def _check_forward_internal(self, dst_device_spec, src_device, dst_device,
x_mode):
x = src_device.send(self.x)
if x_mode == 'array':
pass
elif x_mode == 'non_requires_grad':
x = chainer.Variable(x, requires_grad=False)
elif x_mode == 'requires_grad':
x = chainer.Variable(x, requires_grad=True)
else:
assert False, x_mode
error_expected = ((src_device.xp is chainerx) !=
(dst_device.xp is chainerx)
and x_mode == 'requires_grad')
if error_expected:
with pytest.raises(RuntimeError):
functions.copy(x, dst_device_spec)
return
y = functions.copy(x, dst_device_spec)
assert y.device == dst_device
assert backend.get_device_from_array(y.array) == dst_device
assert y.dtype == self.dtype
numpy.testing.assert_array_equal(_numpy_device.send(y.array), self.x)
def forward(self, x):
if self.device0 != self.device1:
# assume x is on device0
x1 = F.copy(x, self.device1)
z0 = self.first0(x)
z1 = self.first1(x1)
# synchronize
h0 = z0 + F.copy(z1, self.device0)
h1 = z1 + F.copy(z0, self.device1)
y0 = self.second0(F.relu(h0))
y1 = self.second1(F.relu(h1))
y = y0 + F.copy(y1, self.device0)
return y # output is on device0
else:
z0 = self.first0(x)
z1 = self.first1(x)
h = z0 + z1
y0 = self.second0(F.relu(h))
y1 = self.second1(F.relu(h))
y = y0 + y1
return y
def forward(x_data, y_data, train=True):
# Neural net architecture
x_0 = chainer.Variable(cuda.to_gpu(x_data, 0), volatile=not train)
x_1 = chainer.Variable(cuda.to_gpu(x_data, 1), volatile=not train)
t = chainer.Variable(cuda.to_gpu(y_data, 0), volatile=not train)
h1_0 = F.dropout(F.relu(model.gpu0.l1(x_0)), train=train)
h1_1 = F.dropout(F.relu(model.gpu1.l1(x_1)), train=train)
h2_0 = F.dropout(F.relu(model.gpu0.l2(h1_0)), train=train)
h2_1 = F.dropout(F.relu(model.gpu1.l2(h1_1)), train=train)
h3_0 = F.dropout(F.relu(model.gpu0.l3(h2_0)), train=train)
h3_1 = F.dropout(F.relu(model.gpu1.l3(h2_1)), train=train)
# Synchronize
h3_0 += F.copy(h3_1, 0)
h3_1 = F.copy(h3_0, 1)
h4_0 = F.dropout(F.relu(model.gpu0.l4(h3_0)), train=train)
h4_1 = F.dropout(F.relu(model.gpu1.l4(h3_1)), train=train)
h5_0 = F.dropout(F.relu(model.gpu0.l5(h4_0)), train=train)
h5_1 = F.dropout(F.relu(model.gpu1.l5(h4_1)), train=train)
h6_0 = F.relu(model.gpu0.l6(h5_0))
h6_1 = F.relu(model.gpu1.l6(h5_1))
# Synchronize
y = h6_0 + F.copy(h6_1, 0)
return F.softmax_cross_entropy(y, t), F.accuracy(y, t)
def __call__(self, x):
if self._device_id is not None and self.gpu_for_nn_only:
x = F.copy(x, self._device_id)
y = self.l(x)
if self._device_id is not None and self.gpu_for_nn_only:
y = F.copy(y, -1)
return y
def __call__(self, x):
A = [None] * self.layers
R = [None] * (self.layers - 1)
# update layers
for nth in range(self.layers):
if nth == 0 & nth == self.layers - 1:
(_, _) = self.Layer0(x)
elif nth == 0:
(A[1], _) = self.Layer0(x, self.R0)
elif nth == self.layers - 1:
(_, R[nth - 1]) = getattr(self, 'Layer' + str(nth))(
getattr(self, 'A' + str(nth)))
else:
(A[nth + 1], R[nth - 1]) = getattr(self, 'Layer' + str(nth))(
getattr(self, 'A' + str(nth)),
getattr(self, 'R' + str(nth)))
# copy data to device
if isinstance(x.data, numpy.ndarray) or self.devices[nth] is None:
for nth in range(self.layers - 1):
setattr(self, 'A' + str(nth + 1), A[nth + 1])
setattr(self, 'R' + str(nth), R[nth])
else:
for nth in range(self.layers - 1):
setattr(self, 'A' + str(nth + 1),
F.copy(A[nth + 1], self.devices[nth + 1]))
setattr(self, 'R' + str(nth),
F.copy(R[nth], self.devices[nth]))
return self.Layer0.P
def forward(x_data, y_data, train=True):
x_0 = chainer.Variable(cuda.to_gpu(x_data, 0), volatile=not train)
x_1 = chainer.Variable(cuda.to_gpu(x_data, 1), volatile=not train)
t = chainer.Variable(cuda.to_gpu(y_data, 0), volatile=not train)
h1_0 = F.dropout(F.relu(model.gpu0.l1(x_0)), train=train)
h1_1 = F.dropout(F.relu(model.gpu1.l1(x_1)), train=train)
h2_0 = F.dropout(F.relu(model.gpu0.l2(h1_0)), train=train)
h2_1 = F.dropout(F.relu(model.gpu1.l2(h1_1)), train=train)
h3_0 = F.dropout(F.relu(model.gpu0.l3(h2_0)), train=train)
h3_1 = F.dropout(F.relu(model.gpu1.l3(h2_1)), train=train)
# Synchronize
h3_0 += F.copy(h3_1, 0)
h3_1 = F.copy(h3_0, 1)
h4_0 = F.dropout(F.relu(model.gpu0.l4(h3_0)), train=train)
h4_1 = F.dropout(F.relu(model.gpu1.l4(h3_1)), train=train)
h5_0 = F.dropout(F.relu(model.gpu0.l5(h4_0)), train=train)
h5_1 = F.dropout(F.relu(model.gpu1.l5(h4_1)), train=train)
h6_0 = F.relu(model.gpu0.l6(h5_0))
h6_1 = F.relu(model.gpu1.l6(h5_1))
# Synchronize
y = h6_0 + F.copy(h6_1, 0)
return F.softmax_cross_entropy(y, t), F.accuracy(y, t)
def forward(self, x):
if self.gpu0 != self.gpu1:
# assume x is on gpu0
x1 = F.copy(x, self.gpu1)
z0 = self.first0(x)
z1 = self.first1(x1)
# synchronize
h0 = z0 + F.copy(z1, self.gpu0)
h1 = z1 + F.copy(z0, self.gpu1)
y0 = self.second0(F.relu(h0))
y1 = self.second1(F.relu(h1))
y = y0 + F.copy(y1, self.gpu0)
return y # output is on gpu0
else:
z0 = self.first0(x)
z1 = self.first1(x)
h = z0 + z1
y0 = self.second0(F.relu(h))
y1 = self.second1(F.relu(h))
y = y0 + y1
return y
def check_invalid(self, src_device, dst_device_spec):
x = src_device.send(
numpy.random.uniform(-1, 1, (10, 5)).astype(self.dtype))
x_var = chainer.Variable(x)
with pytest.raises(RuntimeError):
functions.copy(x_var, dst_device_spec)
def update_core(self):
optimizer = self.get_optimizer('main')
# it is main wrapper class: au_rcnn_train_chain, space_time_rnn
model_main = optimizer.target
loss_head_module = model_main.loss_head_module
models_others = {k: v for k, v in self._models.items()
if v != model_main.au_rcnn_train_chain}
batch = self.get_iterator('main').next()
in_arrays = self.converter(batch, -1)
images, bboxes, labels = in_arrays
batch_size, T, channel, height, width = images.shape
images = images.reshape(batch_size * T, channel, height, width) # B*T, C, H, W
bboxes = bboxes.reshape(batch_size * T, config.BOX_NUM[self.database], 4) # B*T, 9, 4
labels = chainer.cuda.to_gpu(labels, device=self._devices["main"])
# labels = labels.reshape(batch_size * T, config.BOX_NUM[self.database], -1) # B*T, 9, 12/22
# For reducing memory
for model in six.itervalues(models_others):
model.cleargrads()
model_main.cleargrads()
#
# Split the batch to sub-batches.
#
n = len(self._models)
in_arrays_list = {}
sub_index = self.split_list(list(range(batch_size * T)), n)
for i, key in enumerate(sorted(self._models.keys(), key=lambda e:str(e))): # self._models are all au_rcnn_train_chain includes main gpu
in_arrays_list[key] = (F.copy(images[sub_index[i]], self._devices.get(key, self._devices["main"])),
F.copy(bboxes[sub_index[i]], self._devices.get(key, self._devices["main"])))
# self._models are all au_rcnn_train_chain includes main gpu
with function.force_backprop_mode():
roi_feature_multi_gpu = []
for model_key, au_rcnn_train_chain in sorted(self._models.items(), key=lambda e:str(e[0])):
images, bboxes = in_arrays_list[model_key]
assert int(images.data.device) == au_rcnn_train_chain._device_id
roi_feature = au_rcnn_train_chain(images, bboxes) # shape =(B*T//n, F, D)
roi_feature_multi_gpu.append(F.copy(roi_feature, self._devices["main"]))
roi_feature = F.concat(roi_feature_multi_gpu, axis=0) # multiple batch combine
roi_feature = roi_feature.reshape(batch_size, T, config.BOX_NUM[self.database], roi_feature.shape[-1])
loss = loss_head_module(roi_feature, labels)
model_main.cleargrads()
for model in six.itervalues(self._models):
model.cleargrads()
loss.backward()
for model in six.itervalues(models_others):
model_main.au_rcnn_train_chain.addgrads(model)
optimizer.update()
for model in six.itervalues(models_others):
model.copyparams(model_main.au_rcnn_train_chain) # only the main model will update parameter, so copy to each other models
def __call__(self, x):
z0 = self.mlp1_gpu0(x)
z1 = self.mlp1_gpu1(F.copy(x, 1))
h0 = F.relu(z0 + F.copy(z1, 0))
h1 = F.relu(z1 + F.copy(z0, 1))
y0 = self.mlp2_gpu0(h0)
y1 = self.mlp2_gpu1(h1)
y = y0 + F.copy(y1, 0)
return y
def prepare_images(self, images):
if self.xp != np:
device = images.data.device
images = F.copy(images, -1)
converted_images = [
resnet.prepare(image.data, size=None)
for image in F.separate(images, axis=0)
]
converted_images = F.stack(converted_images, axis=0)
if self.xp != np:
converted_images = F.copy(converted_images, device.id)
return converted_images
def test_double_backward(self, src_backend_config, dst_backend_config):
src_device = src_backend_config.device
dst_device = dst_backend_config.device
if (src_device.xp is chainerx) is not (dst_device.xp is chainerx):
raise unittest.SkipTest(
'ChainerX to non-ChainerX does not support backward.')
x = src_backend_config.get_array(self.x)
gy = dst_backend_config.get_array(self.gy)
ggx = src_backend_config.get_array(self.ggx)
x_var = chainer.Variable(x, requires_grad=True)
y_var = functions.copy(x_var, dst_device)
y_var.grad = gy
gy_var = y_var.grad_var
y_var.backward(enable_double_backprop=True)
assert x_var.grad_var.requires_grad is True
x_var.grad_var.grad = ggx
x_var.grad_var.backward()
assert gy_var.grad_var.device == dst_device
assert (backend.get_device_from_array(
gy_var.grad_var.array) == dst_device)
numpy.testing.assert_array_equal(
_numpy_device.send(gy_var.grad_var.array), self.ggx)
def __call__(self, atom_array, adj, wle_array=None, is_real_node=None):
self.reset_state()
if atom_array.dtype == self.xp.int32:
h = self.embed(atom_array)
else:
# TODO: GraphLinear or GraphMLP can be used.
h = atom_array
h0 = functions.copy(h, cuda.get_device_from_array(h.data).id)
# all Combined NLE processes are done here.
if self.with_wle:
h_s = self.embed_wle(wle_array)
# gated sum
gate_input = self.gate_W1(h) + self.gate_W2(h_s)
gate_coefff = functions.sigmoid(gate_input)
h = (1.0 - gate_coefff) * h + gate_coefff * h_s
additional_kwargs = self.preprocess_addtional_kwargs(
atom_array, adj, wle_array=wle_array, is_real_node=is_real_node)
if self.scale_adj:
adj = rescale_adj(adj)
g_list = []
for step in range(self.n_update_layers):
update_layer_index = 0 if self.weight_tying else step
h = self.update_layers[update_layer_index](h=h,
adj=adj,
**additional_kwargs)
if self.use_batchnorm:
h = self.bnorms[update_layer_index](h)
if self.dropout_ratio > 0.:
h = functions.dropout(h, ratio=self.dropout_ratio)
if self.activation is not None and step < self.n_activation:
h = self.activation(h)
if self.concat_hidden or self.sum_hidden:
g = self.readout_layers[step](h=h,
h0=h0,
is_real_node=is_real_node,
**additional_kwargs)
g_list.append(g)
if self.concat_hidden:
return functions.concat(g_list, axis=1)
else:
if self.sum_hidden:
g = functions.sum(functions.stack(g_list), axis=0)
else:
g = self.readout_layers[0](h=h,
h0=h0,
is_real_node=is_real_node)
return g
def test_double_backward(self, src_backend_config, dst_backend_config):
x = src_backend_config.get_array(self.x)
gy = dst_backend_config.get_array(self.gy)
ggx = src_backend_config.get_array(self.ggx)
dst_device = dst_backend_config.device
x_var = chainer.Variable(x, requires_grad=True)
y_var = functions.copy(x_var, dst_device)
y_var.grad = gy
gy_var = y_var.grad_var
y_var.backward(enable_double_backprop=True)
assert x_var.grad_var.requires_grad is True
x_var.grad_var.grad = ggx
x_var.grad_var.backward()
assert gy_var.grad_var.device == dst_device
assert (backend.get_device_from_array(
gy_var.grad_var.array) == dst_device)
numpy.testing.assert_array_equal(
_numpy_device.send(gy_var.grad_var.array), self.ggx)
def test_double_backward(self, src_backend_config, dst_backend_config):
x = src_backend_config.get_array(self.x)
gy = dst_backend_config.get_array(self.gy)
ggx = src_backend_config.get_array(self.ggx)
dst_device = dst_backend_config.device
x_var = chainer.Variable(x, requires_grad=True)
y_var = functions.copy(x_var, dst_device)
# TODO(niboshi): Remove this workround after Variable.grad.setter is
# fixed so that it calls gy.require_grad() internally.
if dst_backend_config.xp is chainerx:
gy.require_grad()
y_var.grad = gy
gy_var = y_var.grad_var
y_var.backward(enable_double_backprop=True)
assert x_var.grad_var.requires_grad is True
x_var.grad_var.grad = ggx
x_var.grad_var.backward()
assert gy_var.grad_var.device == dst_device
assert (backend.get_device_from_array(
gy_var.grad_var.array) == dst_device)
numpy.testing.assert_array_equal(
_numpy_device.send(gy_var.grad_var.array), self.ggx)
def __call__(self, atom_array, adj):
"""Forward propagation
Args:
atom_array (numpy.ndarray): minibatch of molecular which is
represented with atom IDs (representing C, O, S, ...)
`atom_array[mol_index, atom_index]` represents `mol_index`-th
molecule's `atom_index`-th atomic number
adj (numpy.ndarray): minibatch of adjancency matrix with edge-type
information
Returns:
~chainer.Variable: minibatch of fingerprint
"""
# reset state
self.update_layer.reset_state()
if atom_array.dtype == self.xp.int32:
raise "use vector expression"
else:
h = atom_array
h0 = functions.copy(h, self.gpu_device)
g_list = []
for step in range(self.n_layers):
h = self.update(h, adj, step)
if self.concat_hidden:
g = self.readout(h, h0, step)
g_list.append(g)
if self.concat_hidden:
return functions.concat(g_list, axis=1)
else:
g = self.readout(h, h0, 0)
return g
def check_forward(self, src_id, dst_id):
x_data = _to_gpu(self.x_data, src_id)
x = chainer.Variable(x_data)
y = functions.copy(x, dst_id)
self.assertEqual(self.x_data.dtype, self.dtype)
numpy.testing.assert_array_equal(self.x_data, cuda.to_cpu(y.data))
def __call__(self, atom_array, adj):
# reset state
# self.update_layer.reset_state()
# [layer.reset_state() for layer in self.update_layer]
if atom_array.dtype == self.xp.int32:
h = self.embed(atom_array) # (minibatch, max_num_atoms)
else:
h = atom_array
h0 = functions.copy(h, cuda.get_device_from_array(h.data).id)
g_list = []
h_list = []
for step in range(self.n_layers):
h = self.update(h, adj, step)
if self.dropout_rate != 0.0:
h = functions.dropout(h, ratio=self.dropout_rate)
if self.concat_hidden:
g = self.readout(h, h0, step)
g_list.append(g)
if self.layer_aggr:
h_list.append(h)
if self.concat_hidden:
return functions.concat(g_list, axis=1)
elif self.layer_aggr:
output = self.aggr(h_list)
return self.readout(output, h0, 0)
else:
g = self.readout(h, h0, 0)
return g
def __call__(self, atom_array, adj, is_real_node=None):
"""Forward propagation
Args:
atom_array (numpy.ndarray): minibatch of molecular which is
represented with atom IDs (representing C, O, S, ...)
`atom_array[mol_index, atom_index]` represents `mol_index`-th
molecule's `atom_index`-th atomic number
adj (numpy.ndarray): minibatch of adjancency matrix with edge-type
information
is_real_node (numpy.ndarray): 2-dim array (minibatch, num_nodes).
1 for real node, 0 for virtual node.
If `None`, all node is considered as real node.
Returns:
~chainer.Variable: minibatch of fingerprint
"""
# reset state
self.reset_state()
if atom_array.dtype == self.xp.int32:
h = self.embed(atom_array) # (minibatch, max_num_atoms)
else:
h = atom_array
h0 = functions.copy(h, cuda.get_device_from_array(h.data).id)
g_list = []
for step in range(self.n_layers):
message_layer_index = 0 if self.weight_tying else step
h = self.update_layers[message_layer_index](h, adj)
if self.concat_hidden:
g = self.readout_layers[step](h, h0, is_real_node)
g_list.append(g)
if self.concat_hidden:
return functions.concat(g_list, axis=1)
else:
g = self.readout_layers[0](h, h0, is_real_node)
return g
def check_forward(self, src_id, dst_id):
x_data = _to_gpu(self.x_data, src_id)
x = chainer.Variable(x_data)
y = functions.copy(x, dst_id)
self.assertEqual(self.x_data.dtype, self.dtype)
numpy.testing.assert_array_equal(self.x_data, cuda.to_cpu(y.data))
def predict(self,
x: np.ndarray,
crf_pact_structure: CRFPackageStructure,
is_bin=False):
'''
:param xs:
:param crf_pact_structures:
:return: bin array for multi-label, shape= B x N x D
'''
with chainer.no_backprop_mode():
if not isinstance(x, chainer.Variable):
x = chainer.Variable(x)
xp = chainer.cuda.get_array_module(x)
# return shape = B * N * D , B is batch_size(=1 only), N is one video all nodes count, D is each node output vector
if self.with_crf: # 作废，这句if不会进去
xs = F.expand_dims(x, axis=0)
crf_pact_structures = [crf_pact_structure]
hs = self.structural_rnn(
xs, crf_pact_structures
) # hs shape = B x N x D, B is batch_size
hs = F.copy(hs, -1) # data transfer to cpu
h = hs.data[0]
pred_labels = self.open_crf.predict(
h, crf_pact_structures[0],
is_bin=is_bin) # shape = N x D or N x 1
return np.asarray(
pred_labels,
dtype=xp.int32) # shape =N x D, where D = AU_squeeze_size
else:
return self.structural_rnn.predict(
x, crf_pact_structure,
is_bin=is_bin) # 是binary形式的label. N x D
def __call__(self, atom_array, adj):
# reset state
self.reset_state()
if atom_array.dtype == self.xp.int32:
h = self.embed(atom_array)
else:
h = atom_array
if self.readout_func == 'ggnn':
h0 = functions.copy(h, cuda.get_device_from_array(h.data).id)
readout_layers = [
partial(readout_layer, h0=h0)
for readout_layer in self.readout_layers
]
else:
readout_layers = self.readout_layers
g_list = []
for step in range(self.n_layers):
message_layer_index = 0 if self.weight_tying else step
h = self.update_layers[message_layer_index](h, adj)
if self.concat_hidden:
g = readout_layers[step](h)
g_list.append(g)
if self.concat_hidden:
return functions.concat(g_list, axis=1)
else:
g = readout_layers[0](h)
return g
def __call__(self, atom_array, adj):
"""Forward propagation
Args:
atom_array (numpy.ndarray): minibatch of molecular which is
represented with atom IDs (representing C, O, S, ...)
`atom_array[mol_index, atom_index]` represents `mol_index`-th
molecule's `atom_index`-th atomic number
adj (numpy.ndarray): minibatch of adjancency matrix with edge-type
information
Returns:
~chainer.Variable: minibatch of fingerprint
"""
# reset state
self.update_layer.reset_state()
if atom_array.dtype == numpy.int32 \
or atom_array.dtype == cuda.cupy.int32:
h = self.embed(atom_array) # (minibatch, max_num_atoms)
else:
h = atom_array
h0 = functions.copy(h, cuda.get_device_from_array(h.data).id)
g_list = []
for step in range(self.n_layers):
h = self.update(h, adj, step)
if self.concat_hidden:
g = self.readout(h, h0, step)
g_list.append(g)
if self.concat_hidden:
return functions.concat(g_list, axis=2)
else:
g = self.readout(h, h0, 0)
return g
def __call__(self, sparse_batch, is_real_node=None):
if sparse_batch.x.dtype == self.xp.int32:
h = self.embed(sparse_batch.x) # (minibatch, max_num_atoms)
else:
h = self.first_mlp(sparse_batch.x)
h0 = functions.copy(h, cuda.get_device_from_array(h.data).id)
g_list = []
for step in range(self.n_message_layers):
message_layer_index = 0 if self.weight_tying else step
h = self.update_layers[message_layer_index](
h, sparse_batch.edge_index)
if step != self.n_message_layers - 1:
h = functions.relu(h)
if self.concat_hidden:
g = self.readout_layers[step](h, h0, is_real_node)
g_list.append(g)
if self.node_embedding:
return h
if self.concat_hidden:
return functions.concat(g_list, axis=1)
else:
g = self.readout_layers[0](h, sparse_batch.batch, h0, is_real_node)
return g
def __call__(self, atom_array, adj):
"""Forward propagation
Args:
atom_array (numpy.ndarray): minibatch of molecular which is
represented with atom IDs (representing C, O, S, ...)
`atom_array[mol_index, atom_index]` represents `mol_index`-th
molecule's `atom_index`-th atomic number
adj (numpy.ndarray): minibatch of adjancency matrix with edge-type
information
Returns:
~chainer.Variable: minibatch of fingerprint
"""
# reset state
if atom_array.dtype == self.xp.int32:
h = self.embed(atom_array) # (minibatch, max_num_atoms)
else:
h = atom_array
h0 = functions.copy(h, cuda.get_device_from_array(h.data).id)
g_list = []
for step in range(self.n_layers):
message_layer_index = 0 if self.weight_tying else step
h = self.update_layers[message_layer_index](h, adj)
if self.concat_hidden:
g = self.readout_layers[step](h, h0)
g_list.append(g)
if self.concat_hidden:
return functions.concat(g_list, axis=1)
else:
g = self.readout_layers[0](h, h0)
return g
def __call__(self, xs: chainer.Variable, crf_pact_structures
): # crf_pact_structure is batch of CRFPackageStructure
xp = chainer.cuda.cupy.get_array_module(xs.data) # xs is batch
# return shape = B * N * D , B is batch_size, N is one video all nodes count, D is each node output vector dimension
h = self.structural_rnn(xs, crf_pact_structures)
if self.with_crf:
# open_crf only support CPU mode
# convert_xs = self.bn(self.convert_dim_fc(xs.reshape(-1, xs.shape[-1]))) # note that we remove batch = 1 dimension
# h = F.relu(h.reshape(-1, h.shape[-1]) + convert_xs) # just like ResNet
# h = F.expand_dims(h, 0) # add one batch dimension
h = F.copy(h, -1)
# gt_label is hidden inside crf_pact_structure's sample. this step directly compute loss
loss = self.open_crf(h, crf_pact_structures)
else: # only structural_rnn
ts = self.get_gt_labels(
xp, crf_pact_structures,
is_bin=False) # B x N x 1, and B = 1 forever
ts = chainer.Variable(
ts.reshape(-1)
) # because ts label is 0~L which is one more than ground truth, 0 represent 0,0,0,0,0
h = h.reshape(
-1, h.shape[-1]
) # h must have 0~L which = L+1 including non_AU = 0(also background class)
assert ts.shape[0] == h.shape[0]
loss = F.hinge(h, ts, norm='L2', reduce='mean')
accuracy = F.accuracy(h, ts)
report_dict = {'loss': loss}
if not self.with_crf:
report_dict["accuracy"] = accuracy
chainer.reporter.report(report_dict, self)
return loss
def extract_images(data):
"""Extract image data from array or chainer.Variable"""
if isinstance(data, list):
data = [extract_images(d) for d in data]
if isinstance(data, chainer.Variable):
data = F.copy(data, -1)
data = data.array
if data.ndim > 5:
raise ValueError("invalid data: data.ndim > 5")
elif data.ndim == 5:
data = data[0] # NCHW
channels = data.shape[1]
if channels == 1: # mono
out_shape = list(data.shape)
out_shape[1] = 3
data = np.broadcast_to(data, out_shape)
data.flags.writeable = True
# clip element
data[data <= 0.0] = 0.0
data[data >= 1.0] = 1.0
return data
def __call__(self, atom_array, adj):
# reset state
self.atoms_list = []
self.g_vec_list = []
self.update_layer.reset_state()
if atom_array.dtype == self.xp.int32:
h = self.embed(atom_array) # (minibatch, max_num_atoms)
else:
h = atom_array
h0 = functions.copy(h, cuda.get_device_from_array(h.data).id)
g_list = []
h_list = []
for step in range(self.n_layers):
h = self.update(h, adj, step)
if self.dropout_rate != 0.0:
h = functions.dropout(h, ratio=self.dropout_rate)
if self.concat_hidden:
g = self.readout(h, h0, step)
g_list.append(g)
h_list.append(h)
self.atoms_list.append(h)
g_vec = self.readout(h, h0, step)
self.g_vec_list.append(g_vec)
if self.concat_hidden:
return functions.concat(g_list, axis=1)
else:
g = self.readout(h, h0, 0)
# g = self.att_readout(h, h_list, 0)
g = functions.sum(h, axis=1)
return g
def __call__(self, atom_array, adj, super_node=None, is_real_node=None):
self.reset_state()
if atom_array.dtype == self.xp.int32:
h = self.embed(atom_array)
else:
# TODO: GraphLinear or GraphMLP can be used.
h = atom_array
h0 = functions.copy(h, cuda.get_device_from_array(h.data).id)
if self.with_gwm:
h_s = self.embed_super(super_node)
additional_kwargs = self.preprocess_addtional_kwargs(
atom_array, adj, super_node=super_node, is_real_node=is_real_node)
if self.scale_adj:
adj = rescale_adj(adj)
g_list = []
for step in range(self.n_update_layers):
update_layer_index = 0 if self.weight_tying else step
h_new = self.update_layers[update_layer_index](h=h,
adj=adj,
**additional_kwargs)
if self.with_gwm:
h_new, h_s = self.gwm(h, h_new, h_s, update_layer_index)
h = h_new
if self.use_batchnorm:
h = self.bnorms[update_layer_index](h)
if self.dropout_ratio > 0.:
h = functions.dropout(h, ratio=self.dropout_ratio)
if self.activation is not None and step < self.n_activation:
h = self.activation(h)
if self.concat_hidden or self.sum_hidden:
g = self.readout_layers[step](h=h,
h0=h0,
is_real_node=is_real_node,
**additional_kwargs)
g_list.append(g)
if self.concat_hidden:
return functions.concat(g_list, axis=1)
else:
if self.sum_hidden:
g = functions.sum(functions.stack(g_list), axis=0)
else:
g = self.readout_layers[0](h=h,
h0=h0,
is_real_node=is_real_node)
if self.with_gwm:
g = functions.concat((g, h_s), axis=1)
g = functions.relu(self.linear_for_concat_super(g))
return g
def getDecoderInputEmbeddings(self,
xs): # xs: [arr(l1), ...], still on cpu
x_len = [len(x) for x in xs]
x_section = np.cumsum(x_len[:-1])
vxs = [F.copy(chainer.Variable(x), self.gpu) for x in xs] # to gpu
ex = self.model.decoderEmbed(F.concat(tuple(vxs), axis=0))
exs = F.split_axis(ex, x_section, 0)
return list(exs)
def __call__(self, x):
# assume x is on gpu0
x1 = F.copy(x, self.gpu1)
z0 = self.first0(x)
z1 = self.first1(x1)
# synchronize
h0 = z0 + F.copy(z1, self.gpu0)
h1 = z1 + F.copy(z0, self.gpu1)
y0 = self.second0(F.relu(h0))
y1 = self.second1(F.relu(h1))
# synchronize
y = y0 + F.copy(y1, self.gpu0)
return y # output is on gpu0
def __call__(self, x):
# assume x is on gpu0
x1 = F.copy(x, self.gpu1)
z0 = self.first0(x)
z1 = self.first1(x1)
# synchronize
h0 = z0 + F.copy(z1, self.gpu0)
h1 = z1 + F.copy(z0, self.gpu1)
y0 = self.second0(F.relu(h0))
y1 = self.second1(F.relu(h1))
# synchronize
y = y0 + F.copy(y1, self.gpu0)
return y # output is on gpu0
def __call__(self, x):
# assume x is on GPU 0
x1 = F.copy(x, 1)
z0 = self.first0(x)
z1 = self.first1(x1)
# sync
h0 = z0 + F.copy(z1, 0)
h1 = z1 + F.copy(z0, 1)
y0 = self.second0(F.relu(h0))
y1 = self.second1(F.relu(h1))
# sync
y = y0 + F.copy(y1, 0)
return y
def __call__(self, x):
# assume x is on GPU 0
x1 = F.copy(x, 1)
z0 = self.first0(x)
z1 = self.first1(x1)
# sync
h0 = z0 + F.copy(z1, 0)
h1 = z1 + F.copy(z0, 1)
y0 = self.second0(F.relu(h0))
y1 = self.second1(F.relu(h1))
# sync
y = y0 + F.copy(y1, 0)
return y
def recurse_copy(object, device):
if isinstance(object, tuple) or isinstance(object, list):
object = [recurse_copy(sub_object, device) for sub_object in object]
elif isinstance(object, chainer.Variable) or isinstance(
object, cuda.cupy.ndarray) or isinstance(object, numpy.ndarray):
object = F.copy(object, device)
return object
def copy():
x = rand((1, 2, 3, 4))
y = F.copy(x, -1) - 0
# A dummy layer `-0` is inserted because our implementation of copy plugin
# reuses input tensor as the layer output.
# (a tensor cannot be both input and output of a network)
# when using copy inside a network with multiple layers,
# no need to insert any dummy operations.
return {'input': x}, {'out': y}
def forward(self, x):
# FIXME: Adding with constant and F.resize_images() using 0th device
# Forward FCN
score = super().forward(x)
# Convert score to probability
prob = F.softmax(score, axis=1)
# Increase gradient of the probability
prob = prob - 0.5
prob = self.prob_scale(prob)
prob = F.clip(prob, -0.5, 0.5)
prob = prob + 0.5
# Down sampling
h, w = x.shape[2:4]
down_shape = (h // self.mat_scale, w // self.mat_scale)
prob = F.resize_images(prob, down_shape)
x = F.resize_images(x, down_shape)
# Split into foreground, background and unknown sores
prob_b, _, prob_f = F.split_axis(prob, 3, axis=1) # (n, 1, h, w)
# Copy to CPU
x = F.copy(x, -1)
prob_b = F.copy(prob_b, -1)
prob_f = F.copy(prob_f, -1)
# Compute laplacian
laplacian = _compute_laplacians(x, prob_b, prob_f)
# Matting
alpha = self.matting_link(prob_b, prob_f, laplacian) # (n, 1, h, w)
# Up sampling
alpha = F.resize_images(alpha, (h, w))
# Remove extra channel (n, 1, h, w) -> (n, h, w)
alpha_shape = (alpha.shape[0], alpha.shape[2], alpha.shape[3])
alpha = F.reshape(alpha, alpha_shape)
self.alpha = alpha
return alpha
def check_backward(self, src_id, dst_id):
x_data = _to_gpu(self.x_data, src_id)
x = chainer.Variable(x_data)
y = functions.copy(x, dst_id)
gy = _to_gpu(self.gy, dst_id)
y.grad = gy
y.backward()
x_grad = x.grad
numpy.testing.assert_array_equal(
cuda.to_cpu(x_grad), self.gy)
def __call__(self, atom_array, adj, super_node, is_real_node=None):
"""
Describe a layer
Args:
atom_array (numpy.ndarray): mol-minibatch by node numpy.ndarray,
minibatch of molecular which is represented with atom IDs (representing C, O, S, ...)
atom_array[m, i] = a represents
m-th molecule's i-th node is value a (atomic number)
adj (numpy.ndarray): mol-minibatch by relation-types by node by node numpy.ndarray,
minibatch of multiple relational adjancency matrix with edge-type information
adj[i, j] = b represents
m-th molecule's edge from node i to node j has value b
super_node (numpy.ndarray): 1D array, the supernode hidden state
is_real_node:
Returns:
numpy.ndarray: final molecule representation
"""
if atom_array.dtype == self.xp.int32:
h = self.embed(atom_array) # (minibatch, max_num_atoms)
else:
h = atom_array
# end if-else
h0 = functions.copy(h, cuda.get_device_from_array(h.data).id)
self.gwm.GRU_local.reset_state()
self.gwm.GRU_super.reset_state()
# ebmbed super node
h_s = self.embed_super(super_node)
g_list = []
for step in range(self.n_message_layers):
message_layer_index = 0 if self.weight_tying else step
h2 = self.update_layers[message_layer_index](h, adj)
h, h_s = self.gwm(h, h2, h_s, message_layer_index)
if self.concat_hidden:
g = self.readout_layers[step](h, h0, is_real_node)
g_list.append(g)
if self.concat_hidden:
return functions.concat(g_list, axis=1)
else:
g = self.readout_layers[0](h, h0, is_real_node)
g2 = functions.concat( (g, h_s), axis=1 )
out_g = functions.relu(self.linear_for_concat_super(g2))
return out_g
def __call__(self, atom_array, adj, is_real_node=None):
"""
Describe the whole forwar path
Args:
atom_array (numpy.ndarray): mol-minibatch by node numpy.ndarray,
minibatch of molecular which is represented with atom IDs
(representing C, O, S, ...) atom_array[m, i] = a represents
m-th molecule's i-th node is value a (atomic number)
adj (numpy.ndarray): mol-minibatch by relation-types by node by
node numpy.ndarray,
minibatch of multple relational adjancency matrix with
edge-type information adj[i, j] = b represents
m-th molecule's edge from node i to node j has value b
is_real_node:
Returns:
numpy.ndarray: final molecule representation
"""
if atom_array.dtype == self.xp.int32:
h = self.embed(atom_array) # (minibatch, max_num_atoms)
else:
h = atom_array
h0 = functions.copy(h, cuda.get_device_from_array(h.data).id)
g_list = []
for step in range(self.n_message_layers):
message_layer_index = 0 if self.weight_tying else step
h = self.update_layers[message_layer_index](h, adj)
if self.concat_hidden:
g = self.readout_layers[step](h, h0, is_real_node)
g_list.append(g)
if self.concat_hidden:
return functions.concat(g_list, axis=1)
else:
g = self.readout_layers[0](h, h0, is_real_node)
return g
def test_call_forward_with_device(self):
functions.copy(self.x_data, cuda.DummyDevice)
def f(x):
return functions.copy(x, -1)
def test_check_backward_cpu(self):
x = chainer.Variable(self.x_data)
y = functions.copy(x, -1)
y.grad = self.gy
y.backward()
gradient_check.assert_allclose(x.grad, self.gy, atol=0, rtol=0)
def f(x):
y = functions.copy(x, -1)
return y * y
def test_check_forward_cpu(self):
x = chainer.Variable(self.x_data)
y = functions.copy(x, -1)
gradient_check.assert_allclose(self.x_data, y.data, atol=0, rtol=0)
