def _construct_network(self, cfg):
"""
Builds a X3D model.
Args:
cfg (CfgNode): model building configs, details are in the
comments of the config file.
"""
# Params from configs.
norm_module = get_norm(cfg)
temp_kernel = _TEMPORAL_KERNEL_BASIS[cfg.MODEL.ARCH]
self.model = create_x3d(
# Input clip configs.
input_channel=cfg.DATA.INPUT_CHANNEL_NUM[0],
input_clip_length=cfg.DATA.NUM_FRAMES,
input_crop_size=cfg.DATA.TRAIN_CROP_SIZE,
# Model configs.
model_num_class=cfg.MODEL.NUM_CLASSES,
dropout_rate=cfg.MODEL.DROPOUT_RATE,
width_factor=cfg.X3D.WIDTH_FACTOR,
depth_factor=cfg.X3D.DEPTH_FACTOR,
# Normalization configs.
norm=norm_module,
norm_eps=1e-5,
norm_momentum=0.1,
# Activation configs.
activation=partial(nn.ReLU, inplace=cfg.RESNET.INPLACE_RELU),
# Stem configs.
stem_dim_in=cfg.X3D.DIM_C1,
stem_conv_kernel_size=(temp_kernel[0][0][0], 3, 3),
stem_conv_stride=(1, 2, 2),
# Stage configs.
stage_conv_kernel_size=(
(temp_kernel[1][0][0], 3, 3),
(temp_kernel[2][0][0], 3, 3),
(temp_kernel[3][0][0], 3, 3),
(temp_kernel[4][0][0], 3, 3),
),
stage_spatial_stride=(2, 2, 2, 2),
stage_temporal_stride=(1, 1, 1, 1),
bottleneck=create_x3d_bottleneck_block,
bottleneck_factor=cfg.X3D.BOTTLENECK_FACTOR,
se_ratio=0.0625,
inner_act=Swish,
# Head configs.
head_dim_out=cfg.X3D.DIM_C5,
head_pool_act=partial(nn.ReLU, inplace=cfg.RESNET.INPLACE_RELU),
head_bn_lin5_on=cfg.X3D.BN_LIN5,
head_activation=None,
head_output_with_global_average=False,
)
self.post_act = get_head_act(cfg.MODEL.HEAD_ACT)
def _x3d(
pretrained: bool = False,
progress: bool = True,
checkpoint_path: str = "",
**kwargs: Any,
) -> nn.Module:
model = create_x3d(**kwargs)
if pretrained and len(checkpoint_path) > 0:
checkpoint = load_state_dict_from_url(checkpoint_path, progress=progress)
state_dict = checkpoint["model_state"]
model.load_state_dict(state_dict, strict=True)
return model
def _x3d(
pretrained: bool = False,
progress: bool = True,
checkpoint_path: Optional[str] = None,
**kwargs: Any,
) -> nn.Module:
model = create_x3d(**kwargs)
if pretrained and checkpoint_path is not None:
# All models are loaded onto CPU by default
checkpoint = load_state_dict_from_url(checkpoint_path,
progress=progress,
map_location="cpu")
state_dict = checkpoint["model_state"]
model.load_state_dict(state_dict)
return model
def test_create_x3d(self):
"""
To test different versions of X3D, set the input to:
X3D-XS: (4, 160, 2.0, 2.2, 2.25)
X3D-S: (13, 160, 2.0, 2.2, 2.25)
X3D-M: (16, 224, 2.0, 2.2, 2.25)
X3D-L: (16, 312, 2.0, 5.0, 2.25)
Each of the parameters corresponds to input_clip_length, input_crop_size,
width_factor, depth_factor and bottleneck_factor.
"""
for (
input_clip_length,
input_crop_size,
width_factor,
depth_factor,
bottleneck_factor,
) in [
(4, 160, 2.0, 2.2, 2.25),
]:
model = create_x3d(
input_clip_length=input_clip_length,
input_crop_size=input_crop_size,
model_num_class=400,
dropout_rate=0.5,
width_factor=width_factor,
depth_factor=depth_factor,
norm=nn.BatchNorm3d,
activation=nn.ReLU,
stem_dim_in=12,
stem_conv_kernel_size=(5, 3, 3),
stem_conv_stride=(1, 2, 2),
stage_conv_kernel_size=((3, 3, 3), ) * 4,
stage_spatial_stride=(2, 2, 2, 2),
stage_temporal_stride=(1, 1, 1, 1),
bottleneck=create_x3d_bottleneck_block,
bottleneck_factor=bottleneck_factor,
se_ratio=0.0625,
inner_act=Swish,
head_dim_out=2048,
head_pool_act=nn.ReLU,
head_bn_lin5_on=False,
head_activation=nn.Softmax,
)
# Test forwarding.
for tensor in TestX3d._get_inputs(input_clip_length,
input_crop_size):
if tensor.shape[1] != 3:
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),
)