def valid_inner_epoch(model, data_queue, batch_size):
sum_score = 0
scale = 1. / 255.
xp = utils.get_model_module(model)
valid_x, valid_y = data_queue.get()
perm = np.random.permutation(len(valid_x))
for i in six.moves.range(0, len(valid_x), batch_size):
local_perm = perm[i:i + batch_size]
batch_x = xp.array(valid_x[local_perm], dtype=np.float32) * scale
batch_y = xp.array(valid_y[local_perm], dtype=np.float32) * scale
pred = model(batch_x)
score = iproc.clipped_psnr(pred.data, batch_y)
sum_score += score * len(batch_x)
return sum_score / len(valid_x)
def valid_inner_epoch(model, data_queue, batch_size):
sum_score = 0
xp = model.xp
valid_x, valid_y = data_queue.get()
perm = np.random.permutation(len(valid_x))
with chainer.no_backprop_mode(), chainer.using_config('train', False):
for i in six.moves.range(0, len(valid_x), batch_size):
local_perm = perm[i:i + batch_size]
batch_x = xp.array(valid_x[local_perm], dtype=np.float32) / 255
batch_y = xp.array(valid_y[local_perm], dtype=np.float32) / 255
pred = model(batch_x)
score = iproc.clipped_psnr(pred.data, batch_y)
sum_score += float(score) * len(batch_x)
return sum_score / len(valid_x)
def benchmark(cfg, models, images, sampling_factor, quality):
scores = []
for src in images:
dst = pairwise_transform.scale(np.array(src),
[cfg.downsampling_filter], 1, 1, False)
if quality != 100 or cfg.method != 'scale':
with iproc.array_to_wand(dst) as tmp:
tmp = iproc.jpeg(tmp, sampling_factor, quality)
dst = iproc.wand_to_array(tmp)
dst = Image.fromarray(dst)
if 'noise_scale' in models:
dst = upscale_image(cfg, dst, models['noise_scale'])
else:
if 'noise' in models:
dst = denoise_image(cfg, dst, models['noise'])
if 'scale' in models:
dst = upscale_image(cfg, dst, models['scale'])
score = iproc.clipped_psnr(np.array(dst), np.array(src), a_max=255)
scores.append(score)
return np.mean(scores), np.std(scores) / np.sqrt(len(scores))
