i3d_tf1.py

# Copyright 2017 Google Inc.
#
# 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
#
#     https://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.
# ============================================================================
"""Inception-v1 Inflated 3D ConvNet used for Kinetics CVPR paper.
The model is introduced in:
  Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset
  Joao Carreira, Andrew Zisserman
  https://arxiv.org/pdf/1705.07750v1.pdf.
"""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import sonnet as snt
import tensorflow as tf


class Unit3D(snt.Module):
  """Basic unit containing Conv3D + BatchNorm + non-linearity."""

  def __init__(self, output_channels,
               kernel_shape=(1, 1, 1),
               stride=(1, 1, 1),
               activation_fn=tf.nn.relu,
               use_batch_norm=True,
               use_bias=False,
               name='unit_3d'):
    """Initializes Unit3D module."""
    super(Unit3D, self).__init__(name=name)
    self._output_channels = output_channels
    self._kernel_shape = kernel_shape
    self._stride = stride
    self._use_batch_norm = use_batch_norm
    self._activation_fn = activation_fn
    self._use_bias = use_bias

  def _build(self, inputs, is_training):
    """Connects the module to inputs.
    Args:
      inputs: Inputs to the Unit3D component.
      is_training: whether to use training mode for snt.BatchNorm (boolean).
    Returns:
      Outputs from the module.
    """
    net = snt.Conv3D(output_channels=self._output_channels,
                     kernel_shape=self._kernel_shape,
                     stride=self._stride,
                     padding=snt.pad.same,
                     with_bias=self._use_bias)(inputs)
    if self._use_batch_norm:
      bn = snt.BatchNorm(create_scale=False, create_offset=False)
      net = bn(net, is_training=is_training, test_local_stats=False)
    if self._activation_fn is not None:
      net = self._activation_fn(net)
    return net


class InceptionI3d(snt.Module):
  """Inception-v1 I3D architecture.
  The model is introduced in:
    Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset
    Joao Carreira, Andrew Zisserman
    https://arxiv.org/pdf/1705.07750v1.pdf.
  See also the Inception architecture, introduced in:
    Going deeper with convolutions
    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed,
    Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich.
    http://arxiv.org/pdf/1409.4842v1.pdf.
  """

  # Endpoints of the model in order. During construction, all the endpoints up
  # to a designated `final_endpoint` are returned in a dictionary as the
  # second return value.
  VALID_ENDPOINTS = (
      'Conv3d_1a_7x7',
      'MaxPool3d_2a_3x3',
      'Conv3d_2b_1x1',
      'Conv3d_2c_3x3',
      'MaxPool3d_3a_3x3',
      'Mixed_3b',
      'Mixed_3c',
      'MaxPool3d_4a_3x3',
      'Mixed_4b',
      'Mixed_4c',
      'Mixed_4d',
      'Mixed_4e',
      'Mixed_4f',
      'MaxPool3d_5a_2x2',
      'Mixed_5b',
      'Mixed_5c',
      'Logits',
      'Predictions',
  )

  def __init__(self, num_classes=400, spatial_squeeze=True,
               final_endpoint='Logits', name='inception_i3d'):
    """Initializes I3D model instance.
    Args:
      num_classes: The number of outputs in the logit layer (default 400, which
          matches the Kinetics dataset).
      spatial_squeeze: Whether to squeeze the spatial dimensions for the logits
          before returning (default True).
      final_endpoint: The model contains many possible endpoints.
          `final_endpoint` specifies the last endpoint for the model to be built
          up to. In addition to the output at `final_endpoint`, all the outputs
          at endpoints up to `final_endpoint` will also be returned, in a
          dictionary. `final_endpoint` must be one of
          InceptionI3d.VALID_ENDPOINTS (default 'Logits').
      name: A string (optional). The name of this module.
    Raises:
      ValueError: if `final_endpoint` is not recognized.
    """

    if final_endpoint not in self.VALID_ENDPOINTS:
      raise ValueError('Unknown final endpoint %s' % final_endpoint)

    super(InceptionI3d, self).__init__(name=name)
    self._num_classes = num_classes
    self._spatial_squeeze = spatial_squeeze
    self._final_endpoint = final_endpoint

  def _build(self, inputs, is_training, dropout_keep_prob=1.0):
    """Connects the model to inputs.
    Args:
      inputs: Inputs to the model, which should have dimensions
          `batch_size` x `num_frames` x 224 x 224 x `num_channels`.
      is_training: whether to use training mode for snt.BatchNorm (boolean).
      dropout_keep_prob: Probability for the tf.nn.dropout layer (float in
          [0, 1)).
    Returns:
      A tuple consisting of:
        1. Network output at location `self._final_endpoint`.
        2. Dictionary containing all endpoints up to `self._final_endpoint`,
           indexed by endpoint name.
    Raises:
      ValueError: if `self._final_endpoint` is not recognized.
    """
    if self._final_endpoint not in self.VALID_ENDPOINTS:
      raise ValueError('Unknown final endpoint %s' % self._final_endpoint)

    net = inputs
    end_points = {}
    end_point = 'Conv3d_1a_7x7'
    net = Unit3D(output_channels=64, kernel_shape=[7, 7, 7],
                 stride=[2, 2, 2], name=end_point)._build(net, is_training=is_training)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points
    end_point = 'MaxPool3d_2a_3x3'
    net = tf.nn.max_pool3d(net, ksize=[1, 1, 3, 3, 1], strides=[1, 1, 2, 2, 1],
                           padding=snt.pad.same, name=end_point)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points
    end_point = 'Conv3d_2b_1x1'
    net = Unit3D(output_channels=64, kernel_shape=[1, 1, 1],
                 name=end_point)(net, is_training=is_training)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points
    end_point = 'Conv3d_2c_3x3'
    net = Unit3D(output_channels=192, kernel_shape=[3, 3, 3],
                 name=end_point)(net, is_training=is_training)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points
    end_point = 'MaxPool3d_3a_3x3'
    net = tf.nn.max_pool3d(net, ksize=[1, 1, 3, 3, 1], strides=[1, 1, 2, 2, 1],
                           padding=snt.pad.same, name=end_point)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points

    end_point = 'Mixed_3b'
    with tf.variable_scope(end_point):
      with tf.variable_scope('Branch_0'):
        branch_0 = Unit3D(output_channels=64, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
      with tf.variable_scope('Branch_1'):
        branch_1 = Unit3D(output_channels=96, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_1 = Unit3D(output_channels=128, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_1,
                                                is_training=is_training)
      with tf.variable_scope('Branch_2'):
        branch_2 = Unit3D(output_channels=16, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_2 = Unit3D(output_channels=32, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_2,
                                                is_training=is_training)
      with tf.variable_scope('Branch_3'):
        branch_3 = tf.nn.max_pool3d(net, ksize=[1, 3, 3, 3, 1],
                                    strides=[1, 1, 1, 1, 1], padding=snt.pad.same,
                                    name='MaxPool3d_0a_3x3')
        branch_3 = Unit3D(output_channels=32, kernel_shape=[1, 1, 1],
                          name='Conv3d_0b_1x1')(branch_3,
                                                is_training=is_training)

      net = tf.concat([branch_0, branch_1, branch_2, branch_3], 4)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points

    end_point = 'Mixed_3c'
    with tf.variable_scope(end_point):
      with tf.variable_scope('Branch_0'):
        branch_0 = Unit3D(output_channels=128, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
      with tf.variable_scope('Branch_1'):
        branch_1 = Unit3D(output_channels=128, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_1 = Unit3D(output_channels=192, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_1,
                                                is_training=is_training)
      with tf.variable_scope('Branch_2'):
        branch_2 = Unit3D(output_channels=32, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_2 = Unit3D(output_channels=96, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_2,
                                                is_training=is_training)
      with tf.variable_scope('Branch_3'):
        branch_3 = tf.nn.max_pool3d(net, ksize=[1, 3, 3, 3, 1],
                                    strides=[1, 1, 1, 1, 1], padding=snt.pad.same,
                                    name='MaxPool3d_0a_3x3')
        branch_3 = Unit3D(output_channels=64, kernel_shape=[1, 1, 1],
                          name='Conv3d_0b_1x1')(branch_3,
                                                is_training=is_training)
      net = tf.concat([branch_0, branch_1, branch_2, branch_3], 4)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points

    end_point = 'MaxPool3d_4a_3x3'
    net = tf.nn.max_pool3d(net, ksize=[1, 3, 3, 3, 1], strides=[1, 2, 2, 2, 1],
                           padding=snt.pad.same, name=end_point)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points

    end_point = 'Mixed_4b'
    with tf.variable_scope(end_point):
      with tf.variable_scope('Branch_0'):
        branch_0 = Unit3D(output_channels=192, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
      with tf.variable_scope('Branch_1'):
        branch_1 = Unit3D(output_channels=96, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_1 = Unit3D(output_channels=208, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_1,
                                                is_training=is_training)
      with tf.variable_scope('Branch_2'):
        branch_2 = Unit3D(output_channels=16, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_2 = Unit3D(output_channels=48, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_2,
                                                is_training=is_training)
      with tf.variable_scope('Branch_3'):
        branch_3 = tf.nn.max_pool3d(net, ksize=[1, 3, 3, 3, 1],
                                    strides=[1, 1, 1, 1, 1], padding=snt.pad.same,
                                    name='MaxPool3d_0a_3x3')
        branch_3 = Unit3D(output_channels=64, kernel_shape=[1, 1, 1],
                          name='Conv3d_0b_1x1')(branch_3,
                                                is_training=is_training)
      net = tf.concat([branch_0, branch_1, branch_2, branch_3], 4)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points

    end_point = 'Mixed_4c'
    with tf.variable_scope(end_point):
      with tf.variable_scope('Branch_0'):
        branch_0 = Unit3D(output_channels=160, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
      with tf.variable_scope('Branch_1'):
        branch_1 = Unit3D(output_channels=112, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_1 = Unit3D(output_channels=224, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_1,
                                                is_training=is_training)
      with tf.variable_scope('Branch_2'):
        branch_2 = Unit3D(output_channels=24, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_2 = Unit3D(output_channels=64, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_2,
                                                is_training=is_training)
      with tf.variable_scope('Branch_3'):
        branch_3 = tf.nn.max_pool3d(net, ksize=[1, 3, 3, 3, 1],
                                    strides=[1, 1, 1, 1, 1], padding=snt.pad.same,
                                    name='MaxPool3d_0a_3x3')
        branch_3 = Unit3D(output_channels=64, kernel_shape=[1, 1, 1],
                          name='Conv3d_0b_1x1')(branch_3,
                                                is_training=is_training)
      net = tf.concat([branch_0, branch_1, branch_2, branch_3], 4)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points

    end_point = 'Mixed_4d'
    with tf.variable_scope(end_point):
      with tf.variable_scope('Branch_0'):
        branch_0 = Unit3D(output_channels=128, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
      with tf.variable_scope('Branch_1'):
        branch_1 = Unit3D(output_channels=128, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_1 = Unit3D(output_channels=256, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_1,
                                                is_training=is_training)
      with tf.variable_scope('Branch_2'):
        branch_2 = Unit3D(output_channels=24, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_2 = Unit3D(output_channels=64, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_2,
                                                is_training=is_training)
      with tf.variable_scope('Branch_3'):
        branch_3 = tf.nn.max_pool3d(net, ksize=[1, 3, 3, 3, 1],
                                    strides=[1, 1, 1, 1, 1], padding=snt.pad.same,
                                    name='MaxPool3d_0a_3x3')
        branch_3 = Unit3D(output_channels=64, kernel_shape=[1, 1, 1],
                          name='Conv3d_0b_1x1')(branch_3,
                                                is_training=is_training)
      net = tf.concat([branch_0, branch_1, branch_2, branch_3], 4)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points

    end_point = 'Mixed_4e'
    with tf.variable_scope(end_point):
      with tf.variable_scope('Branch_0'):
        branch_0 = Unit3D(output_channels=112, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
      with tf.variable_scope('Branch_1'):
        branch_1 = Unit3D(output_channels=144, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_1 = Unit3D(output_channels=288, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_1,
                                                is_training=is_training)
      with tf.variable_scope('Branch_2'):
        branch_2 = Unit3D(output_channels=32, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_2 = Unit3D(output_channels=64, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_2,
                                                is_training=is_training)
      with tf.variable_scope('Branch_3'):
        branch_3 = tf.nn.max_pool3d(net, ksize=[1, 3, 3, 3, 1],
                                    strides=[1, 1, 1, 1, 1], padding=snt.pad.same,
                                    name='MaxPool3d_0a_3x3')
        branch_3 = Unit3D(output_channels=64, kernel_shape=[1, 1, 1],
                          name='Conv3d_0b_1x1')(branch_3,
                                                is_training=is_training)
      net = tf.concat([branch_0, branch_1, branch_2, branch_3], 4)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points

    end_point = 'Mixed_4f'
    with tf.variable_scope(end_point):
      with tf.variable_scope('Branch_0'):
        branch_0 = Unit3D(output_channels=256, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
      with tf.variable_scope('Branch_1'):
        branch_1 = Unit3D(output_channels=160, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_1 = Unit3D(output_channels=320, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_1,
                                                is_training=is_training)
      with tf.variable_scope('Branch_2'):
        branch_2 = Unit3D(output_channels=32, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_2 = Unit3D(output_channels=128, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_2,
                                                is_training=is_training)
      with tf.variable_scope('Branch_3'):
        branch_3 = tf.nn.max_pool3d(net, ksize=[1, 3, 3, 3, 1],
                                    strides=[1, 1, 1, 1, 1], padding=snt.pad.same,
                                    name='MaxPool3d_0a_3x3')
        branch_3 = Unit3D(output_channels=128, kernel_shape=[1, 1, 1],
                          name='Conv3d_0b_1x1')(branch_3,
                                                is_training=is_training)
      net = tf.concat([branch_0, branch_1, branch_2, branch_3], 4)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points

    end_point = 'MaxPool3d_5a_2x2'
    net = tf.nn.max_pool3d(net, ksize=[1, 2, 2, 2, 1], strides=[1, 2, 2, 2, 1],
                           padding=snt.pad.same, name=end_point)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points

    end_point = 'Mixed_5b'
    with tf.variable_scope(end_point):
      with tf.variable_scope('Branch_0'):
        branch_0 = Unit3D(output_channels=256, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
      with tf.variable_scope('Branch_1'):
        branch_1 = Unit3D(output_channels=160, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_1 = Unit3D(output_channels=320, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_1,
                                                is_training=is_training)
      with tf.variable_scope('Branch_2'):
        branch_2 = Unit3D(output_channels=32, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_2 = Unit3D(output_channels=128, kernel_shape=[3, 3, 3],
                          name='Conv3d_0a_3x3')(branch_2,
                                                is_training=is_training)
      with tf.variable_scope('Branch_3'):
        branch_3 = tf.nn.max_pool3d(net, ksize=[1, 3, 3, 3, 1],
                                    strides=[1, 1, 1, 1, 1], padding=snt.pad.same,
                                    name='MaxPool3d_0a_3x3')
        branch_3 = Unit3D(output_channels=128, kernel_shape=[1, 1, 1],
                          name='Conv3d_0b_1x1')(branch_3,
                                                is_training=is_training)
      net = tf.concat([branch_0, branch_1, branch_2, branch_3], 4)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points

    end_point = 'Mixed_5c'
    with tf.variable_scope(end_point):
      with tf.variable_scope('Branch_0'):
        branch_0 = Unit3D(output_channels=384, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
      with tf.variable_scope('Branch_1'):
        branch_1 = Unit3D(output_channels=192, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_1 = Unit3D(output_channels=384, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_1,
                                                is_training=is_training)
      with tf.variable_scope('Branch_2'):
        branch_2 = Unit3D(output_channels=48, kernel_shape=[1, 1, 1],
                          name='Conv3d_0a_1x1')(net, is_training=is_training)
        branch_2 = Unit3D(output_channels=128, kernel_shape=[3, 3, 3],
                          name='Conv3d_0b_3x3')(branch_2,
                                                is_training=is_training)
      with tf.variable_scope('Branch_3'):
        branch_3 = tf.nn.max_pool3d(net, ksize=[1, 3, 3, 3, 1],
                                    strides=[1, 1, 1, 1, 1], padding=snt.pad.same,
                                    name='MaxPool3d_0a_3x3')
        branch_3 = Unit3D(output_channels=128, kernel_shape=[1, 1, 1],
                          name='Conv3d_0b_1x1')(branch_3,
                                                is_training=is_training)
      net = tf.concat([branch_0, branch_1, branch_2, branch_3], 4)
    end_points[end_point] = net
    if self._final_endpoint == end_point: return net, end_points

    end_point = 'Logits'
    with tf.variable_scope(end_point):
      net = tf.nn.avg_pool3d(net, ksize=[1, 2, 7, 7, 1],
                             strides=[1, 1, 1, 1, 1], padding=snt.VALID)
      net = tf.nn.dropout(net, dropout_keep_prob)
      logits = Unit3D(output_channels=self._num_classes,
                      kernel_shape=[1, 1, 1],
                      activation_fn=None,
                      use_batch_norm=False,
                      use_bias=True,
                      name='Conv3d_0c_1x1')(net, is_training=is_training)
      if self._spatial_squeeze:
        logits = tf.squeeze(logits, [2, 3], name='SpatialSqueeze')
    averaged_logits = tf.reduce_mean(logits, axis=1)
    end_points[end_point] = averaged_logits
    if self._final_endpoint == end_point: return averaged_logits, end_points

    end_point = 'Predictions'
    predictions = tf.nn.softmax(averaged_logits)
    end_points[end_point] = predictions
    return predictions, end_points