def slow_r50(
pretrained: bool = False, progress: bool = True, **kwargs: Any
) -> nn.Module:
r"""
Slow R50 model architecture [1] with pretrained weights based on 8x8 setting
on the Kinetics dataset. Model with pretrained weights has top1 accuracy of 74.58.
[1] Christoph Feichtenhofer et al, "SlowFast Networks for Video Recognition"
https://arxiv.org/pdf/1812.03982.pdf
Args:
pretrained (bool): If True, returns a model pre-trained on the Kinetics dataset
progress (bool): If True, displays a progress bar of the download to stderr
kwargs: use these to modify any of the other model settings. All the
options are defined in pytorchvideo/models/resnet.py
NOTE: to use the pretrained model, do not modify the model configuration
via the kwargs. Only modify settings via kwargs to initialize a new model
without pretrained weights.
"""
model = create_resnet(
stem_conv_kernel_size=(1, 7, 7),
head_pool_kernel_size=(8, 7, 7),
model_depth=50,
**kwargs,
)
if pretrained:
path = checkpoint_paths["slow_r50"]
checkpoint = load_state_dict_from_url(path, progress=progress)
state_dict = checkpoint["model_state"]
model.load_state_dict(state_dict)
return model
def make_kinetics_resnet():
resnet_model = resnet.create_resnet(input_channel=3,
model_depth=50,
model_num_class=3,
norm=nn.BatchNorm3d,
activation=nn.ReLU)
return resnet_model
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.
"""
# Params from configs.
norm_module = get_norm(cfg)
head_act = get_head_act(cfg.MODEL.HEAD_ACT)
pool_size = _POOL1[cfg.MODEL.ARCH]
num_groups = cfg.RESNET.NUM_GROUPS
spatial_dilations = cfg.RESNET.SPATIAL_DILATIONS
spatial_strides = cfg.RESNET.SPATIAL_STRIDES
temp_kernel = _TEMPORAL_KERNEL_BASIS[cfg.MODEL.ARCH]
stage1_pool = pool_size[0][0] != 1 or len(set(pool_size[0])) > 1
stage_spatial_stride = (
spatial_strides[0][0],
spatial_strides[1][0],
spatial_strides[2][0],
spatial_strides[3][0],
)
if cfg.MODEL.ARCH == "i3d":
stage_conv_a_kernel_size = (
(3, 1, 1),
[(3, 1, 1), (1, 1, 1)],
[(3, 1, 1), (1, 1, 1)],
[(1, 1, 1), (3, 1, 1)],
)
else:
stage_conv_a_kernel_size = (
(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),
)
self.model = create_resnet(
# 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=norm_module,
# 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=(temp_kernel[0][0][0], 7, 7),
stem_conv_stride=(1, 2, 2),
stem_pool=nn.MaxPool3d,
stem_pool_kernel_size=(1, 3, 3),
stem_pool_stride=(1, 2, 2),
# Stage configs.
stage1_pool=nn.MaxPool3d if stage1_pool else None,
stage1_pool_kernel_size=pool_size[0],
stage_conv_a_kernel_size=stage_conv_a_kernel_size,
stage_conv_b_kernel_size=(
(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,
),
stage_conv_b_dilation=(
(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]),
),
stage_spatial_h_stride=stage_spatial_stride,
stage_spatial_w_stride=stage_spatial_stride,
stage_temporal_stride=(1, 1, 1, 1),
bottleneck=create_bottleneck_block,
# Head configs.
head_pool=nn.AvgPool3d,
head_pool_kernel_size=(
cfg.DATA.NUM_FRAMES // pool_size[0][0],
cfg.DATA.TRAIN_CROP_SIZE // 32 // pool_size[0][1],
cfg.DATA.TRAIN_CROP_SIZE // 32 // pool_size[0][2],
),
head_activation=None,
head_output_with_global_average=False,
)
self.post_act = head_act
