def __init__(self, size):
self._cpu_storage = torch.Storage(size)
self._gpu_storages = []
if torch.cuda.is_available():
for device_idx in range(torch.cuda.device_count()):
with torch.cuda.device(device_idx):
self._gpu_storages.append(torch.Storage(size).cuda())
def __init__(self,
in_num,
out_num,
layer_num,
max_link,
storage_size=1024):
input_storage_1 = torch.Storage(storage_size)
input_storage_2 = torch.Storage(storage_size)
self.shared_allocation_1 = _SharedAllocation(input_storage_1)
self.shared_allocation_2 = _SharedAllocation(input_storage_2)
max_in_num = in_num + out_num * max_link
self.final_num_features = max_in_num
self.saved_features = []
self.max_link = max_link
super(_IntermediaBlock, self).__init__()
print('creating intermedia block ...')
self.adapters = []
for i in range(0, layer_num - 1):
if i < max_link:
tmp_in_num = in_num + (i + 1) * out_num
else:
tmp_in_num = max_in_num
print('intermedia layer %d input channel number is %d' %
(i, tmp_in_num))
self.adapters.append(
_EfficientDensenetBottleneck(self.shared_allocation_1,
self.shared_allocation_2,
tmp_in_num, out_num))
self.adapters = nn.ModuleList(self.adapters)
print('intermedia layer output channel number is %d' % out_num)
def __init__(self,
in_num,
neck_size,
growth_rate,
layer_num,
max_link,
storage_size=1024,
requires_skip=True,
is_up=False):
input_storage_1 = torch.Storage(storage_size)
input_storage_2 = torch.Storage(storage_size)
self.shared_allocation_1 = _SharedAllocation(input_storage_1)
self.shared_allocation_2 = _SharedAllocation(input_storage_2)
self.saved_features = []
self.max_link = max_link
self.requires_skip = requires_skip
super(_DenseBlock, self).__init__()
max_in_num = in_num + max_link * growth_rate
self.final_num_features = max_in_num
self.layers = []
#print('layer number is %d' % layer_num)
for i in range(0, layer_num):
if i < max_link:
tmp_in_num = in_num + i * growth_rate
else:
tmp_in_num = max_in_num
#print('layer %d input channel number is %d' % (i, tmp_in_num))
self.layers.append(
_DenseLayer(self.shared_allocation_1, self.shared_allocation_2,
tmp_in_num, neck_size, growth_rate))
self.layers = nn.ModuleList(self.layers)
self.adapters_ahead = []
adapter_in_nums = []
adapter_out_num = in_num
if is_up:
adapter_out_num = adapter_out_num / 2
for i in range(0, layer_num):
if i < max_link:
tmp_in_num = in_num + (i + 1) * growth_rate
else:
tmp_in_num = max_in_num + growth_rate
adapter_in_nums.append(tmp_in_num)
#print('adapter %d input channel number is %d' % (i, adapter_in_nums[i]))
self.adapters_ahead.append(
_EfficientDensenetBottleneck(self.shared_allocation_1,
self.shared_allocation_2,
adapter_in_nums[i],
adapter_out_num))
self.adapters_ahead = nn.ModuleList(self.adapters_ahead)
#print('adapter output channel number is %d' % adapter_out_num)
if requires_skip:
print('creating skip layers ...')
self.adapters_skip = []
for i in range(0, layer_num):
self.adapters_skip.append(
_EfficientDensenetBottleneck(self.shared_allocation_1,
self.shared_allocation_2,
adapter_in_nums[i],
adapter_out_num))
self.adapters_skip = nn.ModuleList(self.adapters_skip)
def __init__(self, num_layers, num_input_features, bn_size, growth_rate, drop_rate, storage_size=1024):
input_storage_1 = torch.Storage(storage_size)
input_storage_2 = torch.Storage(storage_size)
self.final_num_features = num_input_features + (growth_rate * num_layers)
self.shared_allocation_1 = _SharedAllocation(input_storage_1)
self.shared_allocation_2 = _SharedAllocation(input_storage_2)
super(_DenseBlock, self).__init__()
for i in range(num_layers):
layer = _DenseLayer(self.shared_allocation_1, self.shared_allocation_2, num_input_features + i * growth_rate,
growth_rate, bn_size, drop_rate)
self.add_module('denselayer%d' % (i + 1), layer)
def test_backward_computes_backward_pass():
bn_weight = torch.randn(8).cuda()
bn_bias = torch.randn(8).cuda()
bn_running_mean = torch.randn(8).cuda()
bn_running_var = torch.randn(8).abs().cuda()
conv_weight = torch.randn(4, 8, 1, 1).cuda()
input_1 = torch.randn(4, 6, 4, 4).cuda()
input_2 = torch.randn(4, 2, 4, 4).cuda()
layer = nn.Sequential(OrderedDict([
('norm', nn.BatchNorm2d(8)),
('relu', nn.ReLU(inplace=True)),
('conv', nn.Conv2d(8, 4, bias=None, kernel_size=1, stride=1)),
])).cuda()
layer.train()
layer.norm.weight.data.copy_(bn_weight)
layer.norm.bias.data.copy_(bn_bias)
layer.norm.running_mean.copy_(bn_running_mean)
layer.norm.running_var.copy_(bn_running_var)
layer.conv.weight.data.copy_(conv_weight)
input_1_var = Variable(input_1, requires_grad=True)
input_2_var = Variable(input_2, requires_grad=True)
out_var = layer(torch.cat([input_1_var, input_2_var], dim=1))
out_var.sum().backward()
storage_1 = torch.Storage(4 * 8 * 3 * 3).cuda()
storage_2 = torch.Storage(4 * 8 * 3 * 3).cuda()
layer_efficient = _EfficientDensenetBottleneck(
_SharedAllocation(storage_1), _SharedAllocation(storage_2), 8, 4
).cuda()
layer_efficient.train()
layer_efficient.norm_weight.data.copy_(bn_weight)
layer_efficient.norm_bias.data.copy_(bn_bias)
layer_efficient.norm_running_mean.copy_(bn_running_mean)
layer_efficient.norm_running_var.copy_(bn_running_var)
layer_efficient.conv_weight.data.copy_(conv_weight)
input_efficient_1_var = Variable(input_1, requires_grad=True)
input_efficient_2_var = Variable(input_2, requires_grad=True)
out_efficient_var = layer_efficient([input_efficient_1_var, input_efficient_2_var])
out_efficient_var.sum().backward()
# print(input_1_var.grad.data[:, 0], input_efficient_1_var.grad.data[:, 0])
assert(almost_equal(out_var.data, out_efficient_var.data))
assert(almost_equal(layer.norm.running_mean, layer_efficient.norm_running_mean))
assert(almost_equal(layer.norm.running_var, layer_efficient.norm_running_var))
assert(almost_equal(layer.conv.weight.grad.data, layer_efficient.conv_weight.grad.data))
assert(almost_equal(layer.norm.weight.grad.data, layer_efficient.norm_weight.grad.data))
assert(almost_equal(layer.norm.bias.grad.data, layer_efficient.norm_bias.grad.data))
assert(almost_equal(input_1_var.grad.data, input_efficient_1_var.grad.data))
assert(almost_equal(input_2_var.grad.data, input_efficient_2_var.grad.data))
def create_multi_gpu_storage(size=1024):
multi_storage = []
device_cnt = torch.cuda.device_count()
for device_no in range(device_cnt):
with torch.cuda.device(device_no):
multi_storage.append(torch.Storage(size).cuda())
return multi_storage
def test_parse_numpy_int(self, device):
# Only concrete class can be given where "Type[number[_64Bit]]" is expected
self.assertRaisesRegex(RuntimeError, "Overflow", lambda: torch.mean(
torch.randn(1, 1), np.uint64(-1))) # type: ignore[call-overload]
# https://github.com/pytorch/pytorch/issues/29252
for nptype in [np.int16, np.int8, np.uint8, np.int32, np.int64]:
scalar = 3
np_arr = np.array([scalar], dtype=nptype)
np_val = np_arr[0]
# np integral type can be treated as a python int in native functions with
# int parameters:
self.assertEqual(
torch.ones(5).diag(scalar),
torch.ones(5).diag(np_val))
self.assertEqual(
torch.ones([2, 2, 2, 2]).mean(scalar),
torch.ones([2, 2, 2, 2]).mean(np_val))
# numpy integral type parses like a python int in custom python bindings:
self.assertEqual(torch.Storage(np_val).size(),
scalar) # type: ignore[attr-defined]
tensor = torch.tensor([2], dtype=torch.int)
tensor[0] = np_val
self.assertEqual(tensor[0], np_val)
# Original reported issue, np integral type parses to the correct
# PyTorch integral type when passed for a `Scalar` parameter in
# arithmetic operations:
t = torch.from_numpy(np_arr)
self.assertEqual((t + np_val).dtype, t.dtype)
self.assertEqual((np_val + t).dtype, t.dtype)
def get_up_network(env_name, num):
import sys
import os
sys.path.append(
os.path.join(os.path.abspath(os.path.dirname(__file__)),
'PT/policy_transfer/uposi'))
sys.path.append(
os.path.join(os.path.abspath(os.path.dirname(__file__)),
'PT/baselines'))
from a2c_ppo_acktr import algo, utils
from a2c_ppo_acktr.algo import gail
from a2c_ppo_acktr.arguments import get_args
from a2c_ppo_acktr.envs import make_vec_envs
from a2c_ppo_acktr.model import Policy
from a2c_ppo_acktr.storage import RolloutStorage
env_name = env_name[:-5]
if 'Dart' in env_name:
path = f"/home/hza/policy_transfer/PT/trained_models/ppo/UP_{env_name}_{num}.pt"
else:
path = f"/home/hza/policy_transfer/PT/trained_models/ppo/UP_{env_name}_{num}.pt"
result = torch.load(path, map_location=lambda a, b: torch.Storage().cuda())
actor_critic = result[0]
actor_critic.cuda()
ob_rms = result[1]
return UP(actor_critic, ob_rms)
def test_forward_eval_mode_computes_forward_pass():
momentum = 0.1
eps = 1e-5
weight = torch.randn(10).cuda()
bias = torch.randn(10).cuda()
running_mean = torch.randn(10).cuda()
running_var = torch.randn(10).abs().cuda()
input_1 = torch.randn(4, 5).cuda()
input_2 = torch.randn(4, 5).cuda()
storage = torch.Storage(40).cuda()
bn = F.batch_norm(input=Variable(torch.cat([input_1, input_2], dim=1)),
running_mean=running_mean,
running_var=running_var,
weight=Parameter(weight),
bias=Parameter(bias),
training=False,
momentum=momentum,
eps=eps).data
input_efficient = torch.cat([input_1, input_2], dim=1)
func = _EfficientBatchNorm(storage=storage,
running_mean=running_mean,
running_var=running_var,
training=False,
momentum=momentum,
eps=eps)
bn_efficient = func.forward(weight, bias, input_efficient)
assert (almost_equal(bn, bn_efficient))
assert (bn_efficient.storage().data_ptr() == storage.data_ptr())
def test_backward_train_mode_computes_forward_pass():
momentum = 0.1
eps = 1e-5
weight = torch.randn(10).cuda()
bias = torch.randn(10).cuda()
running_mean = torch.randn(10).cuda()
running_var = torch.randn(10).abs().cuda()
weight_efficient = weight.clone()
bias_efficient = bias.clone()
running_mean_efficient = running_mean.clone()
running_var_efficient = running_var.clone()
input_1 = torch.randn(4, 5).cuda()
input_2 = torch.randn(4, 5).cuda()
storage = torch.Storage(40).cuda()
input_var = Variable(torch.cat([input_1, input_2], dim=1),
requires_grad=True)
weight_var = Parameter(weight)
bias_var = Parameter(bias)
bn_var = F.batch_norm(input=input_var,
running_mean=running_mean,
running_var=running_var,
weight=weight_var,
bias=bias_var,
training=True,
momentum=momentum,
eps=eps)
bn = bn_var.data
bn_var.backward(gradient=input_var.data.clone().fill_(1))
input_grad = input_var.grad.data
weight_grad = weight_var.grad.data
bias_grad = bias_var.grad.data
input_efficient = torch.cat([input_1, input_2], dim=1)
input_efficient_orig = input_efficient.clone()
func = _EfficientBatchNorm(storage=storage,
running_mean=running_mean_efficient,
running_var=running_var_efficient,
training=True,
momentum=momentum,
eps=eps)
bn_efficient = func.forward(weight_efficient, bias_efficient,
input_efficient)
grad_out_efficient = bn_efficient.clone().fill_(1)
weight_grad_efficient, bias_grad_efficient, input_grad_efficient = func.backward(
weight_efficient, bias_efficient, input_efficient_orig,
grad_out_efficient)
assert (almost_equal(bn, bn_efficient))
assert (grad_out_efficient.storage().data_ptr() ==
input_grad_efficient.storage().data_ptr())
assert (almost_equal(input_grad, input_grad_efficient))
assert (almost_equal(weight_grad, weight_grad_efficient))
assert (almost_equal(bias_grad, bias_grad_efficient))
def test_forward_training_false_computes_forward_pass():
bn_weight = torch.randn(8).cuda()
bn_bias = torch.randn(8).cuda()
bn_running_mean = torch.randn(8).cuda()
bn_running_var = torch.randn(8).abs().cuda()
conv_weight = torch.randn(4, 8, 3, 3).cuda()
input_1 = torch.randn(4, 6, 4, 4).cuda()
input_2 = torch.randn(4, 2, 4, 4).cuda()
layer = nn.Sequential(OrderedDict([
('norm', nn.BatchNorm2d(8)),
('relu', nn.ReLU(inplace=True)),
('conv', nn.Conv2d(8, 4, bias=None, kernel_size=3, stride=1, padding=1)),
])).cuda()
layer.eval()
layer.norm.weight.data.copy_(bn_weight)
layer.norm.bias.data.copy_(bn_bias)
layer.norm.running_mean.copy_(bn_running_mean)
layer.norm.running_var.copy_(bn_running_var)
layer.conv.weight.data.copy_(conv_weight)
input_1_var = Variable(input_1)
input_2_var = Variable(input_2)
out_var = layer(torch.cat([input_1_var, input_2_var], dim=1))
storage_1 = torch.Storage(4 * 8 * 3 * 3).cuda()
storage_2 = torch.Storage(4 * 8 * 3 * 3).cuda()
layer_efficient = _EfficientDensenetBottleneck(
_SharedAllocation(storage_1), _SharedAllocation(storage_2), 8, 4
).cuda()
layer_efficient.eval()
layer_efficient.norm_weight.data.copy_(bn_weight)
layer_efficient.norm_bias.data.copy_(bn_bias)
layer_efficient.norm_running_mean.copy_(bn_running_mean)
layer_efficient.norm_running_var.copy_(bn_running_var)
layer_efficient.conv_weight.data.copy_(conv_weight)
input_efficient_1_var = Variable(input_1)
input_efficient_2_var = Variable(input_2)
out_efficient_var = layer_efficient([input_efficient_1_var, input_efficient_2_var])
assert(almost_equal(out_var.data, out_efficient_var.data))
def forward(self, x):
shared_alloc = torch.Storage().type(x.storage().type()) if self.efficient else None
for module in self.features.children():
if isinstance(module, _DenseBlock):
x = module(x, shared_alloc)
else:
x = module(x)
x = F.relu(x)
x = F.adaptive_avg_pool2d(x, output_size=1).view(x.size(0), -1)
# x = F.dropout(x, p=0.5, training=self.training)
out = self.classifier(x)
return out
def test_tolist(self, device):
list0D = []
tensor0D = torch.Tensor(list0D)
self.assertEqual(tensor0D.tolist(), list0D)
table1D = [1, 2, 3]
tensor1D = torch.Tensor(table1D)
storage = torch.Storage(table1D)
self.assertEqual(tensor1D.tolist(), table1D)
self.assertEqual(storage.tolist(), table1D)
self.assertEqual(tensor1D.tolist(), table1D)
self.assertEqual(storage.tolist(), table1D)
table2D = [[1, 2], [3, 4]]
tensor2D = torch.Tensor(table2D)
self.assertEqual(tensor2D.tolist(), table2D)
tensor3D = torch.Tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])
tensorNonContig = tensor3D.select(1, 1)
self.assertFalse(tensorNonContig.is_contiguous())
self.assertEqual(tensorNonContig.tolist(), [[3, 4], [7, 8]])
def test_storage_can_be_converted_to_python_object(self):
with enable_torch_dispatch_mode(LoggingTensorMode):
s = torch.Storage()
z = LoggingTensorMode(torch.empty([]))
z.set_(s)