def run_eval(model, eval_loader, args, output_suffix=''):
for sub_dir in ['codes', 'images']:
cur_eval_dir = os.path.join(args.out_dir, output_suffix, sub_dir)
if not os.path.exists(cur_eval_dir):
print("Creating directory %s." % cur_eval_dir)
os.makedirs(cur_eval_dir)
all_losses, all_msssim, all_psnr = [], [], []
start_time = time.time()
for i, (batch, ctx_frames, filenames,id_num) in enumerate(eval_loader):
with torch.no_grad():
batch = Variable(batch.cuda())
original, out_imgs, losses, code_batch = eval_forward(
model, (batch, ctx_frames,id_num), args)
losses, msssim, psnr = finish_batch(
args, filenames, original, out_imgs,
losses, code_batch, output_suffix)
all_losses += losses
all_msssim += msssim
all_psnr += psnr
if i % 10 == 0:
print('\tevaluating iter %d (%f seconds)...' % (
i, time.time() - start_time))
return (np.array(all_losses).mean(axis=0),
np.array(all_msssim).mean(axis=0),
np.array(all_psnr).mean(axis=0))
def run_eval(model, eval_loader, args, output_suffix=''):
for sub_dir in ['codes', 'images']:
cur_eval_dir = os.path.join(args.out_dir, output_suffix, sub_dir)
if not os.path.exists(cur_eval_dir):
print("Creating directory %s." % cur_eval_dir)
os.makedirs(cur_eval_dir)
#all_losses, all_msssim, all_psnr = [], [], []
#start_time = time.time()
for i, (batch, ctx_frames, filenames) in enumerate(eval_loader):
batch = Variable(batch.cuda(), volatile=True)
codes = eval_forward(model, (batch, ctx_frames), args)
save_codes(name='initial' + str(i),
codes=codes,
args=args,
output_suffix=output_suffix)
'''
def run_eval(model, eval_loader, args, output_suffix=''):
with torch.no_grad():
for sub_dir in ['codes', 'images']:
cur_eval_dir = os.path.join(args.out_dir, output_suffix, sub_dir)
if not os.path.exists(cur_eval_dir):
print("Creating directory %s." % cur_eval_dir)
os.makedirs(cur_eval_dir)
all_losses, all_msssim, all_psnr = [], [], []
total_baseline_scores = np.array([0., 0.])
start_time = time.time()
for i, (batch, ctx_frames, filenames) in enumerate(eval_loader):
# f1, f2 = batch[:, :3].numpy(), batch[:, 6:9].numpy()
batch = batch.cuda()
original, out_imgs, losses, code_batch, baseline_scores = eval_forward(
model, (batch, ctx_frames), args)
losses, msssim, psnr = finish_batch(args, filenames, original,
out_imgs, losses, code_batch,
output_suffix)
all_losses += losses
all_msssim += msssim
all_psnr += psnr
if i % 10 == 0:
print('\tevaluating iter %d (%f seconds)...' %
(i, time.time() - start_time))
total_baseline_scores += baseline_scores
return (np.array(all_losses).mean(axis=0),
np.array(all_msssim).mean(axis=0),
np.array(all_psnr).mean(axis=0),
total_baseline_scores / len(eval_loader))