class Model(nn.Module):
def __init__(self, gz, dim):
super(Model, self).__init__()
self.gz = gz
self.conv_model = nn.Sequential()
self.act = nn.ELU # nn.Softplus # nn.ELU
self.conv = nn.Conv2d
self.bn = nn.BatchNorm2d
if dim == 3:
self.conv = nn.Conv3d
self.bn = nn.BatchNorm3d
bn = False
#self.conv_model.add_module("conv1",self.conv(4, 8, 5, stride=2,padding=2))
self.conv_model.add_module("conv1",
self.conv(2, 8, 5, stride=2, padding=2))
if bn:
self.conv_model.add_module("bn1", self.bn(8))
self.conv_model.add_module("act1", self.act())
for x in range(3):
self.conv_model.add_module("conv%d" % (x + 2),
self.conv(8, 8, 5, stride=1, padding=2))
if bn:
self.conv_model.add_module("bn%d" % (x + 2), self.bn(8))
self.conv_model.add_module("act%d" % (x + 2), self.act())
print(self.conv_model)
#create the distnaces / norm generator
self.distyNorm = DistyNorm(dim, norm="euc")
self.disty = Disty(dim, gz, self.distyNorm)
#create the aggregator
self.aggMax = AggMax()
self.aggSum = AggSum()
self.agg = Agg(dim, gz, self.aggSum)
#create the renger function
self.R = torch.nn.Sequential(torch.nn.Linear(2, 4), torch.nn.ELU(),
torch.nn.Linear(4, 4), torch.nn.ELU(),
torch.nn.Linear(4, 2))
def forward(self, goals, inputs):
tmp = self.disty.forward(inputs)
R_tmp = self.R.forward(tmp)
#tmp=tmp.prod(1,True)
inputs_grid = self.agg.forward(R_tmp, inputs)
output = self.conv_model(inputs_grid)
energy = output.mean(3, False).mean(2, False).mean(1, False)
#iterate...
loss = 0
for idx in range(len(inputs)):
dE_dx, = torch.autograd.grad(output.sum(),
inputs[idx]['points'],
create_graph=True)
inputs[idx]['pred'] = dE_dx
loss += (((dE_dx - goals[idx]['points']) /
inputs[idx]['points'].size()[0])**2).sum()
return loss, 0, 0 #g_sum,g_abs_sum
class Model(nn.Module):
def __init__(self, gz, dim):
super(Model, self).__init__()
self.gz = gz
self.conv_model = nn.Sequential()
self.act = nn.ELU
self.conv = nn.Conv2d
self.bn = nn.BatchNorm2d
if dim == 3:
self.conv = nn.Conv3d
self.bn = nn.BatchNorm3d
bn = True
self.conv_model.add_module("conv1",
self.conv(12, 8, 5, stride=2, padding=2))
if bn:
self.conv_model.add_module("bn1", self.bn(8))
self.conv_model.add_module("act1", self.act())
for x in range(6):
self.conv_model.add_module("conv%d" % (x + 2),
self.conv(8, 8, 5, stride=1, padding=2))
if bn:
self.conv_model.add_module("bn%d" % (x + 2), self.bn(8))
self.conv_model.add_module("act%d" % (x + 2), self.act())
print(self.conv_model)
#create the distnaces / norm generator
self.distyNorm = DistyNorm(dim, norm="euc")
self.disty = Disty(dim, gz, self.distyNorm)
#create the aggregator
self.aggMax = AggMax()
self.agg = Agg(dim, gz, self.aggMax)
#create the renger function
self.R = torch.nn.Sequential(torch.nn.Linear(12, 16), torch.nn.ELU(),
torch.nn.Linear(16, 16), torch.nn.ELU(),
torch.nn.Linear(16, 4))
def forward(self, goals, inputs, backgrounds):
inputs_grid = self.agg.forward(
self.R.forward(self.disty.forward(inputs)), inputs)
goals_grid = self.agg.forward(
self.R.forward(self.disty.forward(goals)), goals)
backgrounds_grid = self.agg.forward(
self.R.forward(self.disty.forward(backgrounds)), backgrounds)
print("I", inputs_grid.sum(), "G", goals_grid.sum(), "B",
backgrounds_grid.sum())
#inputs_grid=self.agg.forward(self.disty.forward(inputs),inputs)
#goals_grid=self.agg.forward(self.disty.forward(goals),goals)
scenes = torch.cat((inputs_grid, goals_grid, backgrounds_grid), 1)
output = self.conv_model(scenes)
print("O", output.sum())
#inputs_tensor = torch.cat([ input['points'] for input in inputs ],0) #TODO WHY CANT I DO THIS?!?!?!
#d_output_to_input = torch.cat(torch.autograd.grad(output,inputs_tensor,create_graph=True),0)
#iterate...
if True:
cost = 0
g_sum = 0
g_abs_sum = 0
for idx in range(len(inputs)):
d_output_to_input = torch.autograd.grad(output[idx].mean(),
inputs[idx]['points'],
create_graph=True)[0]
print("D", d_output_to_input.sum())
cost += ((d_output_to_input - goals[idx]['points'])**2).mean()
g_sum += d_output_to_input.sum()
g_abs_sum += d_output_to_input.abs().sum()
cost /= len(inputs)
g_sum /= len(inputs)
g_abs_sum /= len(inputs)
#do it all at once
#goals_tensor = torch.cat([ goal['points'] for goal in goals ],0)
#d_output_to_input = torch.autograd.grad(output.mean(),inputs['points'],create_graph=True)
#d_output_to_input = torch.cat(torch.autograd.grad(output.mean(),inputs['points'],create_graph=True),1)
#cost = ((d_output_to_input-goals_tensor)**2).mean()
#g_sum = d_output_to_input.sum()
#g_abs_sum = d_output_to_input.abs().sum()
return cost, g_sum, g_abs_sum
class Model(nn.Module):
def __init__(self,gz,dim):
super(Model, self).__init__()
self.gz = gz
self.conv_model = nn.Sequential()
self.act = nn.ELU
self.conv = nn.Conv2d
self.bn = nn.BatchNorm2d
if dim==3:
self.conv = nn.Conv3d
self.bn = nn.BatchNorm3d
bn=True
self.conv_model.add_module("conv1",self.conv(4, 32, 5, stride=2,padding=2))
if bn:
self.conv_model.add_module("bn1",self.bn(32))
self.conv_model.add_module("act1",self.act())
for x in range(16):
self.conv_model.add_module("conv%d" % (x+2),self.conv(32, 32, 5, stride=1,padding=2))
if bn:
self.conv_model.add_module("bn%d" % (x+2),self.bn(32))
self.conv_model.add_module("act%d" % (x+2),self.act())
print(self.conv_model)
#create the distnaces / norm generator
self.distyNorm=DistyNorm(dim,norm="euc")
self.disty=Disty(dim,gz,self.distyNorm)
#create the aggregator
self.aggMax=AggMax()
self.agg=Agg(dim,gz,self.aggMax)
#create the renger function
self.R=torch.nn.Sequential(torch.nn.Linear(2,16),torch.nn.ELU(),torch.nn.Linear(16,16),torch.nn.ELU(),torch.nn.Linear(16,4))
def forward(self, goals,inputs):
inputs_grid=self.agg.forward(self.R.forward(self.disty.forward(inputs)),inputs)
scenes=inputs_grid #torch.cat((inputs_grid,goals_grid),1)
output = self.conv_model(scenes)
#inputs_tensor = torch.cat([ input['points'] for input in inputs ],0) #TODO WHY CANT I DO THIS?!?!?!
#d_output_to_input = torch.cat(torch.autograd.grad(output,inputs_tensor,create_graph=True),0)
#iterate...
dpoints=[]
dattrs=[]
if True:
cost=0
g_sum=0
g_abs_sum=0
for idx in range(len(inputs)):
if False:
d_output_to_input = torch.autograd.grad(output[idx].mean(),inputs[idx]['attrs'],create_graph=True)[0]
doutputs.append(d_output_to_input.data.numpy().reshape(1,-1))
cost += ((d_output_to_input-goals[idx]['attrs'])**2).mean()
else:
d_output_to_points,d_output_to_attrs = torch.autograd.grad(output[idx].mean(),[inputs[idx]['points'],inputs[idx]['attrs']],create_graph=True)
dpoints.append(d_output_to_points.data.numpy())
dattrs.append(d_output_to_attrs.data.numpy())
cost += ((d_output_to_points-goals[idx]['force'])**2).mean()+((d_output_to_attrs-goals[idx]['attrs'])**2).mean()
g_sum += d_output_to_points.sum()+d_output_to_attrs.sum()
g_abs_sum += d_output_to_points.abs().sum()+d_output_to_attrs.abs().sum()
cost/=len(inputs)
g_sum/=len(inputs)
g_abs_sum/=len(inputs)
#do it all at once
#goals_tensor = torch.cat([ goal['points'] for goal in goals ],0)
#d_output_to_input = torch.autograd.grad(output.mean(),inputs['points'],create_graph=True)
#d_output_to_input = torch.cat(torch.autograd.grad(output.mean(),inputs['points'],create_graph=True),1)
#cost = ((d_output_to_input-goals_tensor)**2).mean()
#g_sum = d_output_to_input.sum()
#g_abs_sum = d_output_to_input.abs().sum()
return cost,g_sum,g_abs_sum,dpoints,dattrs #np.concatenate(doutputs,axis=1)
class Model(nn.Module):
def __init__(self,gz,dim):
super(Model, self).__init__()
self.gz = gz
self.conv_model = nn.Sequential()
self.act = nn.ELU
self.conv = nn.Conv2d
self.bn = nn.BatchNorm2d
if dim==3:
self.conv = nn.Conv3d
self.bn = nn.BatchNorm3d
bn=True
self.conv_model.add_module("conv1",self.conv(2, 1, 5, stride=1,padding=2))
if bn:
self.conv_model.add_module("bn1",self.bn(1))
self.conv_model.add_module("act1",self.act())
for x in range(2):
self.conv_model.add_module("conv%d" % (x+2),self.conv(1, 1, 5, stride=1,padding=2))
if bn:
self.conv_model.add_module("bn%d" % (x+2),self.bn(1))
self.conv_model.add_module("act%d" % (x+2),self.act())
print(self.conv_model)
#create the distnaces / norm generator
self.distyNorm=DistyNorm(dim,norm="euc")
self.disty=Disty(dim,gz,self.distyNorm)
#create the aggregator
self.aggMax=AggMax()
self.agg=Agg(dim,gz,self.aggMax)
#create the renger function
self.R=torch.nn.Linear(1,1)
def forward(self, goals,inputs):
inputs_grid=self.agg.forward(self.R.forward(self.disty.forward(inputs)),inputs)
goals_grid=self.agg.forward(self.R.forward(self.disty.forward(goals)),goals)
#inputs_grid=self.agg.forward(self.disty.forward(inputs),inputs)
#goals_grid=self.agg.forward(self.disty.forward(goals),goals)
scenes=torch.cat((inputs_grid,goals_grid),1)
output = self.conv_model(scenes)
#inputs_tensor = torch.cat([ input['points'] for input in inputs ],0) #TODO WHY CANT I DO THIS?!?!?!
#d_output_to_input = torch.cat(torch.autograd.grad(output,inputs_tensor,create_graph=True),0)
#goals_tensor = torch.cat([ goal['points'] for goal in goals ],0)
cost=0
g_sum=0
g_abs_sum=0
for idx in range(len(inputs)):
d_output_to_input = torch.autograd.grad(output[idx].mean(),inputs[idx]['points'],create_graph=True)[0]
diff=Variable(goals[idx]['points']-inputs[idx]['points'],requires_grad=False)
#cost+=((d_output_to_input-goals[idx]['points'])**2).mean()
cost+=((d_output_to_input-diff)**2).mean()
g_sum += d_output_to_input.sum()
g_abs_sum += d_output_to_input.abs().sum()
cost/=len(inputs)
g_sum/=len(inputs)
g_abs_sum/=len(inputs)
return cost,g_sum,g_abs_sum