def forward(self,
x,
out_feat_keys: List[str] = None) -> List[torch.Tensor]:
if isinstance(x, MultiDimensionalTensor):
out = get_tunk_forward_interpolated_outputs(
input_type=self.model_config.INPUT_TYPE,
interpolate_out_feat_key_name="res5",
remove_padding_before_feat_key_name="avgpool",
feat=x,
feature_blocks=self._feature_blocks,
# FSDP has its own activation checkpoint method: disable vissl's method here.
use_checkpointing=False,
checkpointing_splits=0,
)
else:
model_input = transform_model_input_data_type(
x, self.model_config.INPUT_TYPE)
out = get_trunk_forward_outputs(
feat=model_input,
out_feat_keys=out_feat_keys,
feature_blocks=self._feature_blocks,
# FSDP has its own activation checkpoint method: disable vissl's method here.
use_checkpointing=False,
checkpointing_splits=0,
)
return out
def forward(
self, x: torch.Tensor, out_feat_keys: List[str] = None
) -> List[torch.Tensor]:
if isinstance(x, MultiDimensionalTensor):
out = get_tunk_forward_interpolated_outputs(
input_type=self.model_config.INPUT_TYPE,
interpolate_out_feat_key_name="res5",
remove_padding_before_feat_key_name="avgpool",
feat=x,
feature_blocks=self._feature_blocks,
feature_mapping=self.feat_eval_mapping,
use_checkpointing=self.use_checkpointing,
checkpointing_splits=self.num_checkpointing_splits,
)
else:
model_input = transform_model_input_data_type(
x, self.model_config.INPUT_TYPE
)
out = get_trunk_forward_outputs(
feat=model_input,
out_feat_keys=out_feat_keys,
feature_blocks=self._feature_blocks,
feature_mapping=self.feat_eval_mapping,
use_checkpointing=self.use_checkpointing,
checkpointing_splits=self.num_checkpointing_splits,
)
return out
