Python tensor.isinf方法代码示例

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def sgdmgc(cost, params, lr=1.0, alpha=0.1, max_magnitude=5.0, infDecay=0.1):
    """SGD with momentum and gradient clipping"""
    grads = T.grad(cost=cost, wrt=params)
    updates = []

    norm = norm_gs(params, grads)
    sqrtnorm = T.sqrt(norm)
    not_finite = T.or_(T.isnan(sqrtnorm), T.isinf(sqrtnorm))
    adj_norm_gs = T.switch(T.ge(sqrtnorm, max_magnitude), max_magnitude / sqrtnorm, 1.)

    for p, g in zip(params, grads):
        v = shared(p.get_value() * 0.)
        g = T.switch(not_finite, infDecay * p, g * adj_norm_gs)
        v_new = v * (1.0 - alpha) - alpha * lr * g
        updates.append((v, v_new))
        updates.append((p, p + v_new ))
    
    return updates, norm 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def gradient_descent(self, params, gparams, learning_rate):
        """Momentum GD with gradient clipping."""
        #grad = T.grad(loss, self.params)
        self.momentum_velocity_ = [0.] * len(gparams)
        grad_norm = T.sqrt(sum(map(lambda x: T.sqr(x).sum(), gparams)))
        updates = OrderedDict()
        not_finite = T.or_(T.isnan(grad_norm), T.isinf(grad_norm))
        scaling_den = T.maximum(5.0, grad_norm)
        for n, (param, grad) in enumerate(zip(params, gparams)):
            grad = T.switch(not_finite, 0.1 * param,
                            grad * (5.0 / scaling_den))
            velocity = self.momentum_velocity_[n]
            update_step = self.momentum * velocity - learning_rate * grad
            self.momentum_velocity_[n] = update_step
            updates[param] = param + update_step
        return updates
        
    ##################### calculate total loss #######################    
    # only loss D 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def gradient_descent(self, loss):
        """Momentum GD with gradient clipping."""
        grad = T.grad(loss, self.params)
        self.momentum_velocity_ = [0.] * len(grad)
        grad_norm = T.sqrt(sum(map(lambda x: T.sqr(x).sum(), grad)))
        updates = OrderedDict()
        not_finite = T.or_(T.isnan(grad_norm), T.isinf(grad_norm))
        scaling_den = T.maximum(5.0, grad_norm)
        for n, (param, grad) in enumerate(zip(self.params, grad)):
            grad = T.switch(not_finite, 0.1 * param,
                            grad * (5.0 / scaling_den))
            velocity = self.momentum_velocity_[n]
            update_step = self.momentum * velocity - self.learning_rate * grad
            self.momentum_velocity_[n] = update_step
            updates[param] = param + update_step
        return updates 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def isvalid(x):
    return T.all(T.logical_not(T.logical_or(T.isnan(x), T.isinf(x)))) 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def gradient_clipping(grads, tparams, clip_c=10):
    g2 = 0.
    for g in grads:
        g2 += (g**2).sum()

    g2 = tensor.sqrt(g2)
    not_finite = tensor.or_(tensor.isnan(g2), tensor.isinf(g2))
    new_grads = []

    for p, g in zip(tparams.values(), grads):
        new_grads.append(tensor.switch(g2 > clip_c,
                                       g * (clip_c / g2),
                                       g))

    return new_grads, not_finite, tensor.lt(clip_c, g2) 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def compute_updates(self, training_cost, params):
        updates = []
        grads = T.grad(training_cost, params)
        grads = OrderedDict(zip(params, grads))
        # Clip stuff
        c = numpy.float32(self.cutoff)
        clip_grads = []

        norm_gs = T.sqrt(sum(T.sum(g ** 2) for p, g in grads.items()))
        normalization = T.switch(T.ge(norm_gs, c), c / norm_gs, np.float32(1.))
        notfinite = T.or_(T.isnan(norm_gs), T.isinf(norm_gs))
        for p, g in grads.items():
            clip_grads.append((p, T.switch(notfinite, numpy.float32(.1) * p, g * normalization)))

        grads = OrderedDict(clip_grads)

        if self.updater == 'adagrad':
            updates = Adagrad(grads, self.lr)
        elif self.updater == 'sgd':
            raise Exception("Sgd not implemented!")
        elif self.updater == 'adadelta':
            updates = Adadelta(grads)
        elif self.updater == 'rmsprop':
            updates = RMSProp(grads, self.lr)
        elif self.updater == 'adam':
            updates = Adam(grads)
        else:
            raise Exception("Updater not understood!")
        return updates 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def sgdgc(cost, params, lr=1.0, max_magnitude=5.0, infDecay=0.1):
    """SGD with gradient clipping"""
    grads = T.grad(cost=cost, wrt=params)
    updates = []

    norm = norm_gs(params, grads)
    sqrtnorm = T.sqrt(norm)
    #not_finite = T.or_(T.isnan(sqrtnorm), T.isinf(sqrtnorm))
    adj_norm_gs = T.switch(T.ge(sqrtnorm, max_magnitude), max_magnitude / sqrtnorm, 1.)

    for p, g in zip(params, grads):
        #g = T.switch(not_finite, infDecay * p, g * adj_norm_gs)
        updates.append((p, p - lr * g * adj_norm_gs))  
    
    return updates, norm 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def adamgc_(cost, params, lr=0.0002, b1=0.1, b2=0.01, e=1e-8, max_magnitude=5.0, infDecay=0.1):
    updates = []
    grads = T.grad(cost, params)
    
    norm = norm_gs(params, grads)
    sqrtnorm = T.sqrt(norm)
    not_finite = T.or_(T.isnan(sqrtnorm), T.isinf(sqrtnorm))
    adj_norm_gs = T.switch(T.ge(sqrtnorm, max_magnitude), max_magnitude / sqrtnorm, 1.)

    i = shared(floatX(0.))
    i_t = i + 1.
    fix1 = 1. - (1. - b1)**i_t
    fix2 = 1. - (1. - b2)**i_t
    lr_t = lr * (T.sqrt(fix2) / fix1)
    for p, g in zip(params, grads):
        g = T.switch(not_finite, infDecay * p, g * adj_norm_gs)
        m = shared(p.get_value() * 0.)
        v = shared(p.get_value() * 0.)
        m_t = (b1 * g) + ((1. - b1) * m) 
        v_t = (b2 * T.sqr(g)) + ((1. - b2) * v)
        g_t = m_t / (T.sqrt(v_t) + e)
        p_t = p - (lr_t * g_t)

        #e_t = shared(p.get_value() * 0.)
        #de_t = (srnd.normal(p.shape, std = 0.05, dtype=theano.config.floatX)*p_t - e_t)*0.05  #*p_t
        #p_t = p_t + de_t
        #updates.append((e_t, e_t + de_t))

        updates.append((m, m_t))
        updates.append((v, v_t))
        updates.append((p, p_t))
    updates.append((i, i_t))
    return updates, norm 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def adamgc(cost, params, lr=0.0002, b1=0.1, b2=0.001, e=1e-8, max_magnitude=5.0, infDecay=0.1):
    updates = []
    grads = T.grad(cost, params)
    
    norm = norm_gs(params, grads)
    sqrtnorm = T.sqrt(norm)
    not_finite = T.or_(T.isnan(sqrtnorm), T.isinf(sqrtnorm))
    adj_norm_gs = T.switch(T.ge(sqrtnorm, max_magnitude), max_magnitude / sqrtnorm, 1.)

    i = shared(floatX(0.))
    i_t = i + 1.
    fix1 = 1. - (1. - b1)**i_t
    fix2 = 1. - (1. - b2)**i_t
    lr_t = lr * (T.sqrt(fix2) / fix1)
    for p, g in zip(params, grads):
        g = T.switch(not_finite, infDecay * p, g * adj_norm_gs)
        m = shared(p.get_value() * 0.)
        v = shared(p.get_value() * 0.)
        m_t = (b1 * g) + ((1. - b1) * m) 
        v_t = (b2 * T.sqr(g)) + ((1. - b2) * v)
        g_t = m_t / (T.sqrt(v_t) + e)
        p_t = p - (lr_t * g_t)
        updates.append((m, m_t))
        updates.append((v, v_t))
        updates.append((p, p_t))
    updates.append((i, i_t))
    return updates, norm

#-------------------------------------------------------------------------------------------------- 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def compute_step(self, parameter, previous_step):
        step_sum = tensor.sum(previous_step)
        not_finite = (tensor.isnan(step_sum) +
                      tensor.isinf(step_sum))
        step = tensor.switch(
            not_finite > 0, (1 - self.scaler) * parameter, previous_step)
        return step, [] 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def updates(self, params, grads, learning_rate, momentum, rescale=5.):
        grad_norm = tensor.sqrt(sum(map(lambda x: tensor.sqr(x).sum(), grads)))
        not_finite = tensor.or_(tensor.isnan(grad_norm),
                                tensor.isinf(grad_norm))
        grad_norm = tensor.sqrt(grad_norm)
        scaling_num = rescale
        scaling_den = tensor.maximum(rescale, grad_norm)
        # Magic constants
        combination_coeff = 0.9
        minimum_grad = 1E-4
        updates = []
        for n, (param, grad) in enumerate(zip(params, grads)):
            grad = tensor.switch(not_finite, 0.1 * param,
                                 grad * (scaling_num / scaling_den))
            old_square = self.running_square_[n]
            new_square = combination_coeff * old_square + (
                1. - combination_coeff) * tensor.sqr(grad)
            old_avg = self.running_avg_[n]
            new_avg = combination_coeff * old_avg + (
                1. - combination_coeff) * grad
            rms_grad = tensor.sqrt(new_square - new_avg ** 2)
            rms_grad = tensor.maximum(rms_grad, minimum_grad)
            memory = self.memory_[n]
            update = momentum * memory - learning_rate * grad / rms_grad
            update2 = momentum * momentum * memory - (
                1 + momentum) * learning_rate * grad / rms_grad
            updates.append((old_square, new_square))
            updates.append((old_avg, new_avg))
            updates.append((memory, update))
            updates.append((param, param + update2))
        return updates 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def remove_nans(x):
    return T.switch(T.isnan(x) + T.isinf(x), 0, x) 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def compute_activation(scaled_time_vector):
        compare = tt.isinf(scaled_time_vector)
        subvector = scaled_time_vector[(1-compare).nonzero()]
        activation_from_time = tt.log(subvector.sum())
        return activation_from_time 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def compute_updates(self, training_cost, params):
        updates = []
         
        grads = T.grad(training_cost, params)
        grads = OrderedDict(zip(params, grads))

        # Gradient clipping
        c = numpy.float32(self.cutoff)
        clip_grads = []
        
        norm_gs = T.sqrt(sum(T.sum(g ** 2) for p, g in grads.items()))
        normalization = T.switch(T.ge(norm_gs, c), c / norm_gs, np.float32(1.))
        notfinite = T.or_(T.isnan(norm_gs), T.isinf(norm_gs))
         
        for p, g in grads.items():
            clip_grads.append((p, T.switch(notfinite, numpy.float32(.1) * p, g * normalization)))
        
        grads = OrderedDict(clip_grads)

        if self.W_emb in grads:
            if self.initialize_from_pretrained_word_embeddings and self.fix_pretrained_word_embeddings:
                assert not self.fix_encoder_parameters
                # Keep pretrained word embeddings fixed
                logger.debug("Will use mask to fix pretrained word embeddings")
                grads[self.W_emb] = grads[self.W_emb] * self.W_emb_pretrained_mask
            elif self.fix_encoder_parameters:
                # If 'fix_encoder_parameters' is on, the word embeddings will be excluded from parameter training set
                logger.debug("Will fix word embeddings to initial embeddings or embeddings from resumed model")
            else:
                logger.debug("Will train all word embeddings")

        optimizer_variables = []
        if self.updater == 'adagrad':
            updates = Adagrad(grads, self.lr)
        elif self.updater == 'sgd':
            raise Exception("Sgd not implemented!")
        elif self.updater == 'adadelta':
            updates = Adadelta(grads)
        elif self.updater == 'rmsprop':
            updates = RMSProp(grads, self.lr)
        elif self.updater == 'adam':
            updates, optimizer_variables = Adam(grads, self.lr)
        else:
            raise Exception("Updater not understood!") 

        return updates, optimizer_variables
  
    # Batch training function. 

# 需要导入模块: from theano import tensor [as 别名]
# 或者: from theano.tensor import isinf [as 别名]
def rmsprop(cost, params, learning_rate, momentum=0.5, rescale=5.):
    
    grads = T.grad(cost=cost, wrt=params)
    
    running_square_ = [theano.shared(np.zeros_like(p.get_value(),dtype=p.dtype), broadcastable=p.broadcastable)
                      for p in params]
    running_avg_ = [theano.shared(np.zeros_like(p.get_value(),dtype=p.dtype), broadcastable=p.broadcastable)
                   for p in params]
    memory_ = [theano.shared(np.zeros_like(p.get_value(),dtype=p.dtype), broadcastable=p.broadcastable)
                       for p in params]
    
    grad_norm = T.sqrt(sum(map(lambda x: T.sqr(x).sum(), grads)))
    not_finite = T.or_(T.isnan(grad_norm), T.isinf(grad_norm))
    grad_norm = T.sqrt(grad_norm)
    scaling_num = rescale
    scaling_den = T.maximum(rescale, grad_norm)
    # Magic constants
    combination_coeff = 0.9
    minimum_grad = 1E-4
    updates = []
    for n, (param, grad) in enumerate(zip(params, grads)):
       grad = T.switch(not_finite, 0.1 * param,
                       grad * (scaling_num / scaling_den))
       old_square = running_square_[n]
       new_square = combination_coeff * old_square + (
           1. - combination_coeff) * T.sqr(grad)
       old_avg = running_avg_[n]
       new_avg = combination_coeff * old_avg + (
           1. - combination_coeff) * grad
       rms_grad = T.sqrt(new_square - new_avg ** 2)
       rms_grad = T.maximum(rms_grad, minimum_grad)
       memory = memory_[n]
       update = momentum * memory - learning_rate * grad / rms_grad

       update2 = momentum * momentum * memory - (
           1 + momentum) * learning_rate * grad / rms_grad
           
       updates.append((old_square, new_square))
       updates.append((old_avg, new_avg))
       updates.append((memory, update))
       updates.append((param, param + update2))
    return updates