-
Notifications
You must be signed in to change notification settings - Fork 0
/
gpt.py
160 lines (139 loc) · 7.32 KB
/
gpt.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
import tensorflow as tf
import time
import numpy as np
from tensorflow import keras
from data import load_mrpc_gpt
from utils import set_soft_gpu
class MultiHeadAttention(keras.layers.Layer):
def __init__(self, model_dim, n_head, dropout_rate, use_bias=False):
super().__init__()
self.model_dim = model_dim
self.n_head = n_head
self.head_dim = model_dim // n_head
self.wq = keras.layers.Dense(n_head * self.head_dim, use_bias=use_bias)
self.wk = keras.layers.Dense(n_head * self.head_dim, use_bias=use_bias)
self.wv = keras.layers.Dense(n_head * self.head_dim, use_bias=use_bias)
self.o_dense = keras.layers.Dense(model_dim)
self.dropout = keras.layers.Dropout(dropout_rate)
def call(self, q, k, v, mask, training):
_q = self.wq(q) # [None q_step model_dim] -> [None q_step n_head*head_dim]
_k, _v = self.wk(k), self.wv(v) # [None q_step model_dim] -> [None q_step n_head*head_dim]
_q = self.split_head(_q) # [None q_step n_head*head_dim] -> [None n_head step head_dim]
_k, _v= self.split_head(_k), self.split_head(_v) # [None q_step n_head*head_dim] -> [None n_head step head_dim]
context = self.scale_dot_product_attention(_q, _k, _v, mask) # -> [None q_step n_head*head_dim]
o = self.o_dense(context) # [None q_step n_head*head_dim] -> [None q_step model_dim]
o = self.dropout(o)
return o
def split_head(self, x):
# input: [None step n_head*head_dim]
# [None step n_head*head_dim] -> [None step n_head head_dim]
x = tf.reshape(x, shape=[x.shape[0], x.shape[1], self.n_head, self.head_dim])
# [None step n_head head_dim] -> [None n_head step head_dim]
x = tf.transpose(x, perm=[0, 2, 1, 3])
return x
def scale_dot_product_attention(self, q, k, v, mask=None):
dk = tf.cast(k.shape[-1], dtype=tf.float32)
# [None n_head q_step head_dim] @ [None n_head step head_dim] -> [None n_head q_step step]
score = tf.matmul(q, k, transpose_b=True) / (tf.sqrt(dk) + 1e-9)
if mask is not None:
score += mask * -1e9
attention = tf.nn.softmax(score, axis=-1)
# [None n_head q_step step] @ [None n_head step head_dim] -> [None n_head q_step head_dim]
context = tf.matmul(attention, v)
# [None n_head q_step head_dim] -> [None q_step n_head head_dim]
context = tf.transpose(context, perm=[0, 2, 1, 3])
# [None q_step n_head head_dim] -> [None q_step n_head*head_dim]
context = tf.reshape(context, shape=[context.shape[0], context.shape[1], -1])
return context
class PositionWiseFFN(keras.layers.Layer):
def __init__(self, model_dim):
super().__init__()
self.o_dense1 = keras.layers.Dense(model_dim*4, activation="relu")
self.o_dense2 = keras.layers.Dense(model_dim)
def call(self, x):
o = self.o_dense1(x)
o = self.o_dense2(o)
return o
class EncoderLayer(keras.layers.Layer):
def __init__(self, model_dim, n_head, dropout_rate, use_bias=False):
super().__init__()
self.ln = [keras.layers.LayerNormalization() for _ in range(2)]
self.mha = MultiHeadAttention(model_dim, n_head, dropout_rate, use_bias)
self.ffn = PositionWiseFFN(model_dim)
self.dropout = keras.layers.Dropout(dropout_rate)
def call(self, xz, mask, training):
attention = self.mha.call(xz, xz, xz, mask, training)
o1 = self.ln[0](self.dropout(attention, training) + xz)
ffn = self.ffn.call(o1)
o2 = self.ln[1](self.dropout(ffn, training) + o1)
return o2
class Encoder(keras.layers.Layer):
def __init__(self, model_dim, n_head, n_layer, dropout_rate, use_bias=False):
super().__init__()
self.l = [EncoderLayer(model_dim, n_head, dropout_rate, use_bias) for _ in range(n_layer)]
def call(self, xz, mask, training):
for l in self.l:
xz = l.call(xz, mask, training)
return xz
class GPT(keras.Model):
def __init__(self, model_dim, n_head, n_layer, dropout_rate, max_length, n_vocab, max_seg=3, padding_idx=0, lr=1e-4):
super().__init__()
self.padding_idx = padding_idx
self.n_vocab = n_vocab
self.max_length = max_length
self.s1s2embedding = keras.layers.Embedding(input_dim=n_vocab, output_dim=model_dim)
self.seg_embedding = keras.layers.Embedding(input_dim=max_seg, output_dim=model_dim)
self.position_emb = self.add_weight(
name="pos_emb", shape=[1, max_length, model_dim], dtype=tf.float32,
initializer=keras.initializers.RandomNormal(0., 0.01)
)
self.ln = keras.layers.LayerNormalization()
self.encoder = Encoder(model_dim, n_head, n_layer, dropout_rate)
self.mlm = keras.layers.Dense(n_vocab)
self.nsp = keras.layers.Dense(2)
self.loss_func = keras.losses.SparseCategoricalCrossentropy(from_logits=True, reduction="none")
self.opt = keras.optimizers.Adam(lr)
def call(self, seq, seg, training=False):
# input ([None step] [None step])
# [None step] -> [None step model_dim]
emb = self.add_emb(seq, seg)
emb = self.ln(emb)
# [None step model_dim] -> [None step model_dim]
z = self.encoder.call(emb, mask=self.mask(seq), training=training)
# [None step model_dim] -> [None step n_vocab]
mlm_logits = self.mlm(z)
# [None step*model_dim] -> [None 2]
nsp_logits = self.nsp(tf.reshape(z, shape=[z.shape[0], -1]))
return mlm_logits, nsp_logits
def train(self, seq, seg, p_seq, label, training=True):
# train: seq[:, :-1] seg[:, :-1] predict: p_seg[:, 1:]
mlm_logits, nsp_logits = self.call(seq, seg, training)
pad_seq = tf.math.not_equal(p_seq, self.padding_idx)
pred_loss = tf.reduce_mean(tf.boolean_mask(self.loss_func(p_seq, mlm_logits), pad_seq))
nsp_loss = tf.reduce_mean(self.loss_func(label, nsp_logits))
loss = pred_loss + 0.2 * nsp_loss
return loss, mlm_logits
def add_emb(self, seq, seg):
emb = self.s1s2embedding(seq) + self.seg_embedding(seg) + self.position_emb
return emb
def mask(self, seq):
mask = 1 - tf.linalg.band_part(tf.ones((self.max_length, self.max_length)), -1, 0)
return mask
def train():
set_soft_gpu(True)
epoch = 100
data, i2v, v2i, max_length = load_mrpc_gpt(min_freq=2)
model = GPT(model_dim=256, n_head=4, n_layer=4, dropout_rate=0.2, max_length=max_length - 1, n_vocab=len(i2v), lr=1e-4)
start_time = time.time()
for e in range(epoch):
for step, (seq, seg, seq_valid_len, label) in enumerate(data):
with tf.GradientTape() as tape:
loss, pred = model.train(seq[:, :-1], seg[:, :-1], seq[:, 1:], label)
grads = tape.gradient(loss, model.trainable_variables)
model.opt.apply_gradients(zip(grads, model.trainable_variables))
if step % 40 == 0:
print("epoch:%d | step:%d | time:%.2f | loss:%.3f"%(e, step, time.time()-start_time, loss))
print("y_true: {}".format(" ".join([i2v[idx] for idx in seq[0].numpy()[:np.sum(seq_valid_len[0])+3]])))
print("y_pred: {}".format(" ".join([" "] + [i2v[idx] for idx in pred[0].numpy().argmax(axis=1)[:np.sum(seq_valid_len[0])+2]])))
if __name__ == "__main__":
train()