Update.
authorFrancois Fleuret <francois@fleuret.org>
Wed, 1 Mar 2017 11:00:50 +0000 (12:00 +0100)
committerFrancois Fleuret <francois@fleuret.org>
Wed, 1 Mar 2017 11:00:50 +0000 (12:00 +0100)
test.py

diff --git a/test.py b/test.py
index de408aa..bf51360 100755 (executable)
--- a/test.py
+++ b/test.py
@@ -4,17 +4,46 @@ import torch
 import torchvision
 from torchvision import datasets
 
-from _ext import mylib
+######################################################################
+
+def sequences_to_image(x):
+    from PIL import Image
+
+    nb_sequences = x.size(0)
+    nb_images_per_sequences = x.size(1)
+    nb_channels = 3
+
+    if x.size(2) != nb_channels:
+        print('Can only handle 3 channel tensors.')
+        exit(1)
+
+    height = x.size(3)
+    width = x.size(4)
+    gap = 1
+    gap_color = (0, 128, 255)
 
-x = torch.ByteTensor(4, 5).fill_(0)
+    result = torch.ByteTensor(nb_channels,
+                              gap + nb_sequences * (height + gap),
+                              gap + nb_images_per_sequences * (width + gap))
 
-print(x.size())
+    result[0].fill_(gap_color[0])
+    result[1].fill_(gap_color[1])
+    result[2].fill_(gap_color[2])
 
-mylib.generate_sequence(8, x)
+    for s in range(0, nb_sequences):
+        for i in range(0, nb_images_per_sequences):
+            result.narrow(1, gap + s * (height + gap), height).narrow(2, gap + i * (width + gap), width).copy_(x[s][i])
+
+    result_numpy = result.cpu().byte().transpose(0, 2).transpose(0, 1).numpy()
+
+    return Image.fromarray(result_numpy, 'RGB')
+
+######################################################################
+
+from _ext import mylib
 
-print(x.size())
+x = torch.ByteTensor()
 
-x = x.float().sub_(128).div_(128)
+mylib.generate_sequence(10, x)
 
-for s in range(0, x.size(0)):
-    torchvision.utils.save_image(x[s], 'example_' + str(s) + '.png')
+sequences_to_image(x).save('sequences.png')