resnet.py

#!/usr/bin/env python3
# Copyright 2019 Christian Henning
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# @title          :mnets/resnet.py
# @author         :ch
# @contact        :henningc@ethz.ch
# @created        :09/20/2019
# @version        :1.0
# @python_version :3.6.8
"""
ResNet
------

This module implements the class of Resnet networks described in section 4.2 of
the following paper:

    "Deep Residual Learning for Image Recognition", He et al., 2015
    https://arxiv.org/abs/1512.03385
"""
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

from mnets.classifier_interface import Classifier
from mnets.mnet_interface import MainNetInterface
from utils.batchnorm_layer import BatchNormLayer
from utils.torch_utils import init_params

class ResNet(Classifier):
    """A resnet with :math:`6n+2` layers with :math:`3n` residual blocks,
    consisting of two layers each.

    Args:
        in_shape (tuple or list): The shape of an input sample in format
            ``HWC``.

            Note
                We assume the Tensorflow format, where the last entry
                denotes the number of channels.
        num_classes (int): The number of output neurons.

            Note:
                The network outputs logits.
        use_bias (bool): Whether layers may have bias terms.

            Note:
                Bias terms are unnecessary in convolutional layers if batch
                normalization is used. However, this option disables bias terms
                altogether (including in the final fully-connected layer).
        num_feature_maps (tuple): A list of 4 integers, each denoting the number
            of feature maps of convolutional layers in a certain group of the
            network architecture. The first entry is the number of feature
            maps of the first convolutional layer, the remaining 3 numbers
            determine the number of feature maps in the consecutive groups
            comprising :math:`2n` convolutional layers each.
        verbose (bool): Allow printing of general information about the
            generated network (such as number of weights).
        n (int): The network will consist of :math:`6n+2` layers. In the
            paper :math:`n` has been chosen to be 3, 5, 7, 9 or 18.
        k (int): The widening factor. Feature maps in the 3 convolutional groups
            will be multiplied by this number. See argument
            ``num_feature_maps``. This argument is typical for wide resnets,
            such as :class:`mnets.wide_resnet.WRN`. Hence, for ``k > 1`` this
            network becomes essentually a wide resnet.
        no_weights (bool): If set to ``True``, no trainable parameters will be
            constructed, i.e., weights are assumed to be produced ad-hoc
            by a hypernetwork and passed to the :meth:`forward` method.

            Note, this also affects the affine parameters of the
            batchnorm layer. I.e., if set to ``True``, then the argument
            ``affine`` of :class:`utils.batchnorm_layer.BatchNormLayer`
            will be set to ``False`` and we expect the batchnorm parameters
            to be passed to the :meth:`forward`.
        init_weights (optional): This option is for convinience reasons.
            The option expects a list of parameter values that are used to
            initialize the network weights. As such, it provides a
            convinient way of initializing a network with a weight draw
            produced by the hypernetwork.
        use_batch_norm: Whether batch normalization should used. It will be
            applied after all convolutional layers (before the activation).
        bn_track_stats: If batch normalization is used, then this option
            determines whether running statistics are tracked in these
            layers or not (see argument ``track_running_stats`` of class
            :class:`utils.batchnorm_layer.BatchNormLayer`).

            If ``False``, then batch statistics are utilized even during
            evaluation. If ``True``, then running stats are tracked. When
            using this network in a continual learning scenario with
            different tasks then the running statistics are expected to be
            maintained externally. The argument ``stats_id`` of the method
            :meth:`utils.batchnorm_layer.BatchNormLayer.forward` can be
            provided using the argument ``condition`` of method
            :meth:`forward`.

            Example:
                To maintain the running stats, one can simply iterate over
                all batch norm layers and checkpoint the current running
                stats (e.g., after learning a task when applying a Continual
                Learning scenario).

                .. code:: python

                    for bn_layer in net.batchnorm_layers:
                        bn_layer.checkpoint_stats()
        distill_bn_stats: If ``True``, then the shapes of the batchnorm
            statistics will be added to the attribute
            :attr:`mnets.mnet_interface.MainNetInterface.\
hyper_shapes_distilled` and the current statistics will be returned by the
            method :meth:`distillation_targets`.

            Note, this attribute may only be ``True`` if ``bn_track_stats``
            is ``True``.
        context_mod_apply_pixel_wise (bool): By default, the context-dependent
            modulation applies a scalar gain and shift to all feature maps in
            the output of a convolutional layer. When activating this option,
            the gain and shift will be a per-pixel parameter in all feature
            maps.

            To be more precise, consider the output of a convolutional layer
            of shape ``[C,H,W]``. By default, there will be ``C`` gain and shift
            parameters for such a layer. Upon activating this option, the
            number of gain and shift parameters for such a layer will increase
            to ``C x H x W``.
        **kwargs: Keyword arguments regarding context modulation. This class
            can process the same context-modulation related arguments as class
            :class:`mnets.mlp.MLP`. Additionally, one may specify the argument
            ``context_mod_apply_pixel_wise``.

            Some additional remarks regarding the handling of keyword arguments:

            - ``use_context_mod``: Context-modulation will be applied after the
              linear computation of each layer (i.e. all hidden layers (conv
              layers) as well as the final FC output layer).

              Similar to Spatial Batch-Normalization, there will be a scalar
              shift and gain applied per feature map for all convolutional
              layers (except if ``context_mod_apply_pixel_wise`` is set).
            - ``context_mod_inputs``: The input is treated like the output of a
              convolutional layer when applying context-dependent modulation.
    """
    def __init__(self, in_shape=(32, 32, 3), num_classes=10, use_bias=True,
                 num_feature_maps=(16, 16, 32, 64), verbose=True, n=5, k=1,
                 no_weights=False, init_weights=None, use_batch_norm=True,
                 bn_track_stats=True, distill_bn_stats=False,
                 context_mod_apply_pixel_wise=False, **kwargs):
        super(ResNet, self).__init__(num_classes, verbose)

        self._in_shape = in_shape
        self._n = n

        ### Parse or set context-mod arguments ###
        rem_kwargs = MainNetInterface._parse_context_mod_args(kwargs)
        if len(rem_kwargs) > 0:
            raise ValueError('Keyword arguments %s unknown.' % str(rem_kwargs))

        self._context_mod_apply_pixel_wise = context_mod_apply_pixel_wise

        self._use_context_mod = kwargs['use_context_mod']
        self._context_mod_inputs = kwargs['context_mod_inputs']
        self._no_last_layer_context_mod = kwargs['no_last_layer_context_mod']
        self._context_mod_no_weights = kwargs['context_mod_no_weights']
        self._context_mod_post_activation = \
            kwargs['context_mod_post_activation']
        self._context_mod_gain_offset = kwargs['context_mod_gain_offset']
        self._context_mod_gain_softplus = kwargs['context_mod_gain_softplus']
        ### Parse or set context-mod arguments - Done ###

        assert(init_weights is None or \
               (not no_weights or not self._context_mod_no_weights))
        self._no_weights = no_weights

        assert(not use_batch_norm or (not distill_bn_stats or bn_track_stats))

        self._use_batch_norm = use_batch_norm
        self._bn_track_stats = bn_track_stats
        self._distill_bn_stats = distill_bn_stats and use_batch_norm

        self._kernel_size = [3, 3]
        if len(num_feature_maps) != 4:
            raise ValueError('Option "num_feature_maps" must be a list of 4 ' +
                             'integers.')
        self._filter_sizes = list(num_feature_maps)
        for i in range(1, 4):
            if k != 1:
                self._filter_sizes[i] = k * num_feature_maps[i]
            if num_feature_maps[i] < num_feature_maps[i-1]:
                raise ValueError('We currently require the number of ' +
                                 'channels to stay constant or to increase, ' +
                                 'in which case we apply zero-padding to the ' +
                                 'shortcut connections.')

        self._has_bias = use_bias
        self._has_fc_out = True
        # We need to make sure that the last 2 entries of `weights` correspond
        # to the weight matrix and bias vector of the last layer.
        self._mask_fc_out = True
        # We don't use any output non-linearity.
        self._has_linear_out = True

        self._param_shapes = []
        self._param_shapes_meta = []
        self._internal_params = None if no_weights and \
            self._context_mod_no_weights else nn.ParameterList()
        self._hyper_shapes_learned = None \
            if not no_weights and not self._context_mod_no_weights else []
        self._hyper_shapes_learned_ref = None if self._hyper_shapes_learned \
            is None else []

        #################################################
        ### Define and initialize context mod weights ###
        #################################################
        self._context_mod_layers = nn.ModuleList() if self._use_context_mod \
            else None

        if self._use_context_mod:
            cm_layer_inds = None # TODO implement
            cm_shapes = [] # Output shape of all layers.
            if self._context_mod_inputs:
                cm_shapes.append([in_shape[2], *in_shape[:2]])
            layer_out_shapes = self._compute_layer_out_sizes()
            if self._no_last_layer_context_mod:
                cm_shapes.extend(layer_out_shapes[:-1])
            else:
                cm_shapes.extend(layer_out_shapes)

            if not context_mod_apply_pixel_wise:
                # Only scalar gain and shift per feature map!
                for i, s in enumerate(cm_shapes):
                    if len(s) == 3:
                        cm_shapes[i] = [s[0], 1, 1]

            self._add_context_mod_layers(cm_shapes, cm_layers=cm_layer_inds)

        ################################################
        ### Define and initialize batch norm weights ###
        ################################################
        self._batchnorm_layers = nn.ModuleList() if use_batch_norm else None

        if use_batch_norm:
            if distill_bn_stats:
                self._hyper_shapes_distilled = []

            for i, s in enumerate(self._filter_sizes):
                if i == 0:
                    num = 1
                else:
                    num = 2*n

                for j in range(num):
                    bn_layer = BatchNormLayer(s, affine=not no_weights,
                        track_running_stats=bn_track_stats)
                    self._batchnorm_layers.append(bn_layer)

                    if distill_bn_stats:
                        self._hyper_shapes_distilled.extend( \
                            [list(p.shape) for p in bn_layer.get_stats(0)])

        # Note, method `_compute_hyper_shapes` doesn't take context-mod into
        # consideration.
        internal_weight_shapes = self._compute_hyper_shapes(no_weights=True)
        self._param_shapes.extend(internal_weight_shapes)
        # It's a bit hacky, as it was post-hoc integrated.
        # `internal_weight_shapes` contains first all batchnorm shapes, then
        # all conv layer shapes and finally the weights of the output layer.
        ii = 0
        if use_batch_norm:
            while True:
                if len(internal_weight_shapes[ii]) == 1:
                    self._param_shapes_meta.append({
                        'name': 'bn_scale' if ii % 2 == 0 else 'bn_shift',
                        'index': -1 if no_weights else \
                            len(self._internal_params)+ii,
                        'layer': -1 # TODO implement
                    })
                    ii += 1
                else:
                    break
        assert len(internal_weight_shapes[ii]) == 4
        while True:
            assert len(internal_weight_shapes[ii]) in [4, 2]
            self._param_shapes_meta.append({
                'name': 'weight',
                'index': -1 if no_weights else len(self._internal_params)+ii,
                'layer': -1 # TODO implement
            })
            if use_bias:
                self._param_shapes_meta.append({
                    'name': 'bias',
                    'index': -1 if no_weights else \
                        len(self._internal_params)+ii+1,
                    'layer': -1 # TODO implement
                })
            if len(internal_weight_shapes[ii]) == 2:
                break
            if use_bias:
                ii += 2
            else:
                ii += 1
        assert len(self._param_shapes) == len(self._param_shapes_meta)

        self._layer_weight_tensors = nn.ParameterList()
        self._layer_bias_vectors = nn.ParameterList()

        ###########################
        ### Print infos to user ###
        ###########################
        # Compute the total number of weights in this network and display
        # them to the user.
        # Note, this complicated calculation is not necessary as we can simply
        # count the number of weights afterwards. But it's an additional sanity
        # check for us.
        fs = self._filter_sizes
        num_weights = np.prod(self._kernel_size) * \
            (in_shape[2] * fs[0] + np.sum([fs[i] * fs[i+1] + \
                (2*n-1) * fs[i+1]**2 for i in range(3)])) + \
            ((fs[0] + 2*n * np.sum([fs[i] for i in range(1, 4)])) \
             if self.has_bias else 0) + \
            fs[-1] * num_classes + (num_classes if self.has_bias else 0)

        cm_num_weights = 2 * MainNetInterface.shapes_to_num_weights(cm_shapes) \
            if self._use_context_mod else 0
        num_weights += cm_num_weights

        if use_batch_norm:
            # The gamma and beta parameters of a batch norm layer are
            # learned as well.
            num_weights += 2 * (fs[0] + \
                                2*n*np.sum([fs[i] for i in range(1, 4)]))

        assert num_weights == self.num_params

        if verbose:
            print('A ResNet with %d layers and %d weights is created' \
                  % (6*n+2, num_weights)
                  + (' (including %d context-mod weights).' % cm_num_weights \
                     if cm_num_weights > 0 else '.')
                  + (' The network uses batchnorm.' if use_batch_norm  else ''))

        if no_weights:
            if self._hyper_shapes_learned is not None:
                prev_len = len(self._hyper_shapes_learned)
                # Context-mod weights might already be included.
                self._hyper_shapes_learned.extend(self._compute_hyper_shapes())
                new_len = len(self._hyper_shapes_learned)
                self._hyper_shapes_learned_ref.extend( \
                    list(range(prev_len, new_len)))

            self._is_properly_setup()
            return

        if use_batch_norm:
            for bn_layer in self._batchnorm_layers:
                self._internal_params.extend(bn_layer.weights)

        ############################################
        ### Define and initialize layer weights ###
        ###########################################
        ### Does not include context-mod or batchnorm weights.
        # First layer.
        self._layer_weight_tensors.append(nn.Parameter(
            torch.Tensor(self._filter_sizes[0], self._in_shape[2],
                *self._kernel_size),
            requires_grad=True))
        if self.has_bias:
            self._layer_bias_vectors.append(nn.Parameter(
                torch.Tensor(self._filter_sizes[0]), requires_grad=True))

        # Each block consists of 2n layers.
        for i in range(1, len(self._filter_sizes)):
            in_filters = self._filter_sizes[i-1]
            out_filters = self._filter_sizes[i]

            for _ in range(2*n):
                self._layer_weight_tensors.append(nn.Parameter(
                    torch.Tensor(out_filters, in_filters, *self._kernel_size),
                    requires_grad=True))
                if self.has_bias:
                    self._layer_bias_vectors.append(nn.Parameter(
                        torch.Tensor(out_filters), requires_grad=True))
                # Note, that the first layer in this block has potentially a
                # different number of input filters.
                in_filters = out_filters

        # After the average pooling, there is one more dense layer.
        self._layer_weight_tensors.append(nn.Parameter(
            torch.Tensor(num_classes, self._filter_sizes[-1]),
            requires_grad=True))
        if self.has_bias:
            self._layer_bias_vectors.append(nn.Parameter( \
                torch.Tensor(num_classes), requires_grad=True))

        # We add the weights interleaved, such that there are always consecutive
        # weight tensor and bias vector per layer. This fulfils the requirements
        # of attribute `mask_fc_out`.
        for i in range(len(self._layer_weight_tensors)):
            self._internal_params.append(self._layer_weight_tensors[i])
            if self.has_bias:
                self._internal_params.append(self._layer_bias_vectors[i])

        ### Initialize weights.
        if init_weights is not None:
            assert len(init_weights) == len(self.weights)
            for i in range(len(init_weights)):
                assert np.all(np.equal(list(init_weights[i].shape),
                                       self.weights[i].shape))
                self.weights[i].data = init_weights[i]
        else:
            for i in range(len(self._layer_weight_tensors)):
                init_params(self._layer_weight_tensors[i],
                    self._layer_bias_vectors[i] if self.has_bias else None)

        self._is_properly_setup()

    def forward(self, x, weights=None, distilled_params=None, condition=None):
        """Compute the output :math:`y` of this network given the input
        :math:`x`.

        Args:
            (....): See docstring of method
                :meth:`mnets.mnet_interface.MainNetInterface.forward`. We
                provide some more specific information below.
            x (torch.Tensor): Batch of flattened input images.

                .. note::
                    We assume the Tensorflow format, where the last entry
                    denotes the number of channels.
            weights (list or dict): If a list of parameter tensors is given and
                context modulation is used (see argument ``use_context_mod`` in
                constructor), then these parameters are interpreted as context-
                modulation parameters if the length of ``weights`` equals
                :code:`2*len(net.context_mod_layers)`. Otherwise, the length is
                expected to be equal to the length of the attribute
                :attr:`mnets.mnet_interface.MainNetInterface.param_shapes`.

                Alternatively, a dictionary can be passed with the possible
                keywords ``internal_weights`` and ``mod_weights``. Each keyword
                is expected to map onto a list of tensors.
                The keyword ``internal_weights`` refers to all weights of this
                network except for the weights of the context-modulation layers.
                The keyword ``mod_weights``, on the other hand, refers
                specifically to the weights of the context-modulation layers.
                It is not necessary to specify both keywords.
            distilled_params: Will be passed as ``running_mean`` and
                ``running_var`` arguments of method
                :meth:`utils.batchnorm_layer.BatchNormLayer.forward` if
                batch normalization is used.
            condition (optional, int or dict): If ``int`` is provided, then this
                argument will be passed as argument ``stats_id`` to the method
                :meth:`utils.batchnorm_layer.BatchNormLayer.forward` if
                batch normalization is used.

                If a ``dict`` is provided instead, the following keywords are
                allowed:

                    - ``bn_stats_id``: Will be handled as ``stats_id`` of the
                      batchnorm layers as described above.
                    - ``cmod_ckpt_id``: Will be passed as argument ``ckpt_id``
                      to the method
                      :meth:`utils.context_mod_layer.ContextModLayer.forward`.

        Returns:
            (torch.Tensor): The output of the network.
        """
        if ((not self._use_context_mod and self._no_weights) or \
                (self._no_weights or self._context_mod_no_weights)) and \
                weights is None:
            raise Exception('Network was generated without weights. ' +
                            'Hence, "weights" option may not be None.')

        ############################################
        ### Extract which weights should be used ###
        ############################################
        # I.e., are we using internally maintained weights or externally given
        # ones or are we even mixing between these groups.
        # FIXME code mostly copied from MLP forward method.
        n_cm = self._num_context_mod_shapes()

        if weights is None:
            weights = self.weights

            if self._use_context_mod:
                cm_weights = weights[:n_cm]
                int_weights = weights[n_cm:]
            else:
                int_weights = weights
        else:
            int_weights = None
            cm_weights = None

            if isinstance(weights, dict):
                assert('internal_weights' in weights.keys() or \
                       'mod_weights' in weights.keys())
                if 'internal_weights' in weights.keys():
                    int_weights = weights['internal_weights']
                if 'mod_weights' in weights.keys():
                    cm_weights = weights['mod_weights']
            else:
                if self._use_context_mod and \
                        len(weights) == n_cm:
                    cm_weights = weights
                else:
                    assert(len(weights) == len(self.param_shapes))
                    if self._use_context_mod:
                        cm_weights = weights[:n_cm]
                        int_weights = weights[n_cm:]
                    else:
                        int_weights = weights

            if self._use_context_mod and cm_weights is None:
                if self._context_mod_no_weights:
                    raise Exception('Network was generated without weights ' +
                        'for context-mod layers. Hence, they must be passed ' +
                        'via the "weights" option.')
                cm_weights = self.weights[:n_cm]
            if int_weights is None:
                if self._no_weights:
                    raise Exception('Network was generated without internal ' +
                        'weights. Hence, they must be passed via the ' +
                        '"weights" option.')
                if self._context_mod_no_weights:
                    int_weights = self.weights
                else:
                    int_weights = self.weights[n_cm:]

            # Note, context-mod weights might have different shapes, as they
            # may be parametrized on a per-sample basis.
            if self._use_context_mod:
                assert len(cm_weights) == self._num_context_mod_shapes()
            int_shapes = self.param_shapes[n_cm:]
            assert(len(int_weights) == len(int_shapes))
            for i, s in enumerate(int_shapes):
                assert(np.all(np.equal(s, list(int_weights[i].shape))))

        ########################
        ### Parse condition ###
        #######################

        bn_cond = None
        cmod_cond = None

        if condition is not None:
            if isinstance(condition, dict):
                assert('bn_stats_id' in condition.keys() or \
                       'cmod_ckpt_id' in condition.keys())
                if 'bn_stats_id' in condition.keys():
                    bn_cond = condition['bn_stats_id']
                if 'cmod_ckpt_id' in condition.keys():
                    cmod_cond = condition['cmod_ckpt_id']
            else:
                bn_cond = condition

        ######################################
        ### Select batchnorm running stats ###
        ######################################
        if self._use_batch_norm:
            # There are 6*n+1 layers that use batch normalization.
            lbw = 2 * (6 * self._n + 1)

            bn_weights = int_weights[:lbw]
            layer_weights = int_weights[lbw:]

            nn = len(self._batchnorm_layers)
            running_means = [None] * nn
            running_vars = [None] * nn
        else:
            layer_weights = int_weights

        if distilled_params is not None:
            if not self._distill_bn_stats:
                raise ValueError('Argument "distilled_params" can only be ' +
                                 'provided if the return value of ' +
                                 'method "distillation_targets()" is not None.')
            shapes = self.hyper_shapes_distilled
            assert(len(distilled_params) == len(shapes))
            for i, s in enumerate(shapes):
                assert(np.all(np.equal(s, list(distilled_params[i].shape))))

            # Extract batchnorm stats from distilled_params
            for i in range(0, len(distilled_params), 2):
                running_means[i//2] = distilled_params[i]
                running_vars[i//2] = distilled_params[i+1]

        elif self._use_batch_norm and self._bn_track_stats and \
                bn_cond is None:
            for i, bn_layer in enumerate(self._batchnorm_layers):
                running_means[i], running_vars[i] = bn_layer.get_stats()

        ###############################################
        ### Extract weight tensors and bias vectors ###
        ###############################################
        w_weights = []
        b_weights = [] if self.has_bias else [None] * len(layer_weights)
        for i, p in enumerate(layer_weights):
            if self.has_bias and i % 2 == 1:
                b_weights.append(p)
            else:
                w_weights.append(p)

        ###########################
        ### Forward Computation ###
        ###########################
        cm_ind = 0
        bn_ind = 0
        layer_ind = 0

        ### Helper function to process convolutional layers.
        def conv_layer(h, stride, shortcut=None):
            """Compute the output of a resnet conv layer including batchnorm,
            context-mod, non-linearity and shortcut.

            The order if the following:

            conv-layer -> context-mod (if pre-activation) -> batch-norm ->
            shortcut -> non-linearity -> context-mod (if post-activation)

            This method increments the indices ``layer_ind``, ``cm_ind`` and
            ``bn_ind``.

            Args:
                h: Input activity.
                stride: Stride of conv. layer (padding is set to 1).
                shortcut: If set, this tensor will be added to the activation
                    before the non-linearity is applied.

            Returns:
                Output of layer.
            """
            nonlocal layer_ind, cm_ind, bn_ind

            h = F.conv2d(h, w_weights[layer_ind], bias=b_weights[layer_ind],
                             stride=stride, padding=1)
            layer_ind += 1

            # Context-dependent modulation (pre-activation).
            if self._use_context_mod and \
                    not self._context_mod_post_activation:
                h = self._context_mod_layers[cm_ind].forward(h,
                    weights=cm_weights[2*cm_ind:2*cm_ind+2],
                    ckpt_id=cmod_cond)
                cm_ind += 1

            # Batch-norm
            if self._use_batch_norm:
                h = self._batchnorm_layers[bn_ind].forward(h,
                    running_mean=running_means[bn_ind],
                    running_var=running_vars[bn_ind],
                    weight=bn_weights[2*bn_ind],
                    bias=bn_weights[2*bn_ind+1], stats_id=bn_cond)
                bn_ind += 1

            # Note, as can be seen in figure 5 of the original paper, the
            # shortcut is performed before the ReLU is applied.
            if shortcut is not None:
                h += shortcut

            # Non-linearity
            h = F.relu(h)

            # Context-dependent modulation (post-activation).
            if self._use_context_mod and self._context_mod_post_activation:
                h = self._context_mod_layers[cm_ind].forward(h,
                    weights=cm_weights[2*cm_ind:2*cm_ind+2],
                    ckpt_id=cmod_cond)
                cm_ind += 1

            return h

        x = x.view(-1, *self._in_shape)
        x = x.permute(0, 3, 1, 2)
        h = x

        # Context-dependent modulation of inputs directly.
        if self._use_context_mod and self._context_mod_inputs:
            h = self._context_mod_layers[cm_ind].forward(h,
                weights=cm_weights[2*cm_ind:2*cm_ind+2], ckpt_id=cmod_cond)
            cm_ind += 1

        ### Initial convolutional layer.
        h = conv_layer(h, 1, shortcut=None)

        ### Three groups, each containing n resnet blocks.
        for i in range(3):
            # Only the first layer in a group may be a strided convolution.
            if i == 0:
                stride = 1
            else:
                stride = 2

            fs = self._filter_sizes[i+1]
            # For each resnet block. A resnet block consists of 2 convolutional
            # layers.
            for j in range(self._n):
                shortcut_h = h
                if j == 0 and fs != self._filter_sizes[i]:
                    # The original paper uses zero padding for added output
                    # feature dimensions. Since we apply a strided conv, we
                    # additionally have to subsample the input.
                    # This implementation is motivated by
                    #    https://git.io/fhcfk
                    # FIXME I guess it is a nicer solution to use 1x1
                    # convolutions to increase/decrease the number of channels.
                    # Note, this would add more layers (and trainable weights)
                    # to the network. Hence, the statement, that this networks
                    # has `6n+2` layers might be invalid.
                    fs_prev = self._filter_sizes[i]
                    pad_left = (fs - fs_prev) // 2
                    pad_right = int(np.ceil((fs - fs_prev) / 2))
                    if stride == 2:
                        shortcut_h = h[:, :, ::2, ::2]
                    shortcut_h = F.pad(shortcut_h,
                        (0, 0, 0, 0, pad_left, pad_right), "constant", 0)


                h = conv_layer(h, stride, shortcut=None)

                stride = 1

                h = conv_layer(h, stride, shortcut=shortcut_h)

        ### Average pool all activities within a feature map.
        h = F.avg_pool2d(h, [h.size()[2], h.size()[3]])
        h = h.view(h.size(0), -1)

        ### Apply final fully-connected layer and compute outputs.
        h = F.linear(h, w_weights[layer_ind], bias=b_weights[layer_ind])

        # Context-dependent modulation in output layer.
        if self._use_context_mod and not self._no_last_layer_context_mod:
            h = self._context_mod_layers[cm_ind].forward(h,
                weights=cm_weights[2*cm_ind:2*cm_ind+2], ckpt_id=cmod_cond)

        return h

    def _compute_hyper_shapes(self, no_weights=None):
        r"""Helper function to compute weight shapes of this network for
        externally maintained weights.

        Returns a list of lists of integers denoting the shape of every
        weight tensor that is not a trainable parameter of this network (i.e.,
        those weight tensors whose shapes are specified in
        :attr:`mnets.mnet_interface.MainNetInterface.hyper_shapes_distilled`).

        If batchnorm layers are used, then the first :math:`2 * (6n+1)` lists
        will denote the shapes of the batchnorm weights
        :math:`[\gamma_1, \beta_1, \gamma_2, ..., \beta_{6n+1}]`.

        The remaining :math:`2 * (6n+2)` entries are weight tensors and bias
        vectors of each convolutional or fully-connected (last two entries)
        layer in this network.

        Args:
            no_weights (optional): If specified, it will overwrite the private
                member :code:`self._no_weights`.

                If set to ``True``, then all weight shapes of the network
                are computed independent of whether they are maintained
                internally or externally.

        Returns:
            A list of lists of integers.
        """
        if no_weights is None:
            no_weights = self._no_weights

        ret = []
        if no_weights is False:
            return ret

        fs = self._filter_sizes
        ks = self._kernel_size
        n = self._n

        if self._use_batch_norm:
            for i, s in enumerate(fs):
                if i == 0:
                    num = 1
                else:
                    num = 2*n

                for _ in range(2*num):
                    ret.append([s])

        f_in = self._in_shape[-1]
        for i, s in enumerate(fs):
            f_out = s
            if i == 0:
                num = 1
            else:
                num = 2*n

            for _ in range(num):
                ret.append([f_out, f_in, *ks])
                if self.has_bias:
                    ret.append([f_out])
                f_in = f_out
        ret.append([self._num_classes, fs[-1]])
        if self.has_bias:
            ret.append([self._num_classes])

        return ret

    def distillation_targets(self):
        """Targets to be distilled after training.

        See docstring of abstract super method
        :meth:`mnets.mnet_interface.MainNetInterface.distillation_targets`.

        This method will return the current batch statistics of all batch
        normalization layers if ``distill_bn_stats`` and ``use_batch_norm``
        were set to ``True`` in the constructor.

        Returns:
            The target tensors corresponding to the shapes specified in
            attribute :attr:`hyper_shapes_distilled`.
        """
        if self.hyper_shapes_distilled is None:
            return None

        ret = []
        for bn_layer in self._batchnorm_layers:
            ret.extend(bn_layer.get_stats())

        return ret

    def _compute_layer_out_sizes(self):
        """Compute the output shapes of all layers in this network.

        This method will compute the output shape of each layer in this network,
        including the output layer, which just corresponds to the number of
        classes.

        Returns:
            (list): A list of shapes (lists of integers). The first entry will
            correspond to the shape of the output of the first convolutional
            layer. The last entry will correspond to the output shape.

            .. note:
                Output shapes of convolutional layers will adhere PyTorch
                convention, i.e., ``[C, H, W]``, where ``C`` denotes the channel
                dimension.
        """
        in_shape = self._in_shape
        fs = self._filter_sizes
        ks = self._kernel_size
        pd = 1 # all paddings are 1.
        assert len(ks) == 2
        assert len(fs) == 4
        n = self._n

        # Note, `in_shape` is in Tensorflow layout.
        assert(len(in_shape) == 3)
        in_shape = [in_shape[2], *in_shape[:2]]

        ret = []

        C, H, W = in_shape

        # Recall the formular for convolutional layers:
        # W_new = (W - K + 2P) // S + 1

        # First conv layer (stride 1).
        C = fs[0]
        H = (H - ks[0] + 2*pd) // 1 + 1
        W = (W - ks[1] + 2*pd) // 1 + 1
        ret.append([C, H, W])

        # First block (no strides).
        C = fs[1]
        H = (H - ks[0] + 2*pd) // 1 + 1
        W = (W - ks[1] + 2*pd) // 1 + 1
        ret.extend([[C, H, W]] * (2*n))

        # Second block (first layer has stride 2).
        C = fs[2]
        H = (H - ks[0] + 2*pd) // 2 + 1
        W = (W - ks[1] + 2*pd) // 2 + 1
        ret.extend([[C, H, W]] * (2*n))

        # Third block (first layer has stride 2).
        C = fs[3]
        H = (H - ks[0] + 2*pd) // 2 + 1
        W = (W - ks[1] + 2*pd) // 2 + 1
        ret.extend([[C, H, W]] * (2*n))

        # Final fully-connected layer (after avg pooling), i.e., output size.
        ret.append([self._num_classes])

        assert len(ret) == 6*n + 2

        return ret

if __name__ == '__main__':
    pass