Dear all,
I have a weird error that I cannot find the source of it. Perhaps someone can help? In particular I am trying to implement the following simple operation of iterating over the elements of a HybridSequential “container”:
class SomeModel(HybridBlock):
def __init__(self, **kwards):
super().__init__(**kwards)
with self.name_scope():
self.layers = gluon.nn.HybridSequential()
for idx in range(4):
self.layers.add(SomeOtherHybridBlock(idx) ) # different idx results different layer
def hybrid_forward(self, F, input):
x = input
for conv in self.layers:
x = x + self.layers(input)
return x
it seems simple, but I am getting the following error:
---------------------------------------------------------------------------
MXNetError Traceback (most recent call last)
<ipython-input-2-e5c3f04ecacb> in <module>
8 net.hybridize()
9 xx = nd.random.uniform(shape=[7,nfilters,F,F])
---> 10 out = net(xx)
~/.local/lib/python3.6/site-packages/mxnet/gluon/block.py in __call__(self, *args)
691 hook(self, args)
692
--> 693 out = self.forward(*args)
694
695 for hook in self._forward_hooks.values():
~/.local/lib/python3.6/site-packages/mxnet/gluon/block.py in forward(self, x, *args)
1146 'Find all contexts = {}'.format(ctx_set))
1147 with ctx:
-> 1148 return self._call_cached_op(x, *args)
1149 with ctx:
1150 try:
~/.local/lib/python3.6/site-packages/mxnet/gluon/block.py in _call_cached_op(self, *args)
979 def _call_cached_op(self, *args):
980 if self._cached_op is None:
--> 981 self._build_cache(*args)
982 assert self._cached_op, "cached op is not None"
983 if self._callback:
~/.local/lib/python3.6/site-packages/mxnet/gluon/block.py in _build_cache(self, *args)
967 flags = [('data_indices', data_indices), ('param_indices', param_indices)] + \
968 self._flags
--> 969 self._cached_op = ndarray.CachedOp(out, flags)
970
971 def _deferred_infer_shape(self, *args):
~/.local/lib/python3.6/site-packages/mxnet/_ctypes/ndarray.py in __init__(self, sym, flags)
134 c_str_array([key for key, _ in flags]),
135 c_str_array([str(val) for _, val in flags]),
--> 136 ctypes.byref(self.handle)))
137
138 def __del__(self):
~/.local/lib/python3.6/site-packages/mxnet/base.py in check_call(ret)
253 """
254 if ret != 0:
--> 255 raise MXNetError(py_str(_LIB.MXGetLastError()))
256
257
MXNetError: [18:27:16] src/core/graph.cc:110: Check failed: it != node2index_.end(): control dep not found in graph
Stack trace:
[bt] (0) /home/dia021/.local/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x65928b) [0x7fa68af2828b]
[bt] (1) /home/dia021/.local/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x8385e68) [0x7fa692c54e68]
[bt] (2) /home/dia021/.local/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x8386fe8) [0x7fa692c55fe8]
[bt] (3) /home/dia021/.local/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x8387b70) [0x7fa692c56b70]
[bt] (4) /home/dia021/.local/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x389a709) [0x7fa68e169709]
[bt] (5) /home/dia021/.local/lib/python3.6/site-packages/mxnet/libmxnet.so(MXCreateCachedOpEx+0x2cf) [0x7fa68e07672f]
[bt] (6) /usr/lib/x86_64-linux-gnu/libffi.so.6(ffi_call_unix64+0x4c) [0x7fa6f55f3dae]
[bt] (7) /usr/lib/x86_64-linux-gnu/libffi.so.6(ffi_call+0x22f) [0x7fa6f55f371f]
[bt] (8) /usr/lib/python3.6/lib-dynload/_ctypes.cpython-36m-x86_64-linux-gnu.so(_ctypes_callproc+0x2b4) [0x7fa6f58075c4]
Complete code example (am trying to implement an attention layer - in 2D) and add it in parallel to the input, in the philosophy of the residual units/blocks):
# Definitions
import mxnet as mx
from mxnet.gluon import HybridBlock
from mxnet import nd, gluon
class Conv2DNormed(HybridBlock):
"""
Convenience wrapper layer for 2D convolution followed by a normalization layer
(either BatchNorm or InstanceNorm).
norm_type: Either BatchNorm (default) or InstanceNorm strings.
axis : axis in normalization (exists only in BatchNorm).
All other keywords are the same as gluon.nn.Conv2D
"""
def __init__(self, channels, kernel_size, strides=(1, 1),
padding=(0, 0), dilation=(1, 1), activation=None,
weight_initializer=None, in_channels=0, _norm_type = 'BatchNorm', axis =1 , groups=1,**kwards):
super().__init__(**kwards)
if (_norm_type == 'BatchNorm'):
self.norm = gluon.nn.BatchNorm
elif (_norm_type == 'SyncBatchNorm'):
self.norm = gluon.contrib.nn.SyncBatchNorm
_prefix = "_SyncBN"
elif (_norm_type == 'InstanceNorm'):
self.norm = gluon.nn.InstanceNorm
elif (_norm_type == 'LayerNorm'):
self.norm = gluon.nn.LayerNorm
else:
raise NotImplementedError
with self.name_scope():
self.conv2d = gluon.nn.Conv2D(channels, kernel_size = kernel_size,
strides= strides,
padding=padding,
dilation= dilation,
activation=activation,
use_bias=False,
weight_initializer = weight_initializer,
groups=groups,
in_channels=0)
self.norm_layer = self.norm(axis=axis)
def hybrid_forward(self,F,_x):
x = self.conv2d(_x)
x = self.norm_layer(x)
return x
def get_norm(name, prefix, ngroups=None):
if (name == 'BatchNorm'):
return gluon.nn.BatchNorm(axis=1,prefix = prefix)
elif (name == 'InstanceNorm'):
return gluon.nn.InstanceNorm(axis=1,prefix = prefix)
elif (name == 'LayerNorm'):
return gluon.nn.LayerNorm(axis=1,prefix = prefix)
elif (name == 'GroupNorm'):
return gluon.nn.GroupNorm(num_groups=ngroups, prefix = prefix)
else:
raise NotImplementedError
class ResNet_v2_block(HybridBlock):
"""
ResNet v2 building block. It is built upon the assumption of ODD kernel
"""
def __init__(self, _nfilters,_kernel_size=(3,3),_dilation_rate=(1,1),
_norm_type='BatchNorm', ngroups=1, **kwards):
super().__init__(**kwards)
self.nfilters = _nfilters
self.kernel_size = _kernel_size
self.dilation_rate = _dilation_rate
with self.name_scope():
# Ensures padding = 'SAME' for ODD kernel selection
p0 = self.dilation_rate[0] * (self.kernel_size[0] - 1)/2
p1 = self.dilation_rate[1] * (self.kernel_size[1] - 1)/2
p = (int(p0),int(p1))
self.BN1 = get_norm(_norm_type, prefix = "1_",ngroups=ngroups )
self.conv1 = gluon.nn.Conv2D(self.nfilters,kernel_size = self.kernel_size,padding=p,dilation=self.dilation_rate,use_bias=False,prefix="_conv1_",groups=ngroups)
self.BN2 = get_norm(_norm_type, prefix = "2_" ,ngroups=ngroups)
self.conv2 = gluon.nn.Conv2D(self.nfilters,kernel_size = self.kernel_size,padding=p,dilation=self.dilation_rate,use_bias=True,prefix="_conv2_",groups=ngroups)
def hybrid_forward(self,F,_input_layer):
x = self.BN1(_input_layer)
x = F.relu(x)
x = self.conv1(x)
x = self.BN2(x)
x = F.relu(x)
x = self.conv2(x)
return x
class ResNet_atrous_unit(HybridBlock):
def __init__(self, _nfilters, _kernel_size=(3,3), _dilation_rates=[3,15,31], _norm_type = 'BatchNorm', **kwards):
super().__init__(**kwards)
with self.name_scope():
self.ResBlocks = gluon.nn.HybridSequential()
self.ResBlocks.add(ResNet_v2_block(_nfilters,_kernel_size,_dilation_rate=(1,1), _norm_type = _norm_type))
for idx,d in enumerate(_dilation_rates):
self.ResBlocks.add(ResNet_v2_block(_nfilters,_kernel_size,_dilation_rate=(d,d), _norm_type = _norm_type))
def hybrid_forward(self,F,_xl):
x = _xl
for conv in self.ResBlocks:
x = x + conv(_xl)
return x
# @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ SOME WEIRD PROBLEM @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
class Attention2D_v3(HybridBlock):
"""
Self spatial Attention layer.
"""
def __init__(self, nkeys, nheads=1, norm = 'BatchNorm',**kwards):
super().__init__(**kwards)
with self.name_scope():
self.query = Conv2DNormed(nkeys,kernel_size=1, _norm_type= norm, groups=nheads)
self.key = Conv2DNormed(nkeys,kernel_size=1, _norm_type= norm, groups=nheads)
self.value = Conv2DNormed(nkeys,kernel_size=1, _norm_type= norm, groups=nheads)
self.act_q = gluon.nn.Swish()
self.act_k = gluon.nn.SELU()
def hybrid_forward(self, F, input):
q = self.act_q(self.query(input))
q = q.reshape([0,0,-1])
k = self.act_k(self.key(input))
k = k.reshape([0,0,-1])
v = F.relu(self.value(input))
v = v.reshape([0,0,-1])
# Spatial attention
att = F.batch_dot(q,k,transpose_b=True)
# Is this scaling good enough?
scale = F.reciprocal(F.norm(att,ord=1,axis=(-1,-2),keepdims=True))
att = F.softmax(F.broadcast_mul(att,scale),axis=-1)
att = F.batch_dot(att,v)
att = att.reshape_like(input)
return att
# @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
class AResNet_v2_block(HybridBlock):
def __init__(self, _nfilters, _kernel_size=(3,3), _dilation_rate=(1,1), _norm_type='BatchNorm', _nkeys=None, nheads=4, ngroups=1, **kwards):
super().__init__(**kwards)
if _nkeys==None:
_nkeys=_nfilters
with self.name_scope():
self.block = ResNet_v2_block(_nfilters,_kernel_size,_dilation_rate= _dilation_rate, _norm_type = _norm_type)
self.att = Attention2D_v3( _nkeys, nheads=nheads, norm= _norm_type)
self.tgamma = self.params.get('tgamma',shape=(1,),init=mx.init.One())
def hybrid_forward(self, F,input, tgamma):
block = self.block(input)
att = self.att(input)
out = F.broadcast_add(block , F.broadcast_mul(att,tgamma))
return out
class AResNet_atrous_unit(HybridBlock):
# PROBLEM HERE
"""
This layer is as Diakogiannis et al 2019 (ResUNet-a) with the addition of an attention layer applied on each dilated conv block.
"""
def __init__(self, _nfilters, _kernel_size=(3,3), _dilation_rates=[1,3,15,31], _nkeys=None, nheads=4, _norm_type = 'BatchNorm', **kwards):
super().__init__(**kwards)
# mxnet doesn't like wrapping things inside a list: it shadows the HybridBlock, remove list
with self.name_scope():
self.AResBlocks = gluon.nn.HybridSequential()
for d in _dilation_rates:
self.AResBlocks.add( AResNet_v2_block(_nfilters,_kernel_size,_dilation_rate=(d,d), nheads=nheads, _nkeys=_nkeys, _norm_type = _norm_type) )
def hybrid_forward(self, F, input):
x = input
for block in self.AResBlocks:
x = x + block(input)
return x
This reproduces the error:
from mxnet import nd
nfilters = 1024
F=8
# @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
net = AResNet_atrous_unit(nfilters,_dilation_rates=[1,3],nheads=nfilters//8) # Doesn't work
# @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
#net = ResNet_atrous_unit(nfilters) # Works fine
#net = AResNet_v2_block(nfilters) # Works fine
net.initialize()
net.hybridize()
xx = nd.random.uniform(shape=[7,nfilters,F,F])
out = net(xx)
Any ideas what I am doing wrong? Thank you in advance.
mxnet version == 1.6.0