Update.
authorFrancois Fleuret <francois@fleuret.org>
Tue, 26 Jul 2022 15:06:43 +0000 (17:06 +0200)
committerFrancois Fleuret <francois@fleuret.org>
Tue, 26 Jul 2022 15:06:43 +0000 (17:06 +0200)
main.py
picoclvr.py

diff --git a/main.py b/main.py
index a83107b..7ce80a3 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -78,8 +78,8 @@ parser.add_argument('--checkpoint_name',
 ##############################
 # picoclvr options
 
-parser.add_argument('--picoclvr_many_colors',
-                    action='store_true', default = False)
+parser.add_argument('--picoclvr_nb_colors',
+                    type = int, default = 5)
 
 parser.add_argument('--picoclvr_height',
                     type = int, default = 12)
@@ -113,22 +113,33 @@ for n in vars(args):
 
 ######################################################################
 
-def produce_results(
-        self,
-        model, nb_samples, nb_tokens_to_generate, starting_input = None,
-        device = 'cpu'
+def autoregression(
+        model,
+        nb_samples, nb_tokens_to_generate, starting_input = None,
+        device = torch.device('cpu')
 ):
-    results = torch.zeros(nb_samples, nb_tokens_to_generate, dtype = torch.int64, device = device)
+    first = 0
+    results = torch.zeros(
+        nb_samples, nb_tokens_to_generate,
+        dtype = torch.int64, device = device
+    )
+
+    if starting_input is not None:
+        first = starting_input.size(1)
+        results = torch.cat((starting_input, results), 1)
+
     for input in results.split(self.batch_size):
-        for s in tqdm.tqdm(range(input.size(1) - 1), desc = 'synth'):
+        for s in tqdm.tqdm(range(first, input.size(1)), desc = 'synth'):
             output = model(input)
             logits = output[:, s]
             if args.synthesis_sampling:
                 dist = torch.distributions.categorical.Categorical(logits = logits)
-                t = dist.sample()
+                t_next = dist.sample()
             else:
-                t = logits.argmax(1)
-            input[:, s + 1] = t
+                t_next = logits.argmax(1)
+            input[:, s] = t_next
+
+    return results
 
 ######################################################################
 
@@ -149,14 +160,14 @@ import picoclvr
 class TaskPicoCLVR(Task):
 
     def __init__(self, batch_size,
-                 height, width, many_colors = False,
+                 height, width, nb_colors = 5,
                  device = torch.device('cpu')):
 
         def generate_descr(nb):
             descr = picoclvr.generate(
                 nb,
                 height = self.height, width = self.width,
-                many_colors = many_colors
+                nb_colors = nb_colors
             )
 
             descr = [ s.strip().split(' ') for s in descr ]
@@ -211,10 +222,10 @@ class TaskPicoCLVR(Task):
             logits = output[0, -1]
             if args.synthesis_sampling:
                 dist = torch.distributions.categorical.Categorical(logits = logits)
-                t = dist.sample()
+                t_next = dist.sample()
             else:
-                t = logits.argmax()
-            t_generated.append(self.id2token[t.item()])
+                t_next = logits.argmax()
+            t_generated.append(self.id2token[t_next.item()])
 
         return ' '.join(t_primer + t_generated)
 
@@ -339,10 +350,10 @@ class TaskWiki103(Task):
                      logits = output[0, -1]
                      if args.synthesis_sampling:
                          dist = torch.distributions.categorical.Categorical(logits = logits)
-                         t = dist.sample()
+                         t_next = dist.sample()
                      else:
-                         t = logits.argmax()
-                     t_generated.append(self.vocab.lookup_token(t))
+                         t_next = logits.argmax()
+                     t_generated.append(self.vocab.lookup_token(t_next))
                      if t_generated[-1] == '<non>': break
 
                  s = ' '.join(t_generated)
@@ -375,18 +386,7 @@ class TaskMNIST(Task):
         return 256
 
     def produce_results(self, n_epoch, model, nb_samples = 64):
-        results = torch.zeros(nb_samples, 28 * 28, dtype = torch.int64, device = self.device)
-        for input in results.split(self.batch_size):
-            for s in tqdm.tqdm(range(input.size(1)), desc = 'synth'):
-                output = model(input)
-                logits = output[:, s]
-                if args.synthesis_sampling:
-                    dist = torch.distributions.categorical.Categorical(logits = logits)
-                    t = dist.sample()
-                else:
-                    t = logits.argmax(1)
-                input[:, s] = t
-
+        results = autoregression(model, nb_samples, 28 * 28, device = self.device)
         image_name = f'result_mnist_{n_epoch:04d}.png'
         torchvision.utils.save_image(1 - results.reshape(-1, 1, 28, 28) / 255.,
                                      image_name, nrow = 16, pad_value = 0.8)
@@ -407,7 +407,7 @@ elif args.data == 'picoclvr':
     task = TaskPicoCLVR(batch_size = args.batch_size,
                         height = args.picoclvr_height,
                         width = args.picoclvr_width,
-                        many_colors = args.picoclvr_many_colors,
+                        nb_colors = args.picoclvr_nb_colors,
                         device = device)
 else:
     raise ValueError(f'Unknown dataset {args.data}.')
index 19517af..2d57505 100755 (executable)
@@ -95,9 +95,9 @@ def all_properties(height, width, nb_squares, square_i, square_j, square_c):
 
 def generate(nb, height, width,
              max_nb_squares = 5, max_nb_properties = 10,
-             many_colors = False):
+             nb_colors = 5):
 
-    nb_colors =  len(color_tokens) - 1 if many_colors else max_nb_squares
+    assert nb_colors >= max_nb_squares and nb_colors <= len(color_tokens) - 1
 
     descr = [ ]