Update.

[mygptrnn.git] / mygpt.py

diff --git a/mygpt.py b/mygpt.py
index b5ac916..eda8685 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -457,77 +457,6 @@ def moving_window(x, dim, win_dim, win_size):
 
 ##############################
 
-# This is one order of magnitude more complicated than I expected, not
-# elegant, slow, hopefully not buggy
-
-
-def flash_back_time_src(N, H, t0, t1, CL, CH, proba, device):
-    # starting flash backs
-    fb_start = (torch.rand(N, CH, t1 - t0, device=device) <= proba).long()
-    fb_start[:, :, -CL:] = 0
-    fb_start[:, :, :CL] = 0
-
-    # Remove series longer than CL
-    fb_body = fb_start.clone()
-    fb_body[:, :, CL + 1 :] -= fb_start[:, :, : -(CL + 1)]
-    fb_body = fb_body.cumsum(dim=2)
-    fb_start = fb_start * (fb_body == 1)
-
-    # Set a origin source time (starting time of the chunck to copy
-    # here) We set it as the current time minus a multiple of CL to be
-    # consistent with the "rolling" caterpillar
-    t = torch.arange(fb_start.size(2), device=fb_start.device)[None, None, :]
-    src_time = fb_start * (
-        t
-        - CL
-        * (
-            1
-            + (
-                torch.rand(fb_start.size(), device=fb_start.device) * (t // CL - 1)
-            ).long()
-        )
-    )
-    src_time[:, :, CL:] -= src_time.clone()[:, :, :-CL]
-    src_time = src_time.cumsum(dim=2)
-
-    src_head = fb_start * torch.randint(H, fb_start.size(), device=fb_start.device)
-    src_head[:, :, CL:] -= src_head.clone()[:, :, :-CL]
-    src_head = src_head.cumsum(dim=2)
-
-    # combine
-    src_delta = fb_start.clone()
-    src_delta[:, :, CL:] -= fb_start[:, :, :-CL]
-    src_delta = src_delta.cumsum(dim=2)
-    src_delta[:, :, CL:] -= CL * fb_start[:, :, :-CL]
-    src_time += src_delta.cumsum(dim=2) - 1
-
-    return src_time, src_head
-
-
-def insert_flash_back(rec_V, V, rec_K, K, t0, t1, CL, proba):
-    N, H, CH = V.size(0), V.size(1), rec_V.size(1)
-
-    fbt, fbh = flash_back_time_src(N, H, t0, t1, CL, CH, proba, rec_V.device)
-
-    fbt_V = fbt[:, :, :, None]
-    fbh_V = fbh[:, :, :, None]
-    t = fbt_V.clamp(min=0)
-    n = torch.arange(V.size(0), device=V.device)[:, None, None, None]
-    d = torch.arange(V.size(3), device=V.device)[None, None, None, :]
-    q = V[:, :, t0:t1][n, fbh_V, t, d]
-    rec_V[:, :, t0:t1] = q * (fbt_V >= 0) + rec_V[:, :, t0:t1] * (fbt_V < 0)
-
-    fbt_K = fbt[:, :, :, None]
-    fbh_K = fbh[:, :, :, None]
-    t = fbt_K.clamp(min=0)
-    n = torch.arange(K.size(0), device=K.device)[:, None, None, None]
-    d = torch.arange(K.size(3), device=K.device)[None, None, None, :]
-    q = K[:, :, t0:t1][n, fbh_K, t, d]
-    rec_K[:, :, t0:t1] = q * (fbt_K >= 0) + rec_K[:, :, t0:t1] * (fbt_K < 0)
-
-
-######################################################################
-
 
 class Caterpillar(nn.Module):
     def __init__(