def test(args):
assert (args.resume is True)
if cfg.MODEL.TEST.FINETUNE:
assert (cfg.MODEL.DECONVBASELINE.BN_GLOBAL_STATS is True)
base_dir = args.save_dir
logging_config(folder=base_dir, name="testing")
save_cfg(dir_path=base_dir, source=cfg.MODEL)
generator_net, loss_net, = construct_modules(args)
if args.dataset == "test":
pd_path = cfg.HKO_PD.RAINY_TEST
elif args.dataset == "valid":
pd_path = cfg.HKO_PD.RAINY_VALID
else:
raise NotImplementedError
hko_benchmark(
ctx=args.ctx[0],
generator_net=generator_net,
loss_net=loss_net,
sample_num=1,
save_dir=os.path.join(base_dir, "iter{}_{}_finetune{}".format(
cfg.MODEL.LOAD_ITER + 1, args.dataset, cfg.MODEL.TEST.FINETUNE)),
finetune=cfg.MODEL.TEST.FINETUNE,
mode=cfg.MODEL.TEST.MODE,
pd_path=pd_path)
def train(args):
base_dir = get_base_dir(args)
### Get modules
generator_net, loss_net = construct_modules(args)
### Prepare data
mnist_iter = MovingMNISTAdvancedIterator(
distractor_num=cfg.MOVINGMNIST.DISTRACTOR_NUM,
initial_velocity_range=(cfg.MOVINGMNIST.VELOCITY_LOWER,
cfg.MOVINGMNIST.VELOCITY_UPPER),
rotation_angle_range=(cfg.MOVINGMNIST.ROTATION_LOWER,
cfg.MOVINGMNIST.ROTATION_UPPER),
scale_variation_range=(cfg.MOVINGMNIST.SCALE_VARIATION_LOWER,
cfg.MOVINGMNIST.SCALE_VARIATION_UPPER),
illumination_factor_range=(cfg.MOVINGMNIST.ILLUMINATION_LOWER,
cfg.MOVINGMNIST.ILLUMINATION_UPPER))
for i in range(cfg.MODEL.TRAIN.MAX_ITER):
seq, flow = mnist_iter.sample(batch_size=cfg.MODEL.TRAIN.BATCH_SIZE,
seqlen=cfg.MOVINGMNIST.IN_LEN +
cfg.MOVINGMNIST.OUT_LEN)
in_seq = seq[:cfg.MOVINGMNIST.IN_LEN, ...]
gt_seq = seq[cfg.MOVINGMNIST.IN_LEN:(cfg.MOVINGMNIST.IN_LEN +
cfg.MOVINGMNIST.OUT_LEN), ...]
# Transform data to NCDHW shape needed for 3D Convolution encoder and normalize
context_nd = mx.nd.array(in_seq) / 255.0
gt_nd = mx.nd.array(gt_seq) / 255.0
context_nd = mx.nd.transpose(context_nd, axes=(1, 2, 0, 3, 4))
gt_nd = mx.nd.transpose(gt_nd, axes=(1, 2, 0, 3, 4))
# Train a step
pred_nd, avg_l2, avg_real_mse, generator_grad_norm =\
train_step(generator_net, loss_net, context_nd, gt_nd)
# Logging
logging.info(
("Iter:{}, L2 Loss:{}, MSE Error:{}, Generator Grad Norm:{}"
).format(i, avg_l2, avg_real_mse, generator_grad_norm))
logging.info("Iter:%d" % i)
if (i + 1) % 100 == 0:
save_gif(context_nd.asnumpy()[0, 0, :, :, :],
os.path.join(base_dir, "input.gif"))
save_gif(gt_nd.asnumpy()[0, 0, :, :, :],
os.path.join(base_dir, "gt.gif"))
save_gif(pred_nd.asnumpy()[0, 0, :, :, :],
os.path.join(base_dir, "pred.gif"))
if cfg.MODEL.SAVE_ITER > 0 and (i + 1) % cfg.MODEL.SAVE_ITER == 0:
generator_net.save_checkpoint(prefix=os.path.join(
base_dir, "generator"),
epoch=i)
def recursive_generation(args):
assert (cfg.MOVINGMNIST.OUT_LEN == 1)
base_dir = get_base_dir(args)
# Get modules
generator_net, = construct_modules(args)
# Prepare data
mnist_iter = MovingMNISTAdvancedIterator(
distractor_num=cfg.MOVINGMNIST.DISTRACTOR_NUM,
initial_velocity_range=(cfg.MOVINGMNIST.VELOCITY_LOWER,
cfg.MOVINGMNIST.VELOCITY_UPPER),
rotation_angle_range=(cfg.MOVINGMNIST.ROTATION_LOWER,
cfg.MOVINGMNIST.ROTATION_UPPER),
scale_variation_range=(cfg.MOVINGMNIST.SCALE_VARIATION_LOWER,
cfg.MOVINGMNIST.SCALE_VARIATION_UPPER),
illumination_factor_range=(cfg.MOVINGMNIST.ILLUMINATION_LOWER,
cfg.MOVINGMNIST.ILLUMINATION_UPPER))
frames = np.empty(shape=(cfg.MODEL.TEST.MAX_ITER,
cfg.MODEL.TRAIN.BATCH_SIZE,
cfg.MOVINGMNIST.TESTING_LEN, 1,
cfg.MOVINGMNIST.IMG_SIZE,
cfg.MOVINGMNIST.IMG_SIZE))
for i in range(cfg.MODEL.TEST.MAX_ITER):
seq, flow = mnist_iter.sample(batch_size=cfg.MODEL.TRAIN.BATCH_SIZE,
seqlen=cfg.MOVINGMNIST.IN_LEN +
cfg.MOVINGMNIST.TESTING_LEN)
in_seq = seq[:cfg.MOVINGMNIST.IN_LEN, ...]
gt_seq = seq[cfg.MOVINGMNIST.IN_LEN:(cfg.MOVINGMNIST.IN_LEN +
cfg.MOVINGMNIST.TESTING_LEN), ...]
# Transform data to NCDHW shape needed for 3D Convolution encoder and normalize
context_nd = mx.nd.array(in_seq) / 255.0
gt_nd = mx.nd.array(gt_seq) / 255.0
context_nd = mx.nd.transpose(context_nd, axes=(1, 2, 0, 3, 4))
gt_nd = mx.nd.transpose(gt_nd, axes=(1, 2, 0, 3, 4))
# Train a step
frames[i] = test_step(generator_net, context_nd)
frames = frames.reshape(
-1,
cfg.MOVINGMNIST.TESTING_LEN,
cfg.MOVINGMNIST.IMG_SIZE,
cfg.MOVINGMNIST.IMG_SIZE,
)
save_gifs(frames, os.path.join(base_dir, "pred"))
def test(args):
base_dir = get_base_dir(args)
logging_config(folder=base_dir, name='testing')
# Get modules
generator_net, loss_net = construct_modules(args)
# Prepare data
mnist_iter = MovingMNISTAdvancedIterator(
distractor_num=cfg.MOVINGMNIST.DISTRACTOR_NUM,
initial_velocity_range=(cfg.MOVINGMNIST.VELOCITY_LOWER,
cfg.MOVINGMNIST.VELOCITY_UPPER),
rotation_angle_range=(cfg.MOVINGMNIST.ROTATION_LOWER,
cfg.MOVINGMNIST.ROTATION_UPPER),
scale_variation_range=(cfg.MOVINGMNIST.SCALE_VARIATION_LOWER,
cfg.MOVINGMNIST.SCALE_VARIATION_UPPER),
illumination_factor_range=(cfg.MOVINGMNIST.ILLUMINATION_LOWER,
cfg.MOVINGMNIST.ILLUMINATION_UPPER))
num_samples, seqlen = mnist_iter.load(file=cfg.MOVINGMNIST.TEST_FILE)
overall_mse = 0
for iter_id in range(num_samples // cfg.MODEL.TRAIN.BATCH_SIZE):
frame_dat, _ = mnist_iter.sample(batch_size=cfg.MODEL.TRAIN.BATCH_SIZE,
seqlen=seqlen,
random=False)
context_nd = mx.nd.array(frame_dat[:cfg.MOVINGMNIST.IN_LEN],
ctx=args.ctx[0]) / 255.0
gt_nd = mx.nd.array(frame_dat[cfg.MOVINGMNIST.IN_LEN:],
ctx=args.ctx[0]) / 255.0
# Transform data to NCDHW shape needed for 3D Convolution encoder
context_nd = mx.nd.transpose(context_nd, axes=(1, 2, 0, 3, 4))
gt_nd = mx.nd.transpose(gt_nd, axes=(1, 2, 0, 3, 4))
pred_nd = test_step(generator_net, context_nd)
overall_mse += mx.nd.mean(mx.nd.square(pred_nd - gt_nd)).asscalar()
print(iter_id, overall_mse / (iter_id + 1))
avg_mse = overall_mse / (num_samples // cfg.MODEL.TRAIN.BATCH_SIZE)
with open(os.path.join(base_dir, 'result.txt'), 'w') as f:
f.write(str(avg_mse))
print(base_dir, avg_mse)
def train(args):
base_dir = get_base_dir(args)
# Get modules
generator_net, loss_net, = construct_modules(args)
# Prepare data
train_hko_iter = HKOIterator(
pd_path=cfg.HKO_PD.RAINY_TRAIN,
sample_mode="random",
seq_len=cfg.HKO.BENCHMARK.IN_LEN + cfg.HKO.BENCHMARK.OUT_LEN)
start_iter = 0 if not cfg.MODEL.RESUME else cfg.MODEL.LOAD_ITER
for i in range(start_iter, cfg.MODEL.TRAIN.MAX_ITER):
frame_dat, mask_dat, datetime_clips, _ = train_hko_iter.sample(
batch_size=cfg.MODEL.TRAIN.BATCH_SIZE)
context_nd = mx.nd.array(
frame_dat[0:cfg.HKO.BENCHMARK.IN_LEN, ...],
ctx=args.ctx[0]) / 255.0
gt_nd = mx.nd.array(
frame_dat[cfg.HKO.BENCHMARK.IN_LEN:(cfg.HKO.BENCHMARK.IN_LEN +
cfg.HKO.BENCHMARK.OUT_LEN)],
ctx=args.ctx[0]) / 255.0
mask_nd = mx.nd.array(
mask_dat[cfg.HKO.BENCHMARK.IN_LEN:(cfg.HKO.BENCHMARK.IN_LEN +
cfg.HKO.BENCHMARK.OUT_LEN)],
ctx=args.ctx[0])
# Transform data to NCDHW shape needed for 3D Convolution encoder and normalize
context_nd = mx.nd.transpose(context_nd, axes=(1, 2, 0, 3, 4))
gt_nd = mx.nd.transpose(gt_nd, axes=(1, 2, 0, 3, 4))
mask_nd = mx.nd.transpose(mask_nd, axes=(1, 2, 0, 3, 4))
# Train a step
pred_nd, avg_l2, avg_real_mse, generator_grad_norm =\
train_step(generator_net, loss_net, context_nd, gt_nd, mask_nd)
if (i + 1) % cfg.MODEL.VALID_ITER == 0:
hko_benchmark(
ctx=args.ctx[0],
generator_net=generator_net,
loss_net=loss_net,
sample_num=i,
save_dir=os.path.join(base_dir, "iter{}_valid".format(i + 1)),
pd_path=cfg.HKO_PD.RAINY_VALID)
# Logging
logging.info((
"Iter:{}, L2 Loss:{}, MSE Error:{}, Generator Grad Norm:{}").format(
i, avg_l2, avg_real_mse, generator_grad_norm))
if cfg.MODEL.SAVE_ITER > 0 and (i + 1) % cfg.MODEL.SAVE_ITER == 0:
generator_net.save_checkpoint(
prefix=os.path.join(base_dir, "generator"), epoch=i)
hko_benchmark(
ctx=args.ctx[0],
generator_net=generator_net,
loss_net=loss_net,
sample_num=i,
save_dir=os.path.join(base_dir, "iter{}_test".format(i + 1)),
pd_path=cfg.HKO_PD.RAINY_TEST)
