def _construct_network(self, cfg):
"""
Builds a single pathway R(2+1)D model.
Args:
cfg (CfgNode): model building configs, details are in the
comments of the config file.
"""
self.model = create_r2plus1d(
# Input clip configs.
input_channel=cfg.DATA.INPUT_CHANNEL_NUM[0],
# Model configs.
model_depth=cfg.RESNET.DEPTH,
model_num_class=cfg.MODEL.NUM_CLASSES,
dropout_rate=cfg.MODEL.DROPOUT_RATE,
# Normalization configs.
norm=get_norm(cfg),
norm_eps=1e-5,
norm_momentum=0.1,
# Activation configs.
activation=partial(nn.ReLU, inplace=cfg.RESNET.INPLACE_RELU),
# Stem configs.
stem_dim_out=cfg.RESNET.WIDTH_PER_GROUP,
stem_conv_kernel_size=(1, 7, 7),
stem_conv_stride=(1, 2, 2),
# Stage configs.
stage_conv_a_kernel_size=(
(1, 1, 1),
(1, 1, 1),
(1, 1, 1),
(1, 1, 1),
),
stage_conv_b_kernel_size=(
(3, 3, 3),
(3, 3, 3),
(3, 3, 3),
(3, 3, 3),
),
stage_conv_b_num_groups=(1, 1, 1, 1),
stage_conv_b_dilation=(
(1, 1, 1),
(1, 1, 1),
(1, 1, 1),
(1, 1, 1),
),
stage_spatial_stride=(2, 2, 2, 2),
stage_temporal_stride=(1, 1, 2, 2),
stage_bottleneck=(
create_2plus1d_bottleneck_block,
create_2plus1d_bottleneck_block,
create_2plus1d_bottleneck_block,
create_2plus1d_bottleneck_block,
),
# Head configs.
head_pool=nn.AvgPool3d,
head_pool_kernel_size=(
cfg.DATA.NUM_FRAMES // 4,
cfg.DATA.TRAIN_CROP_SIZE // 32,
cfg.DATA.TRAIN_CROP_SIZE // 32,
),
head_activation=None,
head_output_with_global_average=False,
)
self.post_act = get_head_act(cfg.MODEL.HEAD_ACT)
def test_create_r2plus1d(self):
"""
Test simple r2plus1d with different inputs.
"""
for input_channel, input_clip_length, input_crop_size in itertools.product(
(3, 2), (4, 8), (56, 64)):
stage_spatial_stride = (2, 2, 2, 2)
stage_temporal_stride = (1, 1, 2, 2)
total_spatial_stride = 2 * np.prod(stage_spatial_stride)
total_temporal_stride = np.prod(stage_temporal_stride)
head_pool_kernel_size = (
input_clip_length // total_temporal_stride,
input_crop_size // total_spatial_stride,
input_crop_size // total_spatial_stride,
)
model = create_r2plus1d(
input_channel=input_channel,
model_depth=50,
model_num_class=400,
dropout_rate=0.0,
norm=nn.BatchNorm3d,
activation=nn.ReLU,
stem_dim_out=8,
stem_conv_kernel_size=(1, 7, 7),
stem_conv_stride=(1, 2, 2),
stage_conv_b_kernel_size=((3, 3, 3), ) * 4,
stage_spatial_stride=stage_spatial_stride,
stage_temporal_stride=stage_temporal_stride,
stage_bottleneck=(
create_bottleneck_block,
create_2plus1d_bottleneck_block,
create_2plus1d_bottleneck_block,
create_2plus1d_bottleneck_block,
),
head_pool=nn.AvgPool3d,
head_pool_kernel_size=head_pool_kernel_size,
head_output_size=(1, 1, 1),
head_activation=nn.Softmax,
)
# Test forwarding.
for tensor in TestR2plus1d._get_inputs(input_channel,
input_clip_length,
input_crop_size):
if tensor.shape[1] != input_channel:
with self.assertRaises(RuntimeError):
out = model(tensor)
continue
out = model(tensor)
output_shape = out.shape
output_shape_gt = (tensor.shape[0], 400)
self.assertEqual(
output_shape,
output_shape_gt,
"Output shape {} is different from expected shape {}".
format(output_shape, output_shape_gt),
)