parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
+##############################
+# Expr options
+
+parser.add_argument("--expr_nb_variables", type=int, default=5)
+
+parser.add_argument("--expr_sequence_length", type=int, default=30)
+
######################################################################
args = parser.parse_args()
"nb_test_samples": 1000,
},
"expr": {
- "nb_epochs": 5,
+ "nb_epochs": 50,
"batch_size": 25,
- "nb_train_samples": 100000,
- "nb_test_samples": 1000,
+ "nb_train_samples": 250000,
+ "nb_test_samples": 10000,
},
}
self,
nb_train_samples,
nb_test_samples,
+ nb_variables,
+ sequence_length,
batch_size,
device=torch.device("cpu"),
):
self.batch_size = batch_size
self.device = device
- train_sequences = expr.generate_sequences(nb_train_samples)
- test_sequences = expr.generate_sequences(nb_test_samples)
+ train_sequences = expr.generate_sequences(
+ nb_train_samples,
+ nb_variables=nb_variables,
+ length=sequence_length,
+ # length=2 * sequence_length,
+ # randomize_length=True,
+ )
+ test_sequences = expr.generate_sequences(
+ nb_test_samples,
+ nb_variables=nb_variables,
+ length=sequence_length,
+ )
self.char2id = dict(
[
(c, n)
- for n, c in enumerate(set("".join(train_sequences + test_sequences)))
+ for n, c in enumerate(
+ set("#" + "".join(train_sequences + test_sequences))
+ )
]
)
- self.id2char = dict([(n, c) for n, c in self.char2id.items()])
- len_max = max([len(x) for x in train_sequences + test_sequences])
+ self.id2char = dict([(n, c) for c, n in self.char2id.items()])
+
+ self.filler, self.space = self.char2id["#"], self.char2id[" "]
+
+ len_max = max([len(x) for x in train_sequences])
self.train_input = torch.cat(
[
torch.tensor(
- [char2id(c) for c in s + " " * (len_max - len(s))]
- for s in train_sequences
+ [
+ [self.char2id[c] for c in s + "#" * (len_max - len(s))]
+ for s in train_sequences
+ ]
)
],
0,
- )
+ ).to(device)
+
+ len_max = max([len(x) for x in test_sequences])
self.test_input = torch.cat(
[
torch.tensor(
- [char2id(c) for c in s + " " * (len_max - len(s))]
- for s in test_sequences
+ [
+ [self.char2id[c] for c in s + "#" * (len_max - len(s))]
+ for s in test_sequences
+ ]
)
],
0,
- )
+ ).to(device)
+
self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
def batches(self, split="train", nb_to_use=-1, desc=None):
for batch in tqdm.tqdm(
input.split(self.batch_size), dynamic_ncols=True, desc=desc
):
+ if split == "train":
+ last = (batch != self.filler).max(0).values.nonzero().max() + 1
+ batch = batch[:, :last]
yield batch
def vocabulary_size(self):
return self.nb_codes
+ def seq2str(self, s):
+ return "".join([self.id2char[k.item()] for k in s])
+
def produce_results(self, n_epoch, model):
- # !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
with torch.autograd.no_grad():
t = model.training
model.eval()
def compute_nb_correct(input):
result = input.clone()
- stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
- ar_mask = (result != input).long()
+ ar_mask = (result == self.space).long().cumsum(dim=1).clamp(max=1)
+ result = (1 - ar_mask) * result + ar_mask * self.filler
masked_inplace_autoregression(
model, self.batch_size, result, ar_mask, device=self.device
)
- errors = ((result != input).long() * ar_mask).reshape(
- -1, 1 + self.nb_digits
- )
- ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
+ nb_total = input.size(0)
+ nb_correct = (input == result).long().min(1).values.sum()
- nb_total = ar_mask.max(1).values.sum()
- nb_correct = nb_total - errors.max(1).values.sum()
+ #######################################################################
+ # Comput predicted vs. true variable values
- return nb_total, nb_correct
+ nb_delta = torch.zeros(5, dtype=torch.int64)
+ nb_missed = 0
- test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
+ values_input = expr.extract_results([self.seq2str(s) for s in input])
+ values_result = expr.extract_results([self.seq2str(s) for s in result])
+
+ for i, r in zip(values_input, values_result):
+ for n, vi in i.items():
+ vr = r.get(n)
+ if vr is None or vr < 0:
+ nb_missed += 1
+ else:
+ d = abs(vr-vi)
+ if d >= nb_delta.size(0):
+ nb_missed += 1
+ else:
+ nb_delta[d] += 1
+
+ ######################################################################
+
+ return nb_total, nb_correct, nb_delta, nb_missed
+
+ test_nb_total, test_nb_correct, test_nb_delta, test_nb_missed = compute_nb_correct(self.test_input[:1000])
log_string(
f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
)
+ nb_total = test_nb_delta.sum() + test_nb_missed
+ for d in range(test_nb_delta.size(0)):
+ log_string(f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%")
+ log_string(f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%")
+
+
##############################################################
# Log a few generated sequences
- input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
+ input = self.test_input[:10]
result = input.clone()
- stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
- ar_mask = (result != input).long()
+ ar_mask = (result == self.space).long().cumsum(dim=1).clamp(max=1)
+ result = (1 - ar_mask) * result + ar_mask * self.filler
for n in range(result.size(0)):
- log_string(
- f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
- )
+ log_string(f"test_before {self.seq2str(result[n])}")
masked_inplace_autoregression(
model, self.batch_size, result, ar_mask, device=self.device
)
+ correct = (1 - ar_mask) * self.space + ar_mask * input
for n in range(result.size(0)):
- log_string(
- f"test_after {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
- )
+ comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
+ log_string(f"test_after {self.seq2str(result[n])} {comment}")
+ log_string(f"correct {self.seq2str(correct[n])}")
##############################################################
model.train(t)
task = TaskExpr(
nb_train_samples=args.nb_train_samples,
nb_test_samples=args.nb_test_samples,
+ nb_variables=args.expr_nb_variables,
+ sequence_length=args.expr_sequence_length,
batch_size=args.batch_size,
device=device,
)