-
Notifications
You must be signed in to change notification settings - Fork 4
/
transformer.py
251 lines (214 loc) · 8.37 KB
/
transformer.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
import numpy as np
from flax import linen as nn
import jax
import jax.numpy as jnp
from typing import Literal, TypedDict
import numpy as np
from tiny_stories import TinyStoriesDataset
from tiny_stories import TOKENIZER_SIZE
from flax.training import train_state
import wandb
import optax
from chebykan_layer import ChebyKAN
from time import perf_counter
D_TYPE = jnp.float32
MAX_LEN = 64
class MLP(nn.Module):
@nn.compact
def __call__(self, x):
d_outer = x.shape[-1]
# 84 is choosen so that the number of parameters matches then KAN layer
x = nn.Dense(features=768, param_dtype=D_TYPE)(x)
x = nn.gelu(x)
x = nn.Dense(features=d_outer, param_dtype=D_TYPE)(x)
return x
class MLPBlock(nn.Module):
@nn.compact
def __call__(self, x):
y = nn.LayerNorm(param_dtype=D_TYPE)(x)
y = MLP()(y)
return x + y
class KANLayer(nn.Module):
@nn.compact
def __call__(self, x):
# y has shape (batch_size, seq_len, d_model) -> (batch_size * seq_len, d_model)
y = x.reshape((-1, x.shape[-1]))
y = ChebyKAN(in_features=x.shape[-1], out_features=x.shape[-1], degree=8)(y)
y = y.reshape(x.shape)
return y
class KANBlock(nn.Module):
@nn.compact
def __call__(self, x):
x_inner = nn.LayerNorm()(x)
x_inner = KANLayer()(x_inner)
return x + x_inner
class SelfAttentionBlock(nn.Module):
d_model: int
n_heads: int
@nn.compact
def __call__(self, x):
# Shape (batch_size, seq_len, d_model)
n_heads, d_model = self.n_heads, self.d_model
assert d_model % n_heads == 0, 'n_heads must divide d_model'
# Shape (batch_size, num_heads, seq_len, seq_len)
mask = jnp.ones((x.shape[0], n_heads, x.shape[1], x.shape[1]))
# Create diagonal mask
mask = jnp.tril(mask)
y = nn.LayerNorm(param_dtype=D_TYPE)(x)
attn = nn.MultiHeadDotProductAttention(
num_heads=n_heads, qkv_features=d_model // n_heads, out_features=d_model, param_dtype=D_TYPE
)(y, mask=mask)
return x + attn
class Transformer(nn.Module):
d_model: int
n_heads: int
n_layers: int
@nn.compact
def __call__(self, x):
# Shape (batch_size, seq_len) -> (batch_size, seq_len, d_model)
x = nn.Embed(num_embeddings=TOKENIZER_SIZE, features=self.d_model, param_dtype=D_TYPE)(x)
pos_emb = nn.Embed(num_embeddings=MAX_LEN, features=self.d_model, param_dtype=D_TYPE)(jnp.arange(MAX_LEN))
x = x + pos_emb
for _ in range(self.n_layers):
x = SelfAttentionBlock(self.d_model, self.n_heads)(x)
x = MLPBlock()(x)
# Shape (batch_size, seq_len, d_model) -> (batch_size, seq_len, vocab_size)
x = nn.Dense(features=TOKENIZER_SIZE, use_bias=False, param_dtype=D_TYPE)(x)
return x
class KANTransformer(nn.Module):
d_model: int
n_heads: int
n_layers: int
@nn.compact
def __call__(self, x):
# Shape (batch_size, seq_len) -> (batch_size, seq_len, d_model)
x = nn.Embed(num_embeddings=TOKENIZER_SIZE, features=self.d_model, param_dtype=D_TYPE)(x)
pos_emb = nn.Embed(num_embeddings=MAX_LEN, features=self.d_model, param_dtype=D_TYPE)(jnp.arange(MAX_LEN))
x = x + pos_emb
for _ in range(self.n_layers):
x = SelfAttentionBlock(self.d_model, self.n_heads)(x)
x = KANBlock()(x)
# Shape (batch_size, seq_len, d_model) -> (batch_size, seq_len, vocab_size)
x = nn.Dense(features=TOKENIZER_SIZE, use_bias=False, param_dtype=D_TYPE)(x)
return x
class KANHybridTransformer(nn.Module):
d_model: int
n_heads: int
n_layers: int
@nn.compact
def __call__(self, x):
assert self.n_layers % 2 == 0, "n_layers must be even"
# Shape (batch_size, seq_len) -> (batch_size, seq_len, d_model)
x = nn.Embed(num_embeddings=TOKENIZER_SIZE, features=self.d_model, param_dtype=D_TYPE)(x)
pos_emb = nn.Embed(num_embeddings=MAX_LEN, features=self.d_model, param_dtype=D_TYPE)(jnp.arange(MAX_LEN))
x = x + pos_emb
for _ in range(self.n_layers // 2):
x = SelfAttentionBlock(self.d_model, self.n_heads)(x)
x = MLPBlock()(x)
x = SelfAttentionBlock(self.d_model, self.n_heads)(x)
x = KANBlock()(x)
# Shape (batch_size, seq_len, d_model) -> (batch_size, seq_len, vocab_size)
x = nn.Dense(features=TOKENIZER_SIZE, use_bias=False, param_dtype=D_TYPE)(x)
return x
def param_count(params):
return sum(x.size for x in jax.tree_util.tree_leaves(params)) / 1e6
class Config(TypedDict):
d_model: int
n_heads: int
n_layers: int
learning_rate: float
max_steps: int
batch_size: int
weight_decay: float
block_type: Literal["MLP", "KAN", "Hybrid"]
def masked_cross_entropy(logits: jnp.ndarray, targets: jnp.ndarray, mask: jnp.ndarray):
# logits shape: (batch_size, seq_len, vocab_size)
# targets shape: (batch_size, seq_len)
# mask shape: (batch_size, seq_len)
# shift everything by 1
logits = logits[:, :-1, :]
targets = targets[:, 1:]
mask = mask[:, 1:]
vocab_size = logits.shape[-1]
one_hot_targets = jax.nn.one_hot(targets, vocab_size)
# one_hot_targets shape: (batch_size, seq_len, vocab_size)
log_probs = jax.nn.log_softmax(logits)
loss = -jnp.sum(log_probs * one_hot_targets, axis=-1)
# loss shape: (batch_size, seq_len)
loss = loss * mask
# Flatten everything divide by the sum of the mask
total_tokens = jnp.sum(mask.flatten())
return jnp.sum(loss.flatten()) / total_tokens
def create_train_state(rng, config):
if config["block_type"] == "MLP":
model = Transformer(
d_model=config['d_model'],
n_heads=config['n_heads'],
n_layers=config['n_layers'],
)
elif config["block_type"] == "KAN":
model = KANTransformer(
d_model=config['d_model'],
n_heads=config['n_heads'],
n_layers=config['n_layers'],
)
else:
model = KANHybridTransformer(
d_model=config['d_model'],
n_heads=config['n_heads'],
n_layers=config['n_layers'],
)
params = model.init(rng, jnp.ones((config['batch_size'], MAX_LEN), dtype=jnp.int32))
optimizer = optax.adamw(learning_rate=config['learning_rate'], weight_decay=config['weight_decay'])
return train_state.TrainState.create(
apply_fn=model.apply,
params=params,
tx=optimizer
)
@jax.jit
def train_step(state: train_state.TrainState, batch: jnp.ndarray, mask: jnp.ndarray):
def loss_fn(params):
logits = state.apply_fn(params, batch)
return masked_cross_entropy(logits, batch, mask)
grad_fn = jax.value_and_grad(loss_fn)
loss, grads = grad_fn(state.params)
state = state.apply_gradients(grads=grads)
return state, loss
rng = jax.random.PRNGKey(0)
config = Config(
d_model=128,
n_heads=8,
n_layers=16,
learning_rate=1e-5,
batch_size=16,
weight_decay=0.001,
block_type="MLP",
)
# ChebyKAN
# Number of parameters: 15.37536
# Number of non-embedding parameters: 2.5013760000000005
# MLP (hidden size 768)
# Number of parameters: 16.176128
# Number of non-embedding parameters: 3.3021439999999984
# Training loop
if __name__ == "__main__":
wandb.init(project="kan-transformer", config=config)
print("Creating model...")
state = create_train_state(rng, config)
print("Number of parameters: ", param_count(state.params))
print("Number of non-embedding parameters:", param_count(state.params) - (config["d_model"] * TOKENIZER_SIZE * 2 + config["d_model"] * MAX_LEN) / 1e6)
for step, (batch, mask) in enumerate(TinyStoriesDataset(max_len=MAX_LEN).create_batches(config['batch_size'])):
step_start_time = perf_counter()
state, loss = train_step(state, batch, mask)
step_end_time = perf_counter()
if step % 50 == 0:
print(f"Step {step}, Loss: {loss}")
wandb.log({"loss": loss})
print(f"Time taken for step: {step_end_time - step_start_time}")
wandb.log({"time": step_end_time - step_start_time})
# save every 1000 steps
if step % 1000 == 0:
print("Saving params...")
model_type = config["block_type"]
np.save(f"checkpoints/params_{model_type}.npy", state.params)
print("Params saved.")