X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=sidebyside;f=picoclvr.py;h=8201f5d8a19eff414db48bf88d6285a1b5a5fc8d;hb=b6a9cc237cdadac2351814f92c20607d46b0f583;hp=26f53ab91f1e44530107bf7f4ed5836f5b2c7071;hpb=dbbed675efe5ae5e7859106519de8a69947f0766;p=mygpt.git

diff --git a/picoclvr.py b/picoclvr.py
index 26f53ab..8201f5d 100755
--- a/picoclvr.py
+++ b/picoclvr.py
@@ -67,6 +67,7 @@ color_names = [
     'azure', 'snow', 'silver', 'gainsboro', 'white_smoke',
 ]
 
+color_id = dict( [ (n, k) for k, n in enumerate(color_names) ] )
 color_tokens = dict( [ (n, c) for n, c in zip(color_names, colors) ] )
 
 ######################################################################
@@ -92,11 +93,11 @@ def all_properties(height, width, nb_squares, square_i, square_j, square_c):
 
 ######################################################################
 
-def generate(nb, height = 6, width = 8,
-             max_nb_squares = 5, max_nb_statements = 10,
-             many_colors = False):
+def generate(nb, height, width,
+             max_nb_squares = 5, max_nb_properties = 10,
+             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 = [ ]
 
@@ -112,14 +113,14 @@ def generate(nb, height = 6, width = 8,
         img = [ 0 ] * height * width
         for k in range(nb_squares): img[square_position[k]] = square_c[k]
 
-        # generates all the true relations
+        # generates all the true properties
 
         s = all_properties(height, width, nb_squares, square_i, square_j, square_c)
 
-        # pick at most max_nb_statements at random
+        # pick at most max_nb_properties at random
 
-        nb_statements = torch.randint(max_nb_statements, (1,)) + 1
-        s = ' <sep> '.join([ s[k] for k in torch.randperm(len(s))[:nb_statements] ] )
+        nb_properties = torch.randint(max_nb_properties, (1,)) + 1
+        s = ' <sep> '.join([ s[k] for k in torch.randperm(len(s))[:nb_properties] ] )
         s += ' <img> ' + ' '.join([ f'{color_names[n]}' for n in img ])
 
         descr += [ s ]
@@ -128,7 +129,10 @@ def generate(nb, height = 6, width = 8,
 
 ######################################################################
 
-def descr2img(descr, height = 6, width = 8):
+def descr2img(descr, height, width):
+
+    if type(descr) == list:
+        return torch.cat([ descr2img(d, height, width) for d in descr ], 0)
 
     def token2color(t):
         try:
@@ -136,25 +140,72 @@ def descr2img(descr, height = 6, width = 8):
         except KeyError:
             return [ 128, 128, 128 ]
 
-    def img_descr(x):
-        u = x.split('<img>', 1)
-        return u[1] if len(u) > 1 else ''
-
-    img = torch.full((len(descr), 3, height, width), 255)
-    d = [ img_descr(x) for x in descr ]
-    d = [ u.strip().split(' ')[:height * width] for u in d ]
-    d = [ u + [ '<unk>' ] * (height * width - len(u)) for u in d ]
-    d = [ [ token2color(t) for t in u ] for u in d ]
-    img = torch.tensor(d).permute(0, 2, 1)
-    img = img.reshape(img.size(0), 3, height, width)
+    d = descr.split('<img>', 1)
+    d = d[-1] if len(d) > 1 else ''
+    d = d.strip().split(' ')[:height * width]
+    d = d + [ '<unk>' ] * (height * width - len(d))
+    d = [ token2color(t) for t in d ]
+    img = torch.tensor(d).permute(1, 0)
+    img = img.reshape(1, 3, height, width)
 
     return img
 
 ######################################################################
 
+def descr2properties(descr, height, width):
+
+    if type(descr) == list:
+        return [ descr2properties(d, height, width) for d in descr ]
+
+    d = descr.split('<img>', 1)
+    d = d[-1] if len(d) > 1 else ''
+    d = d.strip().split(' ')[:height * width]
+
+    seen = {}
+    if len(d) != height * width: return []
+
+    for k, x in enumerate(d):
+        if x != color_names[0]:
+            if x in color_tokens:
+                if x in seen: return []
+            else:
+                return []
+            seen[x] = (color_id[x], k // width, k % width)
+
+    square_infos = tuple(zip(*seen.values()))
+    square_c = torch.tensor(square_infos[0])
+    square_i = torch.tensor(square_infos[1])
+    square_j = torch.tensor(square_infos[2])
+
+    s = all_properties(height, width, len(seen), square_i, square_j, square_c)
+
+    return s
+
+######################################################################
+
+def nb_properties(descr, height, width):
+    if type(descr) == list:
+        return [ nb_properties(d, height, width) for d in descr ]
+
+    d = descr.split('<img>', 1)
+    if len(d) == 0: return 0
+    d = d[0].strip().split('<sep>')
+    d = [ x.strip() for x in d ]
+
+    requested_properties = set(d)
+    all_properties = set(descr2properties(descr, height, width))
+    missing_properties = requested_properties - all_properties
+
+    return (len(requested_properties), len(all_properties), len(missing_properties))
+
+######################################################################
+
 if __name__ == '__main__':
     descr = generate(nb = 5)
 
+    #print(descr2properties(descr))
+    print(nb_properties(descr))
+
     with open('picoclvr_example.txt', 'w') as f:
         for d in descr:
             f.write(f'{d}\n\n')
@@ -166,7 +217,7 @@ if __name__ == '__main__':
     import time
 
     start_time = time.perf_counter()
-    descr = generate(10000)
+    descr = generate(nb = 1000)
     end_time = time.perf_counter()
     print(f'{len(descr) / (end_time - start_time):.02f} samples per second')