def _construct_network(self, cfg):
"""
Builds a single pathway ResNet model.
Args:
cfg (CfgNode): model building configs, details are in the
comments of the config file.
"""
assert cfg.MODEL.ARCH in _POOL1.keys()
pool_size = _POOL1[cfg.MODEL.ARCH]
assert len({len(pool_size), self.num_pathways}) == 1
assert cfg.RESNET.DEPTH in _MODEL_STAGE_DEPTH.keys()
(d2, d3, d4, d5) = _MODEL_STAGE_DEPTH[cfg.RESNET.DEPTH]
num_groups = cfg.RESNET.NUM_GROUPS
width_per_group = cfg.RESNET.WIDTH_PER_GROUP
dim_inner = num_groups * width_per_group
temp_kernel = _TEMPORAL_KERNEL_BASIS[cfg.MODEL.ARCH]
self.s1 = stem_helper.VideoModelStem(
dim_in=cfg.DATA.INPUT_CHANNEL_NUM,
dim_out=[width_per_group],
kernel=[temp_kernel[0][0] + [7, 7]],
stride=[[1, 2, 2]],
padding=[[temp_kernel[0][0][0] // 2, 3, 3]],
)
self.s2 = resnet_helper.ResStage(
dim_in=[width_per_group],
dim_out=[width_per_group * 4],
dim_inner=[dim_inner],
temp_kernel_sizes=temp_kernel[1],
stride=[1],
num_blocks=[d2],
num_groups=[num_groups],
num_block_temp_kernel=cfg.RESNET.NUM_BLOCK_TEMP_KERNEL[0],
nonlocal_inds=cfg.NONLOCAL.LOCATION[0],
nonlocal_group=cfg.NONLOCAL.GROUP[0],
instantiation=cfg.NONLOCAL.INSTANTIATION,
trans_func_name=cfg.RESNET.TRANS_FUNC,
stride_1x1=cfg.RESNET.STRIDE_1X1,
inplace_relu=cfg.RESNET.INPLACE_RELU,
)
for pathway in range(self.num_pathways):
pool = nn.MaxPool3d(
kernel_size=pool_size[pathway],
stride=pool_size[pathway],
padding=[0, 0, 0],
)
self.add_module("pathway{}_pool".format(pathway), pool)
self.s3 = resnet_helper.ResStage(
dim_in=[width_per_group * 4],
dim_out=[width_per_group * 8],
dim_inner=[dim_inner * 2],
temp_kernel_sizes=temp_kernel[2],
stride=[2],
num_blocks=[d3],
num_groups=[num_groups],
num_block_temp_kernel=cfg.RESNET.NUM_BLOCK_TEMP_KERNEL[1],
nonlocal_inds=cfg.NONLOCAL.LOCATION[1],
nonlocal_group=cfg.NONLOCAL.GROUP[1],
instantiation=cfg.NONLOCAL.INSTANTIATION,
trans_func_name=cfg.RESNET.TRANS_FUNC,
stride_1x1=cfg.RESNET.STRIDE_1X1,
inplace_relu=cfg.RESNET.INPLACE_RELU,
)
self.s4 = resnet_helper.ResStage(
dim_in=[width_per_group * 8],
dim_out=[width_per_group * 16],
dim_inner=[dim_inner * 4],
temp_kernel_sizes=temp_kernel[3],
stride=[2],
num_blocks=[d4],
num_groups=[num_groups],
num_block_temp_kernel=cfg.RESNET.NUM_BLOCK_TEMP_KERNEL[2],
nonlocal_inds=cfg.NONLOCAL.LOCATION[2],
nonlocal_group=cfg.NONLOCAL.GROUP[2],
instantiation=cfg.NONLOCAL.INSTANTIATION,
trans_func_name=cfg.RESNET.TRANS_FUNC,
stride_1x1=cfg.RESNET.STRIDE_1X1,
inplace_relu=cfg.RESNET.INPLACE_RELU,
)
self.s5 = resnet_helper.ResStage(
dim_in=[width_per_group * 16],
dim_out=[width_per_group * 32],
dim_inner=[dim_inner * 8],
temp_kernel_sizes=temp_kernel[4],
stride=[2],
num_blocks=[d5],
num_groups=[num_groups],
num_block_temp_kernel=cfg.RESNET.NUM_BLOCK_TEMP_KERNEL[3],
nonlocal_inds=cfg.NONLOCAL.LOCATION[3],
nonlocal_group=cfg.NONLOCAL.GROUP[3],
instantiation=cfg.NONLOCAL.INSTANTIATION,
trans_func_name=cfg.RESNET.TRANS_FUNC,
stride_1x1=cfg.RESNET.STRIDE_1X1,
inplace_relu=cfg.RESNET.INPLACE_RELU,
)
self.head = head_helper.ResNetBasicHead(
dim_in=[width_per_group * 32],
num_classes=cfg.MODEL.NUM_CLASSES,
pool_size=[[
cfg.DATA.NUM_FRAMES // pool_size[0][0],
cfg.DATA.CROP_SIZE // 32 // pool_size[0][1],
cfg.DATA.CROP_SIZE // 32 // pool_size[0][2],
]],
dropout_rate=cfg.MODEL.DROPOUT_RATE,
)
def _construct_network(self, cfg):
"""
Builds a SlowFast model. The first pathway is the Slow pathway and the
second pathway is the Fast pathway.
Args:
cfg (CfgNode): model building configs, details are in the
comments of the config file.
"""
assert cfg.MODEL.ARCH in _POOL1.keys()
pool_size = _POOL1[cfg.MODEL.ARCH]
assert len({len(pool_size), self.num_pathways}) == 1
assert cfg.RESNET.DEPTH in _MODEL_STAGE_DEPTH.keys()
(d2, d3, d4, d5) = _MODEL_STAGE_DEPTH[cfg.RESNET.DEPTH]
num_groups = cfg.RESNET.NUM_GROUPS
width_per_group = cfg.RESNET.WIDTH_PER_GROUP
dim_inner = num_groups * width_per_group
out_dim_ratio = (cfg.SLOWFAST.BETA_INV //
cfg.SLOWFAST.FUSION_CONV_CHANNEL_RATIO)
temp_kernel = _TEMPORAL_KERNEL_BASIS[cfg.MODEL.ARCH]
self.s1 = stem_helper.VideoModelStem(
dim_in=cfg.DATA.INPUT_CHANNEL_NUM,
dim_out=[
width_per_group, width_per_group // cfg.SLOWFAST.BETA_INV
],
kernel=[temp_kernel[0][0] + [7, 7], temp_kernel[0][1] + [7, 7]],
stride=[[1, 2, 2]] * 2,
padding=[
[temp_kernel[0][0][0] // 2, 3, 3],
[temp_kernel[0][1][0] // 2, 3, 3],
],
)
self.s1_fuse = FuseFastToSlow(
width_per_group // cfg.SLOWFAST.BETA_INV,
cfg.SLOWFAST.FUSION_CONV_CHANNEL_RATIO,
cfg.SLOWFAST.FUSION_KERNEL_SZ,
cfg.SLOWFAST.ALPHA,
)
self.s2 = resnet_helper.ResStage(
dim_in=[
width_per_group + width_per_group // out_dim_ratio,
width_per_group // cfg.SLOWFAST.BETA_INV,
],
dim_out=[
width_per_group * 4,
width_per_group * 4 // cfg.SLOWFAST.BETA_INV,
],
dim_inner=[dim_inner, dim_inner // cfg.SLOWFAST.BETA_INV],
temp_kernel_sizes=temp_kernel[1],
stride=[1] * 2,
num_blocks=[d2] * 2,
num_groups=[num_groups] * 2,
num_block_temp_kernel=cfg.RESNET.NUM_BLOCK_TEMP_KERNEL[0],
nonlocal_inds=cfg.NONLOCAL.LOCATION[0],
nonlocal_group=cfg.NONLOCAL.GROUP[0],
instantiation=cfg.NONLOCAL.INSTANTIATION,
trans_func_name=cfg.RESNET.TRANS_FUNC,
)
self.s2_fuse = FuseFastToSlow(
width_per_group * 4 // cfg.SLOWFAST.BETA_INV,
cfg.SLOWFAST.FUSION_CONV_CHANNEL_RATIO,
cfg.SLOWFAST.FUSION_KERNEL_SZ,
cfg.SLOWFAST.ALPHA,
)
for pathway in range(self.num_pathways):
pool = nn.MaxPool3d(
kernel_size=pool_size[pathway],
stride=pool_size[pathway],
padding=[0, 0, 0],
)
self.add_module("pathway{}_pool".format(pathway), pool)
self.s3 = resnet_helper.ResStage(
dim_in=[
width_per_group * 4 + width_per_group * 4 // out_dim_ratio,
width_per_group * 4 // cfg.SLOWFAST.BETA_INV,
],
dim_out=[
width_per_group * 8,
width_per_group * 8 // cfg.SLOWFAST.BETA_INV,
],
dim_inner=[dim_inner * 2, dim_inner * 2 // cfg.SLOWFAST.BETA_INV],
temp_kernel_sizes=temp_kernel[2],
stride=[2] * 2,
num_blocks=[d3] * 2,
num_groups=[num_groups] * 2,
num_block_temp_kernel=cfg.RESNET.NUM_BLOCK_TEMP_KERNEL[1],
nonlocal_inds=cfg.NONLOCAL.LOCATION[1],
nonlocal_group=cfg.NONLOCAL.GROUP[1],
instantiation=cfg.NONLOCAL.INSTANTIATION,
trans_func_name=cfg.RESNET.TRANS_FUNC,
)
self.s3_fuse = FuseFastToSlow(
width_per_group * 8 // cfg.SLOWFAST.BETA_INV,
cfg.SLOWFAST.FUSION_CONV_CHANNEL_RATIO,
cfg.SLOWFAST.FUSION_KERNEL_SZ,
cfg.SLOWFAST.ALPHA,
)
self.s4 = resnet_helper.ResStage(
dim_in=[
width_per_group * 8 + width_per_group * 8 // out_dim_ratio,
width_per_group * 8 // cfg.SLOWFAST.BETA_INV,
],
dim_out=[
width_per_group * 16,
width_per_group * 16 // cfg.SLOWFAST.BETA_INV,
],
dim_inner=[dim_inner * 4, dim_inner * 4 // cfg.SLOWFAST.BETA_INV],
temp_kernel_sizes=temp_kernel[3],
stride=[2] * 2,
num_blocks=[d4] * 2,
num_groups=[num_groups] * 2,
num_block_temp_kernel=cfg.RESNET.NUM_BLOCK_TEMP_KERNEL[2],
nonlocal_inds=cfg.NONLOCAL.LOCATION[2],
nonlocal_group=cfg.NONLOCAL.GROUP[2],
instantiation=cfg.NONLOCAL.INSTANTIATION,
trans_func_name=cfg.RESNET.TRANS_FUNC,
)
self.s4_fuse = FuseFastToSlow(
width_per_group * 16 // cfg.SLOWFAST.BETA_INV,
cfg.SLOWFAST.FUSION_CONV_CHANNEL_RATIO,
cfg.SLOWFAST.FUSION_KERNEL_SZ,
cfg.SLOWFAST.ALPHA,
)
self.s5 = resnet_helper.ResStage(
dim_in=[
width_per_group * 16 + width_per_group * 16 // out_dim_ratio,
width_per_group * 16 // cfg.SLOWFAST.BETA_INV,
],
dim_out=[
width_per_group * 32,
width_per_group * 32 // cfg.SLOWFAST.BETA_INV,
],
dim_inner=[dim_inner * 8, dim_inner * 8 // cfg.SLOWFAST.BETA_INV],
temp_kernel_sizes=temp_kernel[4],
stride=[2] * 2,
num_blocks=[d5] * 2,
num_groups=[num_groups] * 2,
num_block_temp_kernel=cfg.RESNET.NUM_BLOCK_TEMP_KERNEL[3],
nonlocal_inds=cfg.NONLOCAL.LOCATION[3],
nonlocal_group=cfg.NONLOCAL.GROUP[3],
instantiation=cfg.NONLOCAL.INSTANTIATION,
trans_func_name=cfg.RESNET.TRANS_FUNC,
)
self.head = head_helper.ResNetBasicHead(
dim_in=[
width_per_group * 32,
width_per_group * 32 // cfg.SLOWFAST.BETA_INV,
],
num_classes=cfg.MODEL.NUM_CLASSES,
pool_size=[
[
cfg.DATA.NUM_FRAMES // cfg.SLOWFAST.ALPHA //
pool_size[0][0],
cfg.DATA.CROP_SIZE // 32 // pool_size[0][1],
cfg.DATA.CROP_SIZE // 32 // pool_size[0][2],
],
[
cfg.DATA.NUM_FRAMES // pool_size[1][0],
cfg.DATA.CROP_SIZE // 32 // pool_size[1][1],
cfg.DATA.CROP_SIZE // 32 // pool_size[1][2],
],
],
dropout_rate=cfg.MODEL.DROPOUT_RATE,
)
