def test_create_slowfast_with_callable(self):
"""
Test builder `create_slowfast` with callable inputs.
"""
for (norm, activation) in itertools.product(
(nn.BatchNorm3d, None), (nn.ReLU, nn.Sigmoid, None)):
input_clip_length = 32
input_crop_size = 224
input_channel = 3
model = create_slowfast(
slowfast_channel_reduction_ratio=8,
slowfast_conv_channel_fusion_ratio=2,
slowfast_fusion_conv_kernel_size=(7, 1, 1),
slowfast_fusion_conv_stride=(4, 1, 1),
input_channels=(input_channel, ) * 2,
model_depth=18,
model_num_class=400,
dropout_rate=0,
norm=norm,
activation=activation,
)
# Test forwarding.
for tensor in TestSlowFast._get_inputs(input_channel,
input_clip_length,
input_crop_size):
with torch.no_grad():
if tensor[0].shape[1] != input_channel:
with self.assertRaises(RuntimeError):
model(tensor)
continue
model(tensor)
def _slowfast(
pretrained: bool = False,
progress: bool = True,
checkpoint_path: str = "",
**kwargs: Any,
) -> nn.Module:
model = create_slowfast(**kwargs)
if pretrained:
checkpoint = load_state_dict_from_url(checkpoint_path, progress=progress)
state_dict = checkpoint["model_state"]
model.load_state_dict(state_dict)
return model
def _slowfast(
pretrained: bool = False,
progress: bool = True,
checkpoint_path: str = "",
**kwargs: Any,
) -> nn.Module:
model = create_slowfast(**kwargs)
if pretrained:
# 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 _construct_network(self, cfg):
"""
Builds a SlowFast model.
Args:
cfg (CfgNode): model building configs, details are in the
comments of the config file.
"""
# Params from configs.
norm_module = get_norm(cfg)
pool_size = _POOL1[cfg.MODEL.ARCH]
num_groups = cfg.RESNET.NUM_GROUPS
width_per_group = cfg.RESNET.WIDTH_PER_GROUP
spatial_dilations = cfg.RESNET.SPATIAL_DILATIONS
spatial_strides = cfg.RESNET.SPATIAL_STRIDES
temp_kernel = _TEMPORAL_KERNEL_BASIS[cfg.MODEL.ARCH]
self.model = create_slowfast(
# SlowFast configs.
slowfast_channel_reduction_ratio=cfg.SLOWFAST.BETA_INV,
slowfast_conv_channel_fusion_ratio=cfg.SLOWFAST.
FUSION_CONV_CHANNEL_RATIO,
slowfast_fusion_conv_kernel_size=(
cfg.SLOWFAST.FUSION_KERNEL_SZ,
1,
1,
),
slowfast_fusion_conv_stride=(cfg.SLOWFAST.ALPHA, 1, 1),
# Input clip configs.
input_channels=cfg.DATA.INPUT_CHANNEL_NUM,
# Model configs.
model_depth=cfg.RESNET.DEPTH,
model_num_class=cfg.MODEL.NUM_CLASSES,
dropout_rate=cfg.MODEL.DROPOUT_RATE,
# Normalization configs.
norm=norm_module,
# Activation configs.
activation=partial(nn.ReLU, inplace=cfg.RESNET.INPLACE_RELU),
# Stem configs.
stem_dim_outs=(
width_per_group,
width_per_group // cfg.SLOWFAST.BETA_INV,
),
stem_conv_kernel_sizes=(
(temp_kernel[0][0][0], 7, 7),
(temp_kernel[0][1][0], 7, 7),
),
stem_conv_strides=((1, 2, 2), (1, 2, 2)),
stem_pool=nn.MaxPool3d,
stem_pool_kernel_sizes=((1, 3, 3), (1, 3, 3)),
stem_pool_strides=((1, 2, 2), (1, 2, 2)),
# Stage configs.
stage_conv_a_kernel_sizes=(
(
(temp_kernel[1][0][0], 1, 1),
(temp_kernel[2][0][0], 1, 1),
(temp_kernel[3][0][0], 1, 1),
(temp_kernel[4][0][0], 1, 1),
),
(
(temp_kernel[1][1][0], 1, 1),
(temp_kernel[2][1][0], 1, 1),
(temp_kernel[3][1][0], 1, 1),
(temp_kernel[4][1][0], 1, 1),
),
),
stage_conv_b_kernel_sizes=(
((1, 3, 3), (1, 3, 3), (1, 3, 3), (1, 3, 3)),
((1, 3, 3), (1, 3, 3), (1, 3, 3), (1, 3, 3)),
),
stage_conv_b_num_groups=(
(num_groups, num_groups, num_groups, num_groups),
(num_groups, num_groups, num_groups, num_groups),
),
stage_conv_b_dilations=(
(
(1, spatial_dilations[0][0], spatial_dilations[0][0]),
(1, spatial_dilations[1][0], spatial_dilations[1][0]),
(1, spatial_dilations[2][0], spatial_dilations[2][0]),
(1, spatial_dilations[3][0], spatial_dilations[3][0]),
),
(
(1, spatial_dilations[0][1], spatial_dilations[0][1]),
(1, spatial_dilations[1][1], spatial_dilations[1][1]),
(1, spatial_dilations[1][1], spatial_dilations[1][1]),
(1, spatial_dilations[1][1], spatial_dilations[1][1]),
),
),
stage_spatial_strides=(
(
spatial_strides[0][0],
spatial_strides[1][0],
spatial_strides[2][0],
spatial_strides[3][0],
),
(
spatial_strides[0][1],
spatial_strides[1][1],
spatial_strides[2][1],
spatial_strides[3][1],
),
),
stage_temporal_strides=((1, 1, 1, 1), (1, 1, 1, 1)),
bottleneck=create_bottleneck_block,
# Head configs.
head_pool=nn.AvgPool3d,
head_pool_kernel_sizes=(
(
cfg.DATA.NUM_FRAMES // cfg.SLOWFAST.ALPHA //
pool_size[0][0],
cfg.DATA.TRAIN_CROP_SIZE // 32 // pool_size[0][1],
cfg.DATA.TRAIN_CROP_SIZE // 32 // pool_size[0][2],
),
(
cfg.DATA.NUM_FRAMES // pool_size[1][0],
cfg.DATA.TRAIN_CROP_SIZE // 32 // pool_size[1][1],
cfg.DATA.TRAIN_CROP_SIZE // 32 // pool_size[1][2],
),
),
head_activation=None,
head_output_with_global_average=False,
)
self.post_act = get_head_act(cfg.MODEL.HEAD_ACT)
def _construct_network(self, cfg):
"""
Builds a SlowFast model.
Args:
cfg (CfgNode): model building configs, details are in the
comments of the config file.
"""
_MODEL_STAGE_DEPTH = {50: (3, 4, 6, 3), 101: (3, 4, 23, 3)}
# Params from configs.
norm_module = get_norm(cfg)
pool_size = _POOL1[cfg.MODEL.ARCH]
num_groups = cfg.RESNET.NUM_GROUPS
width_per_group = cfg.RESNET.WIDTH_PER_GROUP
spatial_dilations = cfg.RESNET.SPATIAL_DILATIONS
spatial_strides = cfg.RESNET.SPATIAL_STRIDES
temp_kernel = _TEMPORAL_KERNEL_BASIS[cfg.MODEL.ARCH]
num_block_temp_kernel = cfg.RESNET.NUM_BLOCK_TEMP_KERNEL
stage_depth = _MODEL_STAGE_DEPTH[cfg.RESNET.DEPTH]
stage_conv_a_kernel_sizes = [[], []]
for pathway in range(2):
for stage in range(4):
stage_conv_a_kernel_sizes[pathway].append(
((temp_kernel[stage + 1][pathway][0], 1, 1),)
* num_block_temp_kernel[stage][pathway]
+ ((1, 1, 1),)
* (
stage_depth[stage]
- num_block_temp_kernel[stage][pathway]
)
)
# Head from config
# Number of stages = 4
stage_dim_in = cfg.RESNET.WIDTH_PER_GROUP * 2 ** (4 + 1)
head_in_features = stage_dim_in + stage_dim_in // cfg.SLOWFAST.BETA_INV
if cfg.DETECTION.ENABLE:
self.detection_head = create_res_roi_pooling_head(
in_features=head_in_features,
out_features=cfg.MODEL.NUM_CLASSES,
pool=None,
output_size=(1, 1, 1),
dropout_rate=cfg.MODEL.DROPOUT_RATE,
activation=None,
output_with_global_average=False,
pool_spatial=nn.MaxPool2d,
resolution=[cfg.DETECTION.ROI_XFORM_RESOLUTION] * 2,
spatial_scale=1.0 / float(cfg.DETECTION.SPATIAL_SCALE_FACTOR),
sampling_ratio=0,
roi=ROIAlign,
)
head_pool_kernel_sizes = (
(
cfg.DATA.NUM_FRAMES
// cfg.SLOWFAST.ALPHA
// pool_size[0][0],
1,
1,
),
(cfg.DATA.NUM_FRAMES // pool_size[1][0], 1, 1),
)
else:
head_pool_kernel_sizes = (
(
cfg.DATA.NUM_FRAMES
// cfg.SLOWFAST.ALPHA
// pool_size[0][0],
cfg.DATA.TRAIN_CROP_SIZE // 32 // pool_size[0][1],
cfg.DATA.TRAIN_CROP_SIZE // 32 // pool_size[0][2],
),
(
cfg.DATA.NUM_FRAMES // pool_size[1][0],
cfg.DATA.TRAIN_CROP_SIZE // 32 // pool_size[1][1],
cfg.DATA.TRAIN_CROP_SIZE // 32 // pool_size[1][2],
),
)
self.model = create_slowfast(
# SlowFast configs.
slowfast_channel_reduction_ratio=cfg.SLOWFAST.BETA_INV,
slowfast_conv_channel_fusion_ratio=cfg.SLOWFAST.FUSION_CONV_CHANNEL_RATIO,
slowfast_fusion_conv_kernel_size=(
cfg.SLOWFAST.FUSION_KERNEL_SZ,
1,
1,
),
slowfast_fusion_conv_stride=(cfg.SLOWFAST.ALPHA, 1, 1),
# Input clip configs.
input_channels=cfg.DATA.INPUT_CHANNEL_NUM,
# Model configs.
model_depth=cfg.RESNET.DEPTH,
model_num_class=cfg.MODEL.NUM_CLASSES,
dropout_rate=cfg.MODEL.DROPOUT_RATE,
# Normalization configs.
norm=norm_module,
# Activation configs.
activation=partial(nn.ReLU, inplace=cfg.RESNET.INPLACE_RELU),
# Stem configs.
stem_dim_outs=(
width_per_group,
width_per_group // cfg.SLOWFAST.BETA_INV,
),
stem_conv_kernel_sizes=(
(temp_kernel[0][0][0], 7, 7),
(temp_kernel[0][1][0], 7, 7),
),
stem_conv_strides=((1, 2, 2), (1, 2, 2)),
stem_pool=nn.MaxPool3d,
stem_pool_kernel_sizes=((1, 3, 3), (1, 3, 3)),
stem_pool_strides=((1, 2, 2), (1, 2, 2)),
# Stage configs.
stage_conv_a_kernel_sizes=stage_conv_a_kernel_sizes,
stage_conv_b_kernel_sizes=(
((1, 3, 3), (1, 3, 3), (1, 3, 3), (1, 3, 3)),
((1, 3, 3), (1, 3, 3), (1, 3, 3), (1, 3, 3)),
),
stage_conv_b_num_groups=(
(num_groups, num_groups, num_groups, num_groups),
(num_groups, num_groups, num_groups, num_groups),
),
stage_conv_b_dilations=(
(
(1, spatial_dilations[0][0], spatial_dilations[0][0]),
(1, spatial_dilations[1][0], spatial_dilations[1][0]),
(1, spatial_dilations[2][0], spatial_dilations[2][0]),
(1, spatial_dilations[3][0], spatial_dilations[3][0]),
),
(
(1, spatial_dilations[0][1], spatial_dilations[0][1]),
(1, spatial_dilations[1][1], spatial_dilations[1][1]),
(1, spatial_dilations[1][1], spatial_dilations[1][1]),
(1, spatial_dilations[1][1], spatial_dilations[1][1]),
),
),
stage_spatial_strides=(
(
spatial_strides[0][0],
spatial_strides[1][0],
spatial_strides[2][0],
spatial_strides[3][0],
),
(
spatial_strides[0][1],
spatial_strides[1][1],
spatial_strides[2][1],
spatial_strides[3][1],
),
),
stage_temporal_strides=((1, 1, 1, 1), (1, 1, 1, 1)),
bottleneck=create_bottleneck_block,
# Head configs.
head=create_res_basic_head if not self.detection_mode else None,
head_pool=nn.AvgPool3d,
head_pool_kernel_sizes=head_pool_kernel_sizes,
head_activation=None,
head_output_with_global_average=False,
)
self.post_act = get_head_act(cfg.MODEL.HEAD_ACT)
