-
Notifications
You must be signed in to change notification settings - Fork 0
/
test_basics.py
368 lines (263 loc) · 11.5 KB
/
test_basics.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
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
import numpy as np
import torch
import pytest
from torch.nn import functional as F
import torch.nn as nn
import math
from npgpt.tensor import Tensor, Linear, Embedding, SequentialModel, Model, Sigmoid, Tanh, ReLU, NewGELU, Softmax, cross_entropy
def test_matmul():
np.random.seed(35)
ar1 = np.random.random((8, 4, 6, 12))
ar2 = np.random.random((8, 4, 12, 6))
# npt
t_ar1 = Tensor(ar1)
t_ar2 = Tensor(ar2)
npt_out = t_ar1 @ t_ar2
npt_out.backward()
# torch
ttorch1 = torch.tensor(ar1, requires_grad=True)
ttorch2 = torch.tensor(ar2, requires_grad=True)
torch_out = ttorch1 @ ttorch2
torch_out.backward(torch.ones_like(torch_out))
# test
assert np.max(torch_out.detach().numpy()-npt_out.data) < 1e-6
assert np.max(ttorch1.grad.numpy() - t_ar1.grad) < 1e-6
assert np.max(ttorch2.grad.numpy() - t_ar2.grad) < 1e-6
def test_linear():
np.random.seed(35)
Xi = np.ones((64, 48, 10))
W = np.ones((10,48))
npt_Xi = Tensor(Xi)
npt_W = Tensor(W)
npt_out = npt_Xi @ npt_W
npt_out.backward()
torch_Xi = torch.tensor(Xi, requires_grad=True)
torch_W = torch.tensor(W.T, requires_grad=True)
torch_out = F.linear(torch_Xi, torch_W)
torch_out.backward(torch.ones_like(torch_out))
assert np.max(np.abs(torch_out.detach().numpy()-npt_out.data)) < 1e-10
assert np.max(np.abs(torch_Xi.grad.numpy() - npt_Xi.grad)) < 1e-10
assert np.max(np.abs(torch_W.grad.numpy() - npt_W.grad.T)) < 1e-10
def test_linear_2():
np.random.seed(35)
Xi = np.ones((64, 48))
W = np.ones((48,36))
npt_Xi = Tensor(Xi)
npt_W = Tensor(W)
npt_out = npt_Xi @ npt_W
npt_out.backward()
torch_Xi = torch.tensor(Xi, requires_grad=True)
torch_W = torch.tensor(W.T, requires_grad=True)
torch_out = F.linear(torch_Xi, torch_W)
torch_out.backward(torch.ones_like(torch_out))
assert np.max(np.abs(torch_out.detach().numpy()-npt_out.data)) < 1e-10
assert np.max(np.abs(torch_Xi.grad.numpy() - npt_Xi.grad)) < 1e-10
assert np.max(np.abs(torch_W.grad.numpy() - npt_W.grad.T)) < 1e-10
def test_embedding_layer():
np.random.seed(35)
vocab_size = 24
n_embd = 128
lin_out = 54
emb_init_w = np.random.normal(size=(vocab_size, n_embd), scale=0.02)
lin_init_w = np.random.normal(size=(n_embd, lin_out), scale=0.02)
idx = np.array([
[3,2,3,4,1,2,1,1,7,0],
[3,2,3,4,1,4,1,1,7,0],
[9,5,4,2,1,3,0,5,12,1],
[4,2,4,1,6,14,2,5,19,7],
])
# npt
npt_word_embd = Embedding(vocab_size, n_embd)
npt_word_embd.weight.data = emb_init_w
npt_lin = Linear(n_embd, lin_out)
npt_lin.weight.data = lin_init_w
npt_out = npt_word_embd(idx)
npt_out = npt_lin(npt_out)
npt_out = ReLU(npt_out)
npt_out.backward()
# torch
torch_word_embd = nn.Embedding(vocab_size, n_embd)
torch_word_embd.weight = torch.nn.Parameter(torch.tensor(emb_init_w, requires_grad=True))
torch_lin = nn.Linear(n_embd, lin_out, bias=False)
torch_lin.weight = torch.nn.Parameter(torch.tensor(lin_init_w.T, requires_grad=True))
torch_out = torch_word_embd(torch.tensor(idx))
torch_out = torch_lin(torch_out)
torch_out = nn.ReLU()(torch_out)
torch_out.backward(torch.ones_like(torch_out))
assert np.max(np.abs(torch_word_embd.weight.grad.numpy() - npt_word_embd.weight.grad)) < 1e-6
def test_attention_bias():
np.random.seed(35)
BLS = 32
T = 12
ar1 = np.random.random((8, 4, 12, 6))
ar2 = np.random.random((8, 4, 6, 12))
# npt
t_ar1 = Tensor(ar1)
t_ar2 = Tensor(ar2)
# npt_bias = (1-np.tril(np.ones(shape=(1, 1, BLS, BLS)))) #.astype(np.int8)
# npt_bias[npt_bias==1.] = -np.inf
npt_bias = (np.tril(np.ones(shape=(1, 1, BLS, BLS))))
npt_out = t_ar1 @ t_ar2
# npt_out = npt_out + npt_bias[:,:,:T,:T]
npt_out.data *= npt_bias[:,:,:T,:T]
npt_out.data[npt_out.data==0.] = -np.inf
soft_npt_out = npt_out.softmax()
soft_npt_out.backward()
# torch
t_bias = torch.tensor(torch.tril(torch.ones(BLS, BLS)), requires_grad=False).view(1, 1, BLS, BLS)
ttorch1 = torch.tensor(ar1, requires_grad=True)
ttorch2 = torch.tensor(ar2, requires_grad=True)
out_torch = ttorch1 @ ttorch2
out_torch = out_torch.masked_fill(t_bias[:,:,:T,:T] == 0, float('-inf'))
soft_out_torch = F.softmax(out_torch, dim=-1)
soft_out_torch.backward(torch.ones_like(soft_out_torch))
# test
assert np.max(np.abs(ttorch1.grad.numpy() - t_ar1.grad)) < 1e-7
assert np.max(np.abs(ttorch2.grad.numpy() - t_ar2.grad)) < 1e-7
assert np.max(np.abs(soft_out_torch.detach().numpy()-soft_npt_out.data)) < 1e-7
def test_simple_softmax():
np.random.seed(36)
arr = np.random.random((1, 100))
npt_arr = Tensor(arr)
npt_out = npt_arr.softmax()
npt_out.backward()
torch_arr = torch.tensor(arr, requires_grad=True)
torch_out = F.softmax(torch_arr, dim=-1)
torch_out.backward(torch.ones_like(torch_out))
assert np.max(np.abs(torch_out.detach().numpy()-npt_out.data)) < 1e-7
assert np.max(np.abs(torch_arr.grad.numpy() - npt_arr.grad)) < 1e-7
def test_simple_softmax_four_dims():
np.random.seed(36)
arr = np.random.random((64, 5, 33, 100))
npt_arr = Tensor(arr)
npt_out = npt_arr.softmax()
npt_out.backward()
torch_arr = torch.tensor(arr, requires_grad=True)
torch_out = F.softmax(torch_arr, dim=-1)
torch_out.backward(torch.ones_like(torch_out))
assert np.max(np.abs(torch_out.detach().numpy()-npt_out.data)) < 1e-7
assert np.max(np.abs(torch_arr.grad.numpy() - npt_arr.grad)) < 1e-7
def test_softmax():
np.random.seed(36)
ar1 = np.random.random((64, 3, 5, 16))
ar2 = np.random.random((64, 3, 16, 5))
# npt
t_ar1 = Tensor(ar1)
t_ar2 = Tensor(ar2)
npt_out = t_ar1 @ t_ar2
soft_npt_out = npt_out.softmax()
soft_npt_out.backward()
# torch
ttorch1 = torch.tensor(ar1, requires_grad=True)
ttorch2 = torch.tensor(ar2, requires_grad=True)
out_torch = torch.matmul(ttorch1, ttorch2)
soft_out_torch = F.softmax(out_torch, dim=-1)
soft_out_torch.backward(torch.ones_like(soft_out_torch))
# test
assert np.max(np.abs(soft_out_torch.detach().numpy()-soft_npt_out.data)) < 1e-6
assert np.max(np.abs(ttorch1.grad.numpy() - t_ar1.grad)) < 1e-6
assert np.max(np.abs(ttorch2.grad.numpy() - t_ar2.grad)) < 1e-6
def test_log_softmax():
np.random.seed(36)
arr = np.random.random((16, 64))
# npt
t_ar1 = Tensor(arr)
soft_npt_out = t_ar1.log_softmax()
soft_npt_out.backward()
# torch
ttorch1 = torch.tensor(arr, requires_grad=True)
soft_out_torch = F.log_softmax(ttorch1, dim=-1)
soft_out_torch.backward(torch.ones_like(soft_out_torch))
# test
assert np.max(np.abs(soft_out_torch.detach().numpy()-soft_npt_out.data)) < 1e-6
assert np.max(np.abs(ttorch1.grad.numpy() - t_ar1.grad)) < 1e-6
def test_cross_entropy():
np.random.seed(35)
BS = 64
ar1 = np.random.random((BS, 64))
ar2 = np.random.random((64, 12))
ar3 = np.random.randint(0, 12, size=BS)
# npt
t_ar1 = Tensor(ar1)
t_ar2 = Tensor(ar2)
t_ar3 = Tensor(np.eye(12)[ar3])
npt_out = t_ar1 @ t_ar2
npt_out = npt_out.reshape(-1, npt_out.shape[-1])
loss = cross_entropy(npt_out, t_ar3)
loss.backward()
# torch
ttorch1 = torch.tensor(ar1, requires_grad=True)
ttorch2 = torch.tensor(ar2, requires_grad=True)
t_ar3 = torch.from_numpy(ar3)
out_torch = torch.matmul(ttorch1, ttorch2)
soft_out_torch = F.cross_entropy(out_torch.view(-1, out_torch.size(-1)), t_ar3.view(-1))
soft_out_torch.backward(torch.ones_like(soft_out_torch))
# test
assert np.max(np.abs(ttorch1.grad.numpy() - t_ar1.grad)) < 1e-7
assert np.max(np.abs(ttorch2.grad.numpy() - t_ar2.grad)) < 1e-7
def test_self_attention():
np.random.seed(35)
n_head = 3
n_embd = 48
block_size = 6
BS = 64
X = np.random.normal(scale=0.01, size=(BS, 5, n_embd)).astype(np.float32)
attn_W = np.random.random((3*n_embd, n_embd)).astype(np.float32)
proj_W = np.random.random((n_embd, n_embd)).astype(np.float32)
# npt
npt_attn = Linear(n_embd, 3*n_embd)
npt_attn.weight = Tensor(attn_W.T)
npt_bias = (1-np.tril(np.ones(shape=(1, 1, block_size, block_size)))) #.astype(np.int8)
npt_bias[npt_bias==1.] = -np.inf
npt_out_proj = Linear(n_embd, n_embd)
npt_out_proj.weight = Tensor(proj_W.T)
npt_x = Tensor(np.array(X))
B, T, C = npt_x.shape # batch size, sequence length, embedding dimensionality (n_embd)
attn_out = npt_attn(npt_x)
k0, q0, v0 = attn_out.split(3, axis=2)
npt_k = (1*k0.reshape(B, T, n_head, C // n_head)).transpose(0,2,1,3) # (B, nh, T, hs) # reshape directly to correct shape??
# k = (1*k0).reshape(B, self.n_heads, T, C // self.n_heads) #.transpose(0,2,1,3) # (B, nh, T, hs) # reshape directly to correct shape??
npt_q = (1*q0.reshape(B, T, n_head, C // n_head)).transpose(0,2,1,3) # (B, nh, T, hs)
# q = (1*q0).reshape(B, self.n_heads, T, C // self.n_heads) #.transpose(0,2,1,3) # (B, nh, T, hs) # reshape directly to correct shape??
npt_v = (1*v0.reshape(B, T, n_head, C // n_head)).transpose(0,2,1,3) # (B, nh, T, hs)
# v = (1*v0).reshape(B, self.n_heads, T, C // self.n_heads) #.transpose(0,2,1,3) # (B, nh, T, hs) # reshape directly to correct shape??
npt_att = (npt_q @ npt_k.transpose(0,1,3,2)) / np.sqrt(n_embd) #, nograd=True)
# apply forward attention mask
npt_att_b = npt_att + npt_bias[:,:,:T,:T]
npt_soft_att = Softmax(npt_att_b)
out = npt_soft_att @ npt_v # batch, n_heads, seq_len, n_embd // n_heads
trans_out = 1*out.transpose(0,2,1,3)
reshape_out = trans_out.reshape(B,T,C) # B, T, C # reshape directly to correct shape??
# reshape_out = out.reshape(B,T,C) # B, T, C # reshape directly to correct shape??
proj_out = npt_out_proj(reshape_out)
proj_out.backward()
# torch
# init
c_attn = nn.Linear(n_embd, 3 * n_embd, bias=False) # 3*n_embd, n_embd
c_attn.weight = torch.nn.Parameter(torch.tensor(attn_W, requires_grad=True))
c_proj = nn.Linear(n_embd, n_embd, bias=False) # n_embd, n_embd
c_proj.weight = torch.nn.Parameter(torch.tensor(proj_W, requires_grad=True))
bias = torch.tensor(torch.tril(torch.ones(block_size, block_size)), requires_grad=True).view(1, 1, block_size, block_size)
# exec
x = torch.tensor(X, requires_grad=True) #.float() #.type(torch.DoubleTensor)
B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
k, q, v = c_attn(x).split(n_embd, dim=2)
k = k.view(B, T, n_head, C // n_head).transpose(1, 2) # (B, nh, T, hs)
q = q.view(B, T, n_head, C // n_head).transpose(1, 2) # (B, nh, T, hs)
v = v.view(B, T, n_head, C // n_head).transpose(1, 2) # (B, nh, T, hs)
k.retain_grad()
v.retain_grad()
# causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
att0 = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
att0.retain_grad()
att = att0.masked_fill(bias[:,:,:T,:T] == 0, float('-inf'))
att.retain_grad()
soft_att = F.softmax(att, dim=-1)
soft_att.retain_grad()
y1 = soft_att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
y1.retain_grad()
yt = y1.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
y = c_proj(yt)
y.backward(torch.ones_like(y))
assert np.mean(np.abs(y.detach().numpy() - proj_out.data)) < 1e-3
# assert np.max(np.abs(y.detach().numpy() - proj_out.data)) < 1e-8