Python theano.tensor.set_subtensor() Examples

def pad_to_a_multiple(tensor_, k, pad_with):
    """Pad a tensor to make its first dimension a multiple of a number.

    Parameters
    ----------
    tensor_ : :class:`~theano.Variable`
    k : int
        The number, multiple of which the length of tensor is made.
    pad_with : float or int
        The value for padding.

    """
    new_length = (
        tensor.ceil(tensor_.shape[0].astype('float32') / k) * k).astype('int64')
    new_shape = tensor.set_subtensor(tensor_.shape[:1], new_length)
    canvas = tensor.alloc(pad_with, tensor.prod(new_shape)).reshape(
        new_shape, ndim=tensor_.ndim)
    return tensor.set_subtensor(canvas[:tensor_.shape[0]], tensor_) 

def set_output(self):
        padding = self._padding
        input_shape = self._input_shape
        if np.sum(self._padding) > 0:
            padded_input = tensor.alloc(0.0,  # Value to fill the tensor
                                        input_shape[0],
                                        input_shape[1] + 2 * padding[1],
                                        input_shape[2],
                                        input_shape[3] + 2 * padding[3],
                                        input_shape[4] + 2 * padding[4])

            padded_input = tensor.set_subtensor(
                padded_input[:, padding[1]:padding[1] + input_shape[1], :, padding[3]:padding[3] +
                             input_shape[3], padding[4]:padding[4] + input_shape[4]],
                self._prev_layer.output)
        else:
            padded_input = self._prev_layer.output

        self._output = conv3d2d.conv3d(padded_input, self.W.val) + \
            self.b.val.dimshuffle('x', 'x', 0, 'x', 'x') 

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 adadelta(self, param, grad, updates, sample_idx=None, epsilon=1e-6):
        v1 = np.float32(self.decay)
        v2 = np.float32(1.0 - self.decay)
        acc = theano.shared(param.get_value(borrow=False) * 0., borrow=True)
        upd = theano.shared(param.get_value(borrow=False) * 0., borrow=True)
        if sample_idx is None:
            acc_new = acc + grad ** 2
            updates[acc] = acc_new
            grad = T.sqrt(upd + epsilon) * grad
            upd_new = v1 * upd + v2 * grad ** 2
            updates[upd] = upd_new
        else:
            acc_s = acc[sample_idx]
            acc_new = acc_s + grad ** 2
            updates[acc] = T.set_subtensor(acc_s, acc_new)
            upd_s = upd[sample_idx]
            upd_new = v1 * upd_s + v2 * grad ** 2
            updates[upd] = T.set_subtensor(upd_s, upd_new)
            grad = T.sqrt(upd_s + epsilon) * grad
        gradient_scaling = T.cast(T.sqrt(acc_new + epsilon), theano.config.floatX)
        return grad / gradient_scaling 

def adam(self, param, grad, updates, sample_idx=None, epsilon=1e-6):
        v1 = np.float32(self.decay)
        v2 = np.float32(1.0 - self.decay)
        acc = theano.shared(param.get_value(borrow=False) * 0., borrow=True)
        meang = theano.shared(param.get_value(borrow=False) * 0., borrow=True)
        countt = theano.shared(param.get_value(borrow=False) * 0., borrow=True)
        if sample_idx is None:
            acc_new = v1 * acc + v2 * grad ** 2
            meang_new = v1 * meang + v2 * grad
            countt_new = countt + 1
            updates[acc] = acc_new
            updates[meang] = meang_new
            updates[countt] = countt_new
        else:
            acc_s = acc[sample_idx]
            meang_s = meang[sample_idx]
            countt_s = countt[sample_idx]
            acc_new = v1 * acc_s + v2 * grad ** 2
            meang_new = v1 * meang_s + v2 * grad
            countt_new = countt_s + 1.0
            updates[acc] = T.set_subtensor(acc_s, acc_new)
            updates[meang] = T.set_subtensor(meang_s, meang_new)
            updates[countt] = T.set_subtensor(countt_s, countt_new)
        return (meang_new / (1 - v1 ** countt_new)) / (T.sqrt(acc_new / (1 - v1 ** countt_new)) + epsilon) 

def set_output(self):
        output_shape = self._output_shape
        padding = self._padding
        unpool_size = self._unpool_size
        unpooled_output = tensor.alloc(0.0,  # Value to fill the tensor
                                       output_shape[0],
                                       output_shape[1] + 2 * padding[0],
                                       output_shape[2],
                                       output_shape[3] + 2 * padding[1],
                                       output_shape[4] + 2 * padding[2])

        unpooled_output = tensor.set_subtensor(unpooled_output[:, padding[0]:output_shape[
            1] + padding[0]:unpool_size[0], :, padding[1]:output_shape[3] + padding[1]:unpool_size[
                1], padding[2]:output_shape[4] + padding[2]:unpool_size[2]],
                                               self._prev_layer.output)
        self._output = unpooled_output 

def set_output(self):
        padding = self._padding
        input_shape = self._input_shape
        padded_input = tensor.alloc(0.0,  # Value to fill the tensor
                                    input_shape[0],
                                    input_shape[1] + 2 * padding[1],
                                    input_shape[2],
                                    input_shape[3] + 2 * padding[3],
                                    input_shape[4] + 2 * padding[4])

        padded_input = tensor.set_subtensor(padded_input[:, padding[1]:padding[1] + input_shape[
            1], :, padding[3]:padding[3] + input_shape[3], padding[4]:padding[4] + input_shape[4]],
                                            self._prev_layer.output)

        fc_output = tensor.reshape(
            tensor.dot(self._fc_layer.output, self.Wx.val), self._output_shape)
        self._output = conv3d2d.conv3d(padded_input, self.Wh.val) + \
            fc_output + self.b.val.dimshuffle('x', 'x', 0, 'x', 'x') 

def set_output(self):
        padding = self._padding
        input_shape = self._input_shape
        padded_input = tensor.alloc(0.0,  # Value to fill the tensor
                                    input_shape[0],
                                    input_shape[1] + 2 * padding[1],
                                    input_shape[2],
                                    input_shape[3] + 2 * padding[3],
                                    input_shape[4] + 2 * padding[4])

        padded_input = tensor.set_subtensor(padded_input[:, padding[1]:padding[1] + input_shape[
            1], :, padding[3]:padding[3] + input_shape[3], padding[4]:padding[4] + input_shape[4]],
                                            self._prev_layer.output)

        self._output = conv3d2d.conv3d(padded_input, self.W.val) + \
            self.b.val.dimshuffle('x', 'x', 0, 'x', 'x') 

def inpainting_sample_and_noise(self, X, default_input_include_prob=1., default_input_scale=1.):
        # Very hacky! Specifically for inpainting right half of CIFAR-10 given left half
        # assumes X is b01c
        assert X.ndim == 4
        input_space = self.mlp.get_input_space()
        n = input_space.get_total_dimension()
        image_size = input_space.shape[0]
        half_image = int(image_size / 2)
        data_shape = (X.shape[0], image_size, half_image, input_space.num_channels)

        noise = self.theano_rng.normal(size=data_shape, dtype='float32')
        Xg = T.set_subtensor(X[:,:,half_image:,:], noise)
        sampled_part, noise =  self.mlp.dropout_fprop(Xg, default_input_include_prob=default_input_include_prob, default_input_scale=default_input_scale), noise
        sampled_part = sampled_part.reshape(data_shape)
        rval = T.set_subtensor(X[:, :, half_image:, :], sampled_part)
        return rval, noise 

def test_wrong_broadcast(self):
        a = tt.col()
        increment = tt.vector()

        # These symbolic graphs legitimate, as long as increment has exactly
        # one element. So it should fail at runtime, not at compile time.
        rng = numpy.random.RandomState(utt.fetch_seed())

        def rng_randX(*shape):
            return rng.rand(*shape).astype(theano.config.floatX)

        for op in (tt.set_subtensor, tt.inc_subtensor):
            for base in (a[:], a[0]):
                out = op(base, increment)
                f = theano.function([a, increment], out)
                # This one should work
                f(rng_randX(3, 1), rng_randX(1))
                # These ones should not
                self.assertRaises(ValueError,
                                  f, rng_randX(3, 1), rng_randX(2))
                self.assertRaises(ValueError,
                                  f, rng_randX(3, 1), rng_randX(3))
                self.assertRaises(ValueError,
                                  f, rng_randX(3, 1), rng_randX(0)) 

def expand_empty(tensor_var, size):
    """
    Transforms the shape of a tensor from (d1, d2 ... ) to ( d1+size, d2, ..)
    by adding uninitialized memory at the end of the tensor.

    """

    if size == 0:
        return tensor_var
    shapes = [tensor_var.shape[x] for x in xrange(tensor_var.ndim)]
    new_shape = [size + shapes[0]] + shapes[1:]
    empty = tensor.AllocEmpty(tensor_var.dtype)(*new_shape)

    ret = tensor.set_subtensor(empty[:shapes[0]], tensor_var)
    ret.tag.nan_guard_mode_check = False
    return ret 

def test_advset_subtensor1_2d():
    """ Test GPU version of set_subtensor on matrices (uses GpuAdvancedIncSubtensor1_dev20 if compute capability >= 2.0) """
    shp = (10,5)
    shared = cuda.shared_constructor
    xval = numpy.arange(numpy.prod(shp), dtype='float32').reshape(shp) + 1
    idxs = numpy.array([0,2,5,7,3], dtype='int32')
    yval = numpy.ones((len(idxs), shp[1]), dtype='float32')*10
    x = shared(xval, name='x')
    y = T.tensor(dtype='float32', broadcastable=(False,) * len(shp), name='y')
    expr = T.advanced_set_subtensor1(x, y, idxs)
    f = theano.function([y], expr, mode=mode_with_gpu)
    assert sum([isinstance(node.op, cuda.GpuAdvancedIncSubtensor1)
                for node in f.maker.fgraph.toposort()]) == 1
    rval = f(yval)
    rep = xval.copy()
    rep[idxs] = yval
    utt.assert_allclose(rval, rep) 

def adagrad_update(self, cost, learning_rate, eps=1e-8):
        params = [ p if p != self.slices else self.EMB for p in self.params ]
        accumulators = [ theano.shared(numpy.zeros(p.get_value(borrow=True).shape,
                                                dtype=theano.config.floatX))
                         for p in params ]
        gparams = [ T.grad(cost, param) for param in self.params ]
        self.gparams = gparams
        updates = [ ]
        for param, gparam, acc in zip(self.params, gparams, accumulators):
            if param == self.slices:
                acc_slices = acc[self.x.flatten()]
                new_acc_slices = acc_slices + gparam**2
                updates.append( (acc, T.set_subtensor(acc_slices, new_acc_slices)) )
                updates.append( (self.EMB, T.inc_subtensor(param,
                                 - learning_rate * gparam / T.sqrt(new_acc_slices+eps))) )
            else:
                new_acc = acc + gparam**2
                updates.append( (acc, new_acc) )
                updates.append( (param, param - learning_rate * gparam /
                                    T.sqrt(new_acc + eps)) )
        return updates 

def test_advset_subtensor1():
    """ Test GPU version of set_subtensor on vectors (uses GpuAdvancedIncSubtensor1) """
    shp = (10,)
    shared = cuda.shared_constructor
    xval = numpy.arange(shp[0], dtype='float32').reshape(shp) + 1
    idxs = numpy.array([0,2,5,7,3], dtype='int32')
    yval = numpy.ones(len(idxs), dtype='float32')*10
    x = shared(xval, name='x')
    y = T.tensor(dtype='float32', broadcastable=(False,) * len(shp), name='y')
    expr = T.advanced_set_subtensor1(x, y, idxs)
    f = theano.function([y], expr, mode=mode_with_gpu)
    assert sum([isinstance(node.op, cuda.GpuAdvancedIncSubtensor1)
                for node in f.maker.fgraph.toposort()]) == 1
    rval = f(yval)
    rep = xval.copy()
    rep[idxs] = yval
    utt.assert_allclose(rval, rep) 

def rmsprop(self, param, grad, updates, sample_idx=None, epsilon=1e-6):
        v1 = np.float32(self.adapt_params[0])
        v2 = np.float32(1.0 - self.adapt_params[0])
        acc = theano.shared(param.get_value(borrow=False) * 0., borrow=True)
        if sample_idx is None:
            acc_new = v1 * acc + v2 * grad ** 2
            updates[acc] = acc_new
        else:
            acc_s = acc[sample_idx]
            #            acc_new = v1 * acc_s + v2 * grad ** 2 #Faster, but inaccurate when an index occurs multiple times
            #            updates[acc] = T.set_subtensor(acc_s, acc_new) #Faster, but inaccurate when an index occurs multiple times
            updates[acc] = T.inc_subtensor(T.set_subtensor(acc_s, acc_s * v1)[sample_idx],
                                           v2 * grad ** 2)  # Slower, but accurate when an index occurs multiple times
            acc_new = updates[acc][sample_idx]  # Slower, but accurate when an index occurs multiple times
        gradient_scaling = T.cast(T.sqrt(acc_new + epsilon), theano.config.floatX)
        return grad / gradient_scaling 

def adadelta(self, param, grad, updates, sample_idx=None, epsilon=1e-6):
        v1 = np.float32(self.decay)
        v2 = np.float32(1.0 - self.decay)
        acc = theano.shared(param.get_value(borrow=False) * 0., borrow=True)
        upd = theano.shared(param.get_value(borrow=False) * 0., borrow=True)
        if sample_idx is None:
            acc_new = acc + grad ** 2
            updates[acc] = acc_new
            grad = T.sqrt(upd + epsilon) * grad
            upd_new = v1 * upd + v2 * grad ** 2
            updates[upd] = upd_new
        else:
            acc_s = acc[sample_idx]
            acc_new = acc_s + grad ** 2
            updates[acc] = T.set_subtensor(acc_s, acc_new)
            upd_s = upd[sample_idx]
            upd_new = v1 * upd_s + v2 * grad ** 2
            updates[upd] = T.set_subtensor(upd_s, upd_new)
            grad = T.sqrt(upd_s + epsilon) * grad
        gradient_scaling = T.cast(T.sqrt(acc_new + epsilon), theano.config.floatX)
        return grad / gradient_scaling 

def adam(self, param, grad, updates, sample_idx=None, epsilon=1e-6):
        v1 = np.float32(self.decay)
        v2 = np.float32(1.0 - self.decay)
        acc = theano.shared(param.get_value(borrow=False) * 0., borrow=True)
        meang = theano.shared(param.get_value(borrow=False) * 0., borrow=True)
        countt = theano.shared(param.get_value(borrow=False) * 0., borrow=True)
        if sample_idx is None:
            acc_new = v1 * acc + v2 * grad ** 2
            meang_new = v1 * meang + v2 * grad
            countt_new = countt + 1
            updates[acc] = acc_new
            updates[meang] = meang_new
            updates[countt] = countt_new
        else:
            acc_s = acc[sample_idx]
            meang_s = meang[sample_idx]
            countt_s = countt[sample_idx]
            acc_new = v1 * acc_s + v2 * grad ** 2
            meang_new = v1 * meang_s + v2 * grad
            countt_new = countt_s + 1.0
            updates[acc] = T.set_subtensor(acc_s, acc_new)
            updates[meang] = T.set_subtensor(meang_s, meang_new)
            updates[countt] = T.set_subtensor(countt_s, countt_new)
        return (meang_new / (1 - v1 ** countt_new)) / (T.sqrt(acc_new / (1 - v1 ** countt_new)) + epsilon) 

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 temporal_padding(x, padding=(1, 1)):
    """Pad the middle dimension of a 3D tensor
    with "padding" zeros left and right.

    Apologies for the inane API, but Theano makes this
    really hard.
    """
    assert len(padding) == 2
    input_shape = x.shape
    output_shape = (input_shape[0],
                    input_shape[1] + padding[0] + padding[1],
                    input_shape[2])
    output = T.zeros(output_shape)
    result = T.set_subtensor(output[:, padding[0]:x.shape[1] + padding[0], :], x)
    if hasattr(x, '_keras_shape'):
        result._keras_shape = (x._keras_shape[0],
                               x._keras_shape[1] + py_sum(padding),
                               x._keras_shape[2])
    return result 

def depool(X, factor=2):
    """
    Luke perforated upsample: http://www.brml.org/uploads/tx_sibibtex/281.pdf
    """
    output_shape = [
        X.shape[1],
        X.shape[2]*factor,
        X.shape[3]*factor
    ]
    stride = X.shape[2]
    offset = X.shape[3]
    in_dim = stride * offset
    out_dim = in_dim * factor * factor

    upsamp_matrix = T.zeros((in_dim, out_dim))
    rows = T.arange(in_dim)
    cols = rows*factor + (rows/stride * factor * offset)
    upsamp_matrix = T.set_subtensor(upsamp_matrix[rows, cols], 1.)

    flat = T.reshape(X, (X.shape[0], output_shape[0], X.shape[2] * X.shape[3]))

    up_flat = T.dot(flat, upsamp_matrix)
    upsamp = T.reshape(up_flat, (X.shape[0], output_shape[0], output_shape[1], output_shape[2]))

    return upsamp 

def combine_fragments_to_dense_bxcyz(self, tensor, sh):
      """ expected shape: (batch, x, channels, y, z)"""
      ttensor = tensor # be same shape as result, no significant time cost

      output_stride = self.output_stride
      if isinstance(output_stride, list) or isinstance(output_stride, tuple):
          example_stride = np.prod(output_stride)#**3
      else:
          example_stride = output_stride**3
          output_stride  = np.asarray((output_stride,)*3)
      zero = np.array((0), dtype=theano.config.floatX)
      embedding = T.alloc( zero, 1, sh[1]*output_stride[0], sh[2], sh[3]*output_stride[1], sh[4]*output_stride[2]) # first arg. is fill-value (0 in this case) and not an element of the shape
      ix = offset_map(output_stride)
      print "      output_stride",output_stride
      print "      example_stride",example_stride
      for i,(n,m,k) in enumerate(ix):
          embedding = T.set_subtensor(embedding[:,n::output_stride[0],:,m::output_stride[1],k::output_stride[2]], ttensor[i::example_stride])
      return embedding 

def ConvByPattern(x, patterns, mask=None):
    W = np.transpose(patterns, (3, 0, 1, 2))
    out2 = T.nnet.conv2d(x.dimshuffle(0, 3, 1, 2), W, filter_shape=W.shape, border_mode='half')
    if mask is not None:
        ## mask has shape (batchSize, #rows_to_be_masked, nCols)

        ## a subtensor of out2 along the horiz direction
        out2_sub_horiz = out2[:, :, :mask.shape[1], :]
        mask_horiz = mask.dimshuffle(0, 'x', 1, 2)
        out3 = T.set_subtensor(out2_sub_horiz, T.mul(out2_sub_horiz, mask_horiz) )

        ## a subtensor of out3 along the vertical direction
        out3_sub_vertical = out3[:, :, :, :mask.shape[1] ]
        mask_vertical = mask.dimshuffle(0, 'x', 2, 1)
        y = T.set_subtensor(out3_sub_vertical, T.mul(out3_sub_vertical, mask_vertical) )
    else:
	y = out2

    y = y.dimshuffle(0, 2, 3, 1)

    return y/np.prod(patterns.shape[1:3]) 

def adagrad_update(self, cost, learning_rate, eps=1e-8):
        params = [ p if p != self.slices else self.EMB for p in self.params ]
        accumulators = [ theano.shared(numpy.zeros(p.get_value(borrow=True).shape,
                                                dtype=theano.config.floatX))
                         for p in params ]
        gparams = [ T.grad(cost, param) for param in self.params ]
        self.gparams = gparams
        updates = [ ]
        for param, gparam, acc in zip(self.params, gparams, accumulators):
            if param == self.slices:
                acc_slices = acc[self.x.flatten()]
                new_acc_slices = acc_slices + gparam**2
                updates.append( (acc, T.set_subtensor(acc_slices, new_acc_slices)) )
                updates.append( (self.EMB, T.inc_subtensor(param,
                                 - learning_rate * gparam / T.sqrt(new_acc_slices+eps))) )
            else:
                new_acc = acc + gparam**2
                updates.append( (acc, new_acc) )
                updates.append( (param, param - learning_rate * gparam /
                                    T.sqrt(new_acc + eps)) )
        return updates 

def concatenate(tensor_list, axis=0):
    if axis < 0:
        axis += tensor_list[0].ndim

    concat_size = sum(tensor.shape[axis] for tensor in tensor_list)

    output_shape = ()
    for k in range(axis):
        output_shape += (tensor_list[0].shape[k],)
    output_shape += (concat_size,)
    for k in range(axis + 1, tensor_list[0].ndim):
        output_shape += (tensor_list[0].shape[k],)

    out = T.zeros(output_shape)
    offset = 0
    for tensor in tensor_list:
        indices = ()
        for k in range(axis):
            indices += (slice(None),)
        indices += (slice(offset, offset + tensor.shape[axis]),)
        for k in range(axis + 1, tensor_list[0].ndim):
            indices += (slice(None),)

        out = T.set_subtensor(out[indices], tensor)
        offset += tensor.shape[axis]

    return out 

def theano_one_hot(idxs, n):
    z = T.zeros((idxs.shape[0], n))
    one_hot = T.set_subtensor(z[T.arange(idxs.shape[0]), idxs], 1)
    return one_hot 

def categorical_best(tensor):
    """
    tensor should be a tensor of shape (..., categories)
    Return a new tensor of the same shape but one-hot at position of best category
    """
    flat_tensor = tensor.reshape([-1, tensor.shape[-1]])
    argmax_posns = T.argmax(flat_tensor, 1)
    flat_snapped = T.zeros_like(flat_tensor)
    flat_snapped = T.set_subtensor(flat_snapped[T.arange(flat_tensor.shape[0]), argmax_posns], 1.0)
    snapped = flat_snapped.reshape(tensor.shape)
    return snapped 

def pad2dwithchannel(x, size_kernel):
    assert size_kernel[0]>1 or size_kernel[1]>1
    assert size_kernel[0]%2 == 1
    assert size_kernel[1]%2 == 1
    a = (size_kernel[0]-1)/2
    b = (size_kernel[1]-1)/2
    if True:
        n_channels = x.shape[1]
        result_shape = (x.shape[0],x.shape[1]+1,x.shape[2]+2*a,x.shape[3]+2*b)
        result = T.zeros(result_shape, dtype=G.floatX)
        result = T.set_subtensor(result[:,n_channels,:,:], 1.)
        result = T.set_subtensor(result[:,n_channels,a:-a,b:-b], 0.)
        result = T.set_subtensor(result[:,:n_channels,a:-a,b:-b], x)
    else:
        # new code, requires that the minibatch size 'x.tag.test_value.shape[0]' is the same during execution
        # I thought this would be more memory-efficient, but seems not the case in practice
        print 'new code, requires that the minibatch size "x.tag.test_value.shape[0]" is the same during execution' 
        x_shape = x.tag.test_value.shape
        n_channels = x_shape[1]
        result_shape = (x_shape[0],x_shape[1]+1,x_shape[2]+2*a,x_shape[3]+2*b)
        result = np.zeros(result_shape,dtype=G.floatX)
        result[:,n_channels,:,:] = 1.
        result[:,n_channels,a:-a,b:-b] = 0.
        result = T.constant(result)
        result = T.set_subtensor(result[:,:n_channels,a:-a,b:-b], x)
    return result


# Multi-scale conv 

def pad2d(x, n_padding):
    result_shape = (x.shape[0],x.shape[1],x.shape[2]+2*n_padding,x.shape[3]+2*n_padding)
    result = T.zeros(result_shape, dtype=G.floatX)
    result = T.set_subtensor(result[:,:,n_padding:-n_padding,n_padding:-n_padding], x)
    return result


# Pad input, add extra channel 

def upsample2d_perforated(x):
    shape = x.shape
    x = x.reshape((shape[0], shape[1], shape[2], 1, shape[3], 1))
    y = T.zeros((shape[0], shape[1], shape[2], 2, shape[3], 2),dtype=G.floatX)
    x = T.set_subtensor(y[:,:,:,0:1,:,0:1], x)
    x = x.reshape((shape[0], shape[1], shape[2]*2, shape[3]*2))
    return x 

# Pad input 

def fprop(self, var):
        rval = TT.zeros_like(var)
        if self.n >0:
            rval = TT.set_subtensor(rval[self.n:], var[:-self.n])
        elif self.n<0:
            rval = TT.set_subtensor(rval[:self.n], var[-self.n:])
        self.out = rval
        return rval