自己训练一个小模型

自己训练一个小模型

 

Chatgpt都出4o为什么还要自己训练一个小模型呢？

• 如果你能自己训练一个小模型，说明你已经掌握Transformer了

• 有些场景就需要小模型，小而美，就像单片机就是有它的市场

• 训练小模型不是目的，目的是通过训练小模型跑通大模型训练的基本流程

 

想要看懂或写出训练小模型的代码，需要学习前置 Transformer模型、机器学习、线性代数等相关知识

看不懂也没关系，直接运行试试效果

上代码

!pip install numpy requests torch tiktoken matplotlib pandas

 

import os
import requests
import math
import tiktoken
import torch
import torch.nn as nn
from torch.nn import functional as F
​
# Hyperparameters
batch_size = 4  # How many batches per training step
context_length = 16  # Length of the token chunk each batch
d_model = 64  # The size of our model token embeddings
num_blocks = 8  # Number of transformer blocks
num_heads = 4  # Number of heads in Multi-head attention
learning_rate = 1e-3  # 0.001
dropout = 0.1  # Dropout rate
max_iters = 5000  # Total of training iterations <- Change this to smaller number for testing
eval_interval = 50  # How often to evaluate
eval_iters = 20  # Number of iterations to average for evaluation
device = 'cuda' if torch.cuda.is_available() else 'cpu'  # Use GPU if it's available.
TORCH_SEED = 1337
torch.manual_seed(TORCH_SEED)
​
# Load training data
if not os.path.exists('data/sales_textbook.txt'):
    url = 'https://huggingface.co/datasets/goendalf666/sales-textbook_for_convincing_and_selling/raw/main/sales_textbook.txt'
    with open('data/sales_textbook.txt', 'w') as f:
        f.write(requests.get(url).text)
​
with open('data/sales_textbook.txt', 'r', encoding='utf-8') as f:
    text = f.read()
​
# Using TikToken (Same as GPT3) to tokenize the source text
encoding = tiktoken.get_encoding("cl100k_base")
tokenized_text = encoding.encode(text)
max_token_value = max(tokenized_text) + 1  # the maximum value of the tokenized numbers
tokenized_text = torch.tensor(tokenized_text, dtype=torch.long, device=device)  # put tokenized text into tensor
​
# Split train and validation
split_idx = int(len(tokenized_text) * 0.9)
train_data = tokenized_text[:split_idx]
val_data = tokenized_text[split_idx:]
​
​
# Define Feed Forward Network
class FeedForward(nn.Module):
    def __init__(self):
        super().__init__()
        self.d_model = d_model
        self.dropout = dropout
        self.ffn = nn.Sequential(
            nn.Linear(in_features=self.d_model, out_features=self.d_model * 4),
            nn.ReLU(),
            nn.Linear(in_features=self.d_model * 4, out_features=self.d_model),
            nn.Dropout(dropout),
        )
​
    def forward(self, x):
        return self.ffn(x)
​
​
# Define Scaled Dot Product Attention
class Attention(nn.Module):
    def __init__(self, head_size: int):
        super().__init__()
        self.d_model = d_model
        self.head_size = head_size
        self.context_length = context_length
        self.dropout = dropout
​
        self.key_layer = nn.Linear(in_features=self.d_model, out_features=self.head_size, bias=False)
        self.query_layer = nn.Linear(in_features=self.d_model, out_features=self.head_size, bias=False)
        self.value_layer = nn.Linear(in_features=self.d_model, out_features=self.head_size, bias=False)
        self.register_buffer('tril', torch.tril(
            torch.ones((self.context_length, self.context_length))))  # Lower triangular mask
        self.dropout_layer = nn.Dropout(self.dropout)
​
    def forward(self, x):
        B, T, C = x.shape  # Batch size, Time steps(current context_length), Channels(dimensions)
        assert T <= self.context_length
        assert C == self.d_model
        q = self.query_layer(x)
        k = self.key_layer(x)
        v = self.value_layer(x)
​
        # Scaled dot product attention: Q @ K^T / sqrt(d_k)
        weights = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
        # Apply masked attention
        weights = weights.masked_fill(self.tril[:T, :T] == 0, float('-inf'))
        weights = F.softmax(input=weights, dim=-1)
        weights = self.dropout_layer(weights)
​
        # Apply dot product attention: weights @ V
        out = weights @ v
        return out
​
​
class MultiHeadAttention(nn.Module):
    def __init__(self, head_size: int):
        super().__init__()
        self.num_heads = num_heads
        self.head_size = head_size
        self.d_model = d_model
        self.context_length = context_length
        self.dropout = dropout
​
        self.heads = nn.ModuleList([Attention(head_size=self.head_size) for _ in range(self.num_heads)])
        self.projection_layer = nn.Linear(in_features=self.d_model, out_features=self.d_model)
        self.dropout_layer = nn.Dropout(dropout)
​
    def forward(self, x):
        out = torch.cat([h(x) for h in self.heads], dim=-1)
        out = self.projection_layer(out)
        out = self.dropout_layer(out)
        return out
​
​
class TransformerBlock(nn.Module):
​
    def __init__(self, num_heads: int):
        super().__init__()
        self.d_model = d_model
        self.context_length = context_length
        self.head_size = d_model // num_heads  # head size should be divisible by d_model
        self.num_heads = num_heads
        self.dropout = dropout
​
        self.multi_head_attention_layer = MultiHeadAttention(head_size=self.head_size)
        self.feed_forward_layer = FeedForward()
        self.layer_norm_1 = nn.LayerNorm(normalized_shape=self.d_model)
        self.layer_norm_2 = nn.LayerNorm(normalized_shape=self.d_model)
​
    def forward(self, x):
        # Note: The order of the operations is different from the original Transformer paper
        # The order here is: LayerNorm -> Multi-head attention -> LayerNorm -> Feed forward
        x = x + self.multi_head_attention_layer(self.layer_norm_1(x))  # Residual connection
        x = x + self.feed_forward_layer(self.layer_norm_2(x))  # Residual connection
        return x
​
​
class TransformerLanguageModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.d_model = d_model
        self.context_length = context_length
        self.num_heads = num_heads
        self.num_blocks = num_blocks
        self.dropout = dropout
        self.max_token_value = max_token_value
        # Set up token embedding look-up table
        self.token_embedding_lookup_table = nn.Embedding(num_embeddings=self.max_token_value + 1, embedding_dim=self.d_model)
​
        # Run all the transformer blocks
        # Different from original paper, here we add a final layer norm after all the blocks
        self.transformer_blocks = nn.Sequential(*(
                [TransformerBlock(num_heads=self.num_heads) for _ in range(self.num_blocks)] +
                [nn.LayerNorm(self.d_model)]
        ))
        self.language_model_out_linear_layer = nn.Linear(in_features=self.d_model, out_features=self.max_token_value)
​
    def forward(self, idx, targets=None):
        B, T = idx.shape
        """
        # Set up position embedding look-up table
        # following the same approach as the original Transformer paper (Sine and Cosine functions)
        """
        position_encoding_lookup_table = torch.zeros(self.context_length, self.d_model)
        position = torch.arange(0, self.context_length, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, self.d_model, 2).float() * (-math.log(10000.0) / self.d_model))
        position_encoding_lookup_table[:, 0::2] = torch.sin(position * div_term)
        position_encoding_lookup_table[:, 1::2] = torch.cos(position * div_term)
        # change position_encoding_lookup_table from (context_length, d_model) to (T, d_model)
        position_embedding = position_encoding_lookup_table[:T, :].to(device)
        x = self.token_embedding_lookup_table(idx) + position_embedding
        x = self.transformer_blocks(x)
        # The "logits" are the output values of our model before applying softmax
        logits = self.language_model_out_linear_layer(x)
​
        if targets is not None:
            B, T, C = logits.shape
            logits_reshaped = logits.view(B * T, C)
            targets_reshaped = targets.view(B * T)
            loss = F.cross_entropy(input=logits_reshaped, target=targets_reshaped)
        else:
            loss = None
        return logits, loss
​
    def generate(self, idx, max_new_tokens):
        # idx is (B,T) array of indices in the current context
        for _ in range(max_new_tokens):
            # Crop idx to the max size of our positional embeddings table
            idx_crop = idx[:, -self.context_length:]
            # Get predictions
            logits, loss = self(idx_crop)
            # Get the last time step from logits where the dimensions of the logits are (B,T,C)
            logits_last_timestep = logits[:, -1, :]
            # Apply softmax to get probabilities
            probs = F.softmax(input=logits_last_timestep, dim=-1)
            # Sample from the probabilities' distribution.
            idx_next = torch.multinomial(input=probs, num_samples=1)
            # Append the sampled indexes idx_next to idx
            idx = torch.cat((idx, idx_next), dim=1)
        return idx
​
​
# Initialize the model
model = TransformerLanguageModel()
model = model.to(device)
​
​
# Get input embedding batch
def get_batch(split: str):
    data = train_data if split == 'train' else val_data
    idxs = torch.randint(low=0, high=len(data) - context_length, size=(batch_size,))
    x = torch.stack([data[idx:idx + context_length] for idx in idxs]).to(device)
    y = torch.stack([data[idx + 1:idx + context_length + 1] for idx in idxs]).to(device)
    return x, y
​
​
# Calculate loss
@torch.no_grad()
def estimate_loss():
    out = {}
    model.eval()
    for split in ['train', 'valid']:
        losses = torch.zeros(eval_iters)
        for k in range(eval_iters):
            x_batch, y_batch = get_batch(split)
            logits, loss = model(x_batch, y_batch)
            losses[k] = loss.item()
        out[split] = losses.mean()
    model.train()
    return out
​
​
# Use AdamW optimizer
optimizer = torch.optim.AdamW(params=model.parameters(), lr=learning_rate)
tracked_losses = list()
for step in range(max_iters):
    if step % eval_iters == 0 or step == max_iters - 1:
        losses = estimate_loss()
        tracked_losses.append(losses)
        print('Step:', step, 'Training Loss:', round(losses['train'].item(), 3), 'Validation Loss:',
              round(losses['valid'].item(), 3))
​
    xb, yb = get_batch('train')
    logits, loss = model(xb, yb)
    optimizer.zero_grad(set_to_none=True)
    loss.backward()
    optimizer.step()
​
# Save the model state dictionary
torch.save(model.state_dict(), 'model-ckpt.pt')
​
# Generate
model.eval()
start = 'The salesperson'
start_ids = encoding.encode(start)
x = (torch.tensor(start_ids, dtype=torch.long, device=device)[None, ...])
y = model.generate(x, max_new_tokens=100)
print('---------------')
print(encoding.decode(y[0].tolist()))
print('---------------')
​

Step: 0 Training Loss: 11.663 Validation Loss: 11.716
Step: 20 Training Loss: 10.297 Validation Loss: 10.478
Step: 40 Training Loss: 8.867 Validation Loss: 9.022
Step: 60 Training Loss: 7.346 Validation Loss: 7.613
Step: 80 Training Loss: 6.878 Validation Loss: 7.297
Step: 100 Training Loss: 6.659 Validation Loss: 7.208
Step: 120 Training Loss: 6.544 Validation Loss: 7.104
Step: 140 Training Loss: 6.325 Validation Loss: 7.199
Step: 160 Training Loss: 6.34 Validation Loss: 6.684
Step: 180 Training Loss: 6.154 Validation Loss: 6.89
Step: 200 Training Loss: 6.202 Validation Loss: 6.673
Step: 220 Training Loss: 6.045 Validation Loss: 6.761
Step: 240 Training Loss: 5.871 Validation Loss: 6.497
Step: 260 Training Loss: 5.957 Validation Loss: 6.347
Step: 280 Training Loss: 5.679 Validation Loss: 6.389
Step: 300 Training Loss: 5.816 Validation Loss: 6.603
Step: 320 Training Loss: 5.415 Validation Loss: 6.496
Step: 340 Training Loss: 5.32 Validation Loss: 6.1
Step: 360 Training Loss: 5.206 Validation Loss: 6.222
Step: 380 Training Loss: 5.403 Validation Loss: 6.451
Step: 400 Training Loss: 5.317 Validation Loss: 5.937
Step: 420 Training Loss: 5.159 Validation Loss: 6.033
Step: 440 Training Loss: 5.153 Validation Loss: 6.333
Step: 460 Training Loss: 5.232 Validation Loss: 6.001
Step: 480 Training Loss: 5.126 Validation Loss: 6.067
Step: 500 Training Loss: 5.123 Validation Loss: 6.044
Step: 520 Training Loss: 4.966 Validation Loss: 5.676
Step: 540 Training Loss: 4.774 Validation Loss: 6.023
Step: 560 Training Loss: 4.792 Validation Loss: 6.079
Step: 580 Training Loss: 4.743 Validation Loss: 5.722
Step: 600 Training Loss: 4.818 Validation Loss: 5.686
Step: 620 Training Loss: 4.675 Validation Loss: 5.741
Step: 640 Training Loss: 4.805 Validation Loss: 6.014
Step: 660 Training Loss: 4.81 Validation Loss: 5.758
Step: 680 Training Loss: 4.727 Validation Loss: 5.723
Step: 700 Training Loss: 4.737 Validation Loss: 5.792
Step: 720 Training Loss: 4.609 Validation Loss: 5.761
Step: 740 Training Loss: 5.018 Validation Loss: 5.705
Step: 760 Training Loss: 4.906 Validation Loss: 5.721
Step: 780 Training Loss: 4.791 Validation Loss: 5.779
Step: 800 Training Loss: 4.467 Validation Loss: 5.881
Step: 820 Training Loss: 4.443 Validation Loss: 5.502
Step: 840 Training Loss: 4.567 Validation Loss: 5.832
Step: 860 Training Loss: 4.577 Validation Loss: 5.956
Step: 880 Training Loss: 4.55 Validation Loss: 5.583
Step: 900 Training Loss: 4.478 Validation Loss: 5.465
Step: 920 Training Loss: 4.237 Validation Loss: 5.674
Step: 940 Training Loss: 4.462 Validation Loss: 5.427
Step: 960 Training Loss: 4.323 Validation Loss: 5.632
Step: 980 Training Loss: 4.323 Validation Loss: 5.711
Step: 1000 Training Loss: 4.304 Validation Loss: 5.374
Step: 1020 Training Loss: 4.295 Validation Loss: 5.597
Step: 1040 Training Loss: 4.312 Validation Loss: 5.54
Step: 1060 Training Loss: 4.351 Validation Loss: 5.456
Step: 1080 Training Loss: 4.128 Validation Loss: 5.524
Step: 1100 Training Loss: 4.285 Validation Loss: 5.44
Step: 1120 Training Loss: 4.359 Validation Loss: 5.447
Step: 1140 Training Loss: 4.276 Validation Loss: 5.527
Step: 1160 Training Loss: 4.179 Validation Loss: 5.415
Step: 1180 Training Loss: 4.057 Validation Loss: 5.42
Step: 1200 Training Loss: 4.238 Validation Loss: 5.296
Step: 1220 Training Loss: 3.979 Validation Loss: 5.535
Step: 1240 Training Loss: 4.145 Validation Loss: 5.417
Step: 1260 Training Loss: 4.093 Validation Loss: 5.34
Step: 1280 Training Loss: 4.173 Validation Loss: 5.361
Step: 1300 Training Loss: 3.876 Validation Loss: 5.449
Step: 1320 Training Loss: 3.941 Validation Loss: 5.343
Step: 1340 Training Loss: 4.172 Validation Loss: 5.335
Step: 1360 Training Loss: 3.757 Validation Loss: 5.173
Step: 1380 Training Loss: 4.106 Validation Loss: 5.207
Step: 1400 Training Loss: 3.975 Validation Loss: 5.349
Step: 1420 Training Loss: 4.11 Validation Loss: 5.224
Step: 1440 Training Loss: 3.915 Validation Loss: 5.341
Step: 1460 Training Loss: 4.05 Validation Loss: 5.302
Step: 1480 Training Loss: 3.927 Validation Loss: 5.487
Step: 1500 Training Loss: 3.952 Validation Loss: 5.191
Step: 1520 Training Loss: 4.182 Validation Loss: 5.066
Step: 1540 Training Loss: 3.851 Validation Loss: 5.205
Step: 1560 Training Loss: 4.062 Validation Loss: 5.039
Step: 1580 Training Loss: 3.848 Validation Loss: 4.952
Step: 1600 Training Loss: 3.94 Validation Loss: 5.343
Step: 1620 Training Loss: 3.78 Validation Loss: 5.243
Step: 1640 Training Loss: 3.814 Validation Loss: 5.364
Step: 1660 Training Loss: 3.979 Validation Loss: 5.25
Step: 1680 Training Loss: 3.717 Validation Loss: 5.067
Step: 1700 Training Loss: 3.681 Validation Loss: 5.574
Step: 1720 Training Loss: 3.753 Validation Loss: 5.119
Step: 1740 Training Loss: 3.584 Validation Loss: 5.335
Step: 1760 Training Loss: 3.819 Validation Loss: 4.949
Step: 1780 Training Loss: 3.823 Validation Loss: 4.921
Step: 1800 Training Loss: 3.795 Validation Loss: 5.031
Step: 1820 Training Loss: 3.54 Validation Loss: 5.292
Step: 1840 Training Loss: 4.003 Validation Loss: 4.95
Step: 1860 Training Loss: 3.759 Validation Loss: 4.86
Step: 1880 Training Loss: 3.871 Validation Loss: 5.262
Step: 1900 Training Loss: 3.791 Validation Loss: 4.975
Step: 1920 Training Loss: 3.768 Validation Loss: 5.329
Step: 1940 Training Loss: 3.689 Validation Loss: 5.011
Step: 1960 Training Loss: 3.52 Validation Loss: 4.926
Step: 1980 Training Loss: 3.648 Validation Loss: 5.128
Step: 2000 Training Loss: 3.696 Validation Loss: 5.011
Step: 2020 Training Loss: 3.756 Validation Loss: 5.086
Step: 2040 Training Loss: 3.835 Validation Loss: 4.961
Step: 2060 Training Loss: 3.626 Validation Loss: 5.27
Step: 2080 Training Loss: 3.751 Validation Loss: 5.27
Step: 2100 Training Loss: 3.856 Validation Loss: 4.967
Step: 2120 Training Loss: 3.76 Validation Loss: 4.968
Step: 2140 Training Loss: 3.678 Validation Loss: 4.971
Step: 2160 Training Loss: 3.759 Validation Loss: 4.821
Step: 2180 Training Loss: 3.504 Validation Loss: 5.243
Step: 2200 Training Loss: 3.85 Validation Loss: 5.345
Step: 2220 Training Loss: 3.74 Validation Loss: 5.287
Step: 2240 Training Loss: 3.66 Validation Loss: 5.219
Step: 2260 Training Loss: 3.684 Validation Loss: 5.101
Step: 2280 Training Loss: 3.523 Validation Loss: 4.998
Step: 2300 Training Loss: 3.628 Validation Loss: 5.237
Step: 2320 Training Loss: 3.545 Validation Loss: 5.442
Step: 2340 Training Loss: 3.428 Validation Loss: 5.192
Step: 2360 Training Loss: 3.658 Validation Loss: 5.11
Step: 2380 Training Loss: 3.592 Validation Loss: 5.14
Step: 2400 Training Loss: 3.573 Validation Loss: 5.069
Step: 2420 Training Loss: 3.414 Validation Loss: 4.745
Step: 2440 Training Loss: 3.459 Validation Loss: 5.28
Step: 2460 Training Loss: 3.678 Validation Loss: 5.044
Step: 2480 Training Loss: 3.409 Validation Loss: 4.935
Step: 2500 Training Loss: 3.484 Validation Loss: 5.054
Step: 2520 Training Loss: 3.659 Validation Loss: 5.335
Step: 2540 Training Loss: 3.423 Validation Loss: 5.333
Step: 2560 Training Loss: 3.57 Validation Loss: 5.237
Step: 2580 Training Loss: 3.57 Validation Loss: 4.961
Step: 2600 Training Loss: 3.67 Validation Loss: 5.023
Step: 2620 Training Loss: 3.451 Validation Loss: 4.958
Step: 2640 Training Loss: 3.542 Validation Loss: 5.144
Step: 2660 Training Loss: 3.474 Validation Loss: 5.076
Step: 2680 Training Loss: 3.482 Validation Loss: 4.937
Step: 2700 Training Loss: 3.428 Validation Loss: 5.087
Step: 2720 Training Loss: 3.377 Validation Loss: 5.171
Step: 2740 Training Loss: 3.404 Validation Loss: 4.779
Step: 2760 Training Loss: 3.2 Validation Loss: 5.077
Step: 2780 Training Loss: 3.28 Validation Loss: 5.184
Step: 2800 Training Loss: 3.138 Validation Loss: 5.165
Step: 2820 Training Loss: 3.374 Validation Loss: 5.091
Step: 2840 Training Loss: 3.29 Validation Loss: 5.2
Step: 2860 Training Loss: 3.375 Validation Loss: 5.022
Step: 2880 Training Loss: 3.45 Validation Loss: 4.919
Step: 2900 Training Loss: 3.465 Validation Loss: 5.134
Step: 2920 Training Loss: 3.457 Validation Loss: 5.227
Step: 2940 Training Loss: 3.322 Validation Loss: 4.94
Step: 2960 Training Loss: 3.203 Validation Loss: 5.068
Step: 2980 Training Loss: 3.372 Validation Loss: 4.924
Step: 3000 Training Loss: 3.512 Validation Loss: 5.071
Step: 3020 Training Loss: 3.469 Validation Loss: 4.782
Step: 3040 Training Loss: 3.343 Validation Loss: 5.275
Step: 3060 Training Loss: 3.201 Validation Loss: 4.854
Step: 3080 Training Loss: 3.313 Validation Loss: 5.037
Step: 3100 Training Loss: 3.41 Validation Loss: 4.707
Step: 3120 Training Loss: 3.201 Validation Loss: 5.013
Step: 3140 Training Loss: 3.344 Validation Loss: 4.895
Step: 3160 Training Loss: 3.307 Validation Loss: 4.915
Step: 3180 Training Loss: 3.186 Validation Loss: 4.955
Step: 3200 Training Loss: 3.262 Validation Loss: 5.005
Step: 3220 Training Loss: 3.331 Validation Loss: 4.845
Step: 3240 Training Loss: 3.301 Validation Loss: 5.017
Step: 3260 Training Loss: 3.529 Validation Loss: 4.58
Step: 3280 Training Loss: 3.269 Validation Loss: 4.887
Step: 3300 Training Loss: 3.1 Validation Loss: 5.046
Step: 3320 Training Loss: 3.239 Validation Loss: 4.825
Step: 3340 Training Loss: 3.341 Validation Loss: 5.413
Step: 3360 Training Loss: 3.288 Validation Loss: 4.929
Step: 3380 Training Loss: 3.315 Validation Loss: 5.259
Step: 3400 Training Loss: 3.19 Validation Loss: 4.979
Step: 3420 Training Loss: 3.237 Validation Loss: 5.082
Step: 3440 Training Loss: 3.168 Validation Loss: 5.336
Step: 3460 Training Loss: 3.305 Validation Loss: 5.259
Step: 3480 Training Loss: 3.142 Validation Loss: 4.798
Step: 3500 Training Loss: 3.179 Validation Loss: 5.061
Step: 3520 Training Loss: 3.238 Validation Loss: 5.056
Step: 3540 Training Loss: 3.171 Validation Loss: 4.955
Step: 3560 Training Loss: 3.141 Validation Loss: 4.828
Step: 3580 Training Loss: 3.154 Validation Loss: 4.858
Step: 3600 Training Loss: 3.245 Validation Loss: 5.185
Step: 3620 Training Loss: 3.076 Validation Loss: 4.518
Step: 3640 Training Loss: 3.208 Validation Loss: 4.755
Step: 3660 Training Loss: 3.343 Validation Loss: 4.94
Step: 3680 Training Loss: 3.109 Validation Loss: 4.749
Step: 3700 Training Loss: 3.137 Validation Loss: 4.929
Step: 3720 Training Loss: 3.105 Validation Loss: 4.806
Step: 3740 Training Loss: 3.053 Validation Loss: 4.917
Step: 3760 Training Loss: 3.379 Validation Loss: 4.991
Step: 3780 Training Loss: 3.278 Validation Loss: 5.268
Step: 3800 Training Loss: 3.11 Validation Loss: 5.2
Step: 3820 Training Loss: 3.049 Validation Loss: 5.134
Step: 3840 Training Loss: 3.182 Validation Loss: 4.849
Step: 3860 Training Loss: 2.989 Validation Loss: 5.004
Step: 3880 Training Loss: 3.27 Validation Loss: 4.796
Step: 3900 Training Loss: 3.007 Validation Loss: 4.805
Step: 3920 Training Loss: 3.151 Validation Loss: 4.856
Step: 3940 Training Loss: 3.125 Validation Loss: 4.832
Step: 3960 Training Loss: 3.058 Validation Loss: 4.629
Step: 3980 Training Loss: 3.031 Validation Loss: 4.963
Step: 4000 Training Loss: 3.118 Validation Loss: 4.976
Step: 4020 Training Loss: 3.152 Validation Loss: 4.949
Step: 4040 Training Loss: 3.049 Validation Loss: 5.054
Step: 4060 Training Loss: 3.065 Validation Loss: 5.069
Step: 4080 Training Loss: 3.193 Validation Loss: 5.184
Step: 4100 Training Loss: 2.92 Validation Loss: 5.0
Step: 4120 Training Loss: 3.167 Validation Loss: 4.822
Step: 4140 Training Loss: 3.117 Validation Loss: 4.895
Step: 4160 Training Loss: 3.153 Validation Loss: 5.004
Step: 4180 Training Loss: 3.213 Validation Loss: 4.874
Step: 4200 Training Loss: 2.952 Validation Loss: 4.93
Step: 4220 Training Loss: 3.089 Validation Loss: 5.009
Step: 4240 Training Loss: 2.934 Validation Loss: 5.001
Step: 4260 Training Loss: 3.035 Validation Loss: 5.085
Step: 4280 Training Loss: 2.786 Validation Loss: 4.974
Step: 4300 Training Loss: 3.009 Validation Loss: 4.948
Step: 4320 Training Loss: 2.893 Validation Loss: 5.033
Step: 4340 Training Loss: 2.859 Validation Loss: 4.889
Step: 4360 Training Loss: 3.022 Validation Loss: 4.746
Step: 4380 Training Loss: 2.983 Validation Loss: 5.146
Step: 4400 Training Loss: 3.125 Validation Loss: 4.891
Step: 4420 Training Loss: 3.003 Validation Loss: 5.253
Step: 4440 Training Loss: 2.952 Validation Loss: 5.039
Step: 4460 Training Loss: 3.043 Validation Loss: 4.736
Step: 4480 Training Loss: 2.811 Validation Loss: 5.291
Step: 4500 Training Loss: 2.927 Validation Loss: 4.883
Step: 4520 Training Loss: 2.983 Validation Loss: 4.685
Step: 4540 Training Loss: 3.092 Validation Loss: 4.898
Step: 4560 Training Loss: 3.034 Validation Loss: 4.876
Step: 4580 Training Loss: 3.036 Validation Loss: 5.188
Step: 4600 Training Loss: 2.715 Validation Loss: 4.858
Step: 4620 Training Loss: 3.009 Validation Loss: 5.125
Step: 4640 Training Loss: 2.923 Validation Loss: 4.92
Step: 4660 Training Loss: 2.869 Validation Loss: 4.923
Step: 4680 Training Loss: 2.809 Validation Loss: 5.075
Step: 4700 Training Loss: 3.002 Validation Loss: 5.103
Step: 4720 Training Loss: 2.921 Validation Loss: 5.054
Step: 4740 Training Loss: 2.81 Validation Loss: 5.074
Step: 4760 Training Loss: 2.951 Validation Loss: 5.228
Step: 4780 Training Loss: 2.919 Validation Loss: 4.913
Step: 4800 Training Loss: 2.953 Validation Loss: 5.215
Step: 4820 Training Loss: 3.022 Validation Loss: 4.832
Step: 4840 Training Loss: 2.766 Validation Loss: 5.119
Step: 4860 Training Loss: 2.898 Validation Loss: 5.103
Step: 4880 Training Loss: 2.977 Validation Loss: 4.885
Step: 4900 Training Loss: 3.036 Validation Loss: 5.128
Step: 4920 Training Loss: 2.913 Validation Loss: 4.799
Step: 4940 Training Loss: 2.966 Validation Loss: 4.863
Step: 4960 Training Loss: 2.723 Validation Loss: 4.828
Step: 4980 Training Loss: 2.752 Validation Loss: 4.666
Step: 4999 Training Loss: 2.828 Validation Loss: 5.13
---------------
The salesperson, the customer, the salesperson can effectively gather information, and ultimately increasing the likelihood of the sale.
1. Be mindful of reinforcing of-ended questions are identified persuasive and manipulative use can significantly impact, demonstrating genuine interest requires a deeper level that we have successfully suit their approach to share their responses. Some customers, while some the root cause for them, showcasing patterns, sales professionals can meet their concerns and increase their requirements.
When faced with your product or service is not just about attacks but crafted situations
---------------

 

在线查看或执行：https://colab.research.google.com/drive/1hvgnvZhTNqJbfaHIN_29h0p9lqSx1IXU?usp=sharing

 

原创声明：本文为本人原创作品，首发于AI ONES https://wuxiongwei.com，如果转载，请保留本文链接，谢谢。