def from_weights_file(filename, train=False, cuda=False):
filters, blocks, weights = read_weights_file(filename)
net = LeelaModel(filters, blocks)
if not train:
net.eval()
for p in net.parameters():
p.requires_grad = False
parameters = []
for module_name, module in net.named_modules():
class_name = module.__class__.__name__
for typ in ('weight', 'bias', 'running_mean', 'running_var'):
param = getattr(module, typ, None)
if param is not None:
parameters.append((module_name, class_name, typ, param))
for i, w in enumerate(weights):
w = torch.Tensor(w)
module_name, class_name, typ, param = parameters[i]
# print(f"{tuple(w.size())} -- {module_name} - {class_name} - {typ}: {tuple(param.size())}")
if class_name == 'CenteredBatchNorm2d' and typ == 'bias':
# Remember bias so it can be updated to a BatchNorm beta when the running_var is seen
bn_bias_param = param
if class_name == 'CenteredBatchNorm2d' and typ == 'running_var':
# print("Updating bias")
std = torch.sqrt(w + 1e-5)
bn_bias_param.data.div_(std.view_as(param))
param.data.copy_(w.view_as(param))
if cuda:
print("Enabling CUDA!")
net.cuda()
return net
def from_weights_file(filename, train=False, cuda=False, half=False):
if cuda:
torch.backends.cudnn.benchmark = True
filters, blocks, weights = read_weights_file(filename)
net = LeelaModel(filters, blocks)
if not train:
net.eval()
for p in net.parameters():
p.requires_grad = False
parameters = []
for module_name, module in net.named_modules():
class_name = module.__class__.__name__
for typ in ('weight', 'bias', 'mean', 'stddiv'):
param = getattr(module, typ, None)
if param is not None:
parameters.append((module_name, class_name, typ, param))
for i, w in enumerate(weights):
w = torch.Tensor(w)
module_name, class_name, typ, param = parameters[i]
# print(f"{tuple(w.size())} -- {module_name} - {class_name} - {typ}: {tuple(param.size())}")
if class_name == 'Normalization' and typ == 'stddiv':
# print('NStddiv')
w = 1 / torch.sqrt(w + 1e-5)
param.data.copy_(w.view_as(param))
if half:
net.half()
if cuda:
print("Enabling CUDA!")
net.cuda()
net.prenormalize()
return net
def from_weights_file(filename, train=False, cuda=False, half=False):
if cuda:
torch.backends.cudnn.benchmark = True
filters, blocks, weights = read_weights_file(filename)
net = LeelaModel(filters, blocks)
if not train:
net.eval()
for p in net.parameters():
p.requires_grad = False
parameters = []
for module_name, module in net.named_modules():
class_name = module.__class__.__name__
for typ in ('weight', 'bias', 'running_mean', 'running_var'):
param = getattr(module, typ, None)
if param is not None:
if class_name == 'CenteredBatchNorm2d' and typ == 'weight':
continue
parameters.append((module_name, class_name, typ, param))
for i, w in enumerate(weights):
w = torch.Tensor(w)
module_name, class_name, typ, param = parameters[i]
# print(param.shape)
# print(f"{tuple(w.size())} -- {module_name} - {class_name} - {typ}: {tuple(param.size())}")
if class_name == 'CenteredBatchNorm2d' and typ == 'bias':
# Remember bias so it can be updated to a BatchNorm beta when the running_var is seen
bn_bias_param = param
if class_name == 'CenteredBatchNorm2d' and typ == 'running_var':
# print("Updating bias")
std = torch.sqrt(w + 1e-5)
bn_bias_param.detach().div_(std.view_as(param))
param.data.copy_(w.view_as(param))
if half:
net.half()
if cuda:
print("Enabling CUDA!")
net.cuda()
return net
# Simple test to verify saving weights...
# net = load_network()
#
# net.model.save_weights_file('test_weights_2.txt')
#
# import numpy as np
#
# with open('test_weights_2.txt') as f1:
# with open('weights_run1_21754.txt') as f2:
# for l1, l2 in zip(f1, f2):
# l1 = np.array([float(s) for s in l1.strip().split()])
# l2 = np.array([float(s) for s in l2.strip().split()])
# s = sum((l1 - l2) ** 2) / len(l1)
# print(s)
def __init__(self, weights_filename):
filters, blocks, weights = read_weights_file(weights_filename)
cfg = yaml.safe_load(YAMLCFG)
cfg['model']['filters'] = filters
cfg['model']['residual_blocks'] = blocks
cfg['name'] = 'online-{}x{}'.format(filters, blocks)
print(yaml.dump(cfg, default_flow_style=False))
x = [
tf.placeholder(tf.float32, [None, 112, 8 * 8]),
tf.placeholder(tf.float32, [None, 1858]),
tf.placeholder(tf.float32, [None, 1])
]
self.tfp = tfprocess.TFProcess(cfg)
self.tfp.init_net(x)
self.tfp.replace_weights(weights)