def _scatter(link, array, target):
size, num = size_num_grads(link)
ptrs = numpy.zeros(num, dtype=numpy.uint64)
dtypes = numpy.zeros(num, dtype=numpy.int8)
info = numpy.zeros(num + 1, dtype=numpy.int32)
info[0] = 0
i = 0
for _, param in sorted(link.namedparams()):
if param.size == 0:
continue
ptrs[i] = 0 # NULL pointer
d = getattr(param, target)
if d is None:
d = cuda.cupy.zeros(param.shape, dtype=param.dtype)
setattr(param, target, d)
ptrs[i] = d.data.ptr
dtypes[i] = 0 # fp32
if param.dtype == numpy.float16:
dtypes[i] = 1 # fp16
info[i + 1] = info[i] + param.size
i += 1
if i != num:
raise()
info[0] = num
ptrs = cuda.to_gpu(ptrs)
dtypes = cuda.to_gpu(dtypes)
info = cuda.to_gpu(info)
return _memcpy_scatter()(ptrs, dtypes, info, array, size=size)
def test_forward_gpu_train(self):
self.rnn.to_gpu()
with chainer.using_config('use_cudnn', 'always'), \
chainer.using_config('train', True):
self.check_forward(
cuda.to_gpu(self.h),
[cuda.to_gpu(x) for x in self.xs])
def test_double_backward_gpu(self):
self.check_double_backward(
cuda.to_gpu(self.x1),
cuda.to_gpu(self.x2),
cuda.to_gpu(self.gy),
cuda.to_gpu(self.ggx1),
cuda.to_gpu(self.ggx2))
def test_double_backward_gpu(self):
b = None if self.b is None else cuda.to_gpu(self.b)
inputs = [cuda.to_gpu(self.x), cuda.to_gpu(self.W), b]
grad_outputs = cuda.to_gpu([self.gy])
grad_grad_inputs = cuda.to_gpu([self.ggx, self.ggW, self.ggb])
self.check_double_backward(
inputs, grad_outputs, grad_grad_inputs, use_cudnn='never')
def _gather(link, target):
size, num = size_num_grads(link)
ptrs = numpy.empty(num, dtype=numpy.uint64)
dtypes = numpy.empty(num, dtype=numpy.int8)
info = numpy.empty(num + 1, dtype=numpy.int32)
info[0] = 0
i = 0
for _, param in sorted(link.namedparams()):
if param.size == 0:
continue
ptrs[i] = 0 # NULL pointer
d = getattr(param, target)
if d is not None:
ptrs[i] = d.data.ptr
dtypes[i] = 0 # fp32
if param.dtype == numpy.float16:
dtypes[i] = 1 # fp16
info[i + 1] = info[i] + param.size
i += 1
info[0] = num
ptrs = cuda.to_gpu(ptrs)
dtypes = cuda.to_gpu(dtypes)
info = cuda.to_gpu(info)
return _memcpy_gather()(ptrs, dtypes, info, size=size)
def test_backward_chainerx_cuda_nobias(self):
self._skip_if_not_chainerx_supported()
self.check_backward(
backend.to_chainerx(cuda.to_gpu(self.x)),
backend.to_chainerx(cuda.to_gpu(self.W)),
None,
backend.to_chainerx(cuda.to_gpu(self.gy)))
def test_double_backward_gpu(self):
self.check_double_backward(
cuda.to_gpu(self.h),
cuda.to_gpu(self.x),
cuda.to_gpu(self.gy),
cuda.to_gpu(self.ggh),
cuda.to_gpu(self.ggx))
def test_backward_cpu_gpu(self):
# This test compares gradients of CPU and GPU modes.
rng = numpy.random.RandomState()
rng_state = rng.get_state()
# Call CPU mode link and save samples
x = chainer.Variable(self.x)
t = chainer.Variable(self.t)
link = self.create_link(rng)
y, samples = link(x, t, return_samples=True)
y.backward()
assert t.grad is None
gw_cpu = link.W.grad
gx_cpu = x.grad
# Call GPU mode link
rng.set_state(rng_state)
link = self.create_link(rng)
link.to_gpu()
x = chainer.Variable(cuda.to_gpu(self.x))
t = chainer.Variable(cuda.to_gpu(self.t))
y = self.call_link_with_samples(
cuda.to_gpu(samples), lambda: link(x, t))
y.backward()
assert t.grad is None
gw_gpu = link.W.grad
gx_gpu = x.grad
# Compare gradients from CPU and GPU modes
testing.assert_allclose(gx_cpu, gx_gpu, **self.test_backward_options)
testing.assert_allclose(gw_cpu, gw_gpu, **self.test_backward_options)
def test_double_backward_gpu(self):
x1 = cuda.to_gpu(self.x1)
x2 = cuda.to_gpu(self.x2)
gy = cuda.to_gpu(self.gy)
ggx1 = cuda.to_gpu(self.ggx1)
ggx2 = cuda.to_gpu(self.ggx2)
self.check_double_backward(x1, x2, gy, ggx1, ggx2)
def test_forward_output_size_zero(self, backend_config):
if backend_config.use_chainerx:
# TODO(sonots): Support it
if self.dtype == numpy.float16:
raise unittest.SkipTest('ChainerX does not support float16')
with self.assertRaises(Exception):
x = numpy.random.rand(4, 4, 1, 4).astype(self.dtype)
# TODO(sonots): Cleanup to use testing.backend.get_array after
# chainerx.asfortranarray is implemented.
if (backend_config.use_cuda
or (backend_config.use_chainerx
and backend_config.chainerx_device.startswith('cuda:'))):
x = cuda.to_gpu(x)
if backend_config.use_chainerx:
x = chainer.backend.to_chainerx(x)
x = chainer.Variable(x)
with backend_config:
functions.max_pooling_2d(x, 3, stride=2)
with self.assertRaises(Exception):
x = numpy.random.rand(4, 4, 4, 1).astype(self.dtype)
# TODO(sonots): Cleanup to use testing.backend.get_array after
# chainerx.asfortranarray is implemented.
if (backend_config.use_cuda
or (backend_config.use_chainerx
and backend_config.chainerx_device.startswith('cuda:'))):
x = cuda.to_gpu(x)
if backend_config.use_chainerx:
x = chainer.backend.to_chainerx(x)
x = chainer.Variable(x)
with backend_config:
functions.max_pooling_2d(x, 3, stride=2)
def test_batch_double_backward_gpu(self):
x_data = cuda.to_gpu(self.x)
y_grad = cuda.to_gpu(self.gy)
x_grad_grad = cuda.to_gpu(self.ggx)
gradient_check.check_double_backward(
self.det, x_data, y_grad, x_grad_grad,
**self.check_double_backward_options)
def test_backward_gpu(self):
y = self.f(chainer.Variable(cuda.to_gpu(self.x)),
chainer.Variable(cuda.to_gpu(self.x)))
y.grad = cuda.to_gpu(self.gy)
self.f.add_hook(self.h)
y.backward()
expect = r'''^function\tDummyFunction
input data
<variable at 0x[0-9a-f]+>
- device: <CUDA Device 0>
- backend: <(type|class) 'cupy.core.core.ndarray'>
- shape: \(3L?, 5L?\)
- dtype: float32
- statistics: mean=[0-9.\-e]+, std=[0-9.\-e]+
- grad: None
\(removed\)
output gradient
<variable at 0x[0-9a-f]+>
- device: <CUDA Device 0>
- backend: <(type|class) 'cupy.core.core.ndarray'>
- shape: \(3L?, 5L?\)
- dtype: float32
- statistics: mean=[0-9.\-e]+, std=[0-9.\-e]+
- grad: mean=[0-9.\-e]+, std=[0-9.\-e]+$
'''
actual = self.io.getvalue()
self.assertTrue(re.match(expect, actual), actual)
def check_double_backward(
self, inputs, grad_outputs, grad_grad_inputs, backend_config):
# TODO(sonots): Support it
if backend_config.use_chainerx and self.dtype == numpy.float16:
raise unittest.SkipTest('ChainerX does not support float16')
# TODO(sonots): Cleanup to use testing.backend.get_array after
# chainerx.asfortranarray is implemented.
if (backend_config.use_cuda
or (backend_config.use_chainerx
and backend_config.chainerx_device.startswith('cuda:'))):
inputs = cuda.to_gpu(inputs)
grad_outputs = cuda.to_gpu(grad_outputs)
grad_grad_inputs = cuda.to_gpu(grad_grad_inputs)
if not self.c_contiguous:
inputs = _to_fcontiguous(inputs)
grad_outputs = _to_fcontiguous(grad_outputs)
grad_grad_inputs = _to_fcontiguous(grad_grad_inputs)
if backend_config.use_chainerx:
inputs = chainer.backend.to_chainerx(inputs)
grad_outputs = chainer.backend.to_chainerx(grad_outputs)
grad_grad_inputs = chainer.backend.to_chainerx(grad_grad_inputs)
def f(x):
return functions.max_pooling_2d(
x, 3, stride=2, pad=1, cover_all=self.cover_all)
with backend_config:
gradient_check.check_double_backward(
f, inputs, grad_outputs, grad_grad_inputs,
dtype='d',
**self.check_double_backward_options)
def __call__(self, inputs, device=None):
"""Convert DALI arrays to Numpy/CuPy arrays"""
xp = cuda.get_array_module(self.perturbation)
if xp is not cuda.cupy:
self.perturbation = cuda.to_gpu(self.perturbation, device)
outputs = []
for i in range(len(inputs)):
x = inputs[i].as_tensor()
if (isinstance(x, dali.backend_impl.TensorCPU)):
x = np.array(x)
if x.ndim == 2 and x.shape[1] == 1:
x = x.squeeze(axis=1)
if device is not None and device >= 0:
x = cuda.to_gpu(x, device)
elif (isinstance(x, dali.backend_impl.TensorGPU)):
x_cupy = cuda.cupy.empty(shape=x.shape(), dtype=x.dtype())
# Synchronization is necessary here to avoid data corruption
# because DALI and CuPy will use different CUDA streams.
cuda.cupy.cuda.runtime.deviceSynchronize()
# copy data from DALI array to CuPy array
x.copy_to_external(ctypes.c_void_p(x_cupy.data.ptr))
cuda.cupy.cuda.runtime.deviceSynchronize()
x = x_cupy
if self.perturbation is not None:
x = x - self.perturbation
if device is not None and device < 0:
x = cuda.to_cpu(x)
else:
raise ValueError('Unexpected object')
outputs.append(x)
return tuple(outputs)
def test_sqnorm_array_multi_gpu(self):
x0 = cuda.to_gpu(self.x, device=0)
x1 = cuda.to_gpu(self.x, device=1)
a0 = cuda.to_gpu(self.a, device=0)
a1 = cuda.to_gpu(self.a, device=1)
self.assertAlmostEqual(optimizer._sum_sqnorm(
[self.x, self.a, x0, a0, x1, a1]), 8.75 * 3)
def test_backward_chainerx_cuda(self):
# TODO(niboshi): Support it
if self.dtype == numpy.float16:
raise unittest.SkipTest('ChainerX does not support float16')
self.check_backward_chainerx(
backend.to_chainerx(cuda.to_gpu(self.x)),
backend.to_chainerx(cuda.to_gpu(self.t)))
def check_backward(self, inputs, grad_outputs, backend_config):
xp = backend_config.xp
if backend_config.use_cuda:
inputs = cuda.to_gpu(inputs)
grad_outputs = cuda.to_gpu(grad_outputs)
x_data, W_data, b_data = inputs
y_grad, = grad_outputs
if not self.c_contiguous:
x_data = xp.asfortranarray(x_data)
W_data = xp.asfortranarray(W_data)
y_grad = xp.asfortranarray(y_grad)
assert not x_data.flags.c_contiguous
assert not W_data.flags.c_contiguous
assert not y_grad.flags.c_contiguous
if b_data is not None:
b = xp.empty((len(b_data) * 2,), dtype=b_data.dtype)
b[::2] = b_data
b_data = b[::2]
assert not b_data.flags.c_contiguous
args = (x_data, W_data)
if b_data is not None:
args = args + (b_data,)
def f(*args):
return F.deconvolution_2d(
*args, stride=self.stride, pad=self.pad, outsize=self.outsize,
dilate=self.dilate, groups=self.groups)
with backend_config:
gradient_check.check_backward(
f, args, y_grad, **self.check_backward_options)
def to_gpu(self):
"""Make a sampler GPU mode.
"""
if not self.use_gpu:
self.threshold = cuda.to_gpu(self.threshold)
self.values = cuda.to_gpu(self.values)
self.use_gpu = True
def test_forward_gpu(self):
self.link.to_gpu()
x = cuda.to_gpu(self.x)
if self.learn_b:
b = None
else:
b = cuda.to_gpu(self.b)
self.check_forward(x, b, self.y_expected)
def test_forward_gpu(self):
self.link.to_gpu()
x = cuda.to_gpu(self.x)
if self.learn_W:
W = None
else:
W = cuda.to_gpu(self.W)
self.check_forward(x, W, self.y_expected)
def test_backward_non_default_gpu(self):
x0 = chainer.Variable(cuda.to_gpu(self.x0, 1))
x1 = chainer.Variable(cuda.to_gpu(self.x1, 1))
gy = cuda.to_gpu(self.gy, 1)
with cuda.get_device_from_id(0):
y = functions.absolute_error(x0, x1)
y.grad = gy
y.backward()
def test_cupy_array2(self):
x = cuda.to_gpu(self.x, device=self.device_dtype(0))
with x.device:
if not self.c_contiguous:
x = cuda.cupy.asfortranarray(x)
y = cuda.to_gpu(x, device=self.device_dtype(1))
assert isinstance(y, cuda.ndarray)
assert int(y.device) == 1
def test_forward_gpu(self):
if not self.weight_apply:
weight = None
else:
weight = cuda.to_gpu(self.class_weight)
with chainer.using_config('use_cudnn', self.use_cudnn):
self.check_forward(
cuda.to_gpu(self.x), cuda.to_gpu(self.t), weight)
def test_debug_backward_gpu(self):
self.f.forward_gpu = mock.MagicMock(
return_value=(cuda.to_gpu(self.one),))
return_value = tuple(None if x is None else cuda.to_gpu(x)
for x in self.return_value)
input_value = (cuda.to_gpu(self.one),) * len(self.return_value)
self.f.backward_gpu = mock.MagicMock(return_value=return_value)
self.check_debug_backward(*input_value)
def test_double_backward_no_reduction_gpu(self):
with chainer.using_config('use_cudnn', 'always'):
self.check_double_backward(cuda.to_gpu(self.x),
cuda.to_gpu(self.t),
cuda.to_gpu(self.gy),
cuda.to_gpu(self.ggx),
normalize=self.normalize,
reduce='no')
def test_no_gh_backward_gpu(self):
self.check_backward(
cuda.to_gpu(self.c_prev1),
cuda.to_gpu(self.c_prev2),
cuda.to_gpu(self.x1),
cuda.to_gpu(self.x2),
cuda.to_gpu(self.gc),
None)
def test_backward_gpu(self):
self.rnn.to_gpu()
with chainer.using_config('use_cudnn', 'auto'):
self.check_backward(
cuda.to_gpu(self.h),
[cuda.to_gpu(x) for x in self.xs],
cuda.to_gpu(self.gh),
[cuda.to_gpu(gy) for gy in self.gys])
def test_forward_gpu_multi(self):
with cuda.get_device_from_id(0):
self.link.to_gpu()
x1 = cuda.to_gpu(self.x1)
x2 = cuda.to_gpu(self.x2)
x3 = cuda.to_gpu(self.x3)
with cuda.get_device_from_id(1):
self.check_forward(x1, x2, x3)
def test_forward_gpu_multi(self):
with cuda.get_device_from_id(0):
self.link.to_gpu()
c = cuda.to_gpu(self.c)
h = cuda.to_gpu(self.h)
x = cuda.to_gpu(self.x)
with cuda.get_device_from_id(1):
self.check_forward(c, h, x)
def check_backward(self, inputs, grad_outputs, backend_config):
if backend_config.use_cuda:
inputs = cuda.to_gpu(inputs)
grad_outputs = cuda.to_gpu(grad_outputs)
with backend_config:
gradient_check.check_backward(
functions.local_response_normalization, inputs, grad_outputs,
eps=1, dtype=numpy.float64, **self.check_backward_options)
def get_scenes():
data = contrib.get_data()
scenes = visualize_data()
models = morefusion.datasets.YCBVideoModels()
transform = []
points = []
sdf = []
pitch = []
origin = []
grid_target = []
grid_nontarget_empty = []
for instance in data["instances"]:
transform.append(instance["transform_init"].astype(np.float32))
points_i, sdf_i = models.get_sdf(class_id=instance["class_id"])
points.append(cuda.to_gpu(points_i).astype(np.float32))
sdf.append(cuda.to_gpu(sdf_i).astype(np.float32))
pitch.append(instance["pitch"].astype(np.float32))
origin.append(instance["origin"].astype(np.float32))
grid_target.append(instance["grid_target"].astype(np.float32))
grid_nontarget_empty.append(instance["grid_nontarget_empty"].astype(
np.float32))
pitch = cuda.cupy.asarray(pitch)
origin = cuda.cupy.asarray(origin)
grid_target = cuda.cupy.asarray(grid_target)
grid_nontarget_empty = cuda.cupy.asarray(grid_nontarget_empty)
link = morefusion.contrib.IterativeCollisionCheckLink(transform,
sdf_offset=0.02)
link.to_gpu()
optimizer = chainer.optimizers.Adam(alpha=0.01)
optimizer.setup(link)
link.translation.update_rule.hyperparam.alpha *= 0.1
for i in tqdm.trange(100):
transform = morefusion.functions.transformation_matrix(
link.quaternion, link.translation)
transform = cuda.to_cpu(transform.array)
for j, instance in enumerate(data["instances"]):
# cad = models.get_cad(instance["class_id"])
# if hasattr(cad.visual, "to_color"):
# cad.visual = cad.visual.to_color()
scenes["cad"].graph.update(
frame_to=str(instance["id"]),
frame_from="world",
matrix=transform[j],
)
# scenes["cad"].add_geometry(
# cad,
# node_name=str(instance["id"]),
# geom_name=str(instance["id"]),
# transform=transform[j],
# )
yield {"icc": scenes["cad"]}
loss = link(points, sdf, pitch, origin, grid_target,
grid_nontarget_empty)
loss.backward()
optimizer.update()
link.zerograds()
def run(which, steps, which_labels, frames, model, bg_model, face_model,
optimizer, pred_diff, loss_saving, trait, ordered, save_all_results,
twostream, same_frame, record_loss, record_predictions):
print('steps: ', steps)
assert (which in ['train', 'test', 'val'])
assert (trait in ['O', 'C', 'E', 'A', 'S'])
if which == 'train':
which_batch_size = C.TRAIN_BATCH_SIZE
elif which == 'val':
which_batch_size = C.VAL_BATCH_SIZE
elif which == 'test':
which_batch_size = C.TEST_BATCH_SIZE
loss_tmp = []
pd_tmp = np.zeros((steps, 1), dtype=float)
_labs = list(which_labels)
preds = np.zeros((steps, 1), dtype=float)
if not ordered:
shuffle(_labs)
ts = time.time()
for s in tqdm(range(steps)):
labels_selected = _labs[s * which_batch_size:(s + 1) *
which_batch_size]
assert (len(labels_selected) == which_batch_size)
labels_bg, bg_data, frame_num = D.load_data_single(
labels_selected,
which_labels,
frames,
which_data='bg',
resize=True,
ordered=ordered,
twostream=twostream,
same_frame=same_frame,
trait=trait)
labels_face, face_data, _ = D.load_data_single(labels_selected,
which_labels,
frames,
which_data='face',
resize=True,
ordered=ordered,
twostream=twostream,
frame_num=frame_num,
same_frame=same_frame,
trait=trait)
if C.ON_GPU:
bg_data = to_gpu(bg_data, device=C.DEVICE)
face_data = to_gpu(face_data, device=C.DEVICE)
labels = to_gpu(labels_bg, device=C.DEVICE)
with cp.cuda.Device(C.DEVICE):
if which == 'train':
config = True
else:
config = False
with chainer.using_config('train', False):
prediction_bg, bg_activations = bg_model(bg_data)
prediction_face, face_activations = face_model(face_data)
with chainer.using_config('train', config):
if config:
model.cleargrads()
prediction = model(bg_activations, face_activations)
loss = mean_absolute_error(prediction, labels)
if which == 'train':
loss.backward()
optimizer.update()
if record_loss:
loss_tmp.append(float(loss.data))
pd_tmp[s] = U.pred_diff_trait(to_cpu(prediction.data),
to_cpu(labels))
if record_predictions and which == 'test':
preds[s] = to_cpu(prediction.data)
if record_loss:
pred_diff[0] = np.mean(pd_tmp, axis=0)
loss_tmp_mean = np.mean(loss_tmp, axis=0)
loss_saving.append(loss_tmp_mean)
print('E %d. %s loss: ' % (0, which), loss_tmp_mean,
' pred diff %s: ' % trait, pred_diff[0], ' time: ',
time.time() - ts)
U.record_loss_sanity(which, loss_tmp_mean, pred_diff[0])
if which == 'test' and save_all_results:
U.record_loss_all_test(loss_tmp, trait=True)
if record_predictions and which == 'test':
U.record_all_predictions(which, preds)
def to_xp(arrs):
if backend_config.use_cuda:
return cuda.to_gpu(arrs)
else:
return arrs
def test_backward_gpu(self):
b = None if self.b is None else cuda.to_gpu(self.b)
self.check_backward(
cuda.to_gpu(self.x), cuda.to_gpu(self.W), b,
cuda.to_gpu(self.gy), use_cudnn='never')
def test_rtol_gpu(self):
self.check_rtol(cuda.to_gpu(self.x), cuda.to_gpu(self.y))
def test_invalid_reduce_gpu(self):
self.check_invalid_reduce(cuda.to_gpu(self.x), cuda.to_gpu(self.t))
def test_forward_gpu(self):
self.check_forward(cuda.to_gpu(self.x))
def test_double_backward_gpu(self):
self.check_double_backward(cuda.to_gpu(self.x), self.output_shape,
cuda.to_gpu(self.gy), cuda.to_gpu(self.ggx))
def predict_value_model_0(experiment_number,
epoch,
which='val',
time_gap=1,
model_number=0):
use_ccc = True
use_pearson = True
_loss_steps = []
_all_ccc = []
_all_pearson = []
model = Siamese_pilot()
models_path = '/scratch/users/gabras/data/omg_empathy/saving_data/models'
p = os.path.join(
models_path,
'model_%d_experiment_%d' % (model_number, experiment_number),
'epoch_%d' % epoch)
chainer.serializers.load_npz(p, model)
if C.ON_GPU:
model = model.to_gpu(device=C.DEVICE)
for subject in range(10):
_loss_steps_subject = []
all_predictions = []
name = 'Subject_%d_Story_1' % (subject + 1)
path = '/scratch/users/gabras/data/omg_empathy/Validation'
subject_folder = os.path.join(path, 'jpg_participant_662_542', name)
all_frames = os.listdir(subject_folder)
full_name = os.path.join(path, 'Annotations', name + '.csv')
all_labels = np.genfromtxt(full_name,
dtype=np.float32,
skip_header=True)
# num_frames = len(all_frames)
num_frames = 1000
_b = 0
_e = num_frames
for f in range(0, num_frames):
if f == 0:
with cp.cuda.Device(C.DEVICE):
with chainer.using_config('train', False):
prediction = np.array([0.0],
dtype=np.float32) # baseline
labels = np.array([all_labels[f]])
loss = mean_squared_error(prediction, labels)
_loss_steps_subject.append(float(loss.data))
all_predictions.append(0.0)
else:
data_left, data_right = L.get_left_right_consecutively(
which, name, f, time_gap=time_gap)
data_left = np.expand_dims(data_left, 0)
data_right = np.expand_dims(data_right, 0)
labels = np.array([all_labels[f]])
if C.ON_GPU:
data_left = to_gpu(data_left, device=C.DEVICE)
data_right = to_gpu(data_right, device=C.DEVICE)
labels = to_gpu(labels, device=C.DEVICE)
with cp.cuda.Device(C.DEVICE):
with chainer.using_config('train', False):
prediction = model(data_left, data_right)
loss = mean_squared_error(prediction.data[0], labels)
_loss_steps_subject.append(float(to_cpu(loss.data)))
all_predictions.append(float(to_cpu(prediction.data)))
if use_ccc:
ccc_subject = calculate_ccc(all_predictions, all_labels[_b:_e])
_all_ccc.append(ccc_subject)
print('%s, loss: %f, ccc: %f' %
(name, float(np.mean(_loss_steps_subject)), ccc_subject))
if use_pearson:
pearson_subject, p_vals = calculate_pearson(
all_predictions, all_labels[_b:_e])
_all_pearson.append(pearson_subject)
print('%s, loss: %f, pearson: %f' %
(name, float(np.mean(_loss_steps_subject)), pearson_subject))
_loss_steps.append(np.mean(_loss_steps_subject))
print(
'model_%d_experiment_%d_epoch_%d, val_loss: %f, CCC: %f, pearson: %f' %
(model_number, experiment_number, epoch, float(np.mean(_loss_steps)),
float(np.mean(_all_ccc)), float(np.mean(_all_pearson))))
def forward(self, inputs):
# Cause inconsistency between outputs
return inputs[0], cuda.to_gpu(inputs[1])
def test_forward_gpu(self):
self.check_forward(cuda.to_gpu(self.x),
output_shape=self.output_shape[2:])
def test_reference_gpu(self):
self.check_reference(cuda.to_gpu(self.x))
def test_numerical_grad_gpu(self):
xs = tuple(map(cuda.to_gpu, self.xs))
gys = tuple(None if gy is None else cuda.to_gpu(gy) for gy in self.gys)
self.check_invalid_eps(xs, gys, 0)
self.check_invalid_eps(xs, gys, -1.0)
def test_invalid_mixed(self):
y = cuda.to_gpu(self.y)
with self.assertRaises(RuntimeError):
gradient_check.numerical_grad(self.f, (self.x, ), (y, ))
def test_backward_consistency_regression_gpu(self):
self.check_backward_consistency_regression(cuda.to_gpu(self.x),
cuda.to_gpu(self.gy))
def test_identical_gpu(self):
self.check_identical(cuda.to_gpu(self.x))
def test_backward_gpu(self):
self.check_backward(cuda.to_gpu(self.x), cuda.to_gpu(self.gy))
def test_gpu(self):
x = cuda.to_gpu(self.x)
with chainer.using_config('use_cudnn', 'never'):
self._check(x)
with chainer.using_config('use_cudnn', 'always'):
self._check(x)
def test_forward_consistency_regression_gpu(self):
self.check_forward_consistency_regression(cuda.to_gpu(self.x))
def test_double_backward_gpu_no_cudnn(self):
self.check_double_backward(cuda.to_gpu(self.x), cuda.to_gpu(self.gy),
cuda.to_gpu(self.ggx), 'never')
def test_array_supplied_gpu(self):
self.check_array_supplied(cuda.to_gpu(self.x_data),
cuda.to_gpu(self.W_data),
cuda.to_gpu(self.b_data))
def test_backward_gpu(self):
self.check_backward(cuda.to_gpu(self.x), self.output_shape,
cuda.to_gpu(self.gy))
def test_value_check_gpu_cudnn(self):
self.check_value_check(cuda.to_gpu(self.x), cuda.to_gpu(self.t),
'always')
def test_forward_gpu_test(self):
self.rnn.to_gpu()
with chainer.using_config('use_cudnn', 'always'), \
chainer.using_config('train', False):
self.check_forward(cuda.to_gpu(self.h), cuda.to_gpu(self.c),
[cuda.to_gpu(x) for x in self.xs])
def test_forward_consistency_regression_im2col(self):
# Regression test to deconvolution_nd.
if len(self.dims) == 2:
self.check_forward_consistency_regression(
cuda.to_gpu(self.x), cuda.to_gpu(self.W), cuda.to_gpu(self.b),
use_cudnn='never')
def test_double_backward_gpu_non_contiguous(self):
self.check_double_backward(
cuda.cupy.asfortranarray(cuda.to_gpu(self.x)),
cuda.cupy.asfortranarray(cuda.to_gpu(self.gy)),
cuda.cupy.asfortranarray(cuda.to_gpu(self.ggx)))
def test_forward_gpu(self):
self.link.to_gpu()
self.check_forward(cuda.to_gpu(self.xs), cuda.to_gpu(self.ys))
def make_pred_vs_y_plot(experiment_number,
epoch,
which='val',
time_gap=1,
model_number=1,
label_type='discrete'):
use_ccc = True
use_pearson = True
_loss_steps = []
_all_ccc = []
_all_pearson = []
if model_number == 1:
model = Siamese()
elif model_number == 3:
model = Triplet()
elif model_number == 4:
model = TernaryClassifier()
models_path = '/scratch/users/gabras/data/omg_empathy/saving_data/models'
p = os.path.join(
models_path,
'model_%d_experiment_%d' % (model_number, experiment_number),
'epoch_%d' % epoch)
chainer.serializers.load_npz(p, model)
if C.ON_GPU:
model = model.to_gpu(device=C.DEVICE)
for subject in range(10):
_loss_steps_subject = []
previous_prediction = np.array([0.], dtype=np.float32)
all_predictions = []
name = 'Subject_%d_Story_1' % (subject + 1)
path = '/scratch/users/gabras/data/omg_empathy/Validation'
subject_folder = os.path.join(path, 'jpg_participant_662_542', name)
all_frames = os.listdir(subject_folder)
if label_type == 'discrete':
full_name = os.path.join(path, 'Annotations', name + '.csv')
elif label_type == 'smooth':
full_name = os.path.join(path, C.SMOOTH_ANNOTATIONS_PATH,
name + '.csv')
all_labels = np.genfromtxt(full_name,
dtype=np.float32,
skip_header=True)
# num_frames = len(all_frames)
num_frames = 50
# num_frames = 10
if time_gap > 1:
_b = C.OMG_EMPATHY_FRAME_RATE
_e = _b + num_frames
else:
_b = 0
_e = num_frames
# for f in tqdm(range(_b, _e)):
for f in range(_b, _e):
if f == 0:
with cp.cuda.Device(C.DEVICE):
with chainer.using_config('train', False):
prediction = np.array([0.0],
dtype=np.float32) # baseline
labels = np.array([all_labels[f]])
loss = mean_squared_error(prediction, labels)
_loss_steps_subject.append(float(loss.data))
else:
data_left, data_right = L.get_left_right_consecutively(
which, name, f, time_gap=time_gap)
data_left = np.expand_dims(data_left, 0)
data_right = np.expand_dims(data_right, 0)
labels = np.array([all_labels[f]])
if C.ON_GPU:
previous_prediction = to_gpu(previous_prediction,
device=C.DEVICE)
data_left = to_gpu(data_left, device=C.DEVICE)
data_right = to_gpu(data_right, device=C.DEVICE)
labels = to_gpu(labels, device=C.DEVICE)
with cp.cuda.Device(C.DEVICE):
with chainer.using_config('train', False):
if model_number == 3:
pred_1, pred_2 = model(data_left, data_right)
pred_1 = chainer.functions.sigmoid(pred_1)
pred_1 = U.threshold_all(to_cpu(pred_1.data))
prediction = to_gpu(pred_1,
device=C.DEVICE) * pred_2
prediction = previous_prediction + prediction
loss = mean_squared_error(prediction.data[0],
labels)
_loss_steps_subject.append(float(loss.data))
prediction = float(prediction.data)
else:
prediction = model(data_left, data_right)
prediction = previous_prediction + prediction
loss = mean_squared_error(prediction.data[0],
labels)
_loss_steps_subject.append(float(loss.data))
prediction = float(prediction.data)
previous_prediction = to_cpu(previous_prediction)
previous_prediction[0] = float(prediction)
all_predictions.append(prediction)
# if use_ccc:
# ccc_subject = calculate_ccc(all_predictions, all_labels[_b:_e])
# _all_ccc.append(ccc_subject)
# print('%s, loss: %f, ccc: %f' % (name, float(np.mean(_loss_steps_subject)), ccc_subject))
#
# if use_pearson:
# pearson_subject, p_vals = calculate_pearson(all_predictions, all_labels[_b:_e])
# _all_pearson.append(pearson_subject)
# print('%s, loss: %f, pearson: %f' % (name, float(np.mean(_loss_steps_subject)), pearson_subject))
#
# _loss_steps.append(np.mean(_loss_steps_subject))
# save graph
save_graph = True
if save_graph:
labs = all_labels[_b:_e]
diff_arr = labs - all_predictions
diff_matrix = np.zeros((num_frames, num_frames))
for i in range(num_frames):
diff_matrix[i, i] = diff_arr[i]
p = '/scratch/users/gabras/data/omg_empathy/saving_data/logs/val/pred_vs_y'
plots_folder = 'model_%d_experiment_%d' % (model_number,
experiment_number)
plot_path = os.path.join(p, plots_folder)
if not os.path.exists(plot_path):
os.mkdir(plot_path)
fig = plt.figure()
# assert len(labs) == len(all_predictions)
# plt.scatter(labs, all_predictions)
fig, ax = plt.subplots()
im = ax.imshow(diff_matrix)
# We want to show all ticks...
ax.set_xticks(np.arange(len(diff_arr)))
ax.set_yticks(np.arange(len(diff_arr)))
# ... and label them with the respective list entries
# ax.set_xticklabels(farmers)
# ax.set_yticklabels(vegetables)
# Rotate the tick labels and set their alignment.
# plt.setp(ax.get_xticklabels(), rotation=45, ha="right",
# rotation_mode="anchor")
# Loop over data dimensions and create text annotations.
# for i in range(len(diff_arr)):
# for j in range(len(diff_arr)):
# text = ax.text(j, i, str(diff_matrix[i, j])[0:7], ha="center", va="center", color="w")
ax.set_title('difference: labels - predictions')
fig.tight_layout()
plt.savefig(
os.path.join(plot_path, '%s_epoch_%d_.png' % (name, epoch)))
def test_debug_forward_gpu(self):
return_value = tuple(None if x is None else cuda.to_gpu(x)
for x in self.return_value)
self.f.forward_gpu = mock.MagicMock(return_value=return_value)
self.check_debug_forward(cuda.to_gpu(self.one))
def test_differenct_eps_gpu(self):
self.check_different_eps(cuda.to_gpu(self.x), cuda.to_gpu(self.y))
