tinyChessNet/net.py at main · wildercb/tinyChessNet · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
# Now we re create the network with tinygrad
import tinygrad
import torch
from tinygrad.tensor import Tensor
from tinygrad.tensor import *
import tinygrad.nn.optim as optim
from tinygrad.helpers import dtypes
from tinygrad.nn.state import safe_save, safe_load, get_state_dict, load_state_dict
from tinygrad.jit import TinyJit
#!/usr/bin/env python3
import numpy as np
import tinygrad.nn as nn
import torch.nn as nn2
from torch.utils.data import Dataset
from sklearn.metrics import mean_squared_error as MSE
import math
from typing import Union
CUDA=1
# import tinygrad.nn.functional as F
# from torch.utils.data import Dataset
# from torch.nn.utils import clip_grad_norm_
# from torch import optim
class ChessValueDataset(Dataset):
  def __init__(self):
    dat = np.load("processed/dataset_2Mt.npz")
    self.X = dat['arr_0']
    self.Y = dat['arr_1']
    print("loaded", self.X.shape, self.Y.shape)

  def __len__(self):
    return self.X.shape[0]

  def __getitem__(self, idx):
    return (self.X[idx], self.Y[idx])


class tinyChessNet:
    def __init__(self):
        super(tinyChessNet, self).__init__()
        self.a1 = nn.Conv2d(5, 16, kernel_size=3, padding=1)
        # self.bn_a1 = nn.BatchNorm2d(16)
        self.a2 = nn.Conv2d(16, 16, kernel_size=3, padding=1)
        # self.bn_a2 = nn.BatchNorm2d(16)
        self.a3 = nn.Conv2d(16, 32, kernel_size=3, stride=2)
        # self.bn_a3 = nn.BatchNorm2d(32)

        self.b1 = nn.Conv2d(32, 32, kernel_size=3, padding=1)
        # self.bn_b1 = nn.BatchNorm2d(32)
        self.b2 = nn.Conv2d(32, 32, kernel_size=3, padding=1)
        # self.bn_b2 = nn.BatchNorm2d(32)
        self.b3 = nn.Conv2d(32, 64, kernel_size=3, stride=2)
        # self.bn_b3 = nn.BatchNorm2d(64)

        self.c1 = nn.Conv2d(64, 64, kernel_size=2, padding=1)
        # self.bn_c1 = nn.BatchNorm2d(64)
        self.c2 = nn.Conv2d(64, 64, kernel_size=2, padding=1)
        # self.bn_c2 = nn.BatchNorm2d(64)
        self.c3 = nn.Conv2d(64, 128, kernel_size=2, stride=2)
        # self.bn_c3 = nn.BatchNorm2d(128)

        self.d1 = nn.Conv2d(128, 128, kernel_size=1)
        # self.bn_d1 = nn.BatchNorm2d(128)
        self.d2 = nn.Conv2d(128, 128, kernel_size=1)
        # self.bn_d2 = nn.BatchNorm2d(128)
        self.d3 = nn.Conv2d(128, 128, kernel_size=1)
        # self.bn_d3 = nn.BatchNorm2d(128)

        self.last = nn.Linear(128, 1)


    def __call__(self, x):
        x = self.a1(x)
        # x = self.bn_a1(x)
        x = x.relu()
        # x = self.res_block1(x)
        x = self.a2(x)
        # x = self.bn_a2(x)
        x = x.relu()
        x = self.a3(x)
        # x = self.bn_a3(x)
        x = x.relu()
        # 4x4
        x = self.b1(x)
        # x = self.bn_b1(x)
        x = x.relu()
        # x = self.res_block2(x)
        x = self.b2(x)
        # x = self.bn_b2(x)
        x = x.relu()
        x = self.b3(x)
        # x = self.bn_b3(x)
        x = x.relu()
        # 2x2
        x = self.c1(x)
        # x = self.bn_c1(x)
        x = x.relu()
        # x = self.res_block3(x)
        x = self.c2(x)
        # x = self.bn_c2(x)
        x = x.relu()
        x = self.c3(x)
        # x = self.bn_c3(x)
        x = x.relu()
        # 1x128
        x = self.d1(x)
        # x = self.bn_d1(x)
        x = x.relu()
        # x = self.res_block4(x)
        x = self.d2(x)
        # x = self.bn_d2(x)
        x = x.relu()
        x = self.d3(x)
        # x = self.bn_d3(x)
        x = x.relu()
        x = x.reshape(-1, 128)
        x = self.last(x)
        # value output
        return x.tanh()

if __name__ == "__main__":
  device = "cuda"
  chess_dataset = ChessValueDataset()
  train_loader = torch.utils.data.DataLoader(chess_dataset, batch_size=256, shuffle=True)
  model = tinyChessNet()
  optimizer = optim.Adam([model.a1.weight,model.a2.weight,model.a3.weight,model.b1.weight,model.b2.weight,model.b3.weight,model.c1.weight,model.c2.weight,model.c3.weight,model.d1.weight,model.d2.weight,model.d3.weight,model.last.weight], lr = 12e-4)
  # if device == "cuda":
    # model.cuda()
  Tensor.training = True
  for epoch in range(100):
    all_loss = 0
    num_loss = 0
    for batch_idx, (data, target) in enumerate(train_loader):
      target = target.view(-1,1)

      numpy_target = target.numpy()
      tinyTarget = Tensor(numpy_target)
      MutData = data.numpy()
      MutData = MutData.astype(np.float32)
      tinyData = Tensor(MutData)
      batch_size = data.size(0)
      target = tinyTarget
      data = tinyData
      # data, target = data.to(device), target.to(device)
      # data = data.float()
      # target = target.float()
      if batch_size % 256 != 0:
        continue
      # print(data.shape, target.shape)
      optimizer.zero_grad()

      output = model(data)
      loss = Tensor.MSELoss(output,target)
      loss.backward()
      # Gradient clipping
      # max_grad_norm = 1.0  # Set the maximum gradient norm value
      # clip_grad_norm_([model.l1.weight, model.l2.weight], max_grad_norm)
      optimizer.step()
      loss = loss.numpy()
      loss = torch.tensor(loss)
      all_loss += loss.item()
      num_loss += 1
    print("%3d: %f" % (epoch, all_loss/num_loss))
    state_dict = get_state_dict(model)
    safe_save(state_dict, "model.safetensors")
    # torch.save(model, "nets/valueTinyChessNet01.pth")