X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=d21e2648466a606b3067fc680feb7305a6b95781;hb=8ea809c43242d3a2e063692105919a86c3f6fe6b;hp=421aee49f2f0005f7650ea9836fde801fc5598e8;hpb=db7cefe4fefb381e56f1292d5bbe4a18c76afb47;p=picoclvr.git

diff --git a/tasks.py b/tasks.py
index 421aee4..d21e264 100755
--- a/tasks.py
+++ b/tasks.py
@@ -14,10 +14,8 @@ from torch.nn import functional as F
 
 from mygpt import BracketedSequence
 
-try:
-    from graph import save_attention_image
-except ImportError:
-    save_attention_image = None
+# from graph import save_attention_image
+save_attention_image = None
 
 ######################################################################
 
@@ -72,10 +70,168 @@ class Task:
         pass
 
 
+class TaskFromFile(Task):
+    def tensorize(self, pairs, shuffle):
+        len_max = max([len(x[0]) for x in pairs])
+
+        input = torch.cat(
+            [
+                torch.tensor(
+                    [
+                        [self.char2id[c] for c in s[0] + "#" * (len_max - len(s[0]))]
+                        for s in pairs
+                    ]
+                )
+            ],
+            0,
+        ).to("cpu")
+
+        pred_mask = torch.cat(
+            [
+                torch.tensor(
+                    [
+                        [int(c) for c in s[1] + "0" * (len_max - len(s[1]))]
+                        for s in pairs
+                    ]
+                )
+            ],
+            0,
+        ).to("cpu")
+
+        if shuffle:
+            print("SHUFFLING!")
+            i = torch.randperm(input.size(0))
+            input = input[i].contiguous()
+            pred_mask = pred_mask[i].contiguous()
+
+        return input, pred_mask
+
+    # trim all the tensors in the tuple z to remove as much token from
+    # left and right in the first tensor. If z is a tuple, all its
+    # elements are trimed according to the triming for the first
+    def trim(self, z, token="#"):
+        n = self.char2id[token]
+        if type(z) == tuple:
+            x = z[0]
+            i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
+            a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
+            return tuple([t[:, a:b] for t in z])
+        else:
+            i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
+            a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
+            return z[:, a:b]
+
+    def __init__(
+        self,
+        train_filename,
+        test_filename,
+        nb_train_samples,
+        nb_test_samples,
+        batch_size,
+        shuffle=False,
+        device=torch.device("cpu"),
+    ):
+        self.batch_size = batch_size
+        self.device = device
+
+        def read_file(filename, nb=-1):
+            pairs = []
+            with open(filename, "r") as f:
+                while True:
+                    sequence = f.readline().strip()
+                    if not sequence:
+                        break
+                    pred_mask = f.readline().strip()
+                    assert len(sequence) == len(pred_mask)
+                    assert set(pred_mask).issubset({"0", "1", "2"}), f"{set(pred_mask)}"
+                    pairs.append((sequence, pred_mask))
+                    if len(pairs) == nb:
+                        break
+
+            if nb > 0:
+                pairs = pairs[:nb]
+                assert len(pairs) == nb
+
+            return pairs
+
+        train_pairs = read_file(train_filename, nb_train_samples)
+        test_pairs = read_file(test_filename, nb_test_samples)
+
+        symbols = ["#"] + list(
+            set("".join([x[0] for x in train_pairs + test_pairs])) - set(["#"])
+        )
+        self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
+        self.id2char = dict([(n, c) for c, n in self.char2id.items()])
+
+        self.train_input, self.train_pred_masks = self.tensorize(
+            train_pairs, shuffle=shuffle
+        )
+        self.test_input, self.test_pred_masks = self.tensorize(
+            test_pairs, shuffle=shuffle
+        )
+
+    def batches(self, split="train", nb_to_use=-1, desc=None):
+        assert split in {"train", "test"}
+        input = self.train_input if split == "train" else self.test_input
+        if nb_to_use > 0:
+            input = input[:nb_to_use]
+        if desc is None:
+            desc = f"epoch-{split}"
+        for batch in tqdm.tqdm(
+            input.split(self.batch_size), dynamic_ncols=True, desc=desc
+        ):
+            yield self.trim(batch).to(self.device)
+
+    def vocabulary_size(self):
+        return len(self.char2id)
+
+    def tensor2str(self, t):
+        return ["".join([self.id2char[x.item()] for x in s]) for s in t]
+
+    def produce_results(
+        self, n_epoch, model, result_dir, logger, deterministic_synthesis
+    ):
+        correct = self.trim(self.test_input[:1000]).to(self.device)
+        result = correct.clone()
+        pred_mask = self.test_pred_masks[:1000, : result.size(1)].to(self.device)
+        ar_mask = (pred_mask > 0).long()
+        result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
+
+        logger(f"----------------------------------------------------------")
+
+        for e in self.tensor2str(result[:50]):
+            logger(f"test_before {e}")
+
+        masked_inplace_autoregression(
+            model,
+            self.batch_size,
+            result,
+            ar_mask,
+            deterministic_synthesis,
+            device=self.device,
+        )
+
+        logger(f"----------------------------------------------------------")
+
+        for e, c in zip(self.tensor2str(result[:50]), self.tensor2str(correct[:50])):
+            logger(f"test_after  {e}")
+            logger(f"correct     {c}")
+
+        logger(f"----------------------------------------------------------")
+
+        err_mask = (pred_mask == 2).long()
+        nb_total = err_mask.sum().item()
+        nb_correct = ((correct == result).long() * err_mask).sum().item()
+
+        logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
+        logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
+
+
 ####################
 
 import problems
 
+
 class SandBox(Task):
     def __init__(
         self,
@@ -110,13 +266,25 @@ 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(self.train_ar_mask.unique()) == (0, 1)
-            and tuple(self.test_ar_mask.unique()) == (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),
+        }
+
+        if logger is not None:
+            for s, a in zip(self.train_input[:100], self.train_ar_mask[:100]):
+                logger(f"train_sequences {self.problem.seq2str(s)}")
+                a = "".join(["01"[x.item()] for x in a])
+                logger(f"                {a}")
 
     def batches(self, split="train", nb_to_use=-1, desc=None):
         assert split in {"train", "test"}
@@ -150,17 +318,24 @@ 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 = ar_mask.sum().item()
-            nb_correct = ((result == input).long() * ar_mask).sum().item()
+            # nb_total = ar_mask.sum().item()
+            # nb_correct = ((result == input).long() * ar_mask).sum().item()
 
             return nb_total, nb_correct
 
@@ -180,6 +355,41 @@ class SandBox(Task):
             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}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
+        if save_attention_image is not None:
+            for k in range(10):
+                ns = torch.randint(self.test_input.size(0), (1,)).item()
+                input = self.test_input[ns : ns + 1].clone()
+
+                with torch.autograd.no_grad():
+                    t = model.training
+                    model.eval()
+                    # model.record_attention(True)
+                    model(BracketedSequence(input))
+                    model.train(t)
+                    # ram = model.retrieve_attention()
+                    # model.record_attention(False)
+
+                # tokens_output = [c for c in self.problem.seq2str(input[0])]
+                # tokens_input = ["n/a"] + tokens_output[:-1]
+                # for n_head in range(ram[0].size(1)):
+                # filename = os.path.join(
+                # result_dir, f"sandbox_attention_{k}_h{n_head}.pdf"
+                # )
+                # attention_matrices = [m[0, n_head] for m in ram]
+                # save_attention_image(
+                # filename,
+                # tokens_input,
+                # tokens_output,
+                # attention_matrices,
+                # k_top=10,
+                ##min_total_attention=0.9,
+                # token_gap=12,
+                # layer_gap=50,
+                # )
+                # logger(f"wrote {filename}")
+
 
 ######################################################################
 
@@ -335,6 +545,10 @@ class PicoCLVR(Task):
             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
         )
 
+        logger(
+            f"main_test_accuracy {n_epoch} {1-nb_missing_properties/nb_requested_properties}"
+        )
+
     ######################################################################
 
     def produce_results(
@@ -605,6 +819,8 @@ class Maze(Task):
             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}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
         if count is not None:
             proportion_optimal = count.diagonal().sum().float() / count.sum()
             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
@@ -744,6 +960,8 @@ class Snake(Task):
             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}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
 
 ######################################################################
 
@@ -853,6 +1071,8 @@ class Stack(Task):
             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}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
         ##############################################################
         # Log a few generated sequences
         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
@@ -1125,6 +1345,8 @@ class RPL(Task):
                 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}%"
             )
 
+            logger(f"main_test_accuracy {n_epoch} {1-test_nb_errors/test_nb_total}")
+
         test_nb_total, test_nb_errors = compute_nb_errors_output(
             self.test_input[:1000].to(self.device), nb_to_log=10
         )
@@ -1133,9 +1355,11 @@ class RPL(Task):
             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:
-            ns=torch.randint(self.test_input.size(0),(1,)).item()
-            input = self.test_input[ns:ns+1].clone()
+        if save_attention_image is None:
+            logger("no save_attention_image (is pycairo installed?)")
+        else:
+            ns = torch.randint(self.test_input.size(0), (1,)).item()
+            input = self.test_input[ns : ns + 1].clone()
             last = (input != self.t_nul).max(0).values.nonzero().max() + 3
             input = input[:, :last].to(self.device)
 
@@ -1321,6 +1545,8 @@ class Expr(Task):
             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}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
         nb_total = test_nb_delta.sum() + test_nb_missed
         for d in range(test_nb_delta.size(0)):
             logger(
@@ -1367,95 +1593,126 @@ class Expr(Task):
 
 ######################################################################
 
-import world
+import grid
+
+
+class Grid(Task):
+    # Make a tensor from a list of strings
+    def str2tensor(self, descr):
+        token_descr = [s.strip().split(" ") for s in descr]
+        l = max([len(s) for s in token_descr])
+        token_descr = [s + ["#"] * (l - len(s)) for s in token_descr]
+        id_descr = [[self.token2id[u] for u in s] for s in token_descr]
+        return torch.tensor(id_descr, device=self.device)
+
+    # Make a list of strings from a tensor
+    def tensor2str(self, x):
+        return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
+
+    # trim all the tensors in the tuple z to remove as much token from
+    # left and right in the first tensor. If z is a tuple, all its
+    # elements are trimed according to the triming for the first
+    def trim(self, z, token="#"):
+        n = self.token2id[token]
+        if type(z) == tuple:
+            x = z[0]
+            i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
+            a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
+            return tuple([t[:, a:b] for t in z])
+        else:
+            i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
+            a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
+            return z[:, a:b]
 
+    ######################
 
-class World(Task):
     def __init__(
         self,
         nb_train_samples,
         nb_test_samples,
         batch_size,
-        vqae_nb_epochs,
+        size,
+        fraction_play=0.0,
         logger=None,
         device=torch.device("cpu"),
-        device_storage=torch.device("cpu"),
     ):
         super().__init__()
 
-        self.batch_size = batch_size
         self.device = device
+        self.batch_size = batch_size
+        self.grid_factory = grid.GridFactory(size=size)
+        self.fraction_play = fraction_play
 
-        (
-            train_frames,
-            train_action_seq,
-            test_frames,
-            test_action_seq,
-            self.frame2seq,
-            self.seq2frame,
-        ) = world.create_data_and_processors(
-            nb_train_samples,
-            nb_test_samples,
-            mode="first_last",
-            nb_steps=30,
-            nb_epochs=vqae_nb_epochs,
-            logger=logger,
-            device=device,
-            device_storage=device_storage,
-        )
-
-        train_frame_seq = self.frame2seq(train_frames).to(device_storage)
-        test_frame_seq = self.frame2seq(test_frames).to(device_storage)
+        if logger is not None:
+            logger(
+                f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
+            )
 
-        nb_frame_codes = max(train_frame_seq.max(), test_frame_seq.max()) + 1
-        nb_action_codes = max(train_action_seq.max(), test_action_seq.max()) + 1
+        self.train_descr = self.grid_factory.generate_samples(
+            nb=nb_train_samples,
+            fraction_play=fraction_play,
+            progress_bar=lambda r: tqdm.tqdm(r),
+        )
 
-        self.len_frame_seq = train_frame_seq.size(1)
-        self.len_action_seq = train_action_seq.size(1)
-        self.nb_codes = nb_frame_codes + nb_action_codes
+        self.test_descr = self.grid_factory.generate_samples(
+            nb=nb_test_samples, fraction_play=0.0, progress_bar=lambda r: tqdm.tqdm(r)
+        )
 
-        train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -1)
+        if fraction_play > 0:
+            self.play_descr = self.grid_factory.generate_samples(
+                nb=25, fraction_play=1.0, progress_bar=lambda r: tqdm.tqdm(r)
+            )
+        else:
+            self.play_descr = []
 
-        train_action_seq += nb_frame_codes
-        self.train_input = torch.cat(
-            (train_frame_seq[:, 0, :], train_action_seq, train_frame_seq[:, 1, :]), 1
-        )
+        # Build the tokenizer
+        tokens = set()
+        for d in [self.train_descr, self.test_descr, self.play_descr]:
+            for s in d:
+                for t in s.strip().split(" "):
+                    tokens.add(t)
+        # make this set a sorted list to get the same tensors given
+        # the same descr
+        tokens = list(tokens)
+        tokens.sort()
+        tokens = ["#"] + tokens
+        self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
+        self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
+        self.t_nul = self.token2id["#"]
+        self.t_true = self.token2id["true"]
+        self.t_false = self.token2id["false"]
+        self.t_pipe = self.token2id["|"]
 
-        test_frame_seq = test_frame_seq.reshape(test_frame_seq.size(0) // 2, 2, -1)
-        test_action_seq += nb_frame_codes
-        self.test_input = torch.cat(
-            (test_frame_seq[:, 0, :], test_action_seq, test_frame_seq[:, 1, :]), 1
+        # Tokenize the train and test sets
+        self.train_input = self.str2tensor(self.train_descr)
+        self.test_input = self.str2tensor(self.test_descr)
+        self.play_input = (
+            None if len(self.play_descr) == 0 else self.str2tensor(self.play_descr)
         )
 
-    def batches(self, split="train", nb_to_use=-1, desc=None):
+    def batches(self, split="train"):
         assert split in {"train", "test"}
         input = self.train_input if split == "train" else self.test_input
-        if nb_to_use > 0:
-            input = input[:nb_to_use]
-        if desc is None:
-            desc = f"epoch-{split}"
         for batch in tqdm.tqdm(
-            input.split(self.batch_size), dynamic_ncols=True, desc=desc
+            input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
         ):
-            yield batch.to(self.device)
+            yield self.trim(batch)
 
     def vocabulary_size(self):
-        return self.nb_codes
+        return len(self.token2id)
 
     def produce_results(
         self, n_epoch, model, result_dir, logger, deterministic_synthesis
     ):
-        k = torch.arange(
-            2 * self.len_frame_seq + self.len_action_seq, device=self.device
-        )[None, :]
+        correct = self.test_input[:1000]
+        result = correct.clone()
+        ar_mask = torch.logical_or(result == self.t_true, result == self.t_false).long()
+        result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
 
-        input = self.test_input[:64].to(self.device)
-        result = input.clone()
+        logger(f"----------------------------------------------------------")
 
-        ar_mask = (
-            (k >= self.len_frame_seq + self.len_action_seq).long().expand_as(result)
-        )
-        result *= 1 - ar_mask
+        for e in self.tensor2str(result[:10]):
+            logger(f"test_before {e}")
 
         masked_inplace_autoregression(
             model,
@@ -1466,25 +1723,142 @@ class World(Task):
             device=self.device,
         )
 
-        seq_start = input[:, : self.len_frame_seq]
-        seq_end = input[:, self.len_frame_seq + self.len_action_seq :]
-        seq_predicted = result[:, self.len_frame_seq + self.len_action_seq :]
+        logger(f"----------------------------------------------------------")
+
+        for e in self.tensor2str(result[:10]):
+            logger(f"test_after  {e}")
+
+        logger(f"----------------------------------------------------------")
+
+        nb_total = ar_mask.sum().item()
+        nb_correct = ((correct == result).long() * ar_mask).sum().item()
+
+        logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
+        logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
+
+        if self.play_input is not None:
+            result = self.play_input.clone()
+            ar_mask = (result == self.t_pipe).long().cumsum(dim=1).clamp(max=1)
+            result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
+
+            logger(f"----------------------------------------------------------")
+
+            for e in self.tensor2str(result[:10]):
+                logger(f"play_before {e}")
+
+            masked_inplace_autoregression(
+                model,
+                self.batch_size,
+                result,
+                ar_mask,
+                deterministic_synthesis,
+                device=self.device,
+            )
 
-        result = torch.cat(
-            (seq_start[:, None, :], seq_end[:, None, :], seq_predicted[:, None, :]), 1
+            logger(f"----------------------------------------------------------")
+
+            for e in self.tensor2str(result[:10]):
+                logger(f"play_after  {e}")
+
+            logger(f"----------------------------------------------------------")
+
+
+######################################################################
+
+import qmlp
+
+
+class QMLP(Task):
+    ######################
+
+    def __init__(
+        self,
+        nb_train_samples,
+        nb_test_samples,
+        batch_size,
+        result_dir,
+        logger=None,
+        device=torch.device("cpu"),
+    ):
+        super().__init__()
+
+        self.device = device
+        self.batch_size = batch_size
+        self.nb_samples_per_mlp = 256
+
+        if logger is not None:
+            logger(
+                f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
+            )
+
+        seq, q_test_set, test_error = qmlp.generate_sequence_and_test_set(
+            nb_mlps=nb_train_samples + nb_test_samples,
+            nb_samples=self.nb_samples_per_mlp,
+            device=self.device,
+            batch_size=64,
+            nb_epochs=250,
+            nb_mlps_per_batch=1024,
         )
-        result = result.reshape(-1, result.size(-1))
 
-        frames = self.seq2frame(result)
-        image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
-        torchvision.utils.save_image(
-            frames.float() / (world.Box.nb_rgb_levels - 1),
-            image_name,
-            nrow=12,
-            padding=1,
-            pad_value=0.0,
+        self.train_input = seq[:nb_train_samples]
+        self.train_q_test_set = q_test_set[:nb_train_samples]
+        self.train_ref_test_errors = test_error[:nb_train_samples]
+        self.test_input = seq[nb_train_samples:]
+        self.test_q_test_set = q_test_set[nb_train_samples:]
+        self.test_ref_test_errors = test_error[nb_train_samples:]
+
+        filename = os.path.join(result_dir, f"train_errors_ref.dat")
+        with open(filename, "w") as f:
+            for e in self.train_ref_test_errors:
+                f.write(f"{e}\n")
+
+        filename = os.path.join(result_dir, f"test_errors_ref.dat")
+        with open(filename, "w") as f:
+            for e in self.test_ref_test_errors:
+                f.write(f"{e}\n")
+
+        self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
+
+    def batches(self, split="train"):
+        assert split in {"train", "test"}
+        input = self.train_input if split == "train" else self.test_input
+        for batch in tqdm.tqdm(
+            input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
+        ):
+            yield batch
+
+    def vocabulary_size(self):
+        return self.nb_codes
+
+    def produce_results(
+        self, n_epoch, model, result_dir, logger, deterministic_synthesis
+    ):
+        correct = self.test_input[:1000]
+        result = correct.clone()
+        ar_mask = (
+            torch.arange(result.size(1), device=result.device)
+            > self.nb_samples_per_mlp * 3 + 1
+        ).long()[None, :]
+        ar_mask = ar_mask.expand_as(result)
+        result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
+
+        masked_inplace_autoregression(
+            model,
+            self.batch_size,
+            result,
+            ar_mask,
+            deterministic_synthesis,
+            device=self.device,
         )
-        logger(f"wrote {image_name}")
+
+        q_train_set = result[:, : self.nb_samples_per_mlp * 3]
+        q_params = result[:, self.nb_samples_per_mlp * 3 + 1 :]
+        error_test = qmlp.evaluate_q_params(q_params, self.test_q_test_set)
+
+        filename = os.path.join(result_dir, f"test_errors_{n_epoch:04d}.dat")
+        with open(filename, "w") as f:
+            for e in error_test:
+                f.write(f"{e}\n")
 
 
 ######################################################################