def symeig_svd(self, matrix, n_eigenvecs=None):
"""Computes a truncated SVD on `matrix` using symeig
Uses symeig on matrix.T.dot(matrix) or its transpose
Parameters
----------
matrix : 2D-array
n_eigenvecs : int, optional, default is None
if specified, number of eigen[vectors-values] to return
Returns
-------
U : 2D-array
of shape (matrix.shape[0], n_eigenvecs)
contains the right singular vectors
S : 1D-array
of shape (n_eigenvecs, )
contains the singular values of `matrix`
V : 2D-array
of shape (n_eigenvecs, matrix.shape[1])
contains the left singular vectors
"""
# Check that matrix is... a matrix!
if self.ndim(matrix) != 2:
raise ValueError('matrix be a matrix. matrix.ndim is %d != 2' %
self.ndim(matrix))
dim_1, dim_2 = self.shape(matrix)
if dim_1 <= dim_2:
min_dim = dim_1
max_dim = dim_2
else:
min_dim = dim_2
max_dim = dim_1
if n_eigenvecs is None:
n_eigenvecs = max_dim
if min_dim <= n_eigenvecs:
if n_eigenvecs > max_dim:
warnings.warn(
'Trying to compute SVD with n_eigenvecs={0}, which '
'is larger than max(matrix.shape)={1}. Setting '
'n_eigenvecs to {1}'.format(n_eigenvecs, max_dim))
n_eigenvecs = max_dim
# we compute decomposition on the largest of the two to keep more eigenvecs
dim_1, dim_2 = dim_2, dim_1
if dim_1 < dim_2:
U, S = nd.linalg.syevd(dot(matrix, transpose(matrix)))
S = self.sqrt(S)
V = dot(transpose(matrix), U / reshape(S, (1, -1)))
else:
V, S = nd.linalg.syevd(dot(transpose(matrix), matrix))
S = self.sqrt(S)
U = dot(matrix, V) / reshape(S, (1, -1))
U, S, V = U[:, ::-1], S[::-1], transpose(V)[::-1, :]
return U[:, :n_eigenvecs], S[:n_eigenvecs], V[:n_eigenvecs, :]
def forward(self, data, neighbor_data, neighbor_indices, neighbor_indptr,
node_type_mask=None, neighbor_type_mask=None, edge_type_mask=None, seg_indices=None):
"""Map the input features to hidden states + apply pooling + apply FC
Parameters
----------
F
data : Symbol or NDArray
Shape (batch_size, node_num, feat_dim)
neighbor_data : Symbol or NDArray
Shape (batch_size, neighbor_node_num, feat_dim)
data_mask : Symbol or NDArray
Shape (batch_size, node_num, num_set, 1)
neighbor_mask : Symbol or NDArray
Shape (batch_size, neighbor_node_num, num_set, 1)
neighbor_indices : Symbol or NDArray
Shape (nnz, )
neighbor_indptr : Symbol or NDArray
Shape (node_num + 1, )
edge_data : Symbol or NDArray or None
Shape (batch_size, nnz, num_edge_num, 1)
Returns
-------
"""
## TODO does not consider node type
if self._num_node_set is not None:
#print("data", data.shape)
#print("node_type_mask", node_type_mask.shape)
data = self.data_map(data, node_type_mask)
neighbor_data = self.neighbor_mid_map(neighbor_data, neighbor_type_mask)
if self._num_edge_set is not None:
neighbor_data = self.relation_W(neighbor_data) ### (batch_size, neighbor_node_num, mid_units*num_edge_set)
neighbor_data = nd.take(neighbor_data, indices=neighbor_indices, axis=-2) ## (batch_size, nnz, mid_units*num_edge_set)
#print("neighbor_data", neighbor_data.shape)
neighbor_data = nd.reshape(neighbor_data,
shape=(0, 0, self._num_edge_set, self._mid_units)) ## (batch_size, nnz, mid_units*num_edge_set)
#print("neighbor_data", neighbor_data.shape)
#print("edge_data", edge_data.shape)
neighbor_data = nd.reshape(nd.broadcast_mul(neighbor_data, edge_type_mask),
shape=(0, 0, -1))
#print("neighbor_data", neighbor_data.shape)
pool_data = nd.contrib.seg_pool(data=neighbor_data,
indices=seg_indices,
indptr=neighbor_indptr,
pool_type=self._pool_type) # Shape(batch_size, node_num, mid_units*num_edge_set)
if self._num_edge_set is not None:
if self._accum_type == "stack":
pool_data = self._out_act(pool_data)
elif self._accum_type == "sum":
pool_data = self._out_act(nd.sum(nd.reshape(pool_data, shape=(0, 0, self._num_edge_set, self._mid_units )), axis=2))
#out = self.out_layer(nd.concat(pool_data, data, dim=-1))
#out = self.out_layer(pool_data)
return pool_data
def _rearrange(raw, F, upscale_factor):
# (N, C * r^2, H, W) -> (N, C, r^2, H, W)
splitted = F.reshape(raw, shape=(0, -4, -1, upscale_factor**2, 0, 0))
# (N, C, r^2, H, W) -> (N, C, r, r, H, W)
unflatten = F.reshape(splitted, shape=(0, 0, -4, upscale_factor, upscale_factor, 0, 0))
# (N, C, r, r, H, W) -> (N, C, H, r, W, r)
swapped = F.transpose(unflatten, axes=(0, 1, 4, 2, 5, 3))
# (N, C, H, r, W, r) -> (N, C, H*r, W*r)
return F.reshape(swapped, shape=(0, 0, -3, -3))
def _rearrange(raw, F, upscale_factor):
# (N, C * r^2, H, W) -> (N, C, r^2, H, W)
splitted = F.reshape(raw, shape=(0, -4, -1, upscale_factor**2, 0, 0))
# (N, C, r^2, H, W) -> (N, C, r, r, H, W)
unflatten = F.reshape(splitted, shape=(0, 0, -4, upscale_factor, upscale_factor, 0, 0))
# (N, C, r, r, H, W) -> (N, C, H, r, W, r)
swapped = F.transpose(unflatten, axes=(0, 1, 4, 2, 5, 3))
# (N, C, H, r, W, r) -> (N, C, H*r, W*r)
return F.reshape(swapped, shape=(0, 0, -3, -3))
def generate(self, x: nd.NDArray = None, include_intermediate: bool = False, **kwargs) -> \
Union[nd.NDArray, Tuple[nd.NDArray, nd.NDArray]]:
"""
Generate a batch of samples from model. See Section 2.3 in paper.
If x is None, this method generates unconditional samples from the model (as explained in Section 2.3 in the
paper).
If x is provided, this method reconstructs the input to generate the sample. This is not really a true sample
from the model because the model looks at the image it is trying to generate. However, this is useful for seeing
how the model generates a particular image.
:param x: Input to generate images from.
:param include_intermediate: If True, samples from all timesteps (not only the last timestep) are returned.
:return: n x *image_shape array of generated samples. If include_intermediate is True,
then steps x n x *image_shape.
"""
r = nd.zeros((self._batch_size, *self._input_shape), ctx=self._ctx) # reconstruction
h_dec = nd.zeros((self._batch_size, *self._rnn_hidden_shape), ctx=self._ctx)
c_dec = nd.zeros((self._batch_size, *self._rnn_hidden_shape), ctx=self._ctx)
if x is not None:
h_enc = nd.zeros((self._batch_size, *self._rnn_hidden_shape), ctx=self._ctx)
c_enc = nd.zeros((self._batch_size, *self._rnn_hidden_shape), ctx=self._ctx)
encoded_x = self._enc_nn(x)
rs = [] # sample(s) over time
for i in range(self._num_steps):
rs.append(nd.sigmoid(r))
encoded_r = self._enc_nn(rs[-1])
if x is not None:
err = encoded_x - encoded_r
_, (h_enc, c_enc) = self._enc_rnn(nd.concat(encoded_x, err, h_dec, c_dec, dim=1), [h_enc, c_enc])
q = self._q_layer(h_enc)
# convert NxCxHxW to NxF
q = nd.reshape(q, (self._batch_size, -1))
z = self._latent_layer(q)
else:
# sample from prior
p = self._p_layer(h_dec)
p = nd.reshape(p, (self._batch_size, -1))
z = self._latent_layer(p)
dec_z = nd.reshape(z, (self._batch_size, self._num_latent_maps, *self._encoder_output_shape[1:]))
_, (h_dec, c_dec) = self._dec_rnn(nd.concat(dec_z, encoded_r, dim=1), [h_dec, c_dec])
r = r + self._dec_nn(h_dec)
rs.append(nd.sigmoid(r))
if include_intermediate:
samples = nd.stack(*rs, axis=0)
else:
samples = rs[-1]
return samples
def generate(self,
v_q: nd.NDArray,
x_context: nd.NDArray,
v_context: nd.NDArray,
include_intermediate: bool = False,
**kwargs) -> Union[nd.NDArray, Tuple[nd.NDArray, nd.NDArray]]:
"""
Generate a batch of samples from model. See Algorithm S3 in paper.
:param v_q: Query view camera info.
:param x_context: Context frames.
:param v_context: Context camera info.
:param include_intermediate: If True, samples from all timesteps (not only the last timestep) are returned.
:return: n x *image_shape array of generated samples. If include_intermediate is True,
then steps x n x *image_shape.
"""
u = nd.zeros((self._batch_size, *self._upsample_output_shape),
ctx=self._ctx) # canvas (reconstruction)
h_dec = nd.zeros((self._batch_size, *self._rnn_hidden_shape),
ctx=self._ctx)
c_dec = nd.zeros((self._batch_size, *self._rnn_hidden_shape),
ctx=self._ctx)
# reshape camera information so we can concat it to image data
v_q = nd.broadcast_to(
nd.expand_dims(nd.expand_dims(v_q, axis=-1), axis=-1),
(0, 0, *self._downsample_output_shape[1:]))
outs = [] # sample(s) over time
r = self._representation_nn(x_context, v_context)
for i in range(self._num_steps):
outs.append(self._out_layer(u))
# Eq. S11
p = self._p_layer(h_dec)
p = nd.reshape(p, (self._batch_size, -1))
z = self._latent_layer(p)
gen_z = nd.reshape(z, (self._batch_size, self._num_latent_maps,
*self._downsample_output_shape[1:]))
_, (h_dec, c_dec) = self._gen_rnn(nd.concat(gen_z, v_q, r, dim=1),
[h_dec, c_dec])
u = u + self._upsample_nn(h_dec)
outs.append(self._out_layer(u))
if include_intermediate:
samples = nd.stack(*outs, axis=0)
else:
samples = outs[-1]
return nd.clip(samples, a_min=0.0, a_max=1.0)
def forward(self, x):
batch_size, c, h, w = x.shape
if batch_size is None:
batch_size = -1
out_c = c // (self.up_scale * self.up_scale)
out_h = h * self.up_scale
out_w = w * self.up_scale
out = F.reshape(
x, (batch_size, self.up_scale, self.up_scale, out_c, h, w))
out = F.transpose(out, (0, 3, 4, 1, 5, 2))
out = F.reshape(out, (batch_size, out_c, out_h, out_w))
return out
def var(array,W=_W,B=None,square=0,sqrt=0,V=False,order='NCHW',sizz=0):
arrs=array.shape
ashp=W.shape
xi=(-2,-1)
x2=(-2,-1,-3)
sb=(ashp[1],1,1)
WV=ashp[-2:]
print(sb)
mnc=mnd.tile(mnd.reshape(mnd.array([WV[0]*WV[1]]), shape=(1,1,1)),ashp[1])
print(mnc)
if V:
print(W.eval())
print(arrs,ashp)
mul=(mnd.broadcast_mul(array,W))
if V:
print('Wsamp',W[-1,-1])
print('array*w',mul[0,-1])
size=mnd.sum(W,axis=xi,keepdims=True)#shape=(outputs, channel)
if V:
print("sizesamp",size.shape,size)
if B is None:
B=mnd.zeros(W.shape[0:2],dtype=np.float32)#channel
B=mnd.reshape(B,(*B.shape,*[1 for _ in range(len(ashp)-len(B.shape))]))
if sizz==1:
mean=mnd.sum(mul,axis=xi,keepdims=True)/size
else:
mean=mnd.sum(mul,axis=xi,keepdims=True)/mnc
if V:
print("meansamp",mean[0,-1])
if square:
i=mnd.square(mnd.broadcast_add(mnd.broadcast_minus(mul,mean),B))
else:
i=mnd.broadcast_add(mnd.broadcast_minus(mul,mean),B)
di=i/size
if V==2:
print("i",i,"i")
print("di",di,"di")
if V:
print('isamp',i.shape,i[-1,-1,])
out=mnd.sum(mnd.broadcast_add(i,B),axis=x2)
#out=np.rollaxis(np.sum(i+B,axis=x2),-1,1)
#print(out.shape)
if sqrt:
out=mnd.sqrt(out)
out=mnd.swapaxes(out, 3, 1)
#print(out.shape,(arrs[0],ashp[0],arrs[1],arrs[2]))
assert out.shape==(arrs[0],ashp[0],arrs[1],arrs[2])
return(out)
def hybrid_forward(self, F, x, *args, **kwargs):
# batch_size, channels, height, width = x.shape
# assert channels % 2 == 0
# mid_channels = channels // 2
data = F.reshape(x, shape=(0, -4, self.groups, -1, -2))
data = F.swapaxes(data, 1, 2)
data = F.reshape(data, shape=(0, -3, -2))
data_project = F.slice(data,
begin=(None, None, None, None),
end=(None, self.mid_channel, None, None))
data_x = F.slice(data,
begin=(None, self.mid_channel, None, None),
end=(None, None, None, None))
return data_project, data_x
def rgb_to_lab(image_srgb, ctx=None):
if ctx is None:
raise ValueError("ctx can not be None")
if image_srgb is None:
raise ValueError("image_srgb can not be None")
with mx.Context(ctx):
srgb = __check_image(image_srgb)
if nd.max(srgb).asscalar() > 1:
srgb = __normalize_rgb_image(srgb)
srgb_pixels = nd.reshape(srgb, [-1, 3])
linear_mask = nd.cast(srgb_pixels <= 0.04045, dtype='float32')
exponential_mask = nd.cast(srgb_pixels > 0.04045, dtype='float32')
rgb_pixels = (srgb_pixels / 12.92 * linear_mask) + (((srgb_pixels + 0.055) / 1.055) ** 2.4) * exponential_mask
rgb_to_xyz = nd.array([
# X Y Z
[0.412453, 0.212671, 0.019334], # R
[0.357580, 0.715160, 0.119193], # G
[0.180423, 0.072169, 0.950227], # B
])
xyz_pixels = nd.linalg_gemm2(rgb_pixels, rgb_to_xyz)
# https://en.wikipedia.org/wiki/Lab_color_space#CIELAB-CIEXYZ_conversions
# convert to fx = f(X/Xn), fy = f(Y/Yn), fz = f(Z/Zn)
# normalize for D65 white point
xyz_normalized_pixels = nd.multiply(xyz_pixels, nd.array([1 / 0.950456, 1.0, 1 / 1.088754]))
epsilon = 6 / 29
linear_mask = nd.cast(xyz_normalized_pixels <= (epsilon ** 3), dtype='float32')
exponential_mask = nd.cast(xyz_normalized_pixels > (epsilon ** 3), dtype='float32')
fxfyfz_pixels = (xyz_normalized_pixels / (3 * epsilon ** 2) + 4 / 29) * linear_mask + (
xyz_normalized_pixels ** (
1 / 3)) * exponential_mask
# convert to lab
fxfyfz_to_lab = nd.array([
# l a b
[0.0, 500.0, 0.0], # fx
[116.0, -500.0, 200.0], # fy
[0.0, 0.0, -200.0], # fz
])
lab_pixels = nd.linalg_gemm2(fxfyfz_pixels, fxfyfz_to_lab) + nd.array([-16.0, 0.0, 0.0])
return nd.reshape(lab_pixels, srgb.shape)
def verify_loaded_model(net):
"""Run inference using ten random images.
Print both input and output of the model"""
def transform(data, label):
return data.astype(np.float32)/255, label.astype(np.float32)
# Load ten random images from the test dataset
sample_data = mx.gluon.data.DataLoader(mx.gluon.data.vision.MNIST(train=False, transform=transform),
10, shuffle=True)
for data, label in sample_data:
# Display the images
img = nd.transpose(data, (1,0,2,3))
img = nd.reshape(img, (28,10*28,1))
imtiles = nd.tile(img, (1,1,3))
plt.imshow(imtiles.asnumpy())
plt.show()
# Display the predictions
data = nd.transpose(data, (0, 3, 1, 2))
out = net(data.as_in_context(ctx))
predictions = nd.argmax(out, axis=1)
print('Model predictions: ', predictions.asnumpy())
break
def merge(conv_w, gamma, beta, running_mean, running_var):
gamma_over_var = gamma / nd.sqrt(running_var + 1e-5)
gamma_over_var_expanded = nd.reshape(gamma_over_var,
(gamma_over_var.shape[0], 1, 1, 1))
new_w = gamma_over_var_expanded * nd.cast(conv_w, 'float32')
new_b = beta - running_mean * gamma_over_var
return new_w, new_b
def forward(self, x, mask):
"""
Parameters
----------
F
x: Shape(batch_size, num_node, input_dim)
mask: Shape(batch_size, num_node, num_set, 1)
Returns
-------
"""
layer_in_l = [x]
layer_out = None
for i in range(self._layer_num):
if len(layer_in_l) == 1:
layer_in = layer_in_l[0]
else:
layer_in = nd.concat(*layer_in_l, dim=-1)
### TODO assume batch_size=1
x_mW = nd.reshape(self.layers[i](layer_in),
shape=(0, 0, self._num_set, self._units))
layer_out = self._act(nd.sum(nd.broadcast_mul(x_mW, mask),
axis=-2))
layer_in_l.append(layer_out)
return layer_out
def proto_loss(embedding, nc, ns, nq):
embedding = embedding.astype('float64');
cls_data = nd.reshape(embedding[0:nc*ns], (nc, ns, -1)); cls_data.attach_grad()
cls_center = nd.mean(cls_data, axis=1);
data_center_dis = nd.norm(embedding[nc*ns:].expand_dims(axis=1) - cls_center.expand_dims(axis=0),
axis=2) ** 2
# print(nd.max(data_center_dis).asscalar())
weight = nd.zeros((nc*nq, nc), ctx=embedding.context, dtype='float64')
pick_vec = nd.zeros((nc*nq), ctx=embedding.context)
for i in range(0, nc):
weight[i*nq:i*nq+nq, i] = 1
pick_vec[i*nq:i*nq+nq] = i
"""
temp = nd.SoftmaxOutput(-data_center_dis, label)
temp = nd.log(temp) * weight
temp = nd.sum(-temp, axis=1)
predict = nd.argmin(data_center_dis, axis=1)
return -temp * nd.log(temp), predict
"""
temp1 = nd.log_softmax(-data_center_dis, axis=1);
temp2 = nd.pick(temp1, index=pick_vec, axis=1);
temp3 = nd.sum(-temp2);
label = nd.argmin(data_center_dis, axis=1)
return temp3 / (nc * nq), label
def my_loss(data, nc, ns, nq):
data = data.astype('float64')
cls_data = nd.reshape(data[0:nc * ns], (nc, ns, -1))
cls_center = nd.mean(cls_data, axis=1) + 1e-10
data_center_dis = nd.norm(data[nc * ns:].expand_dims(axis=1) -
cls_center.expand_dims(axis=0),
axis=2)**2
weight = nd.zeros((nc * nq, nc), ctx=data.context, dtype='float64')
for i in range(0, nc):
weight[i * nq:i * nq + nq, i] = 1
weight2 = 1 - weight
temp1 = nd.log_softmax(-data_center_dis, axis=1)
temp2 = nd.sum(temp1, axis=1)
temp3 = nd.sum(-temp2)
label = nd.argmin(data_center_dis, axis=1)
return temp3 / (nc * nq), label
loss1 = nd.sum(data_center_dis * weight)
temp = nd.sum(nd.exp(-data_center_dis), axis=1)
loss2 = nd.sum(nd.log(temp))
if loss1 is np.nan or loss2 is np.nan:
raise StopIteration
return (loss1 + loss2) / (nc * nq), label
def __check_image(image):
assert image.shape[-1] == 3
if len(image.shape) not in (3, 4):
raise ValueError("image must be either 3 or 4 dimensions")
shape = list(image.shape)
shape[-1] = 3
return nd.reshape(image, shape)
def test_selective_attention_write():
# filter 3x3
# image 4x5
# batch size 2
ctx = mx.cpu()
attn_block = SelectiveAttentionWrite(filter_size=3,
input_shape=(4, 5),
batch_size=2)
attn_block.collect_params().initialize(ctx=ctx)
h_dec = nd.random.normal(shape=(2, 3), ctx=ctx)
c_dec = nd.random.normal(shape=(2, 3), ctx=ctx)
attn_params = attn_block._attention_params_layer(
nd.concat(h_dec, c_dec, dim=1))
Fx, Fy = attn_block._build_filter(nd, attn_params)
write, _ = attn_block(h_dec, c_dec)
# calculate expected
w = nd.reshape(attn_block._patch_layer(nd.concat(h_dec, c_dec, dim=1)),
(-1, 3, 3))
expected_1 = (
np.dot(np.dot(Fx[0].asnumpy().T, w[0].asnumpy()), Fy[0].asnumpy()) /
np.exp(attn_params[0, 4].asscalar()))
expected_2 = (
np.dot(np.dot(Fx[1].asnumpy().T, w[1].asnumpy()), Fy[1].asnumpy()) /
np.exp(attn_params[1, 4].asscalar()))
expected = np.stack([expected_1.flatten(), expected_2.flatten()], axis=0)
assert np.allclose(expected, write.asnumpy())
def get(self, pred, label):
embedding = nd.L2Normalization(pred, mode='instance')
self.acc = 0
nc = self.nc
ns = self.ns
nq = self.nq
margin = self.margin
s_embedding = embedding.slice_axis(axis=0, begin=0, end=nc * ns)
q_embedding = embedding.slice_axis(axis=0, begin=nc * ns, end=None)
s_cls_data = nd.reshape(s_embedding, (nc, ns, -1))
q_cls_data = nd.reshape(q_embedding, (nc, nq, -1))
s_cls_center = nd.mean(s_cls_data, axis=1)
s_cls_center = nd.L2Normalization(s_cls_center, mode='instance')
temp = q_embedding.expand_dims(axis=1) * s_cls_center.expand_dims(
axis=0)
data_center_dis = nd.sum(temp, axis=2)
cur_label = nd.argmax(data_center_dis, axis=1)
loss = 0
# Calculating loss
for i in range(nc):
temp = data_center_dis[i * nq:(i + 1) * nq, i]
loss += nd.sum(
nd.LeakyReLU(margin - temp, act_type='leaky', slope=0.1))
for i in range(nc):
self.acc += nd.sum(cur_label[nq * i:nq * (i + 1)] == i).asscalar()
self.acc /= (nc * nq)
s_embedding = embedding.slice_axis(axis=0, begin=0, end=nc * ns)
q_embedding = embedding.slice_axis(axis=0, begin=nc * ns, end=None)
s_cls_data = nd.reshape(s_embedding, (nc, ns, -1))
q_cls_data = nd.reshape(q_embedding, (nc, nq, -1))
s_cls_center = nd.mean(s_cls_data, axis=1)
s_cls_center = nd.L2Normalization(s_cls_center, mode='instance')
s_center_broadcast = s_cls_center.expand_dims(axis=1)
s_center_dis = nd.sum(nd.broadcast_mul(q_cls_data, s_center_broadcast),
axis=2)
temp = nd.LeakyReLU(margin - s_center_dis, act_type='leaky', slope=0.1)
loss1 = nd.sum(temp)
return (self.acc, cur_label, loss)
def hybrid_forward(self, F, x, block_channel_mask, *args, **kwargs):
block_channel_mask = F.slice(block_channel_mask,
begin=(None, None),
end=(None, self.channel_number))
block_channel_mask = F.reshape(block_channel_mask,
shape=(1, self.channel_number, 1, 1))
x = F.broadcast_mul(x, block_channel_mask)
return x
def forward(self, signals) -> NDArray:
outputs = []
for signal, RFmapper in zip(signals, self._RFmappers):
for i in range(signal.shape[1]):
outputs.append(RFmapper(signal[:,i,:]))
outputs = stack(*outputs, axis=1)
outputs = reshape(outputs, (outputs.shape[0], outputs.shape[1], *self.output_shape))
return outputs
def calculate_norm(x, y):
assert x.shape == y.shape
ndims = np.product(x.shape)
x = nd.reshape(x, shape=(ndims, ))
y = nd.reshape(y, shape=(ndims, ))
res = x - y
nx = nd.norm(x)
ny = nd.norm(y)
nr = nd.norm(res)
print("saving...")
f = "/home/ryt/data/cmp_"
names = ["nx", "ny", "nr"]
objs = [nx, ny, nr]
for obj in objs:
print(type(obj), obj.shape)
for i in range(3):
nd.save(f + names[i], objs[i])
print('success')
def kr(matrices):
"""Khatri-Rao product of a list of matrices
This can be seen as a column-wise kronecker product.
Parameters
----------
matrices : ndarray list
list of matrices with the same number of columns, i.e.::
for i in len(matrices):
matrices[i].shape = (n_i, m)
Returns
-------
khatri_rao_product: matrix of shape ``(prod(n_i), m)``
where ``prod(n_i) = prod([m.shape[0] for m in matrices])``
i.e. the product of the number of rows of all the matrices in the product.
Notes
-----
Mathematically:
.. math::
\\text{If every matrix } U_k \\text{ is of size } (I_k \\times R),\\\\
\\text{Then } \\left(U_1 \\bigodot \\cdots \\bigodot U_n \\right) \\text{ is of size } (\\prod_{k=1}^n I_k \\times R)
"""
if len(matrices) < 2:
raise ValueError(
'kr requires a list of at least 2 matrices, but {} given.'.format(
len(matrices)))
n_col = shape(matrices[0])[1]
for i, e in enumerate(matrices[1:]):
if not i:
res = matrices[0]
s1, s2 = shape(res)
s3, s4 = shape(e)
if not s2 == s4 == n_col:
raise ValueError(
'All matrices should have the same number of columns.')
res = reshape(
reshape(res, (s1, 1, s2)) * reshape(e, (1, s3, s4)), (-1, n_col))
return res
def forward(self, x):
if not self.pretrained:
input_ = x.reshape(x.shape[0],x.shape[1],x.shape[2]*x.shape[3]*x.shape[4])
else:
input_ = x
output_1 = self.lstm_1(input_)
output_2 = self.lstm_2(output_1)
dense_1 = self.dense(output_2)
output = F.reshape(dense_1,(dense_1.shape[0],self.caption_length,self.caption_length))
return output
def get_minibatch(data_iter):
try:
batch = data_iter.next()
except StopIteration:
data_iter.reset()
batch = data_iter.next()
x = batch.data[0]
x = nd.reshape(x, (x.shape[0], -1))
y = nd.one_hot(batch.label[0], 10)
return x, y
def mxwindow(mna,window):
mnas=mna.shape
mnout=(*mnas[:-2],*window,((mnas[-2]-window[-2])+1),((mnas[-1]-window[-1])+1))
mne2=None
for R in range(window[0]):
j_lim = R + mnout[-2]
for H in range(window[1]):
tdata=mnd.slice(mna, begin=(None,None,R,H), end=(None,None,j_lim,(H + mnout[-1])), step=(None,None,1,1))
if mne2 is None:
mne2=tdata
else:
mne2=mnd.concat(mne2,tdata,dim=1)
return(mnd.expand_dims(mnd.transpose(mnd.reshape(mne2, shape=mnout),axes=(0,5,4,3,2,1)), 3))
def forward(self, is_train, req, in_data, out_data, aux):
arm_cls_preds = in_data[0]
odm_cls_target = in_data[1]
odm_loc_target_mask = in_data[2]
arm_cls_preds = nd.softmax(data=arm_cls_preds)
arm_cls_preds_classes = nd.split(data=arm_cls_preds,axis=1,num_outputs=2)
# arm_cls_preds_bg shape : (batch , h*w*num_anchors[:layers]) 负类【0】
arm_cls_preds_bg = nd.reshape(data=arm_cls_preds_classes[0],shape=(0,-1))
prob_temp = nd.ones_like(arm_cls_preds_bg)*0.99
cond1 = arm_cls_preds_bg >= prob_temp # > 0.99 idx is 1
# print('negative cond1 ------- :',heapq.nlargest(2,arm_cls_preds_bg[0]))
temp1 = nd.ones_like(odm_cls_target)*(-1) ### TODO： 0 还是-1表示背景??
# 如果ARM分类出的负类的置信度大于0.99，将其在ODM的anchor标号中去掉（-1替代），负类转换为背景
odm_cls_target_mask = nd.where(condition=cond1,x=temp1,y=odm_cls_target)
# apply filtering to odm_loc_target_mask
# odm_loc_target_mask_shape: (batch, num_anchors, 4)
arm_cls_preds_bg = nd.reshape(data=arm_cls_preds_bg,shape=(0,-1,1))#(batch , h*w*num_anchors[:layers],1)
# (batch , h*w*num_anchors[:layers] , 4 )
odm_loc_target_mask = nd.reshape(data=odm_loc_target_mask,shape=(0,-1,4))
odm_loc_target_mask = odm_loc_target_mask[:,:,0] #(batch , h*w*num_anchors[:layers])
#(batch , h*w*num_anchors[:layers], 1)
## 取整个batch中 所有行的 第一列，相当于对原来的4个相同label[0 0 0 0 ],[1 1 1 1]变成[0],[1]
odm_loc_target_mask = nd.reshape(data=odm_loc_target_mask,shape=(0,-1,1))
loc_temp = nd.ones_like(odm_loc_target_mask)*0.99
cond2 = arm_cls_preds_bg >= loc_temp
temp2 = nd.zeros_like(odm_loc_target_mask) # 取0
# 如果ARM分类出的负类的置信度大于0.99，将其在ODM的掩码置0
## 实际上不管IOU计算的大小，用AMR的分类结果，如果是大于0.99的负类，不管通过IOU判断的正负类结果如何，都设置为背景
odm_loc_target_bg_mask = nd.where(cond2,temp2,odm_loc_target_mask)
odm_loc_target_bg_mask = nd.concat(*[odm_loc_target_bg_mask]*4,dim=2)
# 还原维度
odm_loc_target_bg_mask = nd.reshape(odm_loc_target_bg_mask,shape=(0,-1))
for ind, val in enumerate([odm_cls_target_mask, odm_loc_target_bg_mask]):
self.assign(out_data[ind], req[ind], val)
def unsorted_1d_segment_sum(input, seg_id, n_segs, dim):
# TODO: support other dimensions
assert dim == 0, 'MXNet only supports segment sum on first dimension'
# Use SPMV to simulate segment sum
ctx = input.context
n_inputs = input.shape[0]
input_shape_suffix = input.shape[1:]
input = input.reshape(n_inputs, -1)
n_range = nd.arange(n_inputs, dtype='int64').as_in_context(input.context)
w_nnz = nd.ones(n_inputs).as_in_context(input.context)
w_nid = nd.stack(seg_id, n_range, axis=0)
w = nd.sparse.csr_matrix((w_nnz, (seg_id, n_range)), (n_segs, n_inputs))
w = w.as_in_context(input.context)
y = nd.dot(w, input)
y = nd.reshape(y, (n_segs, ) + input_shape_suffix)
return y
def _group_norm_func(input_data, num_groups, eps, gamma, beta):
n, c, h, w = input_data.shape
input_data = nd.reshape(data=input_data,
shape=(n, num_groups, c / num_groups, h, w))
# mean
mean = nd.mean(input_data, axis=2, keepdims=True)
# std
temp = nd.square((input_data - mean))
std = nd.sqrt(nd.sum(temp, axis=2, keepdims=True) / (c / num_groups))
input_data = (input_data - mean) / (std + eps)
out = input_data.reshape((n, c, h, w))
gamma = gamma.reshape((c, 1, 1))
beta = beta.reshape((c, 1, 1))
out = out * gamma + beta
return out
def forward(self, X, stride=1):
filters = []
for i in range(self._n_scales):
kernel = (i * 2 + 1, ) * 2
pad = (i, ) * 2
f = nd.Pooling(data=data,
pool_type='max',
kernel=kernel,
stride=(stride, stride),
pad=pad,
cudnn_off=True)
f = nd.reshape(f, (f.shape[0], 1) + f.shape[1:])
filters.append(f)
filters = nd.concat(*filters, dim=1)
weight = nd.softmax(self._get_param(self.weight), axis=1)
filters = nd.mean(filters, axis=1)
# filters = nd.sum(filters * weight, axis=1)
return filters
def verify_reshape_like_dynamic(
xshp, wshp, lhs_begin, lhs_end, rhs_begin, rhs_end):
x_np = np.random.uniform(size=xshp)
w_np = np.random.uniform(size=wshp)
x = nd.array(x_np)
w = nd.array(w_np)
# org op
y = nd.reshape_like(
x, w, lhs_begin=lhs_begin, lhs_end=lhs_end,
rhs_begin=rhs_begin, rhs_end=rhs_end)
# rewrite op
xndims = len(xshp)
lhs_begin = lhs_begin+xndims if lhs_begin < 0 else lhs_begin
lhs_end = lhs_end+xndims if lhs_end< 0 else lhs_end
assert 0 <= lhs_begin < lhs_end <= xndims
wndims = len(wshp)
rhs_begin = rhs_begin+wndims if rhs_begin < 0 else rhs_begin
rhs_end = rhs_end+wndims if rhs_end< 0 else rhs_end
assert 0 <= rhs_begin < rhs_end <= wndims
rshp = xshp[:lhs_begin] + \
wshp[rhs_begin:rhs_end] + xshp[lhs_end:]
batch_axes = [0]
assert len(batch_axes) == 1, \
"Dynamic batch shape fusion only support" + \
"single dimension of batch. Providied: (%s)" % batch_axes
batch_axis = batch_axes[0]
rshp = rshp[:batch_axis] + (-1,) + rshp[batch_axis+1:]
z = nd.reshape(x, shape=rshp)
# compare
assert z.shape == y.shape
zn, zp = get_norm(z)
yn, yp = get_norm(y)
rn = np.linalg.norm(zp-yp)
print(zn, yn, rn)
def get_sample_point():
"""Grabs a single input/label pair from MNIST"""
def transform(data, label):
return data.astype(np.float32) / 255.0, label.astype(np.float32)
# Load ten random images from the test dataset
sample_data = mx.gluon.data.DataLoader(mx.gluon.data.vision.MNIST(
train=False, transform=transform),
1,
shuffle=True)
for data, label in sample_data:
img = nd.transpose(data, (1, 0, 2, 3))
img = nd.reshape(img, (28, 28, 1))
imtiles = nd.tile(img, (1, 1, 3))
plt.imshow(imtiles.asnumpy())
plt.savefig("test_input.png")
data = nd.transpose(data, (0, 3, 1, 2))
data = data.as_in_context(ctx).asnumpy()
label = int(label.asnumpy()[0])
return data, label
