 191 | 192 | # 1024FP-1024FP-10FP 193 | l_dn1 = lab.DenseLayer( 194 | l_nl6, 195 | nonlinearity=lasagne.nonlinearities.identity, 196 | num_units=1024, 197 | method = method) 198 | 199 | l_bn7 = batch_norm.BatchNormLayer( 200 | l_dn1, 201 | epsilon=epsilon, 202 | alpha=alpha) 203 | 204 | l_nl7 = lasagne.layers.NonlinearityLayer( 205 | l_bn7, 206 | nonlinearity = activation) 207 | 208 | l_dn2 = lab.DenseLayer( 209 | l_nl7, 210 | nonlinearity=lasagne.nonlinearities.identity, 211 | num_units=1024, 212 | method = method) 213 | 214 | l_bn8 = batch_norm.BatchNormLayer( 215 | l_dn2, 216 | epsilon=epsilon, 217 | alpha=alpha) 218 | 219 | l_nl8 = lasagne.layers.NonlinearityLayer( 220 | l_bn8, 221 | nonlinearity = activation) 222 | 223 | l_dn3 = lab.DenseLayer( 224 | l_nl8, 225 | nonlinearity=lasagne.nonlinearities.identity, 226 | num_units=10, 227 | method = method) 228 | 229 | l_out = batch_norm.BatchNormLayer( 230 | l_dn3, 231 | epsilon=epsilon, 232 | alpha=alpha) 233 | 234 | train_output = lasagne.layers.get_output(l_out, deterministic=False) 235 | 236 | 237 | # squared hinge loss 238 | loss = T.mean(T.sqr(T.maximum(0.,1.-target*train_output))) 239 | 240 | if method!="FPN": 241 | # W updates 242 | W = lasagne.layers.get_all_params(l_out, binary=True) 243 | W_grads = lab.compute_grads(loss,l_out) 244 | updates = optimizer.adam(loss_or_grads=W_grads, params=W, learning_rate=LR) 245 | updates = lab.clipping_scaling(updates,l_out) 246 | 247 | # other parameters updates 248 | params = lasagne.layers.get_all_params(l_out, trainable=True, binary=False) 249 | updates = OrderedDict(updates.items() + optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR).items()) 250 | 251 | ## update 2nd moment, can get from the adam optimizer also 252 | updates3 = OrderedDict() 253 | acc_tag = lasagne.layers.get_all_params(l_out, acc=True) 254 | idx = 0 255 | beta2 = 0.999 256 | for acc_tag_temp in acc_tag: 257 | updates3[acc_tag_temp]= acc_tag_temp*beta2 + W_grads[idx]*W_grads[idx]*(1-beta2) 258 | idx = idx+1 259 | 260 | updates = OrderedDict(updates.items() + updates3.items()) 261 | else: 262 | params = lasagne.layers.get_all_params(l_out, trainable=True) 263 | updates = optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR) 264 | 265 | test_output = lasagne.layers.get_output(l_out, deterministic=True) 266 | test_loss = T.mean(T.sqr(T.maximum(0.,1.-target*test_output))) 267 | test_err = T.mean(T.neq(T.argmax(test_output, axis=1), T.argmax(target, axis=1)),dtype=theano.config.floatX) 268 | 269 | # Compile a function performing a training step on a mini-batch (by giving the updates dictionary) 270 | # and returning the corresponding training loss: 271 | train_fn = theano.function([input, target, LR], loss, updates=updates) 272 | val_fn = theano.function([input, target], [test_loss, test_err]) 273 | 274 | 275 | ## load data 276 | print('Loading SVHN dataset') 277 | 278 | train_set = SVHN( 279 | which_set= 'splitted_train', 280 | # which_set= 'valid', 281 | path= "${SVHN_LOCAL_PATH}", 282 | axes= ['b', 'c', 0, 1]) 283 | 284 | valid_set = SVHN( 285 | which_set= 'valid', 286 | path= "${SVHN_LOCAL_PATH}", 287 | axes= ['b', 'c', 0, 1]) 288 | 289 | test_set = SVHN( 290 | which_set= 'test', 291 | path= "${SVHN_LOCAL_PATH}", 292 | axes= ['b', 'c', 0, 1]) 293 | 294 | # bc01 format 295 | # print train_set.X.shape 296 | train_set.X = np.reshape(train_set.X,(-1,3,32,32)) 297 | valid_set.X = np.reshape(valid_set.X,(-1,3,32,32)) 298 | test_set.X = np.reshape(test_set.X,(-1,3,32,32)) 299 | 300 | train_set.y = np.array(train_set.y).flatten() 301 | valid_set.y = np.array(valid_set.y).flatten() 302 | test_set.y = np.array(test_set.y).flatten() 303 | 304 | # Onehot the targets 305 | train_set.y = np.float32(np.eye(10)[train_set.y]) 306 | valid_set.y = np.float32(np.eye(10)[valid_set.y]) 307 | test_set.y = np.float32(np.eye(10)[test_set.y]) 308 | 309 | # for hinge loss 310 | train_set.y = 2* train_set.y - 1. 311 | valid_set.y = 2* valid_set.y - 1. 312 | test_set.y = 2* test_set.y - 1. 313 | 314 | 315 | print('Training...') 316 | 317 | # ipdb.set_trace() 318 | lab.train( 319 | name, method, 320 | train_fn,val_fn, 321 | batch_size, 322 | LR_start,LR_decay, 323 | num_epochs, 324 | train_set.X,train_set.y, 325 | valid_set.X,valid_set.y, 326 | test_set.X,test_set.y) 327 | 328 | 329 | if __name__ == "__main__": 330 | parser = ArgumentParser() 331 | parser.add_argument("--method", type=str, dest="method", 332 | default="LAB", help="Method used") 333 | parser.add_argument("--lr_start", type=float, dest="LR_start", 334 | default=0.001, help="Learning rate") 335 | parser.add_argument("--w", type=str, dest="Binarize_weight_only", 336 | default="w", help="true:only binzrize w, false: binarize w and a") 337 | args = parser.parse_args() 338 | 339 | main(**vars(args)) -------------------------------------------------------------------------------- /FNN/cifar10.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | 3 | import sys 4 | import os 5 | import time 6 | 7 | import numpy as np 8 | np.random.seed(1234) 9 | 10 | import theano 11 | import theano.tensor as T 12 | import ipdb 13 | 14 | import lasagne 15 | 16 | import cPickle as pickle 17 | import gzip 18 | 19 | import batch_norm 20 | import lab 21 | import optimizer 22 | 23 | from pylearn2.datasets.zca_dataset import ZCA_Dataset 24 | from pylearn2.utils import serial 25 | 26 | from collections import OrderedDict 27 | from argparse import ArgumentParser 28 | 29 | 30 | def main(method,LR_start,Binarize_weight_only): 31 | 32 | name = "cifar" 33 | print("dataset = "+str(name)) 34 | 35 | print("Binarize_weight_only="+str(Binarize_weight_only)) 36 | 37 | print("Method = "+str(method)) 38 | 39 | # alpha is the exponential moving average factor 40 | alpha = .1 41 | print("alpha = "+str(alpha)) 42 | epsilon = 1e-4 43 | print("epsilon = "+str(epsilon)) 44 | 45 | # Training parameters 46 | batch_size = 50 47 | print("batch_size = "+str(batch_size)) 48 | 49 | num_epochs = 200 50 | print("num_epochs = "+str(num_epochs)) 51 | 52 | print("LR_start = "+str(LR_start)) 53 | LR_decay = 0.5 54 | print("LR_decay="+str(LR_decay)) 55 | 56 | if Binarize_weight_only =="w": 57 | activation = lasagne.nonlinearities.rectify 58 | else: 59 | activation = lab.binary_tanh_unit 60 | print("activation = "+ str(activation)) 61 | 62 | 63 | train_set_size = 45000 64 | print("train_set_size = "+str(train_set_size)) 65 | 66 | print('Loading CIFAR-10 dataset...') 67 | 68 | preprocessor = serial.load("${PYLEARN2_DATA_PATH}/cifar10/pylearn2_gcn_whitened/preprocessor.pkl") 69 | train_set = ZCA_Dataset( 70 | preprocessed_dataset=serial.load("${PYLEARN2_DATA_PATH}/cifar10/pylearn2_gcn_whitened/train.pkl"), 71 | preprocessor = preprocessor, 72 | start=0, stop = train_set_size) 73 | valid_set = ZCA_Dataset( 74 | preprocessed_dataset= serial.load("${PYLEARN2_DATA_PATH}/cifar10/pylearn2_gcn_whitened/train.pkl"), 75 | preprocessor = preprocessor, 76 | start=45000, stop = 50000) 77 | test_set = ZCA_Dataset( 78 | preprocessed_dataset= serial.load("${PYLEARN2_DATA_PATH}/cifar10/pylearn2_gcn_whitened/test.pkl"), 79 | preprocessor = preprocessor) 80 | 81 | # bc01 format 82 | train_set.X = train_set.X.reshape(-1,3,32,32) 83 | valid_set.X = valid_set.X.reshape(-1,3,32,32) 84 | test_set.X = test_set.X.reshape(-1,3,32,32) 85 | 86 | # flatten targets 87 | train_set.y = np.hstack(train_set.y) 88 | valid_set.y = np.hstack(valid_set.y) 89 | test_set.y = np.hstack(test_set.y) 90 | 91 | 92 | # Onehot the targets 93 | train_set.y = np.float32(np.eye(10)[train_set.y]) 94 | valid_set.y = np.float32(np.eye(10)[valid_set.y]) 95 | test_set.y = np.float32(np.eye(10)[test_set.y]) 96 | 97 | # for hinge loss 98 | train_set.y = 2* train_set.y - 1. 99 | valid_set.y = 2* valid_set.y - 1. 100 | test_set.y = 2* test_set.y - 1. 101 | 102 | print('Building the CNN...') 103 | 104 | # Prepare Theano variables for inputs and targets 105 | input = T.tensor4('inputs') 106 | target = T.matrix('targets') 107 | LR = T.scalar('LR', dtype=theano.config.floatX) 108 | 109 | l_in = lasagne.layers.InputLayer( 110 | shape=(None, 3, 32, 32), 111 | input_var=input) 112 | 113 | # 128C3-128C3-P2 114 | l_cnn1 = lab.Conv2DLayer( 115 | l_in, 116 | num_filters=128, 117 | filter_size=(3, 3), 118 | pad=1, 119 | nonlinearity=lasagne.nonlinearities.identity, 120 | method = method) 121 | 122 | l_bn1 = batch_norm.BatchNormLayer( 123 | l_cnn1, 124 | epsilon=epsilon, 125 | alpha=alpha) 126 | 127 | l_nl1 = lasagne.layers.NonlinearityLayer( 128 | l_bn1, 129 | nonlinearity = activation) 130 | 131 | l_cnn2 = lab.Conv2DLayer( 132 | l_nl1, 133 | num_filters=128, 134 | filter_size=(3, 3), 135 | pad=1, 136 | nonlinearity=lasagne.nonlinearities.identity, 137 | method = method) 138 | 139 | l_mp1 = lasagne.layers.MaxPool2DLayer(l_cnn2, pool_size=(2, 2)) 140 | 141 | l_bn2 = batch_norm.BatchNormLayer( 142 | l_mp1, 143 | epsilon=epsilon, 144 | alpha=alpha) 145 | 146 | l_nl2 = lasagne.layers.NonlinearityLayer( 147 | l_bn2, 148 | nonlinearity = activation) 149 | # 256C3-256C3-P2 150 | l_cnn3 = lab.Conv2DLayer( 151 | l_nl2, 152 | num_filters=256, 153 | filter_size=(3, 3), 154 | pad=1, 155 | nonlinearity=lasagne.nonlinearities.identity, 156 | method = method) 157 | 158 | l_bn3 = batch_norm.BatchNormLayer( 159 | l_cnn3, 160 | epsilon=epsilon, 161 | alpha=alpha) 162 | 163 | l_nl3 = lasagne.layers.NonlinearityLayer( 164 | l_bn3, 165 | nonlinearity = activation) 166 | 167 | l_cnn4 = lab.Conv2DLayer( 168 | l_nl3, 169 | num_filters=256, 170 | filter_size=(3, 3), 171 | pad=1, 172 | nonlinearity=lasagne.nonlinearities.identity, 173 | method = method) 174 | 175 | l_mp2 = lasagne.layers.MaxPool2DLayer(l_cnn4, pool_size=(2, 2)) 176 | 177 | l_bn4 = batch_norm.BatchNormLayer( 178 | l_mp2, 179 | epsilon=epsilon, 180 | alpha=alpha) 181 | 182 | l_nl4 = lasagne.layers.NonlinearityLayer( 183 | l_bn4, 184 | nonlinearity = activation) 185 | 186 | # 512C3-512C3-P2 187 | l_cnn5 = lab.Conv2DLayer( 188 | l_nl4, 189 | num_filters=512, 190 | filter_size=(3, 3), 191 | pad=1, 192 | nonlinearity=lasagne.nonlinearities.identity, 193 | method = method) 194 | 195 | l_bn5 = batch_norm.BatchNormLayer( 196 | l_cnn5, 197 | epsilon=epsilon, 198 | alpha=alpha) 199 | 200 | l_nl5 = lasagne.layers.NonlinearityLayer( 201 | l_bn5, 202 | nonlinearity = activation) 203 | 204 | l_cnn6 = lab.Conv2DLayer( 205 | l_nl5, 206 | num_filters=512, 207 | filter_size=(3, 3), 208 | pad=1, 209 | nonlinearity=lasagne.nonlinearities.identity, 210 | method = method) 211 | 212 | l_mp3 = lasagne.layers.MaxPool2DLayer(l_cnn6, pool_size=(2, 2)) 213 | 214 | l_bn6 = batch_norm.BatchNormLayer( 215 | l_mp3, 216 | epsilon=epsilon, 217 | alpha=alpha) 218 | 219 | l_nl6 = lasagne.layers.NonlinearityLayer( 220 | l_bn6, 221 | nonlinearity = activation) 222 | 223 | # print(cnn.output_shape) 224 | 225 | # 1024FP-1024FP-10FP 226 | l_dn1 = lab.DenseLayer( 227 | l_nl6, 228 | nonlinearity=lasagne.nonlinearities.identity, 229 | num_units=1024, 230 | method = method) 231 | 232 | l_bn7 = batch_norm.BatchNormLayer( 233 | l_dn1, 234 | epsilon=epsilon, 235 | alpha=alpha) 236 | 237 | l_nl7 = lasagne.layers.NonlinearityLayer( 238 | l_bn7, 239 | nonlinearity = activation) 240 | 241 | l_dn2 = lab.DenseLayer( 242 | l_nl7, 243 | nonlinearity=lasagne.nonlinearities.identity, 244 | num_units=1024, 245 | method = method) 246 | 247 | l_bn8 = batch_norm.BatchNormLayer( 248 | l_dn2, 249 | epsilon=epsilon, 250 | alpha=alpha) 251 | 252 | l_nl8 = lasagne.layers.NonlinearityLayer( 253 | l_bn8, 254 | nonlinearity = activation) 255 | 256 | l_dn3 = lab.DenseLayer( 257 | l_nl8, 258 | nonlinearity=lasagne.nonlinearities.identity, 259 | num_units=10, 260 | method = method) 261 | 262 | l_out = batch_norm.BatchNormLayer( 263 | l_dn3, 264 | epsilon=epsilon, 265 | alpha=alpha) 266 | 267 | train_output = lasagne.layers.get_output(l_out, deterministic=False) 268 | 269 | # squared hinge loss 270 | loss = T.mean(T.sqr(T.maximum(0.,1.-target*train_output))) 271 | 272 | if method!="FPN": 273 | # W updates 274 | W = lasagne.layers.get_all_params(l_out, binary=True) 275 | W_grads = lab.compute_grads(loss,l_out) 276 | updates = optimizer.adam(loss_or_grads=W_grads, params=W, learning_rate=LR) 277 | updates = lab.clipping_scaling(updates,l_out) 278 | 279 | # other parameters updates 280 | params = lasagne.layers.get_all_params(l_out, trainable=True, binary=False) 281 | updates = OrderedDict(updates.items() + optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR).items()) 282 | 283 | ## update 2nd moment, can get from the adam optimizer also 284 | updates3 = OrderedDict() 285 | acc_tag = lasagne.layers.get_all_params(l_out, acc=True) 286 | idx = 0 287 | beta2 = 0.999 288 | for acc_tag_temp in acc_tag: 289 | updates3[acc_tag_temp]= acc_tag_temp*beta2 + W_grads[idx]*W_grads[idx]*(1-beta2) 290 | idx = idx+1 291 | 292 | updates = OrderedDict(updates.items() + updates3.items()) 293 | else: 294 | params = lasagne.layers.get_all_params(l_out, trainable=True) 295 | updates = optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR) 296 | 297 | test_output = lasagne.layers.get_output(l_out, deterministic=True) 298 | test_loss = T.mean(T.sqr(T.maximum(0.,1.-target*test_output))) 299 | test_err = T.mean(T.neq(T.argmax(test_output, axis=1), T.argmax(target, axis=1)),dtype=theano.config.floatX) 300 | 301 | # Compile a function performing a training step on a mini-batch (by giving the updates dictionary) 302 | # and returning the corresponding training loss: 303 | train_fn = theano.function([input, target, LR], loss, updates=updates) 304 | val_fn = theano.function([input, target], [test_loss, test_err]) 305 | 306 | print('Training...') 307 | 308 | lab.train( 309 | name, method, 310 | train_fn,val_fn, 311 | batch_size, 312 | LR_start,LR_decay, 313 | num_epochs, 314 | train_set.X,train_set.y, 315 | valid_set.X,valid_set.y, 316 | test_set.X,test_set.y) 317 | 318 | 319 | if __name__ == "__main__": 320 | parser = ArgumentParser() 321 | parser.add_argument("--method", type=str, dest="method", 322 | default="LAB", help="Method used") 323 | parser.add_argument("--lr_start", type=float, dest="LR_start", 324 | default=0.02, help="Learning rate") 325 | parser.add_argument("--w", type=str, dest="Binarize_weight_only", 326 | default="w", help="true:only binzrize w, false: binarize w and a") 327 | args = parser.parse_args() 328 | 329 | main(**vars(args)) -------------------------------------------------------------------------------- /RNN/lab.py: -------------------------------------------------------------------------------- 1 | import time 2 | 3 | from collections import OrderedDict 4 | 5 | import numpy as np 6 | np.random.seed(1234) 7 | import ipdb 8 | 9 | import theano 10 | import theano.tensor as T 11 | 12 | import lasagne 13 | from theano.scalar.basic import UnaryScalarOp, same_out_nocomplex 14 | from theano.tensor.elemwise import Elemwise 15 | from theano.sandbox.rng_mrg import MRG_RandomStreams as RandomStreams 16 | class Round3(UnaryScalarOp): 17 | 18 | def c_code(self, node, name, (x,), (z,), sub): 19 | return "%(z)s = round(%(x)s);" % locals() 20 | 21 | def grad(self, inputs, gout): 22 | (gz,) = gout 23 | return gz, 24 | 25 | round3_scalar = Round3(same_out_nocomplex, name='round3') 26 | round3 = Elemwise(round3_scalar) 27 | 28 | def hard_sigmoid(x): 29 | return T.clip((x+1.)/2.,0,1) 30 | 31 | # The neurons' activations binarization function 32 | # It behaves like the sign function during forward propagation 33 | # And like: 34 | # hard_tanh(x) = 2*hard_sigmoid(x)-1 35 | # during back propagation 36 | def binary_tanh_unit(x): 37 | return 2.*round3(hard_sigmoid(x))-1. 38 | 39 | def binary_sigmoid_unit(x): 40 | return round3(hard_sigmoid(x)) 41 | 42 | # The binarization function 43 | def binarization(W,Wacc,method): 44 | 45 | if method == "FPN": 46 | Wb = W 47 | 48 | elif method == "LAB": 49 | L = (T.sqrt(Wacc) + 1e-8) 50 | Wb = hard_sigmoid(W) 51 | Wb = round3(Wb) 52 | Wb = T.cast(T.switch(Wb,1.,-1.), theano.config.floatX) 53 | 54 | alpha = (T.abs_(L*W).sum()/L.sum()).astype('float32') 55 | Wb = alpha*Wb 56 | 57 | return Wb 58 | 59 | ### binarize LSTM 60 | class LSTMLayer(lasagne.layers.LSTMLayer): 61 | 62 | def __init__(self, incoming, num_units, method, **kwargs): 63 | 64 | self.method = method 65 | g_init = 0.08 66 | if self.method!="FPN": 67 | super(LSTMLayer, self).__init__(incoming, num_units, 68 | ingate=lasagne.layers.Gate(W_in=lasagne.init.Uniform((-g_init, g_init)), 69 | W_hid=lasagne.init.Uniform((-g_init, g_init)), 70 | W_cell=lasagne.init.Uniform((-g_init, g_init))), 71 | forgetgate=lasagne.layers.Gate(W_in=lasagne.init.Uniform((-g_init, g_init)), 72 | W_hid=lasagne.init.Uniform((-g_init, g_init)), 73 | W_cell=lasagne.init.Uniform((-g_init, g_init)), 74 | b=lasagne.init.Constant(1.)), 75 | cell=lasagne.layers.Gate(W_in=lasagne.init.Uniform((-g_init, g_init)), 76 | W_hid=lasagne.init.Uniform((-g_init, g_init)), 77 | W_cell=None, nonlinearity=lasagne.nonlinearities.tanh), 78 | outgate=lasagne.layers.Gate(W_in=lasagne.init.Uniform((-g_init, g_init)), 79 | W_hid=lasagne.init.Uniform((-g_init, g_init)), 80 | W_cell=lasagne.init.Uniform((-g_init, g_init))), 81 | **kwargs) 82 | # super(LSTMLayer, self).__init__(incoming, num_units, 83 | # **kwargs) 84 | # add the binary tag to weights 85 | self.params[self.W_in_to_ingate]=set(['binary']) 86 | self.params[self.W_hid_to_ingate]=set(['binary']) 87 | 88 | self.params[self.W_in_to_forgetgate]=set(['binary']) 89 | self.params[self.W_hid_to_forgetgate]=set(['binary']) 90 | 91 | self.params[self.W_in_to_cell]=set(['binary']) 92 | self.params[self.W_hid_to_cell]=set(['binary']) 93 | 94 | self.params[self.W_in_to_outgate]=set(['binary']) 95 | self.params[self.W_hid_to_outgate]=set(['binary']) 96 | 97 | else: 98 | # super(LSTMLayer, self).__init__(incoming, num_units, **kwargs) 99 | super(LSTMLayer, self).__init__(incoming, num_units, 100 | ingate=lasagne.layers.Gate(W_in=lasagne.init.Uniform((-g_init, g_init)), 101 | W_hid=lasagne.init.Uniform((-g_init, g_init)), 102 | W_cell=lasagne.init.Uniform((-g_init, g_init))), 103 | forgetgate=lasagne.layers.Gate(W_in=lasagne.init.Uniform((-g_init, g_init)), 104 | W_hid=lasagne.init.Uniform((-g_init, g_init)), 105 | W_cell=lasagne.init.Uniform((-g_init, g_init)), 106 | b=lasagne.init.Constant(1.)), 107 | cell=lasagne.layers.Gate(W_in=lasagne.init.Uniform((-g_init, g_init)), 108 | W_hid=lasagne.init.Uniform((-g_init, g_init)), 109 | W_cell=None, nonlinearity=lasagne.nonlinearities.tanh), 110 | outgate=lasagne.layers.Gate(W_in=lasagne.init.Uniform((-g_init, g_init)), 111 | W_hid=lasagne.init.Uniform((-g_init, g_init)), 112 | W_cell=lasagne.init.Uniform((-g_init, g_init))), 113 | **kwargs) 114 | 115 | 116 | # initialize 2nd moment 117 | self.acc_W_in_to_ingate = theano.shared(np.zeros((self.W_in_to_ingate.get_value(borrow=True)).shape, dtype='float32')) 118 | self.acc_W_hid_to_ingate = theano.shared(np.zeros((self.W_hid_to_ingate.get_value(borrow=True)).shape, dtype='float32')) 119 | self.acc_W_in_to_forgetgate = theano.shared(np.zeros((self.W_in_to_forgetgate.get_value(borrow=True)).shape, dtype='float32')) 120 | self.acc_W_hid_to_forgetgate = theano.shared(np.zeros((self.W_hid_to_forgetgate.get_value(borrow=True)).shape, dtype='float32')) 121 | self.acc_W_in_to_cell = theano.shared(np.zeros((self.W_in_to_cell.get_value(borrow=True)).shape, dtype='float32')) 122 | self.acc_W_hid_to_cell = theano.shared(np.zeros((self.W_hid_to_cell.get_value(borrow=True)).shape, dtype='float32')) 123 | self.acc_W_in_to_outgate = theano.shared(np.zeros((self.W_in_to_outgate.get_value(borrow=True)).shape, dtype='float32')) 124 | self.acc_W_hid_to_outgate = theano.shared(np.zeros((self.W_hid_to_outgate.get_value(borrow=True)).shape, dtype='float32')) 125 | 126 | self.params[self.acc_W_in_to_ingate]=set(['acc']) 127 | self.params[self.acc_W_hid_to_ingate]=set(['acc']) 128 | 129 | self.params[self.acc_W_in_to_forgetgate]=set(['acc']) 130 | self.params[self.acc_W_hid_to_forgetgate]=set(['acc']) 131 | 132 | self.params[self.acc_W_in_to_cell]=set(['acc']) 133 | self.params[self.acc_W_hid_to_cell]=set(['acc']) 134 | 135 | self.params[self.acc_W_in_to_outgate]=set(['acc']) 136 | self.params[self.acc_W_hid_to_outgate]=set(['acc']) 137 | 138 | def get_output_for(self, input, deterministic=False, **kwargs): 139 | # self.Wb = binarization(self.W,self.H,self.binary,deterministic,self.stochastic,self._srng) 140 | self.bW_in_to_ingate = binarization(self.W_in_to_ingate, self.acc_W_in_to_ingate, self.method) 141 | self.bW_hid_to_ingate = binarization(self.W_hid_to_ingate, self.acc_W_hid_to_ingate, self.method) 142 | 143 | self.bW_in_to_forgetgate = binarization(self.W_in_to_forgetgate, self.acc_W_in_to_forgetgate, self.method) 144 | self.bW_hid_to_forgetgate = binarization(self.W_hid_to_forgetgate, self.acc_W_hid_to_forgetgate, self.method) 145 | 146 | self.bW_in_to_cell = binarization(self.W_in_to_cell, self.acc_W_in_to_cell, self.method) 147 | self.bW_hid_to_cell = binarization(self.W_hid_to_cell, self.acc_W_hid_to_cell, self.method) 148 | 149 | self.bW_in_to_outgate = binarization(self.W_in_to_outgate, self.acc_W_in_to_outgate, self.method) 150 | self.bW_hid_to_outgate = binarization(self.W_hid_to_outgate, self.acc_W_hid_to_outgate, self.method) 151 | 152 | # Wr = self.W 153 | rW_in_to_ingate = self.W_in_to_ingate 154 | rW_hid_to_ingate = self.W_hid_to_ingate 155 | 156 | rW_in_to_forgetgate = self.W_in_to_forgetgate 157 | rW_hid_to_forgetgate = self.W_hid_to_forgetgate 158 | 159 | rW_in_to_cell = self.W_in_to_cell 160 | rW_hid_to_cell = self.W_hid_to_cell 161 | 162 | rW_in_to_outgate = self.W_in_to_outgate 163 | rW_hid_to_outgate = self.W_hid_to_outgate 164 | 165 | #self.W = bW 166 | self.W_in_to_ingate = self.bW_in_to_ingate 167 | self.W_hid_to_ingate = self.bW_hid_to_ingate 168 | 169 | self.W_in_to_forgetgate = self.bW_in_to_forgetgate 170 | self.W_hid_to_forgetgate = self.bW_hid_to_forgetgate 171 | 172 | self.W_in_to_cell = self.bW_in_to_cell 173 | self.W_hid_to_cell = self.bW_hid_to_cell 174 | 175 | self.W_in_to_outgate = self.bW_in_to_outgate 176 | self.W_hid_to_outgate = self.bW_hid_to_outgate 177 | 178 | rvalue = super(LSTMLayer, self).get_output_for(input, **kwargs) 179 | 180 | self.W_in_to_ingate = rW_in_to_ingate 181 | self.W_hid_to_ingate = rW_hid_to_ingate 182 | 183 | self.W_in_to_forgetgate = rW_in_to_forgetgate 184 | self.W_hid_to_forgetgate = rW_hid_to_forgetgate 185 | 186 | self.W_in_to_cell = rW_in_to_cell 187 | self.W_hid_to_cell = rW_hid_to_cell 188 | 189 | self.W_in_to_outgate = rW_in_to_outgate 190 | self.W_hid_to_outgate = rW_hid_to_outgate 191 | 192 | return rvalue 193 | 194 | 195 | 196 | def compute_grads(loss,network): 197 | 198 | layers = lasagne.layers.get_all_layers(network) 199 | grads = [] 200 | 201 | for layer in layers: 202 | 203 | params = layer.get_params(binary=True) 204 | if params: 205 | # print(params[0].name) 206 | grads.append(theano.grad(loss, wrt=layer.bW_in_to_ingate)) 207 | grads.append(theano.grad(loss, wrt=layer.bW_hid_to_ingate)) 208 | 209 | grads.append(theano.grad(loss, wrt=layer.bW_in_to_forgetgate)) 210 | grads.append(theano.grad(loss, wrt=layer.bW_hid_to_forgetgate)) 211 | 212 | grads.append(theano.grad(loss, wrt=layer.bW_in_to_cell)) 213 | grads.append(theano.grad(loss, wrt=layer.bW_hid_to_cell)) 214 | 215 | grads.append(theano.grad(loss, wrt=layer.bW_in_to_outgate)) 216 | grads.append(theano.grad(loss, wrt=layer.bW_hid_to_outgate)) 217 | 218 | return grads 219 | 220 | 221 | def clipping_scaling(updates,network): 222 | 223 | layers = lasagne.layers.get_all_layers(network) 224 | updates = OrderedDict(updates) 225 | 226 | for layer in layers: 227 | params = layer.get_params(binary=True) 228 | for param in params: 229 | updates[param] = T.clip(updates[param], -1.,1.) 230 | return updates 231 | 232 | # Given a dataset and a model, this function trains the model on the dataset for several epochs 233 | def train(name,method,train_fn,val_fn, 234 | batch_size, 235 | SEQ_LENGTH, 236 | N_HIDDEN, 237 | LR_start,LR_decay, 238 | num_epochs, 239 | X_train, 240 | X_val, 241 | X_test): 242 | 243 | def gen_data(pp, batch_size,SEQ_LENGTH, data, return_target=True): 244 | 245 | x = np.zeros((batch_size,SEQ_LENGTH,vocab_size)) ###### 128*100*85 246 | y = np.zeros((batch_size, SEQ_LENGTH)) 247 | 248 | for n in range(batch_size): 249 | # ptr = n 250 | for i in range(SEQ_LENGTH): 251 | x[n,i,char_to_ix[data[pp[n]*SEQ_LENGTH+i]]] = 1. 252 | y[n,i] = char_to_ix[data[pp[n]*SEQ_LENGTH+i+1]] 253 | return x, np.array(y,dtype='int32') 254 | 255 | in_text = X_train+X_val+X_test 256 | chars = list(set(in_text)) 257 | data_size, vocab_size = len(in_text), len(chars) 258 | char_to_ix = { ch:i for i,ch in enumerate(chars) } 259 | ix_to_char = { i:ch for i,ch in enumerate(chars) } 260 | 261 | def train_epoch(X,LR): 262 | 263 | loss = 0 264 | batches = len(X)/batch_size/SEQ_LENGTH 265 | # shuffle 266 | num_seq = len(X)/SEQ_LENGTH 267 | shuffled_ind = range(num_seq) 268 | 269 | np.random.shuffle(shuffled_ind) 270 | for i in range(batches): 271 | # shuffle 272 | tmp_ind = shuffled_ind[i*batch_size:(i+1)*batch_size] 273 | xx,yy = gen_data(tmp_ind,batch_size,SEQ_LENGTH, X) 274 | new_loss, Wg = train_fn(xx,yy,LR) 275 | loss+=new_loss 276 | 277 | loss=loss/batches 278 | 279 | return loss 280 | 281 | # This function tests the model a full epoch (on the whole dataset) 282 | def val_epoch(X): 283 | 284 | # err = 0 285 | loss = 0 286 | batches = len(X)/batch_size/SEQ_LENGTH 287 | 288 | num_seq = len(X)/SEQ_LENGTH 289 | ind = range(num_seq) 290 | for i in range(batches): 291 | tmp_ind = ind[i*batch_size:(i+1)*batch_size] 292 | xx, yy = gen_data(tmp_ind, batch_size, SEQ_LENGTH, X) 293 | new_loss = val_fn(xx,yy) 294 | loss += new_loss 295 | 296 | loss = loss/batches 297 | 298 | return loss 299 | 300 | best_val_loss=100 301 | best_epoch = 1 302 | LR = LR_start 303 | # hello= False 304 | # We iterate over epochs: 305 | for epoch in range(1,num_epochs+1): 306 | start_time = time.time() 307 | train_loss = train_epoch(X_train, LR) 308 | # try_it_out() 309 | 310 | val_loss = val_epoch(X_val) 311 | 312 | # test if validation error went down 313 | if val_loss <= best_val_loss: 314 | 315 | best_val_loss = val_loss 316 | best_epoch = epoch+1 317 | 318 | test_loss = val_epoch(X_test) 319 | 320 | epoch_duration = time.time() - start_time 321 | # Then we print the results for this epoch: 322 | print(" Epoch "+str(epoch)+" of "+str(num_epochs)+" took "+str(epoch_duration)+"s") 323 | print(" LR: "+str(LR)) 324 | print(" training loss: "+str(train_loss)) 325 | print(" validation loss: "+str(val_loss)) 326 | print(" best epoch: "+str(best_epoch)) 327 | print(" test loss: "+str(test_loss)) 328 | 329 | with open("{0}_seq{1}_lr{2}_hid{3}_{4}.txt".format(name, SEQ_LENGTH, LR_start, N_HIDDEN, method), "a") as myfile: 330 | myfile.write("{0} {1:.3f} {2:.3f} {3:.3f} {4:.3f}\n".format(epoch, train_loss, val_loss, test_loss, epoch_duration)) 331 | 332 | # learning rate update scheme 333 | if epoch>10: 334 | LR *= LR_decay 335 | 336 | -------------------------------------------------------------------------------- /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 | {one line to give the program's name and a brief idea of what it does.} 635 | Copyright (C) {year} {name of author} 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 | {project} Copyright (C) {year} {fullname} 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 | --------------------------------------------------------------------------------