Formula: 2 global + 8 local crops; consistency
Usage
Dinmlt(
image,
global_size = 2,
local_size = 8,
tau_s = 0.1,
tau_t = 0.04,
center = NULL
)
Arguments
- image
A V x d matrix of per-crop logits, global crops first.
- global_size
Number of global crops G.
- local_size
Number of local crops; V must equal G + local_size.
- tau_s
Student temperature.
- tau_t
Teacher temperature (sharpening).
- center
Teacher centre vector, or NULL for the global mean.
Value
List with estimate, loss, n_pairs,
teacher, student_entropy, V, d,
method.
Details
Every crop goes through the student, only the GLOBAL crops go through
the teacher, and the loss is the cross-entropy of the student
distribution against the sharpened, centred teacher over all pairs of
different crops. With V views of which G are global that is G(V-1)
terms. Centring plus sharpening is what stops the collapse to a
constant output.
References
Caron et al. (2021), Emerging Properties in
Self-Supervised Vision Transformers, ICCV 2021:9650-9660.
