OCD cosmetics.

[mygpt.git] / main.py

diff --git a/main.py b/main.py
index 7ce80a3..6592204 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -24,9 +24,6 @@ parser = argparse.ArgumentParser(description = 'My own GPT.')
 parser.add_argument('--log_filename',
                     type = str, default = 'train.log')
 
-parser.add_argument('--download',
-                    action='store_true', default = False)
-
 parser.add_argument('--seed',
                     type = int, default = 0)
 
@@ -118,17 +115,18 @@ def autoregression(
         nb_samples, nb_tokens_to_generate, starting_input = None,
         device = torch.device('cpu')
 ):
-    first = 0
     results = torch.zeros(
         nb_samples, nb_tokens_to_generate,
         dtype = torch.int64, device = device
     )
 
-    if starting_input is not None:
+    if starting_input is None:
+        first = 0
+    else:
         first = starting_input.size(1)
         results = torch.cat((starting_input, results), 1)
 
-    for input in results.split(self.batch_size):
+    for input in results.split(args.batch_size):
         for s in tqdm.tqdm(range(first, input.size(1)), desc = 'synth'):
             output = model(input)
             logits = output[:, s]
@@ -172,6 +170,7 @@ class TaskPicoCLVR(Task):
 
             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
@@ -193,6 +192,7 @@ class TaskPicoCLVR(Task):
         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
         t = [ [ self.token2id[u] for u in s ] for s in self.train_descr ]
         self.train_input = torch.tensor(t, device = self.device)
         t = [ [ self.token2id[u] for u in s ] for s in self.test_descr ]
@@ -254,14 +254,14 @@ class TaskPicoCLVR(Task):
         )
         log_string(f'wrote {image_name}')
 
-        nb_missing = sum( [
-            x[2] for x in picoclvr.nb_missing_properties(
-                descr,
-                height = self.height, width = self.width
-            )
-        ] )
+        np = picoclvr.nb_properties(
+            descr,
+            height = self.height, width = self.width
+        )
+
+        nb_requested_properties, _, nb_missing_properties = zip(*np)
 
-        log_string(f'nb_missing {nb_missing / len(descr):.02f}')
+        log_string(f'nb_requested_properties {sum(nb_requested_properties) / len(descr):.02f} nb_missing_properties {sum(nb_missing_properties) / len(descr):.02f}')
 
 ######################################################################
 
@@ -445,7 +445,7 @@ else:
 nb_epochs_finished = 0
 
 if args.no_checkpoint:
-    log_string(f'Not trying to load checkpoint.')
+    log_string(f'not trying to load checkpoint.')
 
 else:
     try:
@@ -453,13 +453,13 @@ else:
         nb_epochs_finished = checkpoint['nb_epochs_finished']
         model.load_state_dict(checkpoint['model_state'])
         optimizer.load_state_dict(checkpoint['optimizer_state'])
-        log_string(f'Checkpoint loaded with {nb_epochs_finished} epochs finished.')
+        log_string(f'checkpoint loaded with {nb_epochs_finished} epochs finished.')
 
     except FileNotFoundError:
-        log_string('Starting from scratch.')
+        log_string('starting from scratch.')
 
     except:
-        log_string('Error when loading the checkpoint.')
+        log_string('error when loading the checkpoint.')
         exit(1)
 
 ######################################################################
@@ -472,7 +472,7 @@ for input in task.batches(split = 'train'):
 token_probas = token_count / token_count.sum()
 h = -torch.xlogy(token_probas, token_probas).sum()
 train_set_perplexity = math.exp(h)
-log_string(f'Train set perplexity {train_set_perplexity}')
+log_string(f'train set perplexity {train_set_perplexity}')
 
 for k in range(nb_epochs_finished, nb_epochs):