X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=graph.py;h=6db9ed7bc53c5a6adc3a3c9047139feac4ed087e;hb=b87078aec53ead1e0a3ca44d4ac46c319bbcd63e;hp=a819283cc28450c28b54a6e5b6b215c408c65243;hpb=3b9ba21fd3d06a20703216cc0a77fe9dc78b079f;p=picoclvr.git

diff --git a/graph.py b/graph.py
index a819283..6db9ed7 100755
--- a/graph.py
+++ b/graph.py
@@ -14,24 +14,23 @@ import cairo
 
 
 def save_attention_image(
-    filename,
+    filename,  # image to save
     tokens_input,
     tokens_output,
-    # An iterable set of BxHxTxT attention matrices
-    attention_matrices,
-    pixel_scale=8,
-    token_gap=15,
-    layer_gap=25,
-    y_eps=0.5,
-    padding=10,
+    attention_matrices,  # list of 2d tensors T1xT2, T2xT3, ..., Tk-1xTk
     # do not draw links with a lesser attention
     min_link_attention=0,
-    # draw only the strongest links necessary to reache
-    # min_total_attention
+    # draw only the strongest links necessary so that their summed
+    # attention is above min_total_attention
     min_total_attention=None,
     # draw only the top k links
     k_top=None,
     curved=True,
+    pixel_scale=8,
+    token_gap=15,
+    layer_gap=25,
+    y_eps=0.5,
+    padding=10,
 ):
     if k_top is not None:
         am = []
@@ -60,7 +59,7 @@ def save_attention_image(
 
     ctx.set_line_width(0.25)
     for d in range(len(attention_matrices)):
-        at = attention_matrices[d]
+        at = attention_matrices[d].to("cpu")
         ni = torch.arange(at.size(0))[:, None].expand_as(at)
         nj = torch.arange(at.size(1))[None, :].expand_as(at)
         at = at.flatten()
@@ -111,7 +110,7 @@ def save_attention_image(
             x_advance,
             y_advance,
         ) = ctx.text_extents(s)
-        ctx.move_to(k * token_gap - width_t / 2, token_gap / 5 - y_bearing)
+        ctx.move_to(k * token_gap - width_t / 2, 2 * token_gap / 5)
         ctx.show_text(s)
 
     for k, t in enumerate(tokens_output):
@@ -147,7 +146,7 @@ def save_attention_image(
 if __name__ == "__main__":
     import mygpt
 
-    tokens_output = ["<wat>", 2, 3, 4, "<end>"]
+    tokens_output = ["<wat>", "-", 3, 4, "<end>"]
     tokens_input = [""] + tokens_output[:-1]
 
     vocabulary_size = 3