def i3d_base(inputs, final_endpoint='Mixed_5c', scope='InceptionV1'):
"""Defines the I3D base architecture.
Note that we use the names as defined in Inception V1 to facilitate checkpoint
conversion from an image-trained Inception V1 checkpoint to I3D checkpoint.
Args:
inputs: A 5-D float tensor of size [batch_size, num_frames, height, width,
channels].
final_endpoint: Specifies the endpoint to construct the network up to. It
can be one of ['Conv2d_1a_7x7', 'MaxPool_2a_3x3', 'Conv2d_2b_1x1',
'Conv2d_2c_3x3', 'MaxPool_3a_3x3', 'Mixed_3b', 'Mixed_3c',
'MaxPool_4a_3x3', 'Mixed_4b', 'Mixed_4c', 'Mixed_4d', 'Mixed_4e',
'Mixed_4f', 'MaxPool_5a_2x2', 'Mixed_5b', 'Mixed_5c']
scope: Optional variable_scope.
Returns:
A dictionary from components of the network to the corresponding activation.
Raises:
ValueError: if final_endpoint is not set to one of the predefined values.
"""
return s3dg.s3dg_base(inputs,
first_temporal_kernel_size=7,
temporal_conv_startat='Conv2d_2c_3x3',
gating_startat=None,
final_endpoint=final_endpoint,
min_depth=16,
depth_multiplier=1.0,
data_format='NDHWC',
scope=scope)
def testBuildAndCheckAllEndPointsUptoMixed5c(self):
batch_size = 5
num_frames = 64
height, width = 224, 224
inputs = tf.random_uniform((batch_size, num_frames, height, width, 3))
_, end_points = s3dg.s3dg_base(inputs,
final_endpoint='Mixed_5c')
endpoints_shapes = {'Conv2d_1a_7x7': [5, 32, 112, 112, 64],
'MaxPool_2a_3x3': [5, 32, 56, 56, 64],
'Conv2d_2b_1x1': [5, 32, 56, 56, 64],
'Conv2d_2c_3x3': [5, 32, 56, 56, 192],
'MaxPool_3a_3x3': [5, 32, 28, 28, 192],
'Mixed_3b': [5, 32, 28, 28, 256],
'Mixed_3c': [5, 32, 28, 28, 480],
'MaxPool_4a_3x3': [5, 16, 14, 14, 480],
'Mixed_4b': [5, 16, 14, 14, 512],
'Mixed_4c': [5, 16, 14, 14, 512],
'Mixed_4d': [5, 16, 14, 14, 512],
'Mixed_4e': [5, 16, 14, 14, 528],
'Mixed_4f': [5, 16, 14, 14, 832],
'MaxPool_5a_2x2': [5, 8, 7, 7, 832],
'Mixed_5b': [5, 8, 7, 7, 832],
'Mixed_5c': [5, 8, 7, 7, 1024]}
self.assertItemsEqual(endpoints_shapes.keys(), end_points.keys())
for endpoint_name, expected_shape in endpoints_shapes.iteritems():
self.assertTrue(endpoint_name in end_points)
self.assertListEqual(end_points[endpoint_name].get_shape().as_list(),
expected_shape)
def testBuildAndCheckAllEndPointsUptoMixed5c(self):
batch_size = 5
num_frames = 64
height, width = 224, 224
inputs = tf.random.uniform((batch_size, num_frames, height, width, 3))
_, end_points = s3dg.s3dg_base(inputs, final_endpoint='Mixed_5c')
endpoints_shapes = {
'Conv2d_1a_7x7': [5, 32, 112, 112, 64],
'MaxPool_2a_3x3': [5, 32, 56, 56, 64],
'Conv2d_2b_1x1': [5, 32, 56, 56, 64],
'Conv2d_2c_3x3': [5, 32, 56, 56, 192],
'MaxPool_3a_3x3': [5, 32, 28, 28, 192],
'Mixed_3b': [5, 32, 28, 28, 256],
'Mixed_3c': [5, 32, 28, 28, 480],
'MaxPool_4a_3x3': [5, 16, 14, 14, 480],
'Mixed_4b': [5, 16, 14, 14, 512],
'Mixed_4c': [5, 16, 14, 14, 512],
'Mixed_4d': [5, 16, 14, 14, 512],
'Mixed_4e': [5, 16, 14, 14, 528],
'Mixed_4f': [5, 16, 14, 14, 832],
'MaxPool_5a_2x2': [5, 8, 7, 7, 832],
'Mixed_5b': [5, 8, 7, 7, 832],
'Mixed_5c': [5, 8, 7, 7, 1024]
}
self.assertItemsEqual(list(endpoints_shapes.keys()),
list(end_points.keys()))
for endpoint_name, expected_shape in six.iteritems(endpoints_shapes):
self.assertTrue(endpoint_name in end_points)
self.assertListEqual(
end_points[endpoint_name].get_shape().as_list(),
expected_shape)
def i3d_base(inputs, final_endpoint='Mixed_5c',
scope='InceptionV1'):
"""Defines the I3D base architecture.
Note that we use the names as defined in Inception V1 to facilitate checkpoint
conversion from an image-trained Inception V1 checkpoint to I3D checkpoint.
Args:
inputs: A 5-D float tensor of size [batch_size, num_frames, height, width,
channels].
final_endpoint: Specifies the endpoint to construct the network up to. It
can be one of ['Conv2d_1a_7x7', 'MaxPool_2a_3x3', 'Conv2d_2b_1x1',
'Conv2d_2c_3x3', 'MaxPool_3a_3x3', 'Mixed_3b', 'Mixed_3c',
'MaxPool_4a_3x3', 'Mixed_4b', 'Mixed_4c', 'Mixed_4d', 'Mixed_4e',
'Mixed_4f', 'MaxPool_5a_2x2', 'Mixed_5b', 'Mixed_5c']
scope: Optional variable_scope.
Returns:
A dictionary from components of the network to the corresponding activation.
Raises:
ValueError: if final_endpoint is not set to one of the predefined values.
"""
return s3dg.s3dg_base(
inputs,
first_temporal_kernel_size=7,
temporal_conv_startat='Conv2d_2c_3x3',
gating_startat=None,
final_endpoint=final_endpoint,
min_depth=16,
depth_multiplier=1.0,
data_format='NDHWC',
scope=scope)
def testTenFrames(self):
batch_size = 5
num_frames = 10
height, width = 224, 224
inputs = tf.random.uniform((batch_size, num_frames, height, width, 3))
mixed_5c, _ = s3dg.s3dg_base(inputs)
self.assertTrue(mixed_5c.op.name.startswith('InceptionV1/Mixed_5c'))
self.assertListEqual(mixed_5c.get_shape().as_list(),
[batch_size, 2, 7, 7, 1024])
def testHalfSizeImages(self):
batch_size = 5
num_frames = 64
height, width = 112, 112
inputs = tf.random.uniform((batch_size, num_frames, height, width, 3))
mixed_5c, _ = s3dg.s3dg_base(inputs)
self.assertTrue(mixed_5c.op.name.startswith('InceptionV1/Mixed_5c'))
self.assertListEqual(mixed_5c.get_shape().as_list(),
[batch_size, 8, 4, 4, 1024])
def testTenFrames(self):
batch_size = 5
num_frames = 10
height, width = 224, 224
inputs = tf.random_uniform((batch_size, num_frames, height, width, 3))
mixed_5c, _ = s3dg.s3dg_base(inputs)
self.assertTrue(mixed_5c.op.name.startswith('InceptionV1/Mixed_5c'))
self.assertListEqual(mixed_5c.get_shape().as_list(),
[batch_size, 2, 7, 7, 1024])
def testHalfSizeImages(self):
batch_size = 5
num_frames = 64
height, width = 112, 112
inputs = tf.random_uniform((batch_size, num_frames, height, width, 3))
mixed_5c, _ = s3dg.s3dg_base(inputs)
self.assertTrue(mixed_5c.op.name.startswith('InceptionV1/Mixed_5c'))
self.assertListEqual(mixed_5c.get_shape().as_list(),
[batch_size, 8, 4, 4, 1024])
def testBuildBaseNetwork(self):
batch_size = 5
num_frames = 64
height, width = 224, 224
inputs = tf.random_uniform((batch_size, num_frames, height, width, 3))
mixed_6c, end_points = s3dg.s3dg_base(inputs)
self.assertTrue(mixed_6c.op.name.startswith('InceptionV1/Mixed_5c'))
self.assertListEqual(mixed_6c.get_shape().as_list(),
[batch_size, 8, 7, 7, 1024])
expected_endpoints = ['Conv2d_1a_7x7', 'MaxPool_2a_3x3', 'Conv2d_2b_1x1',
'Conv2d_2c_3x3', 'MaxPool_3a_3x3', 'Mixed_3b',
'Mixed_3c', 'MaxPool_4a_3x3', 'Mixed_4b', 'Mixed_4c',
'Mixed_4d', 'Mixed_4e', 'Mixed_4f', 'MaxPool_5a_2x2',
'Mixed_5b', 'Mixed_5c']
self.assertItemsEqual(end_points.keys(), expected_endpoints)
def testBuildOnlyUptoFinalEndpointNoGating(self):
batch_size = 5
num_frames = 64
height, width = 224, 224
endpoints = ['Conv2d_1a_7x7', 'MaxPool_2a_3x3', 'Conv2d_2b_1x1',
'Conv2d_2c_3x3', 'MaxPool_3a_3x3', 'Mixed_3b', 'Mixed_3c',
'MaxPool_4a_3x3', 'Mixed_4b', 'Mixed_4c', 'Mixed_4d',
'Mixed_4e', 'Mixed_4f', 'MaxPool_5a_2x2', 'Mixed_5b',
'Mixed_5c']
for index, endpoint in enumerate(endpoints):
with tf.Graph().as_default():
inputs = tf.random_uniform((batch_size, num_frames, height, width, 3))
out_tensor, end_points = s3dg.s3dg_base(
inputs, final_endpoint=endpoint, gating_startat=None)
print(endpoint, out_tensor.op.name)
self.assertTrue(out_tensor.op.name.startswith(
'InceptionV1/' + endpoint))
self.assertItemsEqual(endpoints[:index+1], end_points)
