import torch
import torchvision
-from torchvision import datasets
+import argparse
-from _ext import mylib
+from _ext import flatland
-x = torch.ByteTensor(4, 5).fill_(0)
+######################################################################
-print(x.size())
+parser = argparse.ArgumentParser(
+ description='Dummy test of the flatland sequence generation.',
+ formatter_class=argparse.ArgumentDefaultsHelpFormatter
+)
-mylib.generate_sequence(8, x)
+parser.add_argument('--seed',
+ type = int, default = 0,
+ help = 'Random seed, < 0 is no seeding')
-print(x.size())
+parser.add_argument('--width',
+ type = int, default = 80,
+ help = 'Image width')
-x = x.float().sub_(128).div_(128)
+parser.add_argument('--height',
+ type = int, default = 80,
+ help = 'Image height')
-for s in range(0, x.size(0)):
- torchvision.utils.save_image(x[s], 'example_' + str(s) + '.png')
+parser.add_argument('--nb_shapes',
+ type = int, default = 10,
+ help = 'Image height')
+
+parser.add_argument('--nb_sequences',
+ type = int, default = 1,
+ help = 'How many sequences to generate')
+
+parser.add_argument('--nb_images_per_sequences',
+ type = int, default = 3,
+ help = 'How many images per sequence')
+
+parser.add_argument('--randomize_colors',
+ action='store_true', default=False,
+ help = 'Should the shapes be of different colors')
+
+parser.add_argument('--randomize_shape_size',
+ action='store_true', default=False,
+ help = 'Should the shapes be of different size')
+
+args = parser.parse_args()
+
+if args.seed >= 0:
+ torch.manual_seed(args.seed)
+
+######################################################################
+
+def sequences_to_image(x, gap = 1, gap_color = (0, 128, 255)):
+ 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)
+
+ result = torch.ByteTensor(nb_channels,
+ gap + nb_sequences * (height + gap),
+ gap + nb_images_per_sequences * (width + gap))
+
+ result.copy_(torch.Tensor(gap_color).view(-1, 1, 1).expand_as(result))
+
+ 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')
+
+######################################################################
+
+x = flatland.generate_sequence(args.nb_sequences,
+ args.nb_images_per_sequences,
+ args.height, args.width,
+ args.nb_shapes,
+ args.randomize_colors,
+ args.randomize_shape_size)
+
+sequences_to_image(x, gap = 1, gap_color = (0, 0, 0)).save('sequences.png')