def save_attention_image(
+ # image to save
filename,
tokens_input,
tokens_output,
- # An iterable set of BxHxTxT attention matrices
+ # list of 2d tensors T2xT1, T3xT2, ..., TkxTk-1
attention_matrices,
- pixel_scale=8,
- token_gap=15,
- layer_gap=25,
- y_eps=0.5,
- padding=10,
# 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,
+ # the purely graphical settings
curved=True,
+ pixel_scale=8,
+ token_gap=15,
+ layer_gap=25,
+ y_eps=0.5,
+ padding=10,
):
if k_top is not None:
am = []
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()
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):
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
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)
+ # attention_matrices = [torch.rand(*s) for s in [ (4,5),(3,4),(8,3),(5,8) ]]
save_attention_image(
"attention.pdf",