Update.

[mygptrnn.git] / mygpt.py

diff --git a/mygpt.py b/mygpt.py
index f10f1fe..7aa8578 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -457,7 +457,8 @@ def moving_window(x, dim, win_dim, win_size):
 
 ##############################
 
-# This is one order of magnitude more complicated than I expected
+# 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):
@@ -472,8 +473,9 @@ def flash_back_time_src(N, H, t0, t1, CL, CH, proba, device):
     fb_body = fb_body.cumsum(dim=2)
     fb_start = fb_start * (fb_body == 1)
 
-    # t_s = t0-(t0//L * R)*L
-
+    # 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
@@ -507,32 +509,22 @@ def insert_flash_back(rec_V, V, rec_K, K, t0, t1, CL, proba):
 
     fbt, fbh = flash_back_time_src(N, H, t0, t1, CL, CH, proba, rec_V.device)
 
-    fbt_V = fbt[:, :, :, None].expand_as(rec_V[:, :, t0:t1])
-    fbh_V = fbh[:, :, :, None].expand_as(rec_V[:, :, t0:t1])
+    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].expand_as(
-        rec_V[:, :, t0:t1]
-    )
-    d = torch.arange(V.size(3), device=V.device)[None, None, None, :].expand_as(
-        rec_V[:, :, t0:t1]
-    )
+    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].expand_as(rec_K[:, :, t0:t1])
-    fbh_K = fbh[:, :, :, None].expand_as(rec_K[:, :, t0:t1])
+    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].expand_as(
-        rec_K[:, :, t0:t1]
-    )
-    d = torch.arange(K.size(3), device=K.device)[None, None, None, :].expand_as(
-        rec_K[:, :, t0:t1]
-    )
+    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)
 
-    # print("SANITY", (fbt_K >=0).float().sum()/fbt_K.numel())
-
 
 ######################################################################
 
@@ -560,6 +552,9 @@ class Caterpillar(nn.Module):
         self.caterpillar_height = caterpillar_height
         self.attention_dropout = attention_dropout
 
+        warnings.warn("flash back", RuntimeWarning)
+        self.proba_flashback = 0.1
+
         self.w_G = randw(nb_heads, caterpillar_height, dim_model)
         self.b_G = nn.Parameter(
             torch.full(
@@ -591,6 +586,7 @@ class Caterpillar(nn.Module):
 
         N = bs.x.size(0)
         T = bs.x.size(1)
+        H = self.w_V.size(0)
         DV = self.w_V.size(1)
         DK = self.w_K.size(1)
         DM = self.w_O.size(1)
@@ -659,9 +655,43 @@ class Caterpillar(nn.Module):
         self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3)
         self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
 
-        warnings.warn("flash back", RuntimeWarning)
-        if self.training:
-            insert_flash_back(self.rec_V, V, self.rec_K, K, t0, t1, CL, proba=1e-2 / CL)
+        if self.training and self.proba_flashback:
+            # insert_flash_back(
+            # self.rec_V,
+            # V,
+            # self.rec_K,
+            # K,
+            # t0,
+            # t1,
+            # CL,
+            # proba=self.proba_flashback / CL,
+            # )
+
+            n = torch.arange(N, device=X.device)[:, None, None, None]
+            t = torch.arange(t0, t1, device=X.device)[None, None, :, None]
+            dv = torch.arange(DV)[None, None, None, :]
+            dk = torch.arange(DK)[None, None, None, :]
+
+            u = (
+                torch.rand(N, CH, t1 - t0, 1, device=X.device).mul(t).long() // CL
+            ) * CL
+
+            src_time = t - u - t0
+            src_head = torch.randint(H, (N, CH, t1 - t0, 1), device=X.device)
+
+            mask_V = (torch.rand(N, CH, t1 - t0, DV) <= self.proba_flashback).long()
+            self.rec_V[:, :, t0:t1] = (
+                mask_V * V[n, src_head, src_time, dv]
+                + (1 - mask_V) * self.rec_V[:, :, t0:t1]
+            )
+
+            mask_K = (torch.rand(N, CH, t1 - t0, DK) <= self.proba_flashback).long()
+            self.rec_K[:, :, t0:t1] = (
+                mask_K * K[n, src_head, src_time, dk]
+                + (1 - mask_K) * self.rec_K[:, :, t0:t1]
+            )
+
+        exit(0)
 
         ######################################################################
         # compute the readout