Removed the default image size.

[mygpt.git] / main.py

diff --git a/main.py b/main.py
index 85cf4cf..3bf6b52 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -25,7 +25,7 @@ parser.add_argument('--log_filename',
                     type = str, default = 'train.log')
 
 parser.add_argument('--download',
-                    type = bool, default = False)
+                    action='store_true', default = False)
 
 parser.add_argument('--seed',
                     type = int, default = 0)
@@ -67,11 +67,14 @@ parser.add_argument('--dropout',
                     type = float, default = 0.1)
 
 parser.add_argument('--synthesis_sampling',
-                    type = bool, default = True)
+                    action='store_true', default = True)
 
 parser.add_argument('--checkpoint_name',
                     type = str, default = 'checkpoint.pth')
 
+parser.add_argument('--picoclvr_many_colors',
+                    action='store_true', default = False)
+
 ######################################################################
 
 args = parser.parse_args()
@@ -128,6 +131,9 @@ class TaskPicoCLVR(Task):
             many_colors = many_colors
         )
 
+        # self.test_descr = descr[:nb // 5]
+        # self.train_descr = descr[nb // 5:]
+
         descr = [ s.strip().split(' ') for s in descr ]
         l = max([ len(s) for s in descr ])
         descr = [ s + [ '<unk>' ] * (l - len(s)) for s in descr ]
@@ -156,8 +162,26 @@ class TaskPicoCLVR(Task):
     def vocabulary_size(self):
         return len(self.token2id)
 
+    def generate(self, primer, model, nb_tokens):
+        t_primer = primer.strip().split(' ')
+        t_generated = [ ]
+
+        for j in range(nb_tokens):
+            t = [ [ self.token2id[u] for u in t_primer + t_generated ] ]
+            input = torch.tensor(t, device = self.device)
+            output = model(input)
+            logits = output[0, -1]
+            if args.synthesis_sampling:
+                dist = torch.distributions.categorical.Categorical(logits = logits)
+                t = dist.sample()
+            else:
+                t = logits.argmax()
+            t_generated.append(self.id2token[t.item()])
+
+        return ' '.join(t_primer + t_generated)
+
     def produce_results(self, n_epoch, model, nb_tokens = 50):
-        img = [ ]
+        descr = [ ]
         nb_per_primer = 8
 
         for primer in [
@@ -168,30 +192,18 @@ class TaskPicoCLVR(Task):
         ]:
 
             for k in range(nb_per_primer):
-                t_primer = primer.strip().split(' ')
-                t_generated = [ ]
-
-                for j in range(nb_tokens):
-                    t = [ [ self.token2id[u] for u in t_primer + t_generated ] ]
-                    input = torch.tensor(t, device = self.device)
-                    output = model(input)
-                    logits = output[0, -1]
-                    if args.synthesis_sampling:
-                        dist = torch.distributions.categorical.Categorical(logits = logits)
-                        t = dist.sample()
-                    else:
-                        t = logits.argmax()
-                    t_generated.append(self.id2token[t.item()])
-
-                descr = [ ' '.join(t_primer + t_generated) ]
-                img += [ picoclvr.descr2img(descr) ]
+                descr.append(self.generate(primer, model, nb_tokens))
 
+        img = [ picoclvr.descr2img(d) for d in descr ]
         img = torch.cat(img, 0)
         file_name = f'result_picoclvr_{n_epoch:04d}.png'
         torchvision.utils.save_image(img / 255.,
                                      file_name, nrow = nb_per_primer, pad_value = 0.8)
         log_string(f'wrote {file_name}')
 
+        nb_missing = sum( [ x[2] for x in picoclvr.nb_missing_properties(descr) ] )
+        log_string(f'nb_missing {nb_missing / len(descr):.02f}')
+
 ######################################################################
 
 class TaskWiki103(Task):
@@ -353,7 +365,7 @@ if args.data == 'wiki103':
 elif args.data == 'mnist':
     task = TaskMNIST(batch_size = args.batch_size, device = device)
 elif args.data == 'picoclvr':
-    task = TaskPicoCLVR(batch_size = args.batch_size, device = device)
+    task = TaskPicoCLVR(batch_size = args.batch_size, many_colors = args.picoclvr_many_colors, device = device)
 else:
     raise ValueError(f'Unknown dataset {args.data}.')