Update.

[picoclvr.git] / graph.py

diff --git a/graph.py b/graph.py
index 5bab861..6db9ed7 100755 (executable)
--- a/graph.py
+++ b/graph.py
@@ -11,29 +11,40 @@ import cairo
 
 
 ######################################################################
+
+
 def save_attention_image(
-    filename,
+    filename,  # image to save
     tokens_input,
     tokens_output,
-    attention,
-    n_sample=0,
-    n_head=0,
+    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 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=10,
+    token_gap=15,
     layer_gap=25,
     y_eps=0.5,
     padding=10,
-    min_att=0,
-    k_top=None,
 ):
-    attention = torch.cat(
-        [x[n_sample : n_sample + 1, n_head] for x in attention], dim=0
-    )
-
     if k_top is not None:
-        attention = attention * (
-            attention.sort(dim=-1, descending=True).indices < k_top
-        )
+        am = []
+        for m in attention_matrices:
+            am.append(m * (m.sort(dim=-1, descending=True).indices < k_top))
+        attention_matrices = am
+
+    if min_total_attention is not None:
+        am = []
+        for m in attention_matrices:
+            s = m.sort(dim=-1)
+            m = 1 - (s.values.cumsum(-1) < 1 - min_total_attention).long()
+            b = m.new(m.size()).scatter_(dim=-1, index=s.indices, src=m)
+            am.append(m * b)
 
     surface = cairo.RecordingSurface(cairo.CONTENT_COLOR_ALPHA, None)
 
@@ -46,14 +57,9 @@ def save_attention_image(
 
     x, y = 0, 0
 
-    for d in range(attention.size(0)):
-        if d > 0:
-            for n in range(attention.size(-1)):
-                xc, yc = n * token_gap, -d * layer_gap
-                ctx.arc(xc, yc, token_gap / 10, 0, 2 * math.pi)
-                ctx.fill()
-
-        at = attention[d]
+    ctx.set_line_width(0.25)
+    for d in range(len(attention_matrices)):
+        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()
@@ -62,23 +68,34 @@ def save_attention_image(
         ni = ni.flatten()[o]
         nj = nj.flatten()[o]
         for i, j, a in zip(ni, nj, at):
-            if a > 0 and a >= min_att:
+            if a > 0 and a >= min_link_attention:
                 c = 1 - a.item()
                 ctx.set_source_rgb(c, c, c)
-                ctx.set_line_width(0.5)
-                ctx.move_to(j * token_gap, y - y_eps)
-                ctx.line_to(i * token_gap, y - layer_gap + y_eps)
+                ax, ay = j * token_gap, y - y_eps
+                ctx.move_to(ax, ay)
+                dx, dy = i * token_gap, y - layer_gap + y_eps
+                if curved:
+                    bx, by = ax, ay - layer_gap * 0.5
+                    cx, cy = dx, dy + layer_gap * 0.5
+                    ctx.curve_to(bx, by, cx, cy, dx, dy)
+                else:
+                    ctx.line_to(dx, dy)
                 ctx.stroke()
         y -= layer_gap
 
-    for d in range(1, attention.size(0)):
-        for n in range(attention.size(-1)):
+    for d in range(0, len(attention_matrices) + 1):
+        n = (
+            attention_matrices[0].size(-1)
+            if d == 0
+            else attention_matrices[d - 1].size(-2)
+        )
+        for n in range(n):
             xc, yc = n * token_gap, -d * layer_gap
             ctx.set_source_rgb(1.0, 1.0, 1.0)
-            ctx.arc(xc, yc, token_gap / 10 + 0.5, 0, 2 * math.pi)
+            ctx.arc(xc, yc, token_gap / 10, 0, 2 * math.pi)
             ctx.fill()
             ctx.set_source_rgb(0.0, 0.0, 0.0)
-            ctx.arc(xc, yc, token_gap / 10, 0, 2 * math.pi)
+            ctx.arc(xc, yc, token_gap / 20, 0, 2 * math.pi)
             ctx.fill()
 
     ctx.set_source_rgb(0.0, 0.0, 0.0)
@@ -93,7 +110,7 @@ def save_attention_image(
             x_advance,
             y_advance,
         ) = ctx.text_extents(s)
-        ctx.move_to(k * token_gap - width_t / 2, -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):
@@ -106,7 +123,10 @@ def save_attention_image(
             x_advance,
             y_advance,
         ) = ctx.text_extents(s)
-        ctx.move_to(k * token_gap - width_t / 2, -attention.size(0) * layer_gap)
+        ctx.move_to(
+            k * token_gap - width_t / 2,
+            -token_gap / 5 - len(attention_matrices) * layer_gap,
+        )
         ctx.show_text(s)
 
     x, y, width, height = surface.ink_extents()
@@ -126,8 +146,8 @@ def save_attention_image(
 if __name__ == "__main__":
     import mygpt
 
-    tokens_output = ["bluh", 2, 3, 4, "blih"]
-    tokens_input = ["n/a"] + tokens_output[:-1]
+    tokens_output = ["<wat>", "-", 3, 4, "<end>"]
+    tokens_input = [""] + tokens_output[:-1]
 
     vocabulary_size = 3
     x = torch.randint(vocabulary_size, (1, len(tokens_input)))
@@ -138,7 +158,7 @@ if __name__ == "__main__":
         dim_keys=2,
         dim_hidden=2,
         nb_heads=2,
-        nb_blocks=3,
+        nb_blocks=5,
         dropout=0.1,
         causal=True,
     )
@@ -148,6 +168,16 @@ if __name__ == "__main__":
 
     y1 = model(mygpt.BracketedSequence(x)).x
 
-    attention = model.retrieve_attention()
+    attention_matrices = [m[0, 0] for m in model.retrieve_attention()]
+
+    # attention_matrices = [ torch.rand(3,5), torch.rand(8,3), torch.rand(5,8) ]
+    # for a in attention_matrices: a=a/a.sum(-1,keepdim=True)
 
-    save_attention_image("attention.pdf", tokens_input, tokens_output, attention)
+    save_attention_image(
+        "attention.pdf",
+        tokens_input,
+        tokens_output,
+        attention_matrices,
+        # k_top=2,
+        min_total_attention=0.9,
+    )