def jojo_12(self, input_vars_125, var_1203, var_1159, input_vars_128, input_vars_123, input_vars_124, input_vars_122, input_vars_127, input_vars_130, input_vars_131, input_vars_129):
var_1208 = torch.batch_norm(var_1203, input_vars_122, input_vars_123, input_vars_124, input_vars_125, False, 0.1, 1e-05, True)
var_1210 = torch.add(var_1208, var_1159, alpha=1)
var_1211 = torch.relu_(var_1210)
var_1230 = torch._convolution(var_1211, input_vars_127, None, [1, 1, ], [1, 1, ], [1, 1, ], False, [0, 0, ], 1, False, False, True)
var_1235 = torch.batch_norm(var_1230, input_vars_128, input_vars_129, input_vars_130, input_vars_131, False, 0.1, 1e-05, True)
return var_1211, var_1235
def jojo_14(self, var_1127, input_vars_104, input_vars_111, input_vars_110, input_vars_112, input_vars_106, input_vars_109, input_vars_113, var_1083, input_vars_105, input_vars_107):
var_1132 = torch.batch_norm(var_1127, input_vars_104, input_vars_105, input_vars_106, input_vars_107, False, 0.1, 1e-05, True)
var_1151 = torch._convolution(var_1083, input_vars_109, None, [2, 2, ], [0, 0, ], [1, 1, ], False, [0, 0, ], 1, False, False, True)
var_1156 = torch.batch_norm(var_1151, input_vars_110, input_vars_111, input_vars_112, input_vars_113, False, 0.1, 1e-05, True)
var_1158 = torch.add(var_1132, var_1156, alpha=1)
var_1159 = torch.relu_(var_1158)
return var_1159
def jojo_5(self, input_vars_50, var_586, input_vars_51, input_vars_49,
input_vars_57, input_vars_64, input_vars_52, input_vars_55,
input_vars_61, input_vars_63, input_vars_58, input_vars_56,
input_vars_53, input_vars_65, input_vars_62, input_vars_59):
var_587 = torch.relu_(var_586)
var_606 = torch._convolution(var_587, input_vars_49, None, [
1,
1,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_611 = torch.batch_norm(var_606, input_vars_50, input_vars_51,
input_vars_52, input_vars_53, False, 0.1,
1e-05, True)
var_612 = torch.relu_(var_611)
var_631 = torch._convolution(var_612, input_vars_55, None, [
1,
1,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_636 = torch.batch_norm(var_631, input_vars_56, input_vars_57,
input_vars_58, input_vars_59, False, 0.1,
1e-05, True)
var_638 = torch.add(var_636, var_587, alpha=1)
var_639 = torch.relu_(var_638)
var_658 = torch._convolution(var_639, input_vars_61, None, [
2,
2,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_663 = torch.batch_norm(var_658, input_vars_62, input_vars_63,
input_vars_64, input_vars_65, False, 0.1,
1e-05, True)
return var_663, var_639
def jojo_6(self, input_vars_44, input_vars_34, input_vars_37, var_459,
input_vars_45, var_508, input_vars_31, input_vars_38,
input_vars_41, input_vars_47, input_vars_40, input_vars_43,
input_vars_33, input_vars_35, input_vars_46, input_vars_32,
input_vars_39):
var_510 = torch.add(var_508, var_459, alpha=1)
var_511 = torch.relu_(var_510)
var_530 = torch._convolution(var_511, input_vars_31, None, [
2,
2,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_535 = torch.batch_norm(var_530, input_vars_32, input_vars_33,
input_vars_34, input_vars_35, False, 0.1,
1e-05, True)
var_536 = torch.relu_(var_535)
var_555 = torch._convolution(var_536, input_vars_37, None, [
1,
1,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_560 = torch.batch_norm(var_555, input_vars_38, input_vars_39,
input_vars_40, input_vars_41, False, 0.1,
1e-05, True)
var_579 = torch._convolution(var_511, input_vars_43, None, [
2,
2,
], [
0,
0,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_584 = torch.batch_norm(var_579, input_vars_44, input_vars_45,
input_vars_46, input_vars_47, False, 0.1,
1e-05, True)
return var_560, var_584
def jojo_25(self, input_vars_15, input_vars_14, input_vars_8, input_vars_11, input_vars_10, input_vars_17, input_vars_9, var_681, input_vars_16, input_vars_7, input_vars_13):
var_695 = torch.max_pool2d(var_681, [3, 3, ], [2, 2, ], [1, 1, ], [1, 1, ], False)
var_714 = torch._convolution(var_695, input_vars_7, None, [1, 1, ], [1, 1, ], [1, 1, ], False, [0, 0, ], 1, False, False, True)
var_719 = torch.batch_norm(var_714, input_vars_8, input_vars_9, input_vars_10, input_vars_11, False, 0.1, 1e-05, True)
var_720 = torch.relu_(var_719)
var_739 = torch._convolution(var_720, input_vars_13, None, [1, 1, ], [1, 1, ], [1, 1, ], False, [0, 0, ], 1, False, False, True)
var_744 = torch.batch_norm(var_739, input_vars_14, input_vars_15, input_vars_16, input_vars_17, False, 0.1, 1e-05, True)
var_746 = torch.add(var_744, var_695, alpha=1)
return var_746
def jojo_7(self, input_vars_25, var_431, input_vars_16, input_vars_13,
input_vars_15, input_vars_23, input_vars_19, input_vars_21,
input_vars_22, var_407, input_vars_17, input_vars_14,
input_vars_20):
var_432 = torch.relu_(var_431)
var_451 = torch._convolution(var_432, input_vars_13, None, [
1,
1,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_456 = torch.batch_norm(var_451, input_vars_14, input_vars_15,
input_vars_16, input_vars_17, False, 0.1,
1e-05, True)
var_458 = torch.add(var_456, var_407, alpha=1)
var_459 = torch.relu_(var_458)
var_478 = torch._convolution(var_459, input_vars_19, None, [
1,
1,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_483 = torch.batch_norm(var_478, input_vars_20, input_vars_21,
input_vars_22, input_vars_23, False, 0.1,
1e-05, True)
var_484 = torch.relu_(var_483)
var_503 = torch._convolution(var_484, input_vars_25, None, [
1,
1,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
return var_459, var_503
def jojo_19(self, var_851, input_vars_71, input_vars_70, input_vars_67, input_vars_55, input_vars_57, input_vars_56, var_900, input_vars_62, input_vars_65, input_vars_73, input_vars_61, input_vars_68, input_vars_63, input_vars_69, input_vars_59, input_vars_64, input_vars_58):
var_919 = torch._convolution(var_851, input_vars_55, None, [2, 2, ], [0, 0, ], [1, 1, ], False, [0, 0, ], 1, False, False, True)
var_924 = torch.batch_norm(var_919, input_vars_56, input_vars_57, input_vars_58, input_vars_59, False, 0.1, 1e-05, True)
var_926 = torch.add(var_900, var_924, alpha=1)
var_927 = torch.relu_(var_926)
var_946 = torch._convolution(var_927, input_vars_61, None, [1, 1, ], [1, 1, ], [1, 1, ], False, [0, 0, ], 1, False, False, True)
var_951 = torch.batch_norm(var_946, input_vars_62, input_vars_63, input_vars_64, input_vars_65, False, 0.1, 1e-05, True)
var_952 = torch.relu_(var_951)
var_971 = torch._convolution(var_952, input_vars_67, None, [1, 1, ], [1, 1, ], [1, 1, ], False, [0, 0, ], 1, False, False, True)
var_976 = torch.batch_norm(var_971, input_vars_68, input_vars_69, input_vars_70, input_vars_71, False, 0.1, 1e-05, True)
var_978 = torch.add(var_976, var_927, alpha=1)
var_979 = torch.relu_(var_978)
var_998 = torch._convolution(var_979, input_vars_73, None, [1, 1, ], [1, 1, ], [1, 1, ], False, [0, 0, ], 1, False, False, True)
return var_979, var_998
def _instance_norm(raw, input, running_mean=None, running_var=None, weight=None,
bias=None, use_input_stats=True, momentum=0.1, eps=1e-5):
# TODO: the batch size!=1 view operations
print("WARNING: The Instance Normalization transfers to Caffe using BatchNorm, so the batch size should be 1")
if running_var is not None or weight is not None:
# TODO: the affine=True or track_running_stats=True case
raise NotImplementedError("not implement the affine=True or track_running_stats=True case InstanceNorm")
x= torch.batch_norm(
input, weight, bias, running_mean, running_var,
use_input_stats, momentum, eps,torch.backends.cudnn.enabled)
bottom_blobs = [log.blobs(input)]
layer_name1 = log.add_layer(name='instance_norm')
top_blobs = log.add_blobs([x], name='instance_norm_blob')
layer1 = caffe_net.Layer_param(name=layer_name1, type='BatchNorm',
bottom=bottom_blobs, top=top_blobs)
if running_mean is None or running_var is None:
# not use global_stats, normalization is performed over the current mini-batch
layer1.batch_norm_param(use_global_stats=0,eps=eps)
running_mean=torch.zeros(input.size()[1])
running_var=torch.ones(input.size()[1])
else:
layer1.batch_norm_param(use_global_stats=1, eps=eps)
running_mean_clone = running_mean.clone()
running_var_clone = running_var.clone()
layer1.add_data(running_mean_clone.cpu().numpy(), running_var_clone.cpu().numpy(), np.array([1.0]))
log.cnet.add_layer(layer1)
if weight is not None and bias is not None:
layer_name2 = log.add_layer(name='bn_scale')
layer2 = caffe_net.Layer_param(name=layer_name2, type='Scale',
bottom=top_blobs, top=top_blobs)
layer2.param.scale_param.bias_term = True
layer2.add_data(weight.cpu().data.numpy(), bias.cpu().data.numpy())
log.cnet.add_layer(layer2)
return x
def forward(ctx,
input,
running_mean,
running_var,
weight=None,
bias=None,
training=False,
momentum=0.1,
eps=1e-5):
# Training not currently supported for torchtorchexplain
# if training:
# size = input.size()
# #
# #
# # from operator import mul
# # from functools import reduce
# #
# # if reduce(mul, size[2:], size[0]) == 1
# size_prods = size[0]
# for i in range(len(size) - 2):
# size_prods *= size[i + 2]
# if size_prods == 1:
# raise ValueError('Expected more than 1 value per channel when training, got input size {}'.format(size))
output = torch.batch_norm(input, weight, bias, running_mean,
running_var, training, momentum, eps,
torch.backends.cudnn.enabled)
ctx.save_for_backward(input, output.clone())
ctx.hparams = (weight, bias, running_mean, running_var, eps)
return output
def forward(self, x, opt):
# break up channels of x
batch_size = x.size(0)
in_channels = x.size(1)
x = x.view(-1, 1, x.size(2), x.size(3))
# perform optics individually on each channel
meas = self.optics(x)
# put channels back together
meas = meas.view(batch_size, in_channels, meas.size(2), meas.size(3))
# normalize mask measurements
if opt.normalize_feats:
# 1: normalize to have max of 1
# view is used to get max over the 3D image
# unsqueeze is to get the dimensions correct for elementwise division
meas = meas / meas.view(meas.size(0), -1).max(
dim=1)[0].unsqueeze(1).unsqueeze(2).unsqueeze(3)
# 2: normalize [0, 1] to [-1, 1]
mean = torch.Tensor([0.5, 0.5, 0.5]).cuda(opt.gpu)
var = torch.Tensor([0.25, 0.25, 0.25]).cuda(opt.gpu) # std squared
meas = torch.batch_norm(meas, None, None, mean, var, False, 0, 0,
torch.backends.cudnn.enabled)
# forward through VGG net
output = self.features(meas)
output = output.view(output.size(0), -1)
output = self.classifier(output)
return output
def jojo_20(self, input_vars_45, input_vars_44, var_848, var_799, input_vars_47, input_vars_46, input_vars_49, input_vars_43):
var_850 = torch.add(var_848, var_799, alpha=1)
var_851 = torch.relu_(var_850)
var_870 = torch._convolution(var_851, input_vars_43, None, [2, 2, ], [1, 1, ], [1, 1, ], False, [0, 0, ], 1, False, False, True)
var_875 = torch.batch_norm(var_870, input_vars_44, input_vars_45, input_vars_46, input_vars_47, False, 0.1, 1e-05, True)
var_876 = torch.relu_(var_875)
var_895 = torch._convolution(var_876, input_vars_49, None, [1, 1, ], [1, 1, ], [1, 1, ], False, [0, 0, ], 1, False, False, True)
return var_851, var_895
def forward(self, inp):
out = self.bn(inp)
out = out.view(1, inp.size(0) * self.num_groups, -1)
out = torch.batch_norm(out, None, None, None, None, True, 0, self.eps, True)
out = out.view(inp.size(0), self.num_groups, -1)
out = self.weight * out + self.bias
out = out.view_as(inp)
return out
def _nn_functional_batch_norm(input,
running_mean,
running_var,
weight=None,
bias=None,
training=False,
momentum=0.1,
eps=1e-5):
return torch.batch_norm(input, weight, bias, running_mean, running_var,
training, momentum, eps,
torch.backends.cudnn.enabled)
def forward(self, x):
return torch.batch_norm(
x,
self.weight,
self.bias,
self.running_mean,
self.running_var,
False, # training
self.exponential_average_factor,
self.eps,
False, # cuda_enabled
)
def jojo_4(self, input_vars_69, var_683, input_vars_75, input_vars_71,
input_vars_68, input_vars_70, input_vars_74, input_vars_77,
input_vars_76, var_639, input_vars_73):
var_688 = torch.batch_norm(var_683, input_vars_68, input_vars_69,
input_vars_70, input_vars_71, False, 0.1,
1e-05, True)
var_707 = torch._convolution(var_639, input_vars_73, None, [
2,
2,
], [
0,
0,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_712 = torch.batch_norm(var_707, input_vars_74, input_vars_75,
input_vars_76, input_vars_77, False, 0.1,
1e-05, True)
return var_712, var_688
def jojo_8(self, input_vars_7, input_vars_5, input_vars_4, input_vars_1,
input_vars_2, input_vars_3, input_vars_0):
var_387 = torch._convolution(input_vars_0, input_vars_1, None, [
2,
2,
], [
3,
3,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_392 = torch.batch_norm(var_387, input_vars_2, input_vars_3,
input_vars_4, input_vars_5, False, 0.1,
1e-05, True)
var_393 = torch.relu_(var_392)
var_407 = torch.max_pool2d(var_393, [
3,
3,
], [
2,
2,
], [
1,
1,
], [
1,
1,
], False)
var_426 = torch._convolution(var_407, input_vars_7, None, [
1,
1,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
return var_407, var_426
def backward(ctx, grad_output):
#|z|-rule as proposed in BatchNorm Decomposition for Deep Neural Network Optimisation
input, weight = ctx.saved_tensors
bias, running_mean, running_var, training, momentum, eps = ctx.hparams
output = torch.batch_norm(input, weight, bias, running_mean,
running_var, training, momentum, eps,
torch.backends.cudnn.enabled)
w_in = torch.zeros_like(input)
b_in = torch.zeros_like(input)
for w in range(len(weight)):
w_in[:, w, ...] = abs(weight[w] * input[:, w, ...])
b_in[:, w, ...] = w_in[:, w, ...] + abs(bias[w])
out = w_in / b_in
norm_grad = out * grad_output
# f = (input.grad).cpu()
# mdl = "sf"
# plt.imsave(f"{save_path}{mdl}.png",f)
return norm_grad, None, None, None, None, None, None, None
def jojo_2(self, input_vars_89, input_vars_85, input_vars_87, var_740,
input_vars_86, var_715, input_vars_88):
var_759 = torch._convolution(var_740, input_vars_85, None, [
1,
1,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_764 = torch.batch_norm(var_759, input_vars_86, input_vars_87,
input_vars_88, input_vars_89, False, 0.1,
1e-05, True)
var_766 = torch.add(var_764, var_715, alpha=1)
return var_766
def jojo_1(self, input_vars_95, input_vars_92, input_vars_94, var_786,
input_vars_93):
var_791 = torch.batch_norm(var_786, input_vars_92, input_vars_93,
input_vars_94, input_vars_95, False, 0.1,
1e-05, True)
return var_791
def jojo_5(self, input_vars_170, input_vars_172, var_1411, input_vars_171, input_vars_173):
var_1416 = torch.batch_norm(var_1411, input_vars_170, input_vars_171, input_vars_172, input_vars_173, False, 0.1, 1e-05, True)
return var_1416
def forward(self, x):
assert self.params_set, 'model.set_params(...) must be called before the forward pass'
return torch.batch_norm(x, self.scale, self.bias, self.running_mean,
self.running_var, False, self.momentum,
self.eps, torch.backends.cudnn.enabled)
def jojo_6(self, input_vars_166, input_vars_164, input_vars_167, var_1366, input_vars_165, input_vars_163):
var_1367 = torch.relu_(var_1366.clone())
var_1386 = torch._convolution(var_1367, input_vars_163, None, [1, 1, ], [1, 1, ], [1, 1, ], False, [0, 0, ], 1, False, False, True)
var_1391 = torch.batch_norm(var_1386, input_vars_164, input_vars_165, input_vars_166, input_vars_167, False, 0.1, 1e-05, True)
return var_1391, var_1367
def jojo_11(self, input_vars_135, var_1255, var_1211, input_vars_134, input_vars_136, input_vars_137):
var_1260 = torch.batch_norm(var_1255, input_vars_134, input_vars_135, input_vars_136, input_vars_137, False, 0.1, 1e-05, True)
var_1262 = torch.add(var_1260, var_1211, alpha=1)
return var_1262
def forward(
self,
obs,
deterministic=False,
return_log_prob=False,
pol_idx=None,
optimize_policies=True,
):
"""
Args:
obs (Tensor): Observation(s)
deterministic (bool):
return_log_prob (bool):
pol_idx (int):
optimize_policies (bool):
Returns:
action (Tensor):
pol_info (dict):
"""
h = obs
nbatch = obs.shape[0]
# ############# #
# Shared Layers #
# ############# #
if self.sfc_input is not None:
# h = self.sfc_input(h)
if nbatch > 1:
h = self.sfc_input(h)
else:
h = torch.batch_norm(
h,
self.sfc_input.weight,
self.sfc_input.bias,
self.sfc_input.running_mean,
self.sfc_input.running_var,
True, # TODO: True or False??
self.sfc_input.momentum,
self.sfc_input.eps,
torch.backends.cudnn.enabled)
for ss, fc in enumerate(self._sfcs):
h = fc(h)
if self._mixture_layer_norm:
h = self._sfc_norms[ss](h)
h = self._hidden_activation(h)
# ############## #
# Multi Policies #
# ############## #
hs = [h.clone() for _ in range(self._n_subpolicies)]
# Hidden Layers
if len(self._pfcs) > 0:
for pp in range(self._n_subpolicies):
for ii, fc in enumerate(self._pfcs[pp]):
hs[pp] = fc(hs[pp])
if self._policies_layer_norm:
hs[pp] = self._pfc_norms[pp][ii](hs[pp])
hs[pp] = self._hidden_activation(hs[pp])
subpol_means = \
[self._pol_output_activation(self._pfc_lasts[pp](hs[pp]))
for pp in range(self._n_subpolicies)]
subpols = torch.cat(subpol_means, dim=-1)
if torch.isnan(subpols).any():
raise ValueError('Some subpols are NAN:', subpols)
# ############## #
# Mixing Weigths #
# ############## #
mh = torch.cat([h.clone(), subpols], dim=-1) # N x dZ
if not optimize_policies:
mh = mh.detach()
if len(self._mfcs) > 0:
for mm, mfc in enumerate(self._mfcs):
mh = mfc(mh)
if self._mixture_layer_norm:
mh = self._norm_mfcs[mm](mh)
mh = self._hidden_activation(mh)
# NO nonlinear transformation
mpol_mean = self.mfc_last(mh)
if self.mfc_softmax is not None:
raise NotImplementedError
# mixture_coeff = self.mfc_softmax(mixture_coeff)
# Final Policy
final_pol_inputs = [
ii.unsqueeze(-2) for ii in (subpol_means + [mpol_mean])
]
fph = torch.cat(
final_pol_inputs,
dim=-2,
)
for ff, fpfc in enumerate(self._fpfcs):
fph = fpfc(fph)
if self._final_policy_layer_norm:
fph = self._norm_mfcs[ff](fph)
fph = self._hidden_activation(fph)
means = self._final_pol_output_activation(self.fpfc_last(fph))
log_stds = self._final_pol_output_activation(
self.fpfc_last_log_std(fph))
log_stds = torch.clamp(log_stds, LOG_SIG_MIN, LOG_SIG_MAX)
stds = torch.exp(log_stds)
variances = torch.pow(stds, 2)
if pol_idx is None:
index = self._compo_pol_idx
else:
index = self._pols_idxs[pol_idx]
mean = \
torch.index_select(means, dim=-2, index=index).squeeze(-2)
std = \
torch.index_select(stds, dim=-2, index=index).squeeze(-2)
log_std = \
torch.index_select(log_stds, dim=-2, index=index).squeeze(-2)
variance = \
torch.index_select(variances, dim=-2, index=index).squeeze(-2)
means = \
torch.index_select(means, dim=-2, index=self._pols_idxs).squeeze(-2)
stds = \
torch.index_select(stds, dim=-2, index=self._pols_idxs).squeeze(-2)
log_stds = \
torch.index_select(log_stds, dim=-2, index=self._pols_idxs).squeeze(-2)
variances = \
torch.index_select(variances, dim=-2, index=self._pols_idxs).squeeze(-2)
pre_tanh_value = None
log_prob = None
entropy = None
mean_action_log_prob = None
log_probs = None
pre_tanh_values = None
mixture_coeff = ptu.ones((nbatch, self.n_heads, self.action_dim))
if deterministic:
action = torch.tanh(mean)
actions = torch.tanh(means)
else:
noise = self._normal_dist.sample((nbatch, ))
pre_tanh_value = std * noise + mean
pre_tanh_values = stds * noise.unsqueeze(1) + means
action = torch.tanh(pre_tanh_value)
actions = torch.tanh(pre_tanh_values)
if return_log_prob:
# Log probability: Main Policy
log_prob = -((pre_tanh_value - mean) ** 2) / (2 * variance) \
- torch.log(std) - math.log(math.sqrt(2 * math.pi))
log_prob -= torch.log(1. - action**2 + self._epsilon)
log_prob = log_prob.sum(dim=-1, keepdim=True)
# Log probability: Sub-Policies
log_probs = -((pre_tanh_values - means) ** 2) / (2 * variances) \
- torch.log(stds) - math.log(math.sqrt(2 * math.pi))
log_probs -= torch.log(1. - actions**2 + self._epsilon)
log_probs = log_probs.sum(dim=-1, keepdim=True)
if torch.isnan(action).any():
raise ValueError('ACTION NAN')
if torch.isnan(actions).any():
raise ValueError('ACTION NAN')
info_dict = dict(
mean=mean,
log_std=log_std,
log_prob=log_prob,
entropy=entropy,
std=std,
mean_action_log_prob=mean_action_log_prob,
pre_tanh_value=pre_tanh_value,
# log_mixture_coeff=log_mixture_coeff,
mixing_coeff=mixture_coeff,
pol_actions=actions,
pol_means=means,
pol_stds=stds,
pol_log_stds=log_stds,
pol_log_probs=log_probs,
pol_pre_tanh_values=pre_tanh_values,
)
return action, info_dict
def forward(
self,
obs,
deterministic=False,
return_log_prob=False,
pol_idx=None,
optimize_policies=True,
):
"""
Args:
obs (Tensor): Observation(s)
deterministic (bool):
return_log_prob (bool):
pol_idx (int):
optimize_policies (bool):
Returns:
action (Tensor):
pol_info (dict):
"""
# pol_idx = int(0)
h = obs
nbatch = obs.shape[0]
# ############# #
# Shared Layers #
# ############# #
if self.sfc_input is not None:
# h = self.sfc_input(h)
if nbatch > 1:
h = self.sfc_input(h)
else:
h = torch.batch_norm(
h,
self.sfc_input.weight,
self.sfc_input.bias,
self.sfc_input.running_mean,
self.sfc_input.running_var,
True, # TODO: True or False??
self.sfc_input.momentum,
self.sfc_input.eps,
torch.backends.cudnn.enabled
)
for ss, fc in enumerate(self._sfcs):
h = fc(h)
if self._mixture_layer_norm:
h = self._sfc_norms[ss](h)
h = self._hidden_activation(h)
# ############## #
# Multi Policies #
# ############## #
hs = [h.clone() for _ in range(self._n_subpolicies)]
# Hidden Layers
if len(self._pfcs) > 0:
for pp in range(self._n_subpolicies):
for ii, fc in enumerate(self._pfcs[pp]):
hs[pp] = fc(hs[pp])
if self._policies_layer_norm:
hs[pp] = self._pfc_norms[pp][ii](hs[pp])
hs[pp] = self._hidden_activation(hs[pp])
# Last Mean Layers
means_list = \
[(self._pol_output_activation(self._pfc_lasts[pp](hs[pp]))).unsqueeze(dim=1)
for pp in range(self._n_subpolicies)]
means = torch.cat(means_list, dim=1)
# Last Log-Std Layers
if self.stds is None:
log_stds_list = [
(self._pol_output_activation(
self._pfc_log_std_lasts[pp](hs[pp])
)
).unsqueeze(dim=1)
for pp in range(self._n_subpolicies)]
log_stds = torch.cat(log_stds_list, dim=1)
log_stds = torch.clamp(log_stds, min=LOG_SIG_MIN, max=LOG_SIG_MAX)
stds = torch.exp(log_stds)
variances = stds**2
else:
stds = self.stds
variances = stds**2
log_stds = self.log_std
# ############## #
# Mixing Weigths #
# ############## #
mh = h.clone()
if len(self._mfcs) > 0:
for mm, mfc in enumerate(self._mfcs):
mh = mfc(mh)
if self._mixture_layer_norm:
mh = self._norm_mfcs[mm](mh)
mh = self._hidden_activation(mh)
# NO nonlinear transformation
log_mixture_coeff = \
self.mfc_last(mh).reshape(-1, self._n_subpolicies, self.action_dim)
# log_mixture_coeff = torch.clamp(log_mixture_coeff,
# min=LOG_MIX_COEFF_MIN,
# max=LOG_MIX_COEFF_MAX) # NxK
# if self.mfc_softmax is not None:
mixture_coeff = self.mfc_softmax(log_mixture_coeff)
# mixture_coeff = torch.exp(log_mixture_coeff) \
# / torch.sum(torch.exp(log_mixture_coeff), dim=-1,
# keepdim=True)
if torch.isnan(log_mixture_coeff).any():
raise ValueError('Some mixture coeff(s) is(are) NAN:',
log_mixture_coeff)
if torch.isnan(means).any():
raise ValueError('Some means are NAN:',
means)
if torch.isnan(stds).any():
raise ValueError('Some stds are NAN:',
stds)
if pol_idx is None:
# TODO: CHECK IF NOT PROPAGATING GRADIENTS HERE IS A PROBLEM
# Sample latent variables
z = Multinomial(
logits=log_mixture_coeff.transpose(-2, -1)
).sample().transpose(-2, -1) # NxK
# Choose mixture component corresponding
mean = torch.sum(means*z, dim=-2)
std = torch.sum(stds*z, dim=-2)
log_std = torch.sum(log_stds*z, dim=-2)
variance = torch.sum(variances*z, dim=-2)
else:
index = self._pols_idxs[pol_idx]
mean = \
torch.index_select(means, dim=1, index=index).squeeze(1)
std = \
torch.index_select(stds, dim=1, index=index).squeeze(1)
log_std = \
torch.index_select(log_stds, dim=1, index=index).squeeze(1)
variance = \
torch.index_select(variances, dim=1, index=index).squeeze(1)
pre_tanh_value = None
log_prob = None
entropy = None
mean_action_log_prob = None
log_probs = None
pre_tanh_values = None
if deterministic:
action = torch.tanh(mean)
actions = torch.tanh(means)
else:
noise = self._normal_dist.sample((nbatch,))
pre_tanh_value = std*noise + mean
pre_tanh_values = stds*noise.unsqueeze(1) + means
action = torch.tanh(pre_tanh_value)
actions = torch.tanh(pre_tanh_values)
if return_log_prob:
# temp_pre_tanh_vals = pre_tanh_values
# temp_actions = actions
temp_pre_tanh_vals = pre_tanh_value.unsqueeze(-2).expand((nbatch, self.n_heads, self.action_dim))
temp_actions = action.unsqueeze(-2).expand((nbatch, self.n_heads, self.action_dim))
# Log probability: Sub-Policies | log(x|z)
# log_probs = -((pre_tanh_values - means) ** 2) / (2 * variances) \
temp_log_probs = -((temp_pre_tanh_vals - means) ** 2) / (2 * variances) \
- log_stds - math.log(math.sqrt(2 * math.pi))
# log_probs -= torch.log(1. - temp_actions**2 + self._epsilon)
# Log probability: Main Policy
log_prob = (torch.logsumexp(temp_log_probs.detach() + log_mixture_coeff,
dim=-2, keepdim=True)
- torch.logsumexp(log_mixture_coeff, dim=-2,
keepdim=True)
).squeeze(-2)
log_prob -= torch.log(1. - action**2 + self._epsilon)
log_prob = log_prob.sum(dim=-1, keepdim=True)
log_probs = -((pre_tanh_values - means) ** 2) / (2 * variances) \
- log_stds - math.log(math.sqrt(2 * math.pi))
log_probs = log_probs.sum(dim=-1, keepdim=True)
if (torch.abs(log_probs) > 1e5).any():
print('---MEAN0--')
print(means[:, 0, :])
print('-----')
print('-----')
print('-----')
print('---MEAN1--')
print(means[:, 1, :])
print('-----')
print('-----')
print('-----')
print('--STD---')
print(stds[:, 1, :])
print('-----')
print('-----')
print('-----')
print('--PRE_TANH---')
print(temp_pre_tanh_vals[:, 1, :])
print('-----')
print('-----')
print('-----')
print('--LOG_PROB---')
print(log_probs[:, 1])
raise ValueError
if torch.isnan(log_prob).any():
raise ValueError('LOG_PROB NAN')
if torch.isnan(log_probs).any():
raise ValueError('LOG_PROBS NAN')
if torch.isnan(action).any():
raise ValueError('ACTION NAN')
if torch.isnan(actions).any():
raise ValueError('ACTION NAN')
info_dict = dict(
mean=mean,
log_std=log_std,
log_prob=log_prob,
entropy=entropy,
std=std,
mean_action_log_prob=mean_action_log_prob,
pre_tanh_value=pre_tanh_value,
# log_mixture_coeff=log_mixture_coeff,
mixing_coeff=mixture_coeff,
pol_actions=actions,
pol_means=means,
pol_stds=stds,
pol_log_stds=log_stds,
pol_log_probs=log_probs,
pol_pre_tanh_values=pre_tanh_values,
)
return action, info_dict
def forward_(self, input_vars):
var_407, var_426 = torch.utils.checkpoint.checkpoint(
self.jojo_8, input_vars[7], input_vars[5], input_vars[4],
input_vars[1], input_vars[2], input_vars[3], input_vars[0])
var_431 = torch.batch_norm(var_426, input_vars[8], input_vars[9],
input_vars[10], input_vars[11], False, 0.1,
1e-05, True)
var_459, var_503 = torch.utils.checkpoint.checkpoint(
self.jojo_7, input_vars[25], var_431, input_vars[16],
input_vars[13], input_vars[15], input_vars[23], input_vars[19],
input_vars[21], input_vars[22], var_407, input_vars[17],
input_vars[14], input_vars[20])
var_508 = torch.batch_norm(var_503, input_vars[26], input_vars[27],
input_vars[28], input_vars[29], False, 0.1,
1e-05, True)
var_560, var_584 = torch.utils.checkpoint.checkpoint(
self.jojo_6, input_vars[44], input_vars[34], input_vars[37],
var_459, input_vars[45], var_508, input_vars[31], input_vars[38],
input_vars[41], input_vars[47], input_vars[40], input_vars[43],
input_vars[33], input_vars[35], input_vars[46], input_vars[32],
input_vars[39])
var_586 = torch.add(var_560, var_584, alpha=1)
var_663, var_639 = torch.utils.checkpoint.checkpoint(
self.jojo_5, input_vars[50], var_586, input_vars[51],
input_vars[49], input_vars[57], input_vars[64], input_vars[52],
input_vars[55], input_vars[61], input_vars[63], input_vars[58],
input_vars[56], input_vars[53], input_vars[65], input_vars[62],
input_vars[59])
var_664 = torch.relu_(var_663)
var_683 = torch._convolution(var_664, input_vars[67], None, [
1,
1,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_712, var_688 = torch.utils.checkpoint.checkpoint(
self.jojo_4, input_vars[69], var_683, input_vars[75],
input_vars[71], input_vars[68], input_vars[70], input_vars[74],
input_vars[77], input_vars[76], var_639, input_vars[73])
var_714 = torch.add(var_688, var_712, alpha=1)
var_715 = torch.utils.checkpoint.checkpoint(self.jojo_3, var_714)
var_734 = torch._convolution(var_715, input_vars[79], None, [
1,
1,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_739 = torch.batch_norm(var_734, input_vars[80], input_vars[81],
input_vars[82], input_vars[83], False, 0.1,
1e-05, True)
var_740 = torch.relu_(var_739)
var_766 = torch.utils.checkpoint.checkpoint(
self.jojo_2, input_vars[89], input_vars[85], input_vars[87],
var_740, input_vars[86], var_715, input_vars[88])
var_767 = torch.relu_(var_766)
var_786 = torch._convolution(var_767, input_vars[91], None, [
2,
2,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_791 = torch.utils.checkpoint.checkpoint(self.jojo_1,
input_vars[95],
input_vars[92],
input_vars[94], var_786,
input_vars[93])
var_792 = torch.relu_(var_791)
var_811 = torch._convolution(var_792, input_vars[97], None, [
1,
1,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_816 = torch.batch_norm(var_811, input_vars[98], input_vars[99],
input_vars[100], input_vars[101], False,
0.1, 1e-05, True)
var_835 = torch._convolution(var_767, input_vars[103], None, [
2,
2,
], [
0,
0,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_840 = torch.batch_norm(var_835, input_vars[104], input_vars[105],
input_vars[106], input_vars[107], False,
0.1, 1e-05, True)
var_842 = torch.add(var_816, var_840, alpha=1)
var_843 = torch.relu_(var_842)
var_862 = torch._convolution(var_843, input_vars[109], None, [
1,
1,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_867 = torch.batch_norm(var_862, input_vars[110], input_vars[111],
input_vars[112], input_vars[113], False,
0.1, 1e-05, True)
var_868 = torch.relu_(var_867)
var_887 = torch._convolution(var_868, input_vars[115], None, [
1,
1,
], [
1,
1,
], [
1,
1,
], False, [
0,
0,
], 1, False, False, True)
var_892 = torch.batch_norm(var_887, input_vars[116], input_vars[117],
input_vars[118], input_vars[119], False,
0.1, 1e-05, True)
var_894 = torch.add(var_892, var_843, alpha=1)
var_895 = torch.relu_(var_894)
var_911 = self.layer_0(var_895)
var_914 = torch.flatten(var_911, 1, -1)
var_915 = torch.utils.checkpoint.checkpoint(self.jojo_0,
input_vars[121])
var_918 = torch.addmm(input_vars[122],
var_914,
var_915,
beta=1,
alpha=1)
return var_918
def jojo_13(self, input_vars_117, input_vars_119, input_vars_118, input_vars_116, var_1178):
var_1183 = torch.batch_norm(var_1178, input_vars_116, input_vars_117, input_vars_118, input_vars_119, False, 0.1, 1e-05, True)
var_1184 = torch.relu_(var_1183)
return var_1184
def jojo_15(self, input_vars_100, var_1082, input_vars_97, input_vars_99, input_vars_98, input_vars_101):
var_1083 = torch.relu_(var_1082.clone())
var_1102 = torch._convolution(var_1083, input_vars_97, None, [2, 2, ], [1, 1, ], [1, 1, ], False, [0, 0, ], 1, False, False, True)
var_1107 = torch.batch_norm(var_1102, input_vars_98, input_vars_99, input_vars_100, input_vars_101, False, 0.1, 1e-05, True)
return var_1107, var_1083
def jojo_16(self, input_vars_95, input_vars_94, input_vars_92, var_1075, input_vars_93):
var_1080 = torch.batch_norm(var_1075, input_vars_92, input_vars_93, input_vars_94, input_vars_95, False, 0.1, 1e-05, True)
return var_1080
def jojo_17(self, input_vars_89, var_1031, input_vars_86, input_vars_85, input_vars_87, input_vars_88):
var_1050 = torch._convolution(var_1031, input_vars_85, None, [1, 1, ], [1, 1, ], [1, 1, ], False, [0, 0, ], 1, False, False, True)
var_1055 = torch.batch_norm(var_1050, input_vars_86, input_vars_87, input_vars_88, input_vars_89, False, 0.1, 1e-05, True)
var_1056 = torch.relu_(var_1055)
return var_1056
