Update.
authorFrançois Fleuret <francois@fleuret.org>
Sat, 21 Oct 2023 21:07:54 +0000 (23:07 +0200)
committerFrançois Fleuret <francois@fleuret.org>
Sat, 21 Oct 2023 21:07:54 +0000 (23:07 +0200)
main.py
problems.py
qmlp.py
tasks.py

diff --git a/main.py b/main.py
index 6a1e8c1..6e87cda 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -257,7 +257,7 @@ default_task_args = {
     "degradation": {
         "model": "37M",
         "batch_size": 25,
-        "nb_train_samples": 100000,
+        "nb_train_samples": 250000,
         "nb_test_samples": 10000,
     },
     "mnist": {
index 7aa59be..22b6517 100755 (executable)
@@ -22,37 +22,51 @@ class Problem:
         nb_correct = ((result == input).long() * ar_mask).sum().item()
         return nb_total, nb_correct
 
+
 ####################
 
 
 class ProblemDegradation(Problem):
-    def __init__(self, nb_state_tokens=7, nb_time_steps=10, value_max=100, hard=False):
+    def __init__(self, nb_state_tokens=5, nb_time_steps=12, value_max=25, hard=False):
+        assert value_max // nb_state_tokens >= 2
         self.nb_state_tokens = nb_state_tokens
         self.nb_time_steps = nb_time_steps
         self.value_max = value_max
         self.hard = hard
 
-    def generate_sequences(self,nb):
-
-        x = (torch.rand(nb,self.nb_state_tokens).sort(dim=-1).indices == 0).long() * self.value_max
+    def generate_sequences(self, nb):
+        x = (
+            torch.rand(nb, self.nb_state_tokens).sort(dim=-1).indices == 0
+        ).long() * self.value_max
         seq = [x]
 
-        for t in range(self.nb_time_steps-1):
-            v = torch.rand(x.size()) * (x > 0).float()
-            u = (v.max(dim=-1,keepdim=True).values == v).long()
-            n = (u*x*torch.rand(x.size())).long().sum(dim=-1,keepdim=True) // 2
-            x = x + n * (u.roll(shifts=-1,dims=-1) - 2 * u + u.roll(shifts=1,dims=-1))
+        for t in range(self.nb_time_steps - 1):
+            v = (torch.rand(x.size()).sort(dim=-1).indices + 1) * (x >= 2).long()
+            u = (v.max(dim=-1, keepdim=True).values == v).long()
+            n = (
+                (u * x)
+                .minimum(2 + torch.randint(self.value_max // 4 - 2, x.size()))
+                .sum(dim=-1, keepdim=True)
+            )
+            m = 1 + ((n - 1) * torch.rand(n.size())).long()
+            x = (
+                x
+                + m * u.roll(shifts=-1, dims=-1)
+                - n * u
+                + (n - m) * u.roll(shifts=1, dims=-1)
+            )
             seq.append(x)
 
-        if self.hard: seq.reverse()
+        if self.hard:
+            seq.reverse()
 
-        seq = torch.cat(seq,dim=1)
-        return seq,seq.new_full(seq.size(), 1, dtype=torch.int64)
+        seq = torch.cat(seq, dim=1)
+        return seq, seq.new_full(seq.size(), 1, dtype=torch.int64)
 
     def compute_nb_correct(self, input, ar_mask, result):
         nb_total = result.size(0)
         nb_correct = 0
-        e=result.new_zeros(self.nb_state_tokens)
+        e = result.new_zeros(self.nb_state_tokens)
 
         for seq in result:
             states = list(seq.split(self.nb_state_tokens))
@@ -60,27 +74,38 @@ class ProblemDegradation(Problem):
                 states.reverse()
 
             d = states[0]
-            j=d.sort(descending=True).indices[0]
+            j = d.sort(descending=True).indices[0]
             e.zero_()
-            e[j]=self.value_max
-            if (d-e).abs().sum() == 0:
+            e[j] = self.value_max
+            if (d - e).abs().sum() == 0:
                 nb_errors = 0
-                for k in range(len(states)-1):
-                    d=states[k]-states[k+1]
-                    j=d.sort(descending=True).indices[0]
-                    e.zero_()
-                    e[j]=d[j]
-                    e[(j+1)%e.size(0)]=-d[j]//2
-                    e[(j-1)%e.size(0)]=-d[j]//2
-                    if (d-e).abs().sum() > 0:
+                for k in range(len(states) - 1):
+                    d = states[k + 1] - states[k]
+                    j = d.sort(descending=False).indices[0]
+                    if (
+                        d[j] == 0
+                        or d[j] > self.value_max // 4
+                        or d[(j + 1) % e.size(0)] <= 0
+                        or d[(j + 1) % e.size(0)] >= -d[j]
+                    ):
                         nb_errors += 1
+                    else:
+                        e.zero_()
+                        e[j] = d[j]
+                        e[(j + 1) % e.size(0)] = d[(j + 1) % e.size(0)]
+                        e[(j - 1) % e.size(0)] = -d[(j + 1) % e.size(0)] - d[j]
+                        if (d - e).abs().sum() > 0:
+                            nb_errors += 1
                 if nb_errors == 0:
                     nb_correct += 1
 
         return nb_total, nb_correct
 
     def seq2str(self, seq):
-        return " | ".join( [ " ".join([f"{x:02d}" for x in s ]) for s in seq.split(self.nb_state_tokens) ] )
+        return " | ".join(
+            [" ".join([f"{x:02d}" for x in s]) for s in seq.split(self.nb_state_tokens)]
+        )
+
 
 ####################
 
diff --git a/qmlp.py b/qmlp.py
index b58598a..abebfc1 100755 (executable)
--- a/qmlp.py
+++ b/qmlp.py
@@ -92,23 +92,37 @@ def generate_sets_and_params(
     test_input = dequantize(q_test_input, -1, 1)
 
     if save_as_examples:
-        a = 2 * torch.arange(nb_quantization_levels).float() / (nb_quantization_levels - 1) - 1
-        xf = torch.cat([a[:,None,None].expand(nb_quantization_levels, nb_quantization_levels,1),
-                        a[None,:,None].expand(nb_quantization_levels, nb_quantization_levels,1)], 2)
-        xf = xf.reshape(1,-1,2).expand(min(q_train_input.size(0),10),-1,-1)
+        a = (
+            2
+            * torch.arange(nb_quantization_levels).float()
+            / (nb_quantization_levels - 1)
+            - 1
+        )
+        xf = torch.cat(
+            [
+                a[:, None, None].expand(
+                    nb_quantization_levels, nb_quantization_levels, 1
+                ),
+                a[None, :, None].expand(
+                    nb_quantization_levels, nb_quantization_levels, 1
+                ),
+            ],
+            2,
+        )
+        xf = xf.reshape(1, -1, 2).expand(min(q_train_input.size(0), 10), -1, -1)
         print(f"{xf.size()=} {x.size()=}")
         yf = (
             (
-                (xf[:, None, :, 0] >= rec_support[:xf.size(0), :, None, 0]).long()
-                * (xf[:, None, :, 0] <= rec_support[:xf.size(0), :, None, 1]).long()
-                * (xf[:, None, :, 1] >= rec_support[:xf.size(0), :, None, 2]).long()
-                * (xf[:, None, :, 1] <= rec_support[:xf.size(0), :, None, 3]).long()
+                (xf[:, None, :, 0] >= rec_support[: xf.size(0), :, None, 0]).long()
+                * (xf[:, None, :, 0] <= rec_support[: xf.size(0), :, None, 1]).long()
+                * (xf[:, None, :, 1] >= rec_support[: xf.size(0), :, None, 2]).long()
+                * (xf[:, None, :, 1] <= rec_support[: xf.size(0), :, None, 3]).long()
             )
             .max(dim=1)
             .values
         )
 
-        full_input, full_targets = xf,yf
+        full_input, full_targets = xf, yf
 
         q_full_input = quantize(full_input, -1, 1)
         full_input = dequantize(q_full_input, -1, 1)
@@ -208,8 +222,12 @@ def generate_sets_and_params(
 
 
 def evaluate_q_params(
-        q_params, q_set, batch_size=25, device=torch.device("cpu"), nb_mlps_per_batch=1024,
-        save_as_examples=False,
+    q_params,
+    q_set,
+    batch_size=25,
+    device=torch.device("cpu"),
+    nb_mlps_per_batch=1024,
+    save_as_examples=False,
 ):
     errors = []
     nb_mlps = q_params.size(0)
index 0858282..7a4abbe 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -110,15 +110,20 @@ class SandBox(Task):
 
         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 
-
         # A bit of paranoia never hurts
-        assert (
-            self.nb_codes <= max_nb_codes
-            and self.train_input.min() >= 0
-            and self.test_input.min() >= 0
-            and tuple(x.item() for x in self.train_ar_mask.unique()) in { (0,), (1,), (0,1) }
-            and tuple(x.item() for x in self.test_ar_mask.unique()) in { (0,), (1,), (0,1) }
-        )
+        assert self.nb_codes <= max_nb_codes
+        assert self.train_input.min() >= 0
+        assert self.test_input.min() >= 0
+        assert tuple(x.item() for x in self.train_ar_mask.unique()) in {
+            (0,),
+            (1,),
+            (0, 1),
+        }
+        assert tuple(x.item() for x in self.test_ar_mask.unique()) in {
+            (0,),
+            (1,),
+            (0, 1),
+        }
 
     def batches(self, split="train", nb_to_use=-1, desc=None):
         assert split in {"train", "test"}
@@ -152,16 +157,21 @@ class SandBox(Task):
                 device=self.device,
             )
 
+            log_ground_truth = ar_mask.min() == 0
+
             if logger is not None:
                 for sp, st in zip(result[:10], input[:10]):
                     logger(
                         f"test_sequences {n_epoch} prediction   {self.problem.seq2str(sp)}"
                     )
-                    logger(
-                        f"               {n_epoch} ground truth {self.problem.seq2str(st)}"
-                    )
+                    if log_ground_truth:
+                        logger(
+                            f"               {n_epoch} ground truth {self.problem.seq2str(st)}"
+                        )
 
-            nb_total, nb_correct = self.problem.compute_nb_correct(input, ar_mask, result)
+            nb_total, nb_correct = self.problem.compute_nb_correct(
+                input, ar_mask, result
+            )
 
             # nb_total = ar_mask.sum().item()
             # nb_correct = ((result == input).long() * ar_mask).sum().item()