Update.

[mygpt.git] / main.py

diff --git a/main.py b/main.py
index 5f3e8cf..83227bb 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -126,7 +126,7 @@ def autoregression(
         results = torch.cat((primer, results), 1)
 
     for input in results.split(batch_size):
-        for s in tqdm.tqdm(range(first, input.size(1)), desc = 'synth'):
+        for s in range(first, input.size(1)):
             output = model(input)
             logits = output[:, s]
             if args.synthesis_sampling:
@@ -156,28 +156,32 @@ import picoclvr
 
 class TaskPicoCLVR(Task):
 
+    # Make a tensor from a list of strings
     def tensorize(self, descr):
-        t = [ [ self.token2id[u] for u in s ] for s in descr ]
-        return torch.tensor(t, device = self.device)
+        token_descr = [ s.strip().split(' ') for s in descr ]
+        l = max([ len(s) for s in token_descr ])
+        #token_descr = [ [ '<nul>' ] * (l - len(s)) + s for s in token_descr ]
+        token_descr = [ s + [ '<nul>' ] * (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)
+
+    def trim(self, x, token = '<nul>'):
+        n = self.token2id[token]
+        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 x[:, a:b]
 
     def __init__(self, batch_size,
                  height, width, nb_colors = 5,
                  device = torch.device('cpu')):
 
         def generate_descr(nb):
-            descr = picoclvr.generate(
+            return picoclvr.generate(
                 nb,
                 height = self.height, width = self.width,
                 nb_colors = nb_colors
             )
 
-            descr = [ s.strip().split(' ') for s in descr ]
-            l = max([ len(s) for s in descr ])
-            #descr = [ [ '<nul>' ] * (l - len(s)) + s for s in descr ]
-            descr = [ s + [ '<nul>' ] * (l - len(s)) for s in descr ]
-
-            return descr
-
         self.height = height
         self.width = width
         self.batch_size = batch_size
@@ -188,10 +192,10 @@ class TaskPicoCLVR(Task):
         self.test_descr = generate_descr((nb * 1) // 5)
 
         # Build the tokenizer
-        tokens = set()
+        tokens = { '<nul>' }
         for d in [ self.train_descr, self.test_descr ]:
             for s in d:
-                for t in s: tokens.add(t)
+                for t in s.strip().split(' '): tokens.add(t)
         self.token2id = dict([ (t, n) for n, t in enumerate(tokens) ])
         self.id2token = dict([ (n, t) for n, t in enumerate(tokens) ])
 
@@ -203,13 +207,13 @@ class TaskPicoCLVR(Task):
         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), desc = f'epoch-{split}'):
-            yield batch
+            yield self.trim(batch)
 
     def vocabulary_size(self):
         return len(self.token2id)
 
     def produce_results(self, n_epoch, model):
-        nb_tokens = self.height * self.width + 3
+        nb_tokens_to_generate = self.height * self.width + 3
         result_descr = [ ]
         nb_per_primer = 8
 
@@ -220,15 +224,26 @@ class TaskPicoCLVR(Task):
                 'green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top <img>',
         ]:
 
-            for k in range(nb_per_primer):
-                results = autoregression(
-                    model, self.batch_size,
-                    nb_samples = 1, nb_tokens = nb_tokens,
-                    primer = self.tensorize(primer_descr),
-                    device = self.device
-                )
-                r = ' '.join([ self.id2token[t.item()] for t in results.flatten() ])
-                result_descr.append(r)
+            results = autoregression(
+                model,
+                self.batch_size,
+                nb_samples = nb_per_primer,
+                nb_tokens_to_generate = nb_tokens_to_generate,
+                primer = self.tensorize([ primer_descr ]).expand(nb_per_primer, -1),
+                device = self.device
+            )
+
+            l = [ ' '.join([ self.id2token[t.item()] for t in r ]) for r in results ]
+            result_descr += l
+
+        np = picoclvr.nb_properties(
+            result_descr,
+            height = self.height, width = self.width
+        )
+
+        nb_requested_properties, _, nb_missing_properties = zip(*np)
+
+        log_string(f'nb_requested_properties {sum(nb_requested_properties) / len(result_descr):.02f} nb_missing_properties {sum(nb_missing_properties) / len(result_descr):.02f}')
 
         img = [
             picoclvr.descr2img(d, height = self.height, width = self.width)
@@ -243,15 +258,6 @@ class TaskPicoCLVR(Task):
         )
         log_string(f'wrote {image_name}')
 
-        np = picoclvr.nb_properties(
-            result_descr,
-            height = self.height, width = self.width
-        )
-
-        nb_requested_properties, _, nb_missing_properties = zip(*np)
-
-        log_string(f'nb_requested_properties {sum(nb_requested_properties) / len(result_descr):.02f} nb_missing_properties {sum(nb_missing_properties) / len(result_descr):.02f}')
-
 ######################################################################
 
 class TaskWiki103(Task):
@@ -284,6 +290,7 @@ class TaskWiki103(Task):
 
         self.vocab.set_default_index(self.vocab[ '<unk>' ])
 
+    # makes a tensor from a list of list of tokens
     def tensorize(self, s):
         a = max(len(x) for x in s)
         return torch.tensor([ self.vocab(x + [ '<nul>' ] * (a - len(x))) for x in s ])
@@ -463,7 +470,6 @@ for input in task.batches(split = 'train'):
 token_probas = token_count / token_count.sum()
 entropy = -torch.xlogy(token_probas, token_probas).sum()
 train_set_perplexity = math.exp(entropy)
-#log_string(f'train set perplexity {train_set_perplexity}')
 
 for k in range(nb_epochs_finished, nb_epochs):