测算sample gpt

在这里插入图片描述

测算代码


import pandas as pd
import matplotlib.pyplot as pltloss=pd.read_pickle("loss_8.pkl")
plt.plot(loss)
loss=pd.read_pickle("loss_16.pkl")
plt.plot(loss)
loss=pd.read_pickle("loss_4_8.pkl")
plt.plot(loss)
loss=pd.read_pickle("loss_8_8.pkl")
plt.plot(loss)loss=pd.read_pickle("loss_16_8.pkl")
plt.plot(loss)loss=pd.read_pickle("loss_8_4.pkl")
plt.plot(loss)
loss=pd.read_pickle("loss_16_4.pkl")
plt.plot(loss)
loss=pd.read_pickle("loss_8_32.pkl")
plt.plot(loss)
plt.legend(["8_256","16_256","4_8_256","8_8_512","16_8_512","8_4_512","16_4_512","256_8_32"])
plt.show()
# 总结规律后发现  参数维度越大越好，层数太多不好,8 层几个头影响不大。

在深度学习模型的设计中，参数的数量和层数是两个重要的因素。参数的维度通常与模型的学习能力有关，参数越多，模型能够捕捉的模式就越复杂。然而，这并不意味着参数越多越好，因为过多的参数可能会导致过拟合，即模型在训练数据上表现良好，但在未见过的新数据上表现不佳。
层数的多少影响模型的深度，理论上，更深的模型可以捕捉更复杂的特征，但同样存在过拟合的风险。此外，层数过多还会导致梯度消失或梯度爆炸问题，使得模型难以训练。
关于8层和几个头的影响，这通常指的是Transformer模型中的结构。在Transformer模型中，每层的输出可以被分割成多个“头”，每个头关注不同的信息。这种结构被称为多头注意力机制。8层的Transformer模型是一个相对较深的模型，而头的数量决定了模型关注不同部分输入信息的能力。一般来说，头的数量不会对模型性能产生太大的影响，但过多的头可能会导致计算资源的浪费。
总结来说，参数维度越大，模型的学习能力越强，但过大的参数维度可能会导致过拟合。层数的增加可以提高模型捕捉复杂特征的能力，但同时也增加了过拟合和训练难度。在Transformer模型中，8层的深度可以捕捉较为复杂的特征，而头的数量对模型性能的影响相对较小。在设计模型时，需要根据具体任务和数据集的特点，合理选择参数维度、层数和头的数量，以达到最佳的性能。

网络结构

import mathimport numpy as np
import paddle
import paddle.nn as nn
import paddle.nn.functional as Fclass MaskMultiHeadAttention(nn.Layer):def __init__(self, hidden_size, num_heads):super(MaskMultiHeadAttention, self).__init__()assert hidden_size % num_heads == 0, "Hidden size must be divisible by the number of heads."self.hidden_size = hidden_sizeself.num_heads = num_headsself.head_size = hidden_size // num_heads# Query, Key, Value matricesself.query_linear = nn.Linear(hidden_size, hidden_size, bias_attr=False)self.key_linear = nn.Linear(hidden_size, hidden_size, bias_attr=False)self.value_linear = nn.Linear(hidden_size, hidden_size, bias_attr=False)# Output matrixself.output_linear = nn.Linear(hidden_size, hidden_size,bias_attr=False)def forward(self, x, state=None, seq_len=None):batch_size = x.shape[0]# Compute Query, Key, Value for all heads in parallelquery = self.query_linear(x).reshape([batch_size, -1, self.num_heads, self.head_size]).transpose([0, 2, 1, 3])key = self.query_linear(x).reshape([batch_size, -1, self.num_heads, self.head_size]).transpose([0, 2, 1, 3])value = self.value_linear(x).reshape([batch_size, -1, self.num_heads, self.head_size]).transpose([0, 2, 1, 3])# Compute attention scoresscores = (F.relu(query) + F.relu(-query)) / (self.head_size ** 0.5 + 0.000000000001)key = (F.relu(key) + F.relu(-key)) / (self.head_size ** 0.5 + 0.000000000001)+0.00000000001key = key/paddle.sum(key, axis=-1, keepdim=True)if state is None:state = 0state = paddle.cumsum(scores, -2) + 0.0000000000000001 + statescores = scores / statestate = state[:, :, -1:]out = scores * value*key# Concatenate and transform to get the final outputout = out.transpose([0, 2, 1, 3]).reshape([batch_size, -1, self.hidden_size])out = self.output_linear(out)return out, stateclass FeedForward(nn.Layer):def __init__(self, hidden_size):super(FeedForward, self).__init__()self.ffn1 = nn.Linear(hidden_size, hidden_size*2)self.ffn2 = nn.Linear(hidden_size*2, hidden_size)self.gate = nn.Linear(hidden_size, hidden_size*2)self.relu = nn.Silu()def forward(self, x):x1 = self.ffn1(x)x2 = self.relu(self.gate(x))x = x1*x2x = self.ffn2(x)return xclass RMSNorm(nn.Layer):def __init__(self,dim, eps: float = 1e-6):super(RMSNorm,self).__init__()self.eps = epsself.fc=paddle.create_parameter(shape=[dim],dtype='float32',default_initializer=nn.initializer.Constant(value=1.0))def norm(self, x):return x * paddle.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)def forward(self, x):output = self.norm(x)return output*self.fcclass GPTDecoderLayer(nn.Layer):def __init__(self, hidden_size, num_heads):super(GPTDecoderLayer, self).__init__()self.self_attention = MaskMultiHeadAttention(hidden_size, num_heads)self.ffn = FeedForward(hidden_size)# self.norm1 = nn.LayerNorm(hidden_size)self.norm1 = RMSNorm(hidden_size)def forward(self, x, state=None, seq_len=None):x1, state = self.self_attention(x, state, seq_len=None)  # Self-Attention with residual connectionx = x1 + xx = self.norm1(x)x = self.ffn(x) + x  # Feed-Forward with residual connectionx = self.norm1(x)return x, stateclass PositionalEncoding(nn.Layer):def __init__(self, d_model, max_len=5000):super(PositionalEncoding, self).__init__()# Create a long enough Paddle array to hold position encodings for the maximum sequence lengthposition = paddle.arange(max_len).unsqueeze(1).astype("float32")# Create a constant 'pe' matrix with the same size as the embedding matrixdiv_term = paddle.exp(paddle.arange(0, d_model, 2) * (-math.log(10000.0) / d_model))pe = paddle.zeros([max_len, d_model])pe[:, 0::2] = paddle.sin(position * div_term)pe[:, 1::2] = paddle.cos(position * div_term)self.pe = pe.unsqueeze(0)  # Shape: [1, max_len, d_model]# Register 'pe' as a buffer (non-trainable parameter)def forward(self, x, seq_len=None):# x is of shape [batch_size, seq_len, d_model]if seq_len is None:seq_len = x.shape[1]return x + self.pe[:, :seq_len, :]else:return x + self.pe[:, seq_len - 1:seq_len, :]# %%def sinusoidal_position_embedding(max_len, output_dim):# (max_len, 1)position = paddle.arange(0, max_len, dtype="float32").unsqueeze(-1)# (output_dim//2)ids = paddle.arange(0, output_dim // 2, dtype="float32")  # 即公式里的i, i的范围是 [0,d/2]theta = 10000 ** (-2 * ids / output_dim)# (max_len, output_dim//2)embeddings = position * theta  # 即公式里的：pos / (10000^(2i/d))sin_embeddings = paddle.sin(embeddings)cos_embeddings = paddle.cos(embeddings)return sin_embeddings, cos_embeddingsdef rope(q, sin_em, cos_em, seq_len=None):if seq_len is None:sin_em = sin_em[:q.shape[2]]cos_em = cos_em[:q.shape[2]]else:sin_em = sin_em[seq_len - 1:seq_len]cos_em = cos_em[seq_len - 1:seq_len]q1 = q.reshape([q.shape[0], q.shape[1], q.shape[2], -1, 2])[..., 1]q2 = q.reshape([q.shape[0], q.shape[1], q.shape[2], -1, 2])[..., 0]# 奇数负值*sin_em+偶数正值*cos_em  奇数正值*cos_em+偶数正值*sin_emq3 = paddle.stack([-q1 * sin_em + q2 * cos_em, q1 * cos_em + q2 * sin_em], -1)q = q3.reshape(q.shape)  # reshape后就是正负交替了return qclass GPT(nn.Layer):def __init__(self, vocab_size, hidden_size, num_heads, num_layers):super(GPT, self).__init__()self.embedding = nn.Embedding(vocab_size, hidden_size)self.decoder_layers = nn.LayerList([GPTDecoderLayer(hidden_size, num_heads) for _ in range(num_layers)])self.fc = nn.Linear(hidden_size, vocab_size, bias_attr=False)self.sin_em, self.cos_em = sinusoidal_position_embedding(50000, hidden_size//num_heads//2)def forward(self, x, state=None, seq_len=None):x = self.embedding(x)# x = self.position_embedding(x, seq_len)if state is None:state = [None] * len(self.decoder_layers)i = 0for decoder_layer in self.decoder_layers:x = rope(x.reshape([x.shape[0], x.shape[1], -1, self.sin_em.shape[1] * 2]).transpose([0, 2, 1, 3]),self.sin_em,self.cos_em, seq_len).transpose([0, 2, 1, 3]).reshape(x.shape) + xx, state[i] = decoder_layer(x, state[i])i += 1out = self.fc(x)return out, statedef check_rope():q = paddle.randn((8, 12, 10, 32))sin_em, cos_em = sinusoidal_position_embedding(50000, 32)rope(q, sin_em, cos_em)def check_mask_multi_head_attention():x = paddle.randn([4, 10, 20])mha = MaskMultiHeadAttention(20, 5)out = mha(x)def check_positional_encoding():x = paddle.randn([4, 10, 20])pe = PositionalEncoding(20)out = pe(x)def check_gpt_decoder_layer():x = paddle.randn([4, 10, 20])dl = GPTDecoderLayer(20, 5)out = dl(x)def check_gpt():x = paddle.randint(4, 10, [4, 10])gpt = GPT(10, 20, 5, 2)out = gpt(x)def check_lisa_adam_w():gpt = GPT(10, 20, 5, 2)# 初始化模型和超参数nl = len(gpt.parameters())T = 100K = 10gamma = 2eta0 = 0.001pnl = gamma / np.ones(len(gpt.parameters())) * gamma / nlpnl[0] = 1pnl[-1] = 1criterion = nn.CrossEntropyLoss()optimizer = paddle.optimizer.AdamW(parameters=gpt.parameters(), learning_rate=eta0)# 训练模型for i in range(T // K):# 冻结所有层for l in gpt.parameters():l.trainable = False# 激活层采样for l, p_l in enumerate(pnl):print(gpt.parameters()[l].shape)if paddle.rand([1]) < p_l and len(gpt.parameters()[l].shape) > 1:w_lora_l = paddle.norm(paddle.linalg.svd(gpt.parameters()[l])[0], p=2) + paddle.norm(paddle.linalg.svd(gpt.parameters()[l])[2], p=2)  # LoRA权重范数w_lisa_l = paddle.norm(gpt.parameters()[l], p=2)  # LISA权重范数p_l_1 = w_lora_l / w_lisa_loptimizer.set_lr(optimizer.get_lr() * p_l_1.item())# 激活该层gpt.parameters()[l].trainable = True# 训练激活层for _ in range(K):x = paddle.randint(4, 10, [4, 10])out,_ = gpt(x)loss = criterion(out, paddle.randint(0, 10, [4, 10]))optimizer.clear_grad()loss.backward()optimizer.step()# 输出模型参数# print(gpt.state_dict())def check_all():# check_lisa_adam_w()check_rope()check_mask_multi_head_attention()check_positional_encoding()check_gpt_decoder_layer()check_gpt()if __name__ == '__main__':# 检查所有模块check_all()

训练

import json
import numpy as np
import paddle
import  paddle.nn as nn
from new_attention import GPT
import pandas as pd
with open("唐诗.json","r",encoding="utf-8") as f:data=f.read()
data=json.loads(data)
data=[i[4].split() for i in data if len(i[4].split())>3 ]
data=np.hstack(data)
data = [i for i in data if len("".join(i.split()))==24 and "a" not in i ]
data = [i for i in data if len("".join(i.split()))==24 and "f" not in i ]
data = [i for i in data if len("".join(i.split()))==24 and "e" not in i ]
data = [i for i in data if len("".join(i.split()))==24 and "h" not in i ]
data = [i for i in data if len("".join(i.split()))==24 and "X" not in i ]
data = [i for i in data if len("".join(i.split()))==24 and "“" not in i ]
data = [i for i in data if len("".join(i.split()))==24 and '□' not in i ]
data = [i for i in data if len("".join(i.split()))==24 and '《' not in i ]
data = [i for i in data if len("".join(i.split()))==24 and '》' not in i ]
voc=sorted(set(np.hstack([list(set(list(i))) for i in data])))
data_set=[[voc.index(j) for j in i]  for i  in data]from tqdm import tqdmbatch_size = 600
epochs = 10
model = GPT(len(voc), 256, 8,32)
# model.load_dict(paddle.load("gpt.pdparams"))
loss_func = nn.CrossEntropyLoss()
opt = paddle.optimizer.Adam(parameters=model.parameters(), learning_rate=0.0003)
bar = tqdm(range(epochs))
epochs_loss=[]
for epoch in bar:np.random.shuffle(data_set)loss_list=[]for i in range(0, len(data_set), batch_size):j = batch_size + ibatch_data = paddle.to_tensor(data_set[i:j]).astype('int64')out,_ = model(batch_data[:, :-1])loss = loss_func(out, batch_data[:, 1:])loss_list.append(loss.item())bar.set_description("epoch:{}_____loss:{}".format(epoch,np.mean(loss_list)))opt.clear_grad()loss.backward()opt.step()epochs_loss.append(np.mean(loss_list))paddle.save(model.state_dict(), "gpt.pdparams")pd.to_pickle(epochs_loss,"loss_8_32.pkl")