#!/usr/bin/env python
+# Any copyright is dedicated to the Public Domain.
+# https://creativecommons.org/publicdomain/zero/1.0/
+
+# Written by Francois Fleuret <francois@fleuret.org>
+
import math, os, tqdm
import torch, torchvision
from torch import nn
from torch.nn import functional as F
+from mygpt import BracketedSequence
+
+try:
+ from graph import save_attention_image
+except ImportError:
+ save_attention_image = None
+
######################################################################
batches,
dynamic_ncols=True,
desc=progress_bar_desc,
- # total=input.size(0) // batch_size,
+ total=(input.size(0) + batch_size - 1) // batch_size,
)
with torch.autograd.no_grad():
source = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
marker2 = torch.full((nb, 1), 11)
result = operators.bmm(source[:, :, None]).squeeze(-1)
- print(f"{nb_operators.dtype=} {marker1.dtype=}")
sequences = torch.cat((nb_operators, marker1, source, marker2, result), 1)
- print(f"{sequences.size()=}")
ar_mask = (sequences == 11).long()
ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
return sequences, ar_mask
)
],
0,
- ).to(self.device)
+ )
def seq2str(self, seq):
return " ".join([self.id2token[i] for i in seq])
max_input=9,
prog_len=6,
nb_runs=5,
+ no_prog=False,
+ logger=None,
device=torch.device("cpu"),
):
super().__init__()
self.batch_size = batch_size
self.device = device
+ self.no_prog = no_prog
train_sequences = [
rpl.generate(
nb_starting_values=nb_starting_values,
+ nb_result_values_max=4 * nb_starting_values,
max_input=max_input,
prog_len=prog_len,
nb_runs=nb_runs,
test_sequences = [
rpl.generate(
nb_starting_values=nb_starting_values,
+ nb_result_values_max=4 * nb_starting_values,
max_input=max_input,
prog_len=prog_len,
nb_runs=nb_runs,
symbols = list(filter(lambda x: type(x) is str, symbols))
symbols.sort()
symbols += [str(n) for n in range(val_max + 1)]
- print(f"{val_max=}")
self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
self.id2token = dict([(n, c) for c, n in self.token2id.items()])
- self.t_nul, self.t_prog = self.token2id["<nul>"], self.token2id["<prog>"]
+ self.t_nul = self.token2id["<nul>"]
+ self.t_input = self.token2id["<in>"]
+ self.t_output = self.token2id["<out>"]
+ self.t_prog = self.token2id["<prg>"]
+ self.t_end = self.token2id["<end>"]
self.train_input = self.tensorize(train_sequences)
self.test_input = self.tensorize(test_sequences)
+ if no_prog:
+ # Excise the program from every train and test example
+ k = torch.arange(self.train_input.size(1), device=self.train_input.device)[
+ None, :
+ ]
+ p = (
+ ((self.train_input == self.t_prog).long() * k)
+ .max(1, keepdim=True)
+ .values
+ )
+ self.train_input = (
+ self.train_input * (k <= p).long()
+ + self.t_end * (k == p + 1).long()
+ + self.t_nul * (k > p + 1).long()
+ )
+ k = torch.arange(self.test_input.size(1), device=self.test_input.device)[
+ None, :
+ ]
+ p = (
+ ((self.test_input == self.t_prog).long() * k)
+ .max(1, keepdim=True)
+ .values
+ )
+ self.test_input = (
+ self.test_input * (k <= p).long()
+ + self.t_end * (k == p + 1).long()
+ + self.t_nul * (k > p + 1).long()
+ )
+
+ if logger is not None:
+ logger(f"value_max {val_max}")
+ for x in self.train_input[:25]:
+ end = (x != self.t_nul).nonzero().max().item() + 1
+ seq = [self.id2token[i.item()] for i in x[:end]]
+ s = " ".join(seq)
+ logger(f"example_seq {s}")
+
self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
def batches(self, split="train", nb_to_use=-1, desc=None):
input.split(self.batch_size), dynamic_ncols=True, desc=desc
):
last = (batch != self.t_nul).max(0).values.nonzero().max() + 3
- batch = batch[:, :last]
+ batch = batch[:, :last].to(self.device)
yield batch
def vocabulary_size(self):
def produce_results(
self, n_epoch, model, result_dir, logger, deterministic_synthesis
):
- def compute_nb_errors(input, nb_to_log=0):
+ # --------------------------------------------------------------------
+ def compute_nb_errors_prog(input, nb_to_log=0):
result = input.clone()
s = (result == self.t_prog).long()
ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
)
sum_nb_total, sum_nb_errors = 0, 0
- for x, y in zip(input, result):
- seq = [self.id2token[i.item()] for i in y]
+ for one_input, one_result in zip(input, result):
+ seq = [self.id2token[i.item()] for i in one_result]
nb_total, nb_errors, prog, stacks = rpl.compute_nb_errors(seq)
sum_nb_total += 1
sum_nb_errors += 0 if nb_errors == 0 else 1
if nb_to_log > 0:
- gt_seq = [self.id2token[i.item()] for i in x]
+ gt_seq = [self.id2token[i.item()] for i in one_input]
_, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
gt_prog = " ".join([str(x) for x in gt_prog])
prog = " ".join([str(x) for x in prog])
- logger(f"PROG [{gt_prog}] PREDICTED [{prog}]")
+ comment = "*" if nb_errors == 0 else "-"
+ logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
for start_stack, target_stack, result_stack, correct in stacks:
- comment = " CORRECT" if correct else ""
+ comment = "*" if correct else "-"
start_stack = " ".join([str(x) for x in start_stack])
target_stack = " ".join([str(x) for x in target_stack])
result_stack = " ".join([str(x) for x in result_stack])
logger(
- f" [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]{comment}"
+ f" {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
)
nb_to_log -= 1
return sum_nb_total, sum_nb_errors
- test_nb_total, test_nb_errors = compute_nb_errors(
- self.test_input[:1000], nb_to_log=10
+ # --------------------------------------------------------------------
+ def compute_nb_errors_output(input, nb_to_log=0):
+ result = input.clone()
+ k = torch.arange(result.size(1), device=result.device)[None, :]
+ last_output_idx = (
+ ((result == self.t_output) * k).max(dim=1, keepdim=True).values
+ )
+ first_prog_idx = (
+ ((result == self.t_prog) * k).max(dim=1, keepdim=True).values
+ )
+ ar_mask = (k > last_output_idx).long() * (k < first_prog_idx).long()
+ result = (1 - ar_mask) * result + ar_mask * self.t_nul
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis,
+ device=self.device,
+ )
+
+ sum_nb_total, sum_nb_errors = 0, 0
+ for one_input, one_result, i, j in zip(
+ input, result, last_output_idx, first_prog_idx
+ ):
+ seq = [self.id2token[i.item()] for i in one_result]
+ sum_nb_total += 1
+ correct = (one_input - one_result).abs().max() == 0
+ sum_nb_errors += 0 if correct else 1
+ if nb_to_log > 0:
+ result_stack = [
+ self.id2token[i.item()] for i in one_result[i : j + 1]
+ ]
+ target_stack = [
+ self.id2token[i.item()] for i in one_input[i : j + 1]
+ ]
+ comment = "*" if correct else "-"
+ result_stack = " ".join([str(x) for x in result_stack])
+ target_stack = " ".join([str(x) for x in target_stack])
+ logger(
+ f"output_test {comment} [{target_stack}] PREDICTED [{result_stack}]"
+ )
+ nb_to_log -= 1
+
+ return sum_nb_total, sum_nb_errors
+
+ # --------------------------------------------------------------------
+
+ if not self.no_prog:
+ test_nb_total, test_nb_errors = compute_nb_errors_prog(
+ self.test_input[:1000].to(self.device), nb_to_log=10
+ )
+
+ logger(
+ f"accuracy_prog_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
+ )
+
+ test_nb_total, test_nb_errors = compute_nb_errors_output(
+ self.test_input[:1000].to(self.device), nb_to_log=10
)
logger(
- f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
+ f"accuracy_output_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
)
+ if save_attention_image is not None:
+ input = self.test_input[:1]
+ result = input.clone()
+ s = (result == self.t_prog).long()
+ ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
+ result = (1 - ar_mask) * result + ar_mask * self.t_nul
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis,
+ device=self.device,
+ )
+
+ with torch.autograd.no_grad():
+ t = model.training
+ model.eval()
+ model.record_attention(True)
+ model(BracketedSequence(result))
+ model.train(t)
+ ram = model.retrieve_attention()
+ model.record_attention(False)
+
+ tokens_output = [self.id2token[i.item()] for i in result[0]]
+ tokens_input = ["n/a"] + tokens_output[:-1]
+ for n_head in range(ram[0].size(1)):
+ filename = f"rpl_attention_{n_epoch}_h{n_head}.pdf"
+ attention_matrices = [m[0, n_head] for m in ram]
+ save_attention_image(
+ filename,
+ tokens_input,
+ tokens_output,
+ attention_matrices,
+ token_gap=12,
+ layer_gap=50,
+ k_top=10,
+ # min_total_attention=0.9,
+ )
+ logger(f"wrote {filename}")
+
######################################################################