######################################################################
+
+
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)
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()
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)
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):
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()
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)))
dim_keys=2,
dim_hidden=2,
nb_heads=2,
- nb_blocks=3,
+ nb_blocks=5,
dropout=0.1,
causal=True,
)
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,
+ )