forked from tinygrad/tinygrad
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtest_nn.py
executable file
·290 lines (229 loc) · 10 KB
/
test_nn.py
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
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
#!/usr/bin/env python
import unittest
import numpy as np
from extra.utils import WINDOWS
from tinygrad.helpers import getenv
from tinygrad.jit import TinyJit
from tinygrad.tensor import Tensor, Device
from tinygrad.nn import BatchNorm2d, Conv1d, ConvTranspose1d, Conv2d, ConvTranspose2d, Linear, GroupNorm, LayerNorm, LayerNorm2d, Embedding, InstanceNorm
import torch
import pytest
pytestmark = [pytest.mark.exclude_cuda]
class TestNN(unittest.TestCase):
def test_batchnorm2d(self, training=False):
szs = [4, 8, 16, 32]
for sz in szs:
# create in tinygrad
Tensor.training = training
bn = BatchNorm2d(sz, eps=1e-5, track_running_stats=training)
bn.weight = Tensor.randn(sz)
bn.bias = Tensor.randn(sz)
bn.running_mean = Tensor.randn(sz)
bn.running_var = Tensor.randn(sz)
bn.running_var.numpy()[bn.running_var.numpy() < 0] = 0
# create in torch
with torch.no_grad():
tbn = torch.nn.BatchNorm2d(sz).eval()
tbn.training = training
tbn.weight[:] = torch.tensor(bn.weight.numpy())
tbn.bias[:] = torch.tensor(bn.bias.numpy())
tbn.running_mean[:] = torch.tensor(bn.running_mean.numpy())
tbn.running_var[:] = torch.tensor(bn.running_var.numpy())
np.testing.assert_allclose(bn.running_mean.numpy(), tbn.running_mean.detach().numpy(), rtol=1e-5, atol=1e-6)
np.testing.assert_allclose(bn.running_var.numpy(), tbn.running_var.detach().numpy(), rtol=1e-5, atol=1e-6)
# trial
inn = Tensor.randn(2, sz, 3, 3)
# in tinygrad
outt = bn(inn)
# in torch
toutt = tbn(torch.tensor(inn.cpu().numpy()))
# close
np.testing.assert_allclose(outt.numpy(), toutt.detach().numpy(), rtol=5e-4, atol=1e-6)
np.testing.assert_allclose(bn.running_mean.numpy(), tbn.running_mean.detach().numpy(), rtol=1e-5, atol=1e-6)
np.testing.assert_allclose(bn.running_var.numpy(), tbn.running_var.detach().numpy(), rtol=1e-5, atol=1e-6)
def test_batchnorm2d_training(self):
self.test_batchnorm2d(True)
def test_linear(self):
def _test_linear(x):
# create in tinygrad
model = Linear(in_dim, out_dim)
z = model(x)
# create in torch
with torch.no_grad():
torch_layer = torch.nn.Linear(in_dim, out_dim).eval()
torch_layer.weight[:] = torch.tensor(model.weight.numpy(), dtype=torch.float32)
torch_layer.bias[:] = torch.tensor(model.bias.numpy(), dtype=torch.float32)
torch_x = torch.tensor(x.cpu().numpy(), dtype=torch.float32)
torch_z = torch_layer(torch_x)
# test
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-4, rtol=1e-5)
BS, T, in_dim, out_dim = 4, 2, 8, 16
_test_linear(Tensor.randn(BS, in_dim))
_test_linear(Tensor.randn(BS, T, in_dim)) # test with more dims
def test_conv1d(self):
BS, C1, W = 4, 16, 224
C2, K, S, P = 64, 7, 2, 1
# create in tinygrad
layer = Conv1d(C1, C2, kernel_size=K, stride=S, padding=P)
# create in torch
with torch.no_grad():
torch_layer = torch.nn.Conv1d(C1, C2, kernel_size=K, stride=S, padding=P).eval()
torch_layer.weight[:] = torch.tensor(layer.weight.numpy(), dtype=torch.float32)
torch_layer.bias[:] = torch.tensor(layer.bias.numpy(), dtype=torch.float32)
# test
x = Tensor.uniform(BS, C1, W)
z = layer(x)
torch_x = torch.tensor(x.cpu().numpy())
torch_z = torch_layer(torch_x)
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-4, rtol=1e-5)
def test_conv2d(self):
BS, C1, H, W = 4, 16, 224, 224
C2, K, S, P = 64, 7, 2, 1
# create in tinygrad
layer = Conv2d(C1, C2, kernel_size=K, stride=S, padding=P)
# create in torch
with torch.no_grad():
torch_layer = torch.nn.Conv2d(C1, C2, kernel_size=K, stride=S, padding=P).eval()
torch_layer.weight[:] = torch.tensor(layer.weight.numpy(), dtype=torch.float32)
torch_layer.bias[:] = torch.tensor(layer.bias.numpy(), dtype=torch.float32)
# test
x = Tensor.uniform(BS, C1, H, W)
z = layer(x)
torch_x = torch.tensor(x.cpu().numpy())
torch_z = torch_layer(torch_x)
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-4, rtol=1e-5)
@unittest.skipIf(getenv("CI", "") != "" and (WINDOWS or Device.DEFAULT == "WEBGPU"), "runs out of memory in CI")
def test_conv_transpose1d(self):
BS, C1, W = 4, 16, 224
C2, K, S, P = 64, 7, 2, 1
# create in tinygrad
layer = ConvTranspose1d(C1, C2, kernel_size=K, stride=S, padding=P)
# create in torch
with torch.no_grad():
torch_layer = torch.nn.ConvTranspose1d(C1, C2, kernel_size=K, stride=S, padding=P).eval()
torch_layer.weight[:] = torch.tensor(layer.weight.numpy(), dtype=torch.float32)
torch_layer.bias[:] = torch.tensor(layer.bias.numpy(), dtype=torch.float32)
# test
x = Tensor.uniform(BS, C1, W)
z = layer(x)
torch_x = torch.tensor(x.cpu().numpy())
torch_z = torch_layer(torch_x)
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-4, rtol=1e-5)
@unittest.skipIf(getenv("CI", "") != "" and (WINDOWS or Device.DEFAULT == "WEBGPU"), "runs out of memory in CI")
def test_conv_transpose2d(self):
BS, C1, H, W = 4, 16, 224, 224
C2, K, S, P = 64, 7, 2, 1
# create in tinygrad
layer = ConvTranspose2d(C1, C2, kernel_size=K, stride=S, padding=P)
# create in torch
with torch.no_grad():
torch_layer = torch.nn.ConvTranspose2d(C1, C2, kernel_size=K, stride=S, padding=P).eval()
torch_layer.weight[:] = torch.tensor(layer.weight.numpy(), dtype=torch.float32)
torch_layer.bias[:] = torch.tensor(layer.bias.numpy(), dtype=torch.float32)
# test
x = Tensor.uniform(BS, C1, H, W)
z = layer(x)
torch_x = torch.tensor(x.cpu().numpy())
torch_z = torch_layer(torch_x)
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-4, rtol=1e-5)
def test_groupnorm(self):
BS, H, W, C, G = 20, 10, 10, 6, 3
# create in tinygrad
layer = GroupNorm(G, C)
# create in torch
with torch.no_grad():
torch_layer = torch.nn.GroupNorm(G, C).eval()
torch_layer.weight[:] = torch.tensor(layer.weight.numpy(), dtype=torch.float32)
torch_layer.bias[:] = torch.tensor(layer.bias.numpy(), dtype=torch.float32)
# test
x = Tensor.randn(BS, C, H, W)
z = layer(x)
torch_x = torch.tensor(x.cpu().numpy())
torch_z = torch_layer(torch_x)
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-3, rtol=5e-3)
def test_layernorm(self):
N, C, H, W = 20, 5, 10, 10
# create in tinygrad
layer = LayerNorm([H, W])
# create in torch
with torch.no_grad():
torch_layer = torch.nn.LayerNorm([H, W]).eval()
torch_layer.weight[:] = torch.tensor(layer.weight.numpy(), dtype=torch.float32)
torch_layer.bias[:] = torch.tensor(layer.bias.numpy(), dtype=torch.float32)
# test
x = Tensor.randn(N, C, H, W)
z = layer(x)
torch_x = torch.tensor(x.cpu().numpy())
torch_z = torch_layer(torch_x)
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-3, rtol=5e-3)
def test_layernorm_2d(self):
N, C, H, W = 20, 5, 10, 10
# create in tinygrad
layer = LayerNorm2d(C)
# create in torch
with torch.no_grad():
torch_layer = torch.nn.LayerNorm([C]).eval()
torch_layer.weight[:] = torch.tensor(layer.weight.numpy(), dtype=torch.float32)
torch_layer.bias[:] = torch.tensor(layer.bias.numpy(), dtype=torch.float32)
# test
x = Tensor.randn(N, C, H, W)
z = layer(x)
torch_x = torch.tensor(x.cpu().numpy())
torch_z = torch_layer(torch_x.permute(0,2,3,1)).permute(0,3,1,2)
def test_instancenorm_2d(self):
N, C, H, W = 20, 5, 10, 10
# create in tinygrad
layer = InstanceNorm(C)
# create in torch
with torch.no_grad():
torch_layer = torch.nn.InstanceNorm2d(C, affine=True).eval()
torch_layer.weight[:] = torch.tensor(layer.weight.numpy(), dtype=torch.float32)
torch_layer.bias[:] = torch.tensor(layer.bias.numpy(), dtype=torch.float32)
# test
x = Tensor.randn(N, C, H, W)
z = layer(x)
torch_x = torch.tensor(x.cpu().numpy())
torch_z = torch_layer(torch_x)
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-3, rtol=5e-3)
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-3, rtol=5e-3)
def test_instancenorm_3d(self):
N, C, D, H, W = 20, 5, 3, 10, 10
# create in tinygrad
layer = InstanceNorm(C)
# create in torch
with torch.no_grad():
torch_layer = torch.nn.InstanceNorm3d(C, affine=True).eval()
torch_layer.weight[:] = torch.tensor(layer.weight.numpy(), dtype=torch.float32)
torch_layer.bias[:] = torch.tensor(layer.bias.numpy(), dtype=torch.float32)
# test
x = Tensor.randn(N, C, D, H, W)
z = layer(x)
torch_x = torch.tensor(x.cpu().numpy())
torch_z = torch_layer(torch_x)
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-3, rtol=5e-3)
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-3, rtol=5e-3)
def test_embedding(self):
B, T, C, VS = 4, 10, 20, 28
# create in tinygrad
layer = Embedding(VS, C)
with torch.no_grad():
torch_layer = torch.nn.Embedding(VS, C).eval()
torch_layer.weight[:] = torch.tensor(layer.weight.numpy(), dtype=torch.float32)
# test
x = Tensor(np.random.randint(0, VS, (B, T)).astype(np.float32))
z = layer(x)
torch_x = torch.tensor(x.cpu().numpy().astype(np.int32))
torch_z = torch_layer(torch_x)
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=1e-8, rtol=1e-8)
# test with jit enabled
@TinyJit
def layer_jit(x):
return layer(x).realize()
for _ in range(3):
x = Tensor(np.random.randint(0, VS, (B, T)).astype(np.float32))
z = layer_jit(x)
torch_x = torch.tensor(x.cpu().numpy().astype(np.int32))
torch_z = torch_layer(torch_x)
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=1e-8, rtol=1e-8)
if __name__ == '__main__':
unittest.main()