volatile=True)
# forward (recontruction)
outputs = model.forward(inputs)
# forward (sampling)
samples = []
for k in range(num_samples):
indices = np.random.choice(num_dists, args.batch)
sample = model.forward(inputs,
mean=to_var(means[indices]),
log_var=to_var(log_vars[indices]))
samples.append(sample)
# visualize
outputs = visualize(inputs, outputs)
targets = visualize(inputs, targets)
samples = [visualize(inputs, sample) for sample in samples]
inputs = visualize(inputs)
# logger
logger.image_summary('{0}-inputs'.format(split), inputs, step)
logger.image_summary('{0}-outputs'.format(split),
zip(inputs, outputs), step)
logger.image_summary('{0}-targets'.format(split),
zip(inputs, targets), step)
for k, sample in enumerate(samples):
logger.image_summary('{0}-samples-{1}'.format(split, k + 1),
zip(inputs, sample), step)
def analyze_reprs(max_dims=16, threshold=.5, bound=8., step=.2):
reprs_path = os.path.join('exp', args.exp, 'reprs')
mkdir(reprs_path, clean=True)
images_path = os.path.join(reprs_path, 'images')
mkdir(images_path, clean=True)
# statistics
x, ym, yv = [], [], []
for k in range(means.shape[1]):
x.extend([k, k])
ym.extend([np.min(means[:, k]), np.max(means[:, k])])
yv.extend([np.min(log_vars[:, k]), np.max(log_vars[:, k])])
plt.figure()
plt.bar(x, ym, .5, color='b')
plt.xlabel('dimension')
plt.ylabel('mean')
plt.savefig(os.path.join(images_path, 'means.png'), bbox_inches='tight')
plt.figure()
plt.bar(x, yv, .5, color='b')
plt.xlabel('dimension')
plt.ylabel('log(var)')
plt.savefig(os.path.join(images_path, 'vars.png'), bbox_inches='tight')
# dimensions
values = np.arange(-bound, bound + step, step)
magnitudes = np.max(np.abs(means), axis=0)
indices = np.argsort(-magnitudes)
dimensions = [k for k in indices if magnitudes[k] > threshold][:max_dims]
print('==> dominated dimensions = {0}'.format(dimensions))
for split in ['train', 'test']:
inputs, targets = iter(loaders[split]).next()
inputs, targets = to_var(inputs, volatile=True), to_var(targets,
volatile=True)
outputs, z = model.forward(inputs, returns='z')
for dim in tqdm(dimensions):
repr = to_np(z).copy()
samples = []
for val in tqdm(values, leave=False):
repr[:, dim] = val
sample = model.forward(inputs, z=to_var(repr, volatile=True))
samples.append(visualize(inputs, sample))
for k in range(args.batch):
images = [sample[k] for sample in samples]
image_path = os.path.join(
images_path, '{0}-{1}-{2}.gif'.format(split, k, dim))
save_images(images,
image_path,
duration=.1,
channel_first=True)
inputs = visualize(inputs)
for k in range(args.batch):
image_path = os.path.join(images_path,
'{0}-{1}.png'.format(split, k))
save_image(inputs[k], image_path, channel_first=True)
# visualization
with open(os.path.join(reprs_path, 'index.html'), 'w') as fp:
print('<h3>statistics</h3>', file=fp)
print('<img src="{0}">'.format(os.path.join('images', 'means.png')),
file=fp)
print('<img src="{0}">'.format(os.path.join('images', 'vars.png')),
file=fp)
print('<h3>inputs</h3>', file=fp)
print('<table border="1" style="table-layout: fixed;">', file=fp)
for split in ['train', 'test']:
print('<tr>', file=fp)
for k in range(args.batch):
image_path = os.path.join('images',
'{0}-{1}.png'.format(split, k))
print(
'<td halign="center" style="word-wrap: break-word;" valign="top">',
file=fp)
print(
'<img src="{0}" style="width:128px;">'.format(image_path),
file=fp)
print('</td>', file=fp)
print('</tr>', file=fp)
print('</table>', file=fp)
for dim in dimensions:
print('<h3>dimension [{0}]</h3>'.format(dim), file=fp)
print('<table border="1" style="table-layout: fixed;">', file=fp)
for split in ['train', 'test']:
print('<tr>', file=fp)
for k in range(args.batch):
image_path = os.path.join(
'images', '{0}-{1}-{2}.gif'.format(split, k, dim))
print(
'<td halign="center" style="word-wrap: break-word;" valign="top">',
file=fp)
print('<img src="{0}" style="width:128px;">'.format(
image_path),
file=fp)
print('</td>', file=fp)
print('</tr>', file=fp)
print('</table>', file=fp)
print('==> snapshot "{0}" loaded (epoch {1})'.format(
args.resume, epoch))
else:
raise FileNotFoundError('no snapshot found at "{0}"'.format(
args.resume))
model.train(False)
for k in trange(args.samples // args.batch):
outputs = model.forward(batch_size=args.batch)
outputs = outputs.data.cpu().numpy()
for i, output in enumerate(outputs):
index = k * args.batch + i
scipy.misc.imsave(
os.path.join(demo_path, '{:08d}.png'.format(index)),
visualize(output))
np.save(os.path.join(demo_path, '{:08d}.npy'.format(index)),
output)
# visualization
with open(os.path.join(demo_path, 'index.html'), 'w') as fp:
print('<table border="1" style="table-layout: fixed;">', file=fp)
for k in range(0, args.samples, 8):
print('<tr>', file=fp)
for index in range(k, k + 8):
print(
'<td halign="center" style="word-wrap: break-word;" valign="top">',
file=fp)
print('<img src="{0}" style="width:128px;">'.format(
'{:08d}.png'.format(index)),
file=fp)