Create AppleAI-converter-colab.py
Browse files- AppleAI-converter-colab.py +74 -0
AppleAI-converter-colab.py
ADDED
@@ -0,0 +1,74 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
os.system('pip install mlx')
|
3 |
+
import mlx as mx
|
4 |
+
import mlx.nn as mx_nn
|
5 |
+
import torch
|
6 |
+
import torch.nn as nn
|
7 |
+
import numpy as np
|
8 |
+
|
9 |
+
device = torch.device('cuda')
|
10 |
+
|
11 |
+
CONFIG = {
|
12 |
+
"model_path": "NeoChess/chessy_model.pth",
|
13 |
+
"backup_model_path": "NeoChess/chessy_modelt-1.pth",
|
14 |
+
}
|
15 |
+
|
16 |
+
class NN1(nn.Module):
|
17 |
+
def __init__(self):
|
18 |
+
super().__init__()
|
19 |
+
self.embedding = nn.Embedding(13, 64)
|
20 |
+
self.attention = nn.MultiheadAttention(embed_dim=64, num_heads=16)
|
21 |
+
self.neu = 512
|
22 |
+
self.neurons = nn.Sequential(
|
23 |
+
nn.Linear(4096, self.neu),
|
24 |
+
nn.ReLU(),
|
25 |
+
nn.Linear(self.neu, self.neu),
|
26 |
+
nn.ReLU(),
|
27 |
+
nn.Linear(self.neu, self.neu),
|
28 |
+
nn.ReLU(),
|
29 |
+
nn.Linear(self.neu, self.neu),
|
30 |
+
nn.ReLU(),
|
31 |
+
nn.Linear(self.neu, self.neu),
|
32 |
+
nn.ReLU(),
|
33 |
+
nn.Linear(self.neu, self.neu),
|
34 |
+
nn.ReLU(),
|
35 |
+
nn.Linear(self.neu, self.neu),
|
36 |
+
nn.ReLU(),
|
37 |
+
nn.Linear(self.neu, self.neu),
|
38 |
+
nn.ReLU(),
|
39 |
+
nn.Linear(self.neu, self.neu),
|
40 |
+
nn.ReLU(),
|
41 |
+
nn.Linear(self.neu, self.neu),
|
42 |
+
nn.ReLU(),
|
43 |
+
nn.Linear(self.neu, self.neu),
|
44 |
+
nn.ReLU(),
|
45 |
+
nn.Linear(self.neu, self.neu),
|
46 |
+
nn.ReLU(),
|
47 |
+
nn.Linear(self.neu, self.neu),
|
48 |
+
nn.ReLU(),
|
49 |
+
nn.Linear(self.neu, 64),
|
50 |
+
nn.ReLU(),
|
51 |
+
nn.Linear(64, 4)
|
52 |
+
)
|
53 |
+
|
54 |
+
def forward(self, x):
|
55 |
+
x = self.embedding(x)
|
56 |
+
x = x.permute(1, 0, 2)
|
57 |
+
attn_output, _ = self.attention(x, x, x)
|
58 |
+
x = attn_output.permute(1, 0, 2).contiguous()
|
59 |
+
x = x.view(x.size(0), -1)
|
60 |
+
x = self.neurons(x)
|
61 |
+
return x
|
62 |
+
|
63 |
+
model = NN1().to(device)
|
64 |
+
try:
|
65 |
+
model.load_state_dict(torch.load(CONFIG['model_path'], map_location=device))
|
66 |
+
print(f"Loaded model from {CONFIG['model_path']}")
|
67 |
+
except FileNotFoundError:
|
68 |
+
try:
|
69 |
+
model.load_state_dict(torch.load(CONFIG["backup_model_path"], map_location=device))
|
70 |
+
print(f"Loaded backup model from {CONFIG['backup_model_path']}")
|
71 |
+
except FileNotFoundError:
|
72 |
+
print("No model file found, starting from scratch.")
|
73 |
+
weights = {k: v.detach().cpu().numpy() for k, v in model.state_dict().items()}
|
74 |
+
np.savez("chessy_model_mlx.npz", **weights)
|