OCDC

[mygpt.git] / main.py

diff --git a/main.py b/main.py
index 339d185..427a83a 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 +156,26 @@ import picoclvr
 
 class TaskPicoCLVR(Task):
 
-    def descr2tensor(self, descr):
-        t = [ [ self.token2id[u] for u in s ] for s in descr ]
-        return torch.tensor(t, device = self.device)
+    # Make a tensor from a list of strings
+    def tensorize(self, descr):
+        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 __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 = [ [ '<unk>' ] * (l - len(s)) + s for s in descr ]
-            descr = [ s + [ '<unk>' ] * (l - len(s)) for s in descr ]
-
-            return descr
-
         self.height = height
         self.width = width
         self.batch_size = batch_size
@@ -188,16 +186,16 @@ 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) ])
 
         # Tokenize the train and test sets
-        self.train_input = descr2tensor(self.train_descr)
-        self.test_input = descr2tensor(self.test_descr)
+        self.train_input = self.tensorize(self.train_descr)
+        self.test_input = self.tensorize(self.test_descr)
 
     def batches(self, split = 'train'):
         assert split in { 'train', 'test' }
@@ -208,14 +206,6 @@ class TaskPicoCLVR(Task):
     def vocabulary_size(self):
         return len(self.token2id)
 
-    def generate(self, primer_descr, model, nb_tokens):
-        results = autoregression(
-            model, self.batch_size,
-            nb_samples = 1, nb_tokens = nb_tokens, primer = descr2tensor(primer_descr),
-            device = self.device
-        )
-        return ' '.join([ self.id2token[t.item()] for t in results.flatten() ])
-
     def produce_results(self, n_epoch, model):
         nb_tokens = self.height * self.width + 3
         result_descr = [ ]
@@ -229,10 +219,20 @@ class TaskPicoCLVR(Task):
         ]:
 
             for k in range(nb_per_primer):
-                result_descr.append(self.generate(primer_descr, model, nb_tokens))
+                results = autoregression(
+                    model, self.batch_size,
+                    nb_samples = 1, nb_tokens_to_generate = 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)
+
+        img = [
+            picoclvr.descr2img(d, height = self.height, width = self.width)
+            for d in result_descr
+        ]
 
-        img = [ picoclvr.descr2img(d, height = self.height, width = self.width)
-                for d in result_descr ]
         img = torch.cat(img, 0)
         image_name = f'result_picoclvr_{n_epoch:04d}.png'
         torchvision.utils.save_image(
@@ -276,15 +276,16 @@ class TaskWiki103(Task):
 
         self.vocab = torchtext.vocab.build_vocab_from_iterator(
             yield_tokens(),
-            specials = [ '<unk>', '<non>' ],
+            specials = [ '<unk>', '<nul>' ],
             min_freq = self.min_freq
         )
 
         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 + [ '<non>' ] * (a - len(x))) for x in s ])
+        return torch.tensor([ self.vocab(x + [ '<nul>' ] * (a - len(x))) for x in s ])
 
     def yield_batches(self, ds):
         s = [ ]
@@ -342,7 +343,7 @@ class TaskWiki103(Task):
                      else:
                          t_next = logits.argmax()
                      t_generated.append(self.vocab.lookup_token(t_next))
-                     if t_generated[-1] == '<non>': break
+                     if t_generated[-1] == '<nul>': break
 
                  s = ' '.join(t_generated)