From 96d7800f71df1937b8dda30a0202aaab6201c55e Mon Sep 17 00:00:00 2001
From: Marcin Chrzanowski <m@m-chrzan.xyz>
Date: Mon, 3 May 2021 13:12:31 +0200
Subject: Implement counting with 135 outputs

---
 src/counting_big.py | 36 ++++++++++++++++++++++++++++++++++++
 src/net_types.py    |  4 +++-
 2 files changed, 39 insertions(+), 1 deletion(-)
 create mode 100644 src/counting_big.py

diff --git a/src/counting_big.py b/src/counting_big.py
new file mode 100644
index 0000000..a5fb80c
--- /dev/null
+++ b/src/counting_big.py
@@ -0,0 +1,36 @@
+import torch.nn.functional as F
+import torch
+
+def target_transform(labels):
+    first_shape_count = None
+    first_shape = None
+    second_shape = None
+    for i in range(6):
+        if labels[i] > 0:
+            if first_shape is None:
+                first_shape = i
+                first_shape_count = labels[i]
+            else:
+                second_shape = i
+                break
+
+    first_offset = 15 - (6 - first_shape) * (5 - first_shape) / 2
+    pair_offset = first_offset + (second_shape - first_shape) - 1
+    offset = int(pair_offset * 9 + (first_shape_count - 1))
+    new_labels = torch.zeros(135)
+    new_labels[offset] = 1
+    return new_labels
+
+
+def loss_function(output, target):
+    return F.binary_cross_entropy(output, target)
+
+def count_correct(output, target):
+    output = torch.argmax(output, dim=1)
+    target = torch.argmax(target, dim=1)
+    return torch.sum(output == target).tolist()
+
+def finalizer(x):
+    return torch.softmax(x, 1)
+
+outputs = 135
diff --git a/src/net_types.py b/src/net_types.py
index 14925c8..1f308a4 100644
--- a/src/net_types.py
+++ b/src/net_types.py
@@ -1,7 +1,9 @@
 import classification
 import counting_small
+import counting_big
 
 types = {
     'classification': classification,
-    'counting-small': counting_small
+    'counting-small': counting_small,
+    'counting-big': counting_big,
 }
-- 
cgit v1.2.3