累计注意力大模型

不稳定

import paddle


class HeadLoss(paddle.nn.Layer):
    def __init__(self):
        super(HeadLoss, self).__init__()
        self.loss = paddle.nn.CrossEntropyLoss()

    def forward(self, x_list, label):
        loss = 0
        h = x_list[0].shape[-1]
        p = len(x_list) + 1
        acc_data = 1
        for i, out in enumerate(x_list):
            i += 1
            one_label = (label % h ** (p - i) // h ** ((p - i) - 1)).astype("int64")
            loss += self.loss(out.reshape([-1, h]), one_label.reshape([-1]))
            with paddle.no_grad():
                acc_data *= (paddle.argmax(out, -1) == one_label).numpy()

        return loss, acc_data.mean()


class HiddenHead(paddle.nn.Layer):
    def __init__(self, voc_size=19, hidden_size=512):
        super(HiddenHead, self).__init__()

        self.hidden_size = hidden_size
        p = 0
        while True:
            voc_size //= hidden_size
            if voc_size == 0:
                break
            else:
                p += 1
        self.head = paddle.nn.LayerList(
            [paddle.nn.Linear(hidden_size, hidden_size, bias_attr=False) for _ in range(p + 1)])

    def forward(self, head_add_x):
        x0_list = []
        for i, head in enumerate(self.head):
            i += 1
            x0_list.append(head(head_add_x))
        return x0_list

    def sample(self, head_add_x):
        h = head_add_x.shape[-1]
        x0_res = 0

        p = len(self.head) + 1

        for i, head in enumerate(self.head):
            i += 1
            arg_max = head(head_add_x)
            arg_max = paddle.argmax(arg_max, -1)
            x0_res += arg_max * h ** ((p - i) - 1)
        return x0_res


class EmAdd(paddle.nn.Layer):
    def __init__(self, voc_size=9999, hidden_size=256):
        super(EmAdd, self).__init__()
        self.hidden_size = hidden_size
        p = 0
        while True:
            voc_size //= hidden_size
            if voc_size == 0:
                break
            else:
                p += 1
        self.em = paddle.nn.LayerList([paddle.nn.Embedding(hidden_size, hidden_size) for _ in range(p + 1)])
        # self.em_zero = paddle.nn.LayerList([paddle.nn.Embedding(hidden_size, hidden_size,padding_idx=0)
        # for _ in range(p + 1)])

    def forward(self, em_add_x):

        add = 0
        p = len(self.em) + 1
        mask = paddle.zeros(em_add_x.shape)
        for i, em in enumerate(self.em):
            # mask 是累加不等于0
            i += 1
            x0 = em_add_x % self.hidden_size ** (p - i) // self.hidden_size ** ((p - i) - 1)
            mask += x0
            mask = mask != 0
            mask = mask.astype("int")
            x0 = em(x0) * mask.unsqueeze(-1)
            add = paddle.sin(x0 + add)

        return add


class ReNet(paddle.nn.Layer):
    def __init__(self, voc_size, hidden_dim, num_head):
        super(ReNet, self).__init__()
        self.group_norm = paddle.nn.GroupNorm
        self.em = EmAdd(voc_size, hidden_dim * num_head)
        self.head = num_head

    def forward(self, x):
        x = self.em(x)
        seq_len = x.shape[1]
        bsz = x.shape[0]
        x = x.reshape([bsz, seq_len, self.head, -1]).transpose([0, 2, 1, 3])

        q = paddle.sin(x)
        k = paddle.sin(q + x)
        v = paddle.sin(k + x)
        attention= self.ParallelRetention(seq_len,q, k, v)
        return
    def sample_forward(self, x):
        x = self.em(x)
        seq_len = x.shape[1]
        bsz = x.shape[0]
        x = x.reshape([bsz, seq_len, self.head, -1]).transpose([0, 2, 1, 3])

        q = paddle.sin(x)
        k = paddle.sin(q + x)
        v = paddle.sin(k + x)
        attention,state = self.RecurrentRetention(q, k, v,paddle.zeros(q.shape),"")
        return x

    def ParallelRetention(self,seq_len,
                          q,  # bsz*num_head*len*qk_dim
                          k,  # bsz*num_head*len*qk_dim
                          v,  # bsz*num_head*len*v_dim
                          ):
        
        
        decay_mask = paddle.triu(paddle.ones([seq_len, seq_len])).T  # num_head*len*len
        retention = q @ k.transpose([0, 1, 3, 2])
        retention = retention * decay_mask
        output = retention @ v

        current_kv = k.unsqueeze(-1) * v.unsqueeze(-2)
        output = paddle.sum(q.unsqueeze(-1) * current_kv, -2)
        output = self.group_norm(self.head, self.head)(output)
        return output
def demo():
        
    import numpy as np
    # 定义矩阵 A，B，C，D
    kup = np.array([[5, 2],
                  [3, 4]])
    
    vup = np.array([[5, 6],
                  [7, 8]])
    
    knext = np.array([[9, 10],
                  [11, 12]])
    
    vnext = np.array([[13, 14],
                  [15, 16]])
    
    # 重新排列顺序为：(kup @ (vup + vnext)) + (knext @ (vup + vnext))
    # 在kup 和 vup 完全 在s 
    left = np.dot(kup, vup + vnext) + np.dot(knext, vup + vnext)
    
    # 右边的等式：(A + C) @ (B + D)
    right = np.dot(kup + knext, vup + vnext)
    
    # 检查左边和右边是否相等
    if np.array_equal(left, right):
        print("(kup @ (vup + vnext)) + (knext @ (vup + vnext))")
    else:
        print("不相等")
        
if __name__ == '__main__':
    net = ReNet(12935, 128, 8)
    # net(paddle.randint(1, 123, [3, 23]))
    # net.sample_forward(paddle.randint(1, 123, [3, 23]))
    s=10
    h=12
    a=paddle.randint(1, 123, [s,h]).astype("float32")
    b=paddle.randint(1, 123, [s,h]).astype("float32")
    c=paddle.randint(1, 123, [s,h]).astype("float32")
    print()




#  A1*A4+A2*A5+A3*A6  ,A1*B4+A2*B5+A3*B6
#  B1*A4+B2*A5+B3*A6  ,B1*B4+B2*B5+B3*B6

#  A7(A1*A4+A2*A5+A3*A6) + B7*(A1*B4+A2*B5+A3*B6),  A8(A1*A4+A2*A5+A3*A6) + B8*(A1*B4+A2*B5+A3*B6)
#  A7*(B1*A4+B2*A5+B3*A6)+B7*(B1*B4+B2*B5+B3*B6) ,A8*(B1*A4+B2*A5+B3*A6)+B8*(B1*B4+B2*B5+B3*B6)


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

稳定

import paddle
import numpy as np


class HeadLoss(paddle.nn.Layer):
    def __init__(self):
        super(HeadLoss, self).__init__()
        self.loss = paddle.nn.CrossEntropyLoss()

    def forward(self, x_list, label):
        loss = 0
        h = x_list[0].shape[-1]
        p = len(x_list) + 1
        acc_data = 1
        for i, out in enumerate(x_list):
            i += 1
            one_label = (label % h ** (p - i) // h ** ((p - i) - 1)).astype("int64")
            loss += self.loss(out.reshape([-1, h]), one_label.reshape([-1]))
            with paddle.no_grad():
                acc_data *= (paddle.argmax(out, -1) == one_label).numpy()

        return loss, acc_data.mean()


class HiddenHead(paddle.nn.Layer):
    def __init__(self, voc_size=19, hidden_size=512):
        super(HiddenHead, self).__init__()

        self.hidden_size = hidden_size
        p = 0
        while True:
            voc_size //= hidden_size
            if voc_size == 0:
                break
            else:
                p += 1
        self.head = paddle.nn.LayerList(
            [paddle.nn.Linear(hidden_size, hidden_size, bias_attr=False) for _ in range(p + 1)])

    def forward(self, head_add_x):
        x0_list = []
        for i, head in enumerate(self.head):
            i += 1
            x0_list.append(head(head_add_x))
        return x0_list

    def sample(self, head_add_x):
        h = head_add_x.shape[-1]
        x0_res = 0

        p = len(self.head) + 1

        for i, head in enumerate(self.head):
            i += 1
            arg_max = head(head_add_x)
            arg_max = paddle.argmax(arg_max, -1)
            x0_res += arg_max * h ** ((p - i) - 1)
        return x0_res


class EmAdd(paddle.nn.Layer):
    def __init__(self, voc_size=9999, hidden_size=256):
        super(EmAdd, self).__init__()
        self.hidden_size = hidden_size
        p = 0
        while True:
            voc_size //= hidden_size
            if voc_size == 0:
                break
            else:
                p += 1
        self.em = paddle.nn.LayerList([paddle.nn.Embedding(hidden_size, hidden_size) for _ in range(p + 1)])
        # self.em_zero = paddle.nn.LayerList([paddle.nn.Embedding(hidden_size, hidden_size,padding_idx=0)
        # for _ in range(p + 1)])

    def forward(self, em_add_x):

        add = 0
        p = len(self.em) + 1
        mask = paddle.zeros(em_add_x.shape)
        for i, em in enumerate(self.em):
            # mask 是累加不等于0
            i += 1
            x0 = em_add_x % self.hidden_size ** (p - i) // self.hidden_size ** ((p - i) - 1)
            mask += x0
            mask = mask != 0
            mask = mask.astype("int")
            x0 = em(x0) * mask.unsqueeze(-1)
            add = paddle.sin(x0 + add)

        return add


class ReNet(paddle.nn.Layer):
    def __init__(self, voc_size, hidden_dim, num_head_dim, n_layers):
        super(ReNet, self).__init__()
        self.em = EmAdd(voc_size, hidden_dim * num_head_dim)
        self.qk_list = paddle.nn.LayerList(
            [paddle.nn.Linear(num_head_dim, hidden_dim, bias_attr=False) for _ in range(n_layers)])

        self.head = num_head_dim
        self.out_layer = HiddenHead(voc_size, hidden_dim)


    def forward(self, x, state):
        x = self.em(x)
        q = paddle.sin(x)
        k = paddle.sin(x + q)
        seq_len = q.shape[1]
        bsz = q.shape[0]
        q = q.reshape([bsz, seq_len, self.head, -1]).transpose([0, 2, 1, 3])
        k = k.reshape([bsz, seq_len, self.head, -1]).transpose([0, 2, 1, 3])
        mask = paddle.triu(paddle.ones([seq_len, seq_len])).T

        qk = state + paddle.sum(paddle.sin(q @ k.transpose([0, 1, 3, 2]) * mask), -1)
        state = qk[:, :, -1:]
        new_qk=0
        for one_k in self.qk_list:
            new_qk += one_k(qk.transpose([0, 2, 1]))
            new_qk=paddle.sin(new_qk)

        out = self.out_layer(new_qk)
        return out, state

    def sample(self, x, state):
        x = self.em(x)
        q = paddle.sin(x)
        k = paddle.sin(x + q)
        seq_len = q.shape[1]
        bsz = q.shape[0]
        q = q.reshape([bsz, seq_len, self.head, -1]).transpose([0, 2, 1, 3])
        k = k.reshape([bsz, seq_len, self.head, -1]).transpose([0, 2, 1, 3])
        mask = paddle.triu(paddle.ones([seq_len, seq_len])).T

        qk = state + paddle.sum(paddle.sin(q @ k.transpose([0, 1, 3, 2]) * mask), -1)
        state = qk[:, :, -1:]
        new_qk = 0
        for one_k in self.qk_list:
            new_qk += one_k(qk.transpose([0, 2, 1]))
            new_qk = paddle.sin(new_qk)

        out = self.out_layer.sample(new_qk)
        return out, state



def emheading_train_and_sample():
    print("*" * 100)
    net = ReNet(12935, 80, 80, 8)
    # net.eval()
    x = paddle.to_tensor([
        np.random.randint(1, 124, 100),
        np.random.randint(1, 124, 100),
    ], dtype='int64')

    xx = x

    # 模拟训练

    loss_f = HeadLoss()
    opt = paddle.optimizer.Adam(parameters=net.parameters(), learning_rate=0.0003)

    for i in range(80):
        out, state = net(x[:, :-1], 0)
        loss, av_ac = loss_f(out, x[:, 1:])

        print(i, loss.item(), av_ac)

        opt.clear_grad()
        loss.backward()
        opt.step()

    # 解码，验证
    net.eval()

    out, _ = net.sample(xx[:, :-1], 0)
    print((out == xx[:, 1:]).numpy().mean())
    out, _ = net.sample(xx[:, :-2], 0)
    print((out[:, -1:] == xx[:, -2:-1]).numpy().mean())
    out, _ = net.sample(xx[:, :-1], 0)
    print((out[:, -1:] == xx[:, -1:]).numpy().mean())
    for i in range(50):
        print("超长依赖检验")
        out, _ = net.sample(xx[:, :i + 30], 0)
        print((out[:, -1:] == xx[:, i + 30:i + 31]).numpy().mean())

    out, _ = net(xx[:, :-1], 0)
    loss, av_ac = loss_f(out, xx[:, 1:])
    print(loss.item(), av_ac)


# 进行模型训练和预测
if __name__ == '__main__':
    emheading_train_and_sample()

# if __name__ == '__main__':
#     # demo()
#     net = ReNet(12935, 12, 8,8)
#     net(paddle.randint(1, 123, [3, 23]), 0)

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