ELF2LeelaZ.py

### usage python ELF2LeelaZ.py > weights.txt

import torch

m = torch.load("pretrained-go-19x19-v0.bin")

opt = m["options"]

num_block = opt['num_block']
dim = opt['dim']
bn = opt['bn']
leaky_relu = opt['leaky_relu']


wei = m["state_dict"]
#print(wei.keys())
#print(wei["pi_linear.weight"].size()) # 362 x 722

def print_tensor(wei, name):
    w = wei[name].view(-1)
    #print(wei[name].size()) # to check size 

    s = [str(w[i].item()) for i in range(w.size(0))]
    print(" ".join(s))

def print_conv(wei, name):
    #name = name + "_conv"
    print_tensor(wei, name + ".0.weight")
    print_tensor(wei, name + ".0.bias")
    #print_tensor(wei, name + ".1.weight")
    #print_tensor(wei, name + ".1.bias")
    print_tensor(wei, name + ".1.running_mean")
    print_tensor(wei, name + ".1.running_var")
    
def print_resnet(wei, idx):
    name1 = "resnet.module.resnet.%d.conv_lower" % idx
    name2 = "resnet.module.resnet.%d.conv_upper" % idx
    print_conv(wei, name1)
    print_conv(wei, name2)

    
print("1") # version
print_conv(wei, 'init_conv')
    
for i in range(num_block):
    print_resnet(wei, i)

print_conv(wei, 'pi_final_conv')
print_tensor(wei, "pi_linear.weight")
print_tensor(wei, "pi_linear.bias")

print_conv(wei, 'value_final_conv')
print_tensor(wei, "value_linear1.weight")
print_tensor(wei, "value_linear1.bias")
print_tensor(wei, "value_linear2.weight")
print_tensor(wei, "value_linear2.bias")

whats the issues