Python theano.tensor.minimum() Examples

def ctc_update_log_p(skip_idxs, zeros, active, log_p_curr, log_p_prev):
    active_skip_idxs = skip_idxs[(skip_idxs < active).nonzero()]
    active_next = T.cast(T.minimum(
        T.maximum(
            active + 1,
            T.max(T.concatenate([active_skip_idxs, [-1]])) + 2 + 1
        ), log_p_curr.shape[0]), 'int32')

    common_factor = T.max(log_p_prev[:active])
    p_prev = T.exp(log_p_prev[:active] - common_factor)
    _p_prev = zeros[:active_next]
    # copy over
    _p_prev = T.set_subtensor(_p_prev[:active], p_prev)
    # previous transitions
    _p_prev = T.inc_subtensor(_p_prev[1:], _p_prev[:-1])
    # skip transitions
    _p_prev = T.inc_subtensor(_p_prev[active_skip_idxs + 2], p_prev[active_skip_idxs])
    updated_log_p_prev = T.log(_p_prev) + common_factor

    log_p_next = T.set_subtensor(
        zeros[:active_next],
        log_p_curr[:active_next] + updated_log_p_prev
    )
    return active_next, log_p_next 

def ctc_update_log_p(skip_idxs, zeros, active, log_p_curr, log_p_prev):
    active_skip_idxs = skip_idxs[(skip_idxs < active).nonzero()]
    active_next = T.cast(T.minimum(
        T.maximum(
            active + 1,
            T.max(T.concatenate([active_skip_idxs, [-1]])) + 2 + 1
        ), log_p_curr.shape[0]), 'int32')

    common_factor = T.max(log_p_prev[:active])
    p_prev = T.exp(log_p_prev[:active] - common_factor)
    _p_prev = zeros[:active_next]
    # copy over
    _p_prev = T.set_subtensor(_p_prev[:active], p_prev)
    # previous transitions
    _p_prev = T.inc_subtensor(_p_prev[1:], _p_prev[:-1])
    # skip transitions
    _p_prev = T.inc_subtensor(_p_prev[active_skip_idxs + 2], p_prev[active_skip_idxs])
    updated_log_p_prev = T.log(_p_prev) + common_factor

    log_p_next = T.set_subtensor(
        zeros[:active_next],
        log_p_curr[:active_next] + updated_log_p_prev
    )
    return active_next, log_p_next 

def ctc_update_log_p(skip_idxs, zeros, active, log_p_curr, log_p_prev):
    active_skip_idxs = skip_idxs[(skip_idxs < active).nonzero()]
    active_next = T.cast(T.minimum(
        T.maximum(
            active + 1,
            T.max(T.concatenate([active_skip_idxs, [-1]])) + 2 + 1
        ), log_p_curr.shape[0]), 'int32')

    common_factor = T.max(log_p_prev[:active])
    p_prev = T.exp(log_p_prev[:active] - common_factor)
    _p_prev = zeros[:active_next]
    # copy over
    _p_prev = T.set_subtensor(_p_prev[:active], p_prev)
    # previous transitions
    _p_prev = T.inc_subtensor(_p_prev[1:], _p_prev[:-1])
    # skip transitions
    _p_prev = T.inc_subtensor(_p_prev[active_skip_idxs + 2], p_prev[active_skip_idxs])
    updated_log_p_prev = T.log(_p_prev) + common_factor

    log_p_next = T.set_subtensor(
        zeros[:active_next],
        log_p_curr[:active_next] + updated_log_p_prev
    )
    return active_next, log_p_next 

def ctc_update_log_p(skip_idxs, zeros, active, log_p_curr, log_p_prev):
    active_skip_idxs = skip_idxs[(skip_idxs < active).nonzero()]
    active_next = T.cast(T.minimum(
        T.maximum(
            active + 1,
            T.max(T.concatenate([active_skip_idxs, [-1]])) + 2 + 1
        ), log_p_curr.shape[0]), 'int32')

    common_factor = T.max(log_p_prev[:active])
    p_prev = T.exp(log_p_prev[:active] - common_factor)
    _p_prev = zeros[:active_next]
    # copy over
    _p_prev = T.set_subtensor(_p_prev[:active], p_prev)
    # previous transitions
    _p_prev = T.inc_subtensor(_p_prev[1:], _p_prev[:-1])
    # skip transitions
    _p_prev = T.inc_subtensor(_p_prev[active_skip_idxs + 2], p_prev[active_skip_idxs])
    updated_log_p_prev = T.log(_p_prev) + common_factor

    log_p_next = T.set_subtensor(
        zeros[:active_next],
        log_p_curr[:active_next] + updated_log_p_prev
    )
    return active_next, log_p_next 

def rprop_core(params, gradients, rprop_increase=1.01, rprop_decrease=0.99, rprop_min_step=0, rprop_max_step=100,
               learning_rate=0.01):
    """
    Rprop optimizer.
    See http://sci2s.ugr.es/keel/pdf/algorithm/articulo/2003-Neuro-Igel-IRprop+.pdf.
    """
    for param, grad in zip(params, gradients):
        grad_tm1 = theano.shared(np.zeros_like(param.get_value()), name=param.name + '_grad')
        step_tm1 = theano.shared(np.zeros_like(param.get_value()) + learning_rate, name=param.name+ '_step')

        test = grad * grad_tm1
        same = T.gt(test, 0)
        diff = T.lt(test, 0)
        step = T.minimum(rprop_max_step, T.maximum(rprop_min_step, step_tm1 * (
            T.eq(test, 0) +
            same * rprop_increase +
            diff * rprop_decrease)))
        grad = grad - diff * grad
        yield param, param - T.sgn(grad) * step
        yield grad_tm1, grad
        yield step_tm1, step 

def ctc_update_log_p(skip_idxs, zeros, active, log_p_curr, log_p_prev):
    active_skip_idxs = skip_idxs[(skip_idxs < active).nonzero()]
    active_next = T.cast(T.minimum(
        T.maximum(
            active + 1,
            T.max(T.concatenate([active_skip_idxs, [-1]])) + 2 + 1
        ), log_p_curr.shape[0]), 'int32')

    common_factor = T.max(log_p_prev[:active])
    p_prev = T.exp(log_p_prev[:active] - common_factor)
    _p_prev = zeros[:active_next]
    # copy over
    _p_prev = T.set_subtensor(_p_prev[:active], p_prev)
    # previous transitions
    _p_prev = T.inc_subtensor(_p_prev[1:], _p_prev[:-1])
    # skip transitions
    _p_prev = T.inc_subtensor(_p_prev[active_skip_idxs + 2], p_prev[active_skip_idxs])
    updated_log_p_prev = T.log(_p_prev) + common_factor

    log_p_next = T.set_subtensor(
        zeros[:active_next],
        log_p_curr[:active_next] + updated_log_p_prev
    )
    return active_next, log_p_next 

def ctc_update_log_p(skip_idxs, zeros, active, log_p_curr, log_p_prev):
    active_skip_idxs = skip_idxs[(skip_idxs < active).nonzero()]
    active_next = T.cast(T.minimum(
        T.maximum(
            active + 1,
            T.max(T.concatenate([active_skip_idxs, [-1]])) + 2 + 1
        ), log_p_curr.shape[0]), 'int32')

    common_factor = T.max(log_p_prev[:active])
    p_prev = T.exp(log_p_prev[:active] - common_factor)
    _p_prev = zeros[:active_next]
    # copy over
    _p_prev = T.set_subtensor(_p_prev[:active], p_prev)
    # previous transitions
    _p_prev = T.inc_subtensor(_p_prev[1:], _p_prev[:-1])
    # skip transitions
    _p_prev = T.inc_subtensor(_p_prev[active_skip_idxs + 2], p_prev[active_skip_idxs])
    updated_log_p_prev = T.log(_p_prev) + common_factor

    log_p_next = T.set_subtensor(
        zeros[:active_next],
        log_p_curr[:active_next] + updated_log_p_prev
    )
    return active_next, log_p_next 

def ctc_update_log_p(skip_idxs, zeros, active, log_p_curr, log_p_prev):
    active_skip_idxs = skip_idxs[(skip_idxs < active).nonzero()]
    active_next = T.cast(T.minimum(
        T.maximum(
            active + 1,
            T.max(T.concatenate([active_skip_idxs, [-1]])) + 2 + 1
        ), log_p_curr.shape[0]), 'int32')

    common_factor = T.max(log_p_prev[:active])
    p_prev = T.exp(log_p_prev[:active] - common_factor)
    _p_prev = zeros[:active_next]
    # copy over
    _p_prev = T.set_subtensor(_p_prev[:active], p_prev)
    # previous transitions
    _p_prev = T.inc_subtensor(_p_prev[1:], _p_prev[:-1])
    # skip transitions
    _p_prev = T.inc_subtensor(_p_prev[active_skip_idxs + 2], p_prev[active_skip_idxs])
    updated_log_p_prev = T.log(_p_prev) + common_factor

    log_p_next = T.set_subtensor(
        zeros[:active_next],
        log_p_curr[:active_next] + updated_log_p_prev
    )
    return active_next, log_p_next 

def ctc_update_log_p(skip_idxs, zeros, active, log_p_curr, log_p_prev):
    active_skip_idxs = skip_idxs[(skip_idxs < active).nonzero()]
    active_next = T.cast(T.minimum(
        T.maximum(
            active + 1,
            T.max(T.concatenate([active_skip_idxs, [-1]])) + 2 + 1
        ), log_p_curr.shape[0]), 'int32')

    common_factor = T.max(log_p_prev[:active])
    p_prev = T.exp(log_p_prev[:active] - common_factor)
    _p_prev = zeros[:active_next]
    # copy over
    _p_prev = T.set_subtensor(_p_prev[:active], p_prev)
    # previous transitions
    _p_prev = T.inc_subtensor(_p_prev[1:], _p_prev[:-1])
    # skip transitions
    _p_prev = T.inc_subtensor(_p_prev[active_skip_idxs + 2], p_prev[active_skip_idxs])
    updated_log_p_prev = T.log(_p_prev) + common_factor

    log_p_next = T.set_subtensor(
        zeros[:active_next],
        log_p_curr[:active_next] + updated_log_p_prev
    )
    return active_next, log_p_next 

def relu(x, alpha=0., max_value=None):
    _assert_has_capability(T.nnet, 'relu')
    x = T.nnet.relu(x, alpha)
    if max_value is not None:
        x = T.minimum(x, max_value)
    return x 

def relu(x, alpha=0., max_value=None):
    _assert_has_capability(T.nnet, 'relu')
    x = T.nnet.relu(x, alpha)
    if max_value is not None:
        x = T.minimum(x, max_value)
    return x 

def _log_add(x, y):
    maxx = tt.maximum(x, y)
    minn = tt.minimum(x, y)
    return maxx + tt.log(1 + tt.exp(minn - maxx)) 

def logadd_advanced(x, y):
    maxx = tt.maximum(x, y)
    minn = tt.minimum(x, y)
    return maxx + tt.log(1 + tt.exp(minn - maxx)) 

def safe_exp(x):
    return tt.exp(tt.minimum(x, epsinv).astype(theano.config.floatX)) 

def q_loss(self, y_true, y_pred):
        # assume clip_delta is 1.0
        # along with sum accumulator.
        diff = y_true - y_pred
        _quad = T.minimum(abs(diff), 1.0)
        _lin = abs(diff) - _quad
        loss = 0.5 * _quad ** 2 + _lin
        loss = T.sum(loss)

        return loss 

def relu(x, alpha=0., max_value=None):
    assert hasattr(T.nnet, 'relu'), ('It looks like like your version of '
                                     'Theano is out of date. '
                                     'Install the latest version with:\n'
                                     'pip install git+git://github.com/Theano/Theano.git --upgrade --no-deps')
    x = T.nnet.relu(x, alpha)
    if max_value is not None:
        x = T.minimum(x, max_value)
    return x 

def relu(x, alpha=0., max_value=None):
    _assert_has_capability(T.nnet, 'relu')
    x = T.nnet.relu(x, alpha)
    if max_value is not None:
        x = T.minimum(x, max_value)
    return x 

def get_activation(act=None):
        def compose(a, b):
            c = lambda z: b(a(z))
            c.__theanets_name__ = '%s(%s)' % (b.__theanets_name__, a.__theanets_name__)
            return c
        if '+' in act:
            return functools.reduce(
                compose, (get_activation(a) for a in act.split('+')))
        options = {
            'tanh': T.tanh,
            'linear': lambda z: z,
            'logistic': T.nnet.sigmoid,
            'sigmoid': T.nnet.sigmoid,
            'hard_sigmoid': T.nnet.hard_sigmoid,
            'softplus': T.nnet.softplus,
            'softmax': softmax,
            'theano_softmax': T.nnet.softmax,

            # shorthands
            'relu': lambda z: T.nnet.relu(z),
            'leaky_relu': lambda z: T.nnet.relu(z, 0.01),
            'trel': lambda z: z * (z > 0) * (z < 1),
            'trec': lambda z: z * (z > 1),
            'tlin': lambda z: z * (abs(z) > 1),

            # modifiers
            'rect:max': lambda z: T.minimum(1, z),
            'rect:min': lambda z: T.maximum(0, z),

            # normalization
            'norm:dc': lambda z: (z.T - z.mean(axis=1)).T,
            'norm:max': lambda z: (z.T / T.maximum(1e-10, abs(z).max(axis=1))).T,
            'norm:std': lambda z: (z.T / T.maximum(1e-10, T.std(z, axis=1))).T,
            }
        for k, v in options.items():
            v.__theanets_name__ = k
        try:
            return options[act]
        except KeyError:
            raise KeyError('unknown activation %r' % act) 

def minimum(x, y):
    return T.minimum(x, y) 

def minimum(x, y):
    return T.minimum(x, y) 

def relu(x, alpha=0., max_value=None):
    _assert_has_capability(T.nnet, 'relu')
    x = T.nnet.relu(x, alpha)
    if max_value is not None:
        x = T.minimum(x, max_value)
    return x 

def minimum(x, y):
    return T.minimum(x, y) 

def relu(x, alpha=0., max_value=None):
    _assert_has_capability(T.nnet, 'relu')
    x = T.nnet.relu(x, alpha)
    if max_value is not None:
        x = T.minimum(x, max_value)
    return x 

def relu(x, alpha=0., max_value=None):
    _assert_has_capability(T.nnet, 'relu')
    x = T.nnet.relu(x, alpha)
    if max_value is not None:
        x = T.minimum(x, max_value)
    return x 

def minimum(x, y):
    return T.minimum(x, y) 

def relu(x, alpha=0., max_value=None):
    _assert_has_capability(T.nnet, 'relu')
    x = T.nnet.relu(x, alpha)
    if max_value is not None:
        x = T.minimum(x, max_value)
    return x 

def minimum(x, y):
    return T.minimum(x, y) 

def relu(x, alpha=0., max_value=None):
    _assert_has_capability(T.nnet, 'relu')
    x = T.nnet.relu(x, alpha)
    if max_value is not None:
        x = T.minimum(x, max_value)
    return x 

def minimum(x, y):
    return T.minimum(x, y) 

def minimum(x, y):
    return T.minimum(x, y)