hacktricks/todo/llm-training-data-preparation/6.-pre-training-and-loading-models.md

# 6. Pre-training & Loading models

## Text Generation

In order to train a model we will need that model to be able to generate new tokens. Then we will compare the generated tokens with the expected ones in order to train the model into **learning the tokens it needs to generate**.

As in the previous examples we already predicted some tokens, it's possible to reuse that function for this purpose.

{% hint style="success" %}
The goal of this sixth phase is very simple: **Train the model from scratch**. For this the previous LLM architecture will be used with some loops going over the data sets using the defined loss functions and optimizer to train all the parameters of the model.
{% endhint %}

## Text Evaluation

In order to perform a correct training it's needed to measure check the predictions obtained for the expected token. The goal of the training is to maximize the likelihood of the correct token, which involves increasing its probability relative to other tokens.

In order to maximize the probability of the correct token, the weights of the model must be modified to that probability is maximised. The updates of the weights is done via **backpropagation**. This requires a **loss function to maximize**. In this case, the function will be the **difference between the performed prediction and the desired one**.

However, instead of working with the raw predictions, it will work with a logarithm with base n. So if the current prediction of the expected token was 7.4541e-05, the natural logarithm (base _e_) of **7.4541e-05** is approximately **-9.5042**.\
Then, for each entry with a context length of 5 tokens for example, the model will need to predict 5 tokens, being the first 4 tokens the last one of the input and the fifth the predicted one. Therefore, for each entry we will have 5 predictions in that case (even if the first 4 ones were in the input the model doesn't know this) with 5 expected token and therefore 5 probabilities to maximize.

Therefore, after performing the natural logarithm to each prediction, the **average** is calculated, the **minus symbol removed** (this is called _cross entropy loss_) and thats the **number to reduce as close to 0 as possible** because the natural logarithm of 1 is 0:

<figure><img src="../../.gitbook/assets/image (10) (1).png" alt="" width="563"><figcaption><p><a href="https://camo.githubusercontent.com/3c0ab9c55cefa10b667f1014b6c42df901fa330bb2bc9cea88885e784daec8ba/68747470733a2f2f73656261737469616e72617363686b612e636f6d2f696d616765732f4c4c4d732d66726f6d2d736372617463682d696d616765732f636830355f636f6d707265737365642f63726f73732d656e74726f70792e776562703f313233">https://camo.githubusercontent.com/3c0ab9c55cefa10b667f1014b6c42df901fa330bb2bc9cea88885e784daec8ba/68747470733a2f2f73656261737469616e72617363686b612e636f6d2f696d616765732f4c4c4d732d66726f6d2d736372617463682d696d616765732f636830355f636f6d707265737365642f63726f73732d656e74726f70792e776562703f313233</a></p></figcaption></figure>

Another way to measure how good the model is is called perplexity. **Perplexity** is a metric used to evaluate how well a probability model predicts a sample. In language modelling, it represents the **model's uncertainty** when predicting the next token in a sequence.\
For example, a perplexity value of 48725, means that when needed to predict a token it's unsure about which among 48,725 tokens in the vocabulary is the good one.

## Pre-Train Example

This is the initial code proposed in [https://github.com/rasbt/LLMs-from-scratch/blob/main/ch05/01\_main-chapter-code/ch05.ipynb](https://github.com/rasbt/LLMs-from-scratch/blob/main/ch05/01\_main-chapter-code/ch05.ipynb) some times slightly modify

<details>

<summary>Previous code used here but already explained in previous sections</summary>

```python
"""
This is code explained before so it won't be exaplained
"""

import tiktoken
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader


class GPTDatasetV1(Dataset):
    def __init__(self, txt, tokenizer, max_length, stride):
        self.input_ids = []
        self.target_ids = []

        # Tokenize the entire text
        token_ids = tokenizer.encode(txt, allowed_special={"<|endoftext|>"})

        # Use a sliding window to chunk the book into overlapping sequences of max_length
        for i in range(0, len(token_ids) - max_length, stride):
            input_chunk = token_ids[i:i + max_length]
            target_chunk = token_ids[i + 1: i + max_length + 1]
            self.input_ids.append(torch.tensor(input_chunk))
            self.target_ids.append(torch.tensor(target_chunk))

    def __len__(self):
        return len(self.input_ids)

    def __getitem__(self, idx):
        return self.input_ids[idx], self.target_ids[idx]


def create_dataloader_v1(txt, batch_size=4, max_length=256,
                         stride=128, shuffle=True, drop_last=True, num_workers=0):
    # Initialize the tokenizer
    tokenizer = tiktoken.get_encoding("gpt2")

    # Create dataset
    dataset = GPTDatasetV1(txt, tokenizer, max_length, stride)

    # Create dataloader
    dataloader = DataLoader(
        dataset, batch_size=batch_size, shuffle=shuffle, drop_last=drop_last, num_workers=num_workers)

    return dataloader


class MultiHeadAttention(nn.Module):
    def __init__(self, d_in, d_out, context_length, dropout, num_heads, qkv_bias=False):
        super().__init__()
        assert d_out % num_heads == 0, "d_out must be divisible by n_heads"

        self.d_out = d_out
        self.num_heads = num_heads
        self.head_dim = d_out // num_heads  # Reduce the projection dim to match desired output dim

        self.W_query = nn.Linear(d_in, d_out, bias=qkv_bias)
        self.W_key = nn.Linear(d_in, d_out, bias=qkv_bias)
        self.W_value = nn.Linear(d_in, d_out, bias=qkv_bias)
        self.out_proj = nn.Linear(d_out, d_out)  # Linear layer to combine head outputs
        self.dropout = nn.Dropout(dropout)
        self.register_buffer('mask', torch.triu(torch.ones(context_length, context_length), diagonal=1))

    def forward(self, x):
        b, num_tokens, d_in = x.shape

        keys = self.W_key(x)  # Shape: (b, num_tokens, d_out)
        queries = self.W_query(x)
        values = self.W_value(x)

        # We implicitly split the matrix by adding a `num_heads` dimension
        # Unroll last dim: (b, num_tokens, d_out) -> (b, num_tokens, num_heads, head_dim)
        keys = keys.view(b, num_tokens, self.num_heads, self.head_dim)
        values = values.view(b, num_tokens, self.num_heads, self.head_dim)
        queries = queries.view(b, num_tokens, self.num_heads, self.head_dim)

        # Transpose: (b, num_tokens, num_heads, head_dim) -> (b, num_heads, num_tokens, head_dim)
        keys = keys.transpose(1, 2)
        queries = queries.transpose(1, 2)
        values = values.transpose(1, 2)

        # Compute scaled dot-product attention (aka self-attention) with a causal mask
        attn_scores = queries @ keys.transpose(2, 3)  # Dot product for each head

        # Original mask truncated to the number of tokens and converted to boolean
        mask_bool = self.mask.bool()[:num_tokens, :num_tokens]

        # Use the mask to fill attention scores
        attn_scores.masked_fill_(mask_bool, -torch.inf)

        attn_weights = torch.softmax(attn_scores / keys.shape[-1]**0.5, dim=-1)
        attn_weights = self.dropout(attn_weights)

        # Shape: (b, num_tokens, num_heads, head_dim)
        context_vec = (attn_weights @ values).transpose(1, 2)

        # Combine heads, where self.d_out = self.num_heads * self.head_dim
        context_vec = context_vec.reshape(b, num_tokens, self.d_out)
        context_vec = self.out_proj(context_vec)  # optional projection

        return context_vec


class LayerNorm(nn.Module):
    def __init__(self, emb_dim):
        super().__init__()
        self.eps = 1e-5
        self.scale = nn.Parameter(torch.ones(emb_dim))
        self.shift = nn.Parameter(torch.zeros(emb_dim))

    def forward(self, x):
        mean = x.mean(dim=-1, keepdim=True)
        var = x.var(dim=-1, keepdim=True, unbiased=False)
        norm_x = (x - mean) / torch.sqrt(var + self.eps)
        return self.scale * norm_x + self.shift


class GELU(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x):
        return 0.5 * x * (1 + torch.tanh(
            torch.sqrt(torch.tensor(2.0 / torch.pi)) *
            (x + 0.044715 * torch.pow(x, 3))
        ))


class FeedForward(nn.Module):
    def __init__(self, cfg):
        super().__init__()
        self.layers = nn.Sequential(
            nn.Linear(cfg["emb_dim"], 4 * cfg["emb_dim"]),
            GELU(),
            nn.Linear(4 * cfg["emb_dim"], cfg["emb_dim"]),
        )

    def forward(self, x):
        return self.layers(x)


class TransformerBlock(nn.Module):
    def __init__(self, cfg):
        super().__init__()
        self.att = MultiHeadAttention(
            d_in=cfg["emb_dim"],
            d_out=cfg["emb_dim"],
            context_length=cfg["context_length"],
            num_heads=cfg["n_heads"],
            dropout=cfg["drop_rate"],
            qkv_bias=cfg["qkv_bias"])
        self.ff = FeedForward(cfg)
        self.norm1 = LayerNorm(cfg["emb_dim"])
        self.norm2 = LayerNorm(cfg["emb_dim"])
        self.drop_shortcut = nn.Dropout(cfg["drop_rate"])

    def forward(self, x):
        # Shortcut connection for attention block
        shortcut = x
        x = self.norm1(x)
        x = self.att(x)   # Shape [batch_size, num_tokens, emb_size]
        x = self.drop_shortcut(x)
        x = x + shortcut  # Add the original input back

        # Shortcut connection for feed-forward block
        shortcut = x
        x = self.norm2(x)
        x = self.ff(x)
        x = self.drop_shortcut(x)
        x = x + shortcut  # Add the original input back

        return x


class GPTModel(nn.Module):
    def __init__(self, cfg):
        super().__init__()
        self.tok_emb = nn.Embedding(cfg["vocab_size"], cfg["emb_dim"])
        self.pos_emb = nn.Embedding(cfg["context_length"], cfg["emb_dim"])
        self.drop_emb = nn.Dropout(cfg["drop_rate"])

        self.trf_blocks = nn.Sequential(
            *[TransformerBlock(cfg) for _ in range(cfg["n_layers"])])

        self.final_norm = LayerNorm(cfg["emb_dim"])
        self.out_head = nn.Linear(cfg["emb_dim"], cfg["vocab_size"], bias=False)

    def forward(self, in_idx):
        batch_size, seq_len = in_idx.shape
        tok_embeds = self.tok_emb(in_idx)
        pos_embeds = self.pos_emb(torch.arange(seq_len, device=in_idx.device))
        x = tok_embeds + pos_embeds  # Shape [batch_size, num_tokens, emb_size]
        x = self.drop_emb(x)
        x = self.trf_blocks(x)
        x = self.final_norm(x)
        logits = self.out_head(x)
        return logits
```

</details>

```python
# Download contents to train the data with
import os
import urllib.request

file_path = "the-verdict.txt"
url = "https://raw.githubusercontent.com/rasbt/LLMs-from-scratch/main/ch02/01_main-chapter-code/the-verdict.txt"

if not os.path.exists(file_path):
    with urllib.request.urlopen(url) as response:
        text_data = response.read().decode('utf-8')
    with open(file_path, "w", encoding="utf-8") as file:
        file.write(text_data)
else:
    with open(file_path, "r", encoding="utf-8") as file:
        text_data = file.read()

total_characters = len(text_data)
tokenizer = tiktoken.get_encoding("gpt2")
total_tokens = len(tokenizer.encode(text_data))

print("Data downloaded")
print("Characters:", total_characters)
print("Tokens:", total_tokens)

# Model initialization
GPT_CONFIG_124M = {
    "vocab_size": 50257,   # Vocabulary size
    "context_length": 256, # Shortened context length (orig: 1024)
    "emb_dim": 768,        # Embedding dimension
    "n_heads": 12,         # Number of attention heads
    "n_layers": 12,        # Number of layers
    "drop_rate": 0.1,      # Dropout rate
    "qkv_bias": False      # Query-key-value bias
}

torch.manual_seed(123)
model = GPTModel(GPT_CONFIG_124M)
model.eval()
print ("Model initialized")


# Functions to transform from tokens to ids and from to ids to tokens
def text_to_token_ids(text, tokenizer):
    encoded = tokenizer.encode(text, allowed_special={'<|endoftext|>'})
    encoded_tensor = torch.tensor(encoded).unsqueeze(0) # add batch dimension
    return encoded_tensor

def token_ids_to_text(token_ids, tokenizer):
    flat = token_ids.squeeze(0) # remove batch dimension
    return tokenizer.decode(flat.tolist())



# Define loss functions
def calc_loss_batch(input_batch, target_batch, model, device):
    input_batch, target_batch = input_batch.to(device), target_batch.to(device)
    logits = model(input_batch)
    loss = torch.nn.functional.cross_entropy(logits.flatten(0, 1), target_batch.flatten())
    return loss


def calc_loss_loader(data_loader, model, device, num_batches=None):
    total_loss = 0.
    if len(data_loader) == 0:
        return float("nan")
    elif num_batches is None:
        num_batches = len(data_loader)
    else:
        # Reduce the number of batches to match the total number of batches in the data loader
        # if num_batches exceeds the number of batches in the data loader
        num_batches = min(num_batches, len(data_loader))
    for i, (input_batch, target_batch) in enumerate(data_loader):
        if i < num_batches:
            loss = calc_loss_batch(input_batch, target_batch, model, device)
            total_loss += loss.item()
        else:
            break
    return total_loss / num_batches


# Apply Train/validation ratio and create dataloaders
train_ratio = 0.90
split_idx = int(train_ratio * len(text_data))
train_data = text_data[:split_idx]
val_data = text_data[split_idx:]

torch.manual_seed(123)

train_loader = create_dataloader_v1(
    train_data,
    batch_size=2,
    max_length=GPT_CONFIG_124M["context_length"],
    stride=GPT_CONFIG_124M["context_length"],
    drop_last=True,
    shuffle=True,
    num_workers=0
)

val_loader = create_dataloader_v1(
    val_data,
    batch_size=2,
    max_length=GPT_CONFIG_124M["context_length"],
    stride=GPT_CONFIG_124M["context_length"],
    drop_last=False,
    shuffle=False,
    num_workers=0
)


# Sanity checks
if total_tokens * (train_ratio) < GPT_CONFIG_124M["context_length"]:
    print("Not enough tokens for the training loader. "
          "Try to lower the `GPT_CONFIG_124M['context_length']` or "
          "increase the `training_ratio`")

if total_tokens * (1-train_ratio) < GPT_CONFIG_124M["context_length"]:
    print("Not enough tokens for the validation loader. "
          "Try to lower the `GPT_CONFIG_124M['context_length']` or "
          "decrease the `training_ratio`")

print("Train loader:")
for x, y in train_loader:
    print(x.shape, y.shape)

print("\nValidation loader:")
for x, y in val_loader:
    print(x.shape, y.shape)

train_tokens = 0
for input_batch, target_batch in train_loader:
    train_tokens += input_batch.numel()

val_tokens = 0
for input_batch, target_batch in val_loader:
    val_tokens += input_batch.numel()

print("Training tokens:", train_tokens)
print("Validation tokens:", val_tokens)
print("All tokens:", train_tokens + val_tokens)


# Indicate the device to use
if torch.cuda.is_available():
    device = torch.device("cuda")
elif torch.backends.mps.is_available():
    device = torch.device("mps")
else:
    device = torch.device("cpu")

print(f"Using {device} device.")

model.to(device) # no assignment model = model.to(device) necessary for nn.Module classes



# Pre-calculate losses without starting yet
torch.manual_seed(123) # For reproducibility due to the shuffling in the data loader

with torch.no_grad(): # Disable gradient tracking for efficiency because we are not training, yet
    train_loss = calc_loss_loader(train_loader, model, device)
    val_loss = calc_loss_loader(val_loader, model, device)

print("Training loss:", train_loss)
print("Validation loss:", val_loss)


# Functions to train the data
def train_model_simple(model, train_loader, val_loader, optimizer, device, num_epochs,
                       eval_freq, eval_iter, start_context, tokenizer):
    # Initialize lists to track losses and tokens seen
    train_losses, val_losses, track_tokens_seen = [], [], []
    tokens_seen, global_step = 0, -1

    # Main training loop
    for epoch in range(num_epochs):
        model.train()  # Set model to training mode
        
        for input_batch, target_batch in train_loader:
            optimizer.zero_grad() # Reset loss gradients from previous batch iteration
            loss = calc_loss_batch(input_batch, target_batch, model, device)
            loss.backward() # Calculate loss gradients
            optimizer.step() # Update model weights using loss gradients
            tokens_seen += input_batch.numel()
            global_step += 1

            # Optional evaluation step
            if global_step % eval_freq == 0:
                train_loss, val_loss = evaluate_model(
                    model, train_loader, val_loader, device, eval_iter)
                train_losses.append(train_loss)
                val_losses.append(val_loss)
                track_tokens_seen.append(tokens_seen)
                print(f"Ep {epoch+1} (Step {global_step:06d}): "
                      f"Train loss {train_loss:.3f}, Val loss {val_loss:.3f}")

        # Print a sample text after each epoch
        generate_and_print_sample(
            model, tokenizer, device, start_context
        )

    return train_losses, val_losses, track_tokens_seen


def evaluate_model(model, train_loader, val_loader, device, eval_iter):
    model.eval()
    with torch.no_grad():
        train_loss = calc_loss_loader(train_loader, model, device, num_batches=eval_iter)
        val_loss = calc_loss_loader(val_loader, model, device, num_batches=eval_iter)
    model.train()
    return train_loss, val_loss


def generate_and_print_sample(model, tokenizer, device, start_context):
    model.eval()
    context_size = model.pos_emb.weight.shape[0]
    encoded = text_to_token_ids(start_context, tokenizer).to(device)
    with torch.no_grad():
        token_ids = generate_text(
            model=model, idx=encoded,
            max_new_tokens=50, context_size=context_size
        )
    decoded_text = token_ids_to_text(token_ids, tokenizer)
    print(decoded_text.replace("\n", " "))  # Compact print format
    model.train()


# Start training!
import time
start_time = time.time()

torch.manual_seed(123)
model = GPTModel(GPT_CONFIG_124M)
model.to(device)
optimizer = torch.optim.AdamW(model.parameters(), lr=0.0004, weight_decay=0.1)

num_epochs = 10
train_losses, val_losses, tokens_seen = train_model_simple(
    model, train_loader, val_loader, optimizer, device,
    num_epochs=num_epochs, eval_freq=5, eval_iter=5,
    start_context="Every effort moves you", tokenizer=tokenizer
)

end_time = time.time()
execution_time_minutes = (end_time - start_time) / 60
print(f"Training completed in {execution_time_minutes:.2f} minutes.")



# Show graphics with the training process
import matplotlib.pyplot as plt
from matplotlib.ticker import MaxNLocator
import math
def plot_losses(epochs_seen, tokens_seen, train_losses, val_losses):
    fig, ax1 = plt.subplots(figsize=(5, 3))
    ax1.plot(epochs_seen, train_losses, label="Training loss")
    ax1.plot(
        epochs_seen, val_losses, linestyle="-.", label="Validation loss"
    )
    ax1.set_xlabel("Epochs")
    ax1.set_ylabel("Loss")
    ax1.legend(loc="upper right")
    ax1.xaxis.set_major_locator(MaxNLocator(integer=True))
    ax2 = ax1.twiny()
    ax2.plot(tokens_seen, train_losses, alpha=0)
    ax2.set_xlabel("Tokens seen")
    fig.tight_layout()
    plt.show()

    # Compute perplexity from the loss values
    train_ppls = [math.exp(loss) for loss in train_losses]
    val_ppls = [math.exp(loss) for loss in val_losses]
    # Plot perplexity over tokens seen
    plt.figure()
    plt.plot(tokens_seen, train_ppls, label='Training Perplexity')
    plt.plot(tokens_seen, val_ppls, label='Validation Perplexity')
    plt.xlabel('Tokens Seen')
    plt.ylabel('Perplexity')
    plt.title('Perplexity over Training')
    plt.legend()
    plt.show()

epochs_tensor = torch.linspace(0, num_epochs, len(train_losses))
plot_losses(epochs_tensor, tokens_seen, train_losses, val_losses)


torch.save({
    "model_state_dict": model.state_dict(),
    "optimizer_state_dict": optimizer.state_dict(),
    },
"/tmp/model_and_optimizer.pth"
)
```

Let's see an explanation step by step

### Functions to transform text <--> ids

These are some simple functions that can be used to transform from texts from the vocabulary to ids and backwards. This is needed at the begging of the handling of the text and at the end fo the predictions:

```python
# Functions to transform from tokens to ids and from to ids to tokens
def text_to_token_ids(text, tokenizer):
    encoded = tokenizer.encode(text, allowed_special={'<|endoftext|>'})
    encoded_tensor = torch.tensor(encoded).unsqueeze(0) # add batch dimension
    return encoded_tensor

def token_ids_to_text(token_ids, tokenizer):
    flat = token_ids.squeeze(0) # remove batch dimension
    return tokenizer.decode(flat.tolist())
```

### Generate text functions

In a previos section a function that just got the **most probable token** after getting the logits. However, this will mean that for each entry the same output is always going to be generated which makes it very deterministic.

The following `generate_text` function, will apply the `top-k` , `temperature` and `multinomial` concepts.

* The **`top-k`** means that we will start reducing to `-inf` all the probabilities of all the tokens expect of the top k tokens. So, if k=3, before making a decision only the 3 most probably tokens will have a probability different from `-inf`.
* The **`temperature`** means that every probability will be divided by the temperature value. A value of `0.1` will improve the highest probability compared with the lowest one, while a temperature of `5` for example will make it more flat. This helps to improve to variation in responses we would like the LLM to have.
* After applying the temperature, a **`softmax`** function is applied again to make all the reminding tokens have a total probability of 1.
* Finally, instead of choosing the token with the biggest probability, the function **`multinomial`** is applied to **predict the next token according to the final probabilities**. So if token 1 had a 70% of probabilities, token 2 a 20% and token 3 a 10%, 70% of the times token 1 will be selected, 20% of the times it will be token 2 and 10% of the times will be 10%.

```python
# Generate text function
def generate_text(model, idx, max_new_tokens, context_size, temperature=0.0, top_k=None, eos_id=None):

    # For-loop is the same as before: Get logits, and only focus on last time step
    for _ in range(max_new_tokens):
        idx_cond = idx[:, -context_size:]
        with torch.no_grad():
            logits = model(idx_cond)
        logits = logits[:, -1, :]

        # New: Filter logits with top_k sampling
        if top_k is not None:
            # Keep only top_k values
            top_logits, _ = torch.topk(logits, top_k)
            min_val = top_logits[:, -1]
            logits = torch.where(logits < min_val, torch.tensor(float("-inf")).to(logits.device), logits)

        # New: Apply temperature scaling
        if temperature > 0.0:
            logits = logits / temperature

            # Apply softmax to get probabilities
            probs = torch.softmax(logits, dim=-1)  # (batch_size, context_len)

            # Sample from the distribution
            idx_next = torch.multinomial(probs, num_samples=1)  # (batch_size, 1)

        # Otherwise same as before: get idx of the vocab entry with the highest logits value
        else:
            idx_next = torch.argmax(logits, dim=-1, keepdim=True)  # (batch_size, 1)

        if idx_next == eos_id:  # Stop generating early if end-of-sequence token is encountered and eos_id is specified
            break

        # Same as before: append sampled index to the running sequence
        idx = torch.cat((idx, idx_next), dim=1)  # (batch_size, num_tokens+1)

    return idx
```

{% hint style="info" %}
There is a common alternative to `top-k` called [**`top-p`**](https://en.wikipedia.org/wiki/Top-p\_sampling), also known as nucleus sampling, which instead of getting k samples with the most probability, it **organizes** all the resulting **vocabulary** by probabilities and **sums** them from the highest probability to the lowest until a **threshold is reached**.

Then, **only those words** of the vocabulary will be considered according to their relative probabilities&#x20;

This allows to not need to select a number of `k` samples, as the optimal k might be different on each case, but **only a threshold**.

_Note that this improvement isn't included in the previous code._
{% endhint %}

{% hint style="info" %}
Another way to improve the generated text is by using **Beam search** instead of the greedy search sued in this example.\
Unlike greedy search, which selects the most probable next word at each step and builds a single sequence, **beam search keeps track of the top 𝑘 k highest-scoring partial sequences** (called "beams") at each step. By exploring multiple possibilities simultaneously, it balances efficiency and quality, increasing the chances of **finding a better overall** sequence that might be missed by the greedy approach due to early, suboptimal choices.

_Note that this improvement isn't included in the previous code._
{% endhint %}

### Loss functions

The **`calc_loss_batch`** function calculates the cross entropy of the a prediction of a single batch.\
The **`calc_loss_loader`** gets the cross entropy of all the batches and calculates the **average cross entropy**.

```python
# Define loss functions
def calc_loss_batch(input_batch, target_batch, model, device):
    input_batch, target_batch = input_batch.to(device), target_batch.to(device)
    logits = model(input_batch)
    loss = torch.nn.functional.cross_entropy(logits.flatten(0, 1), target_batch.flatten())
    return loss

def calc_loss_loader(data_loader, model, device, num_batches=None):
    total_loss = 0.
    if len(data_loader) == 0:
        return float("nan")
    elif num_batches is None:
        num_batches = len(data_loader)
    else:
        # Reduce the number of batches to match the total number of batches in the data loader
        # if num_batches exceeds the number of batches in the data loader
        num_batches = min(num_batches, len(data_loader))
    for i, (input_batch, target_batch) in enumerate(data_loader):
        if i < num_batches:
            loss = calc_loss_batch(input_batch, target_batch, model, device)
            total_loss += loss.item()
        else:
            break
    return total_loss / num_batches
```

{% hint style="info" %}
**Gradient clipping** is a technique used to enhance **training stability** in large neural networks by setting a **maximum threshold** for gradient magnitudes. When gradients exceed this predefined `max_norm`, they are scaled down proportionally to ensure that updates to the model’s parameters remain within a manageable range, preventing issues like exploding gradients and ensuring more controlled and stable training.

_Note that this improvement isn't included in the previous code._

Check the following example:
{% endhint %}

<figure><img src="../../.gitbook/assets/image (6) (1).png" alt=""><figcaption></figcaption></figure>

### Loading Data

The functions `create_dataloader_v1` and `create_dataloader_v1` were already discussed in a previous section.

From here note how it's defined that 90% of the text is going to be used for training while the 10% will be used for validation and both sets are stored in 2 different data loaders.\
Note that some times part of the data set is also left for a testing set to evaluate better the performance of the model.

Both data loaders are using the same batch size, maximum length and stride and num workers (0 in this case).\
The main differences are the data used by each, and the the validators is not dropping the last neither shuffling the data is it's not needed for validation purposes.

Also the fact that **stride is as big as the context length**, means that there won't be overlapping between contexts used to train the data (reduces overfitting but also the training data set).

Moreover, note that the batch size in this case it 2 to divide the data in 2 batches, the main goal of this is to allow parallel processing and reduce the consumption per batch.

```python
train_ratio = 0.90
split_idx = int(train_ratio * len(text_data))
train_data = text_data[:split_idx]
val_data = text_data[split_idx:]

torch.manual_seed(123)

train_loader = create_dataloader_v1(
    train_data,
    batch_size=2,
    max_length=GPT_CONFIG_124M["context_length"],
    stride=GPT_CONFIG_124M["context_length"],
    drop_last=True,
    shuffle=True,
    num_workers=0
)

val_loader = create_dataloader_v1(
    val_data,
    batch_size=2,
    max_length=GPT_CONFIG_124M["context_length"],
    stride=GPT_CONFIG_124M["context_length"],
    drop_last=False,
    shuffle=False,
    num_workers=0
)
```

## Sanity Checks

The goal is to check there are enough tokens for training, shapes are the expected ones and get some info about the number of tokens used for training and for validation:

```python
# Sanity checks
if total_tokens * (train_ratio) < GPT_CONFIG_124M["context_length"]:
    print("Not enough tokens for the training loader. "
          "Try to lower the `GPT_CONFIG_124M['context_length']` or "
          "increase the `training_ratio`")

if total_tokens * (1-train_ratio) < GPT_CONFIG_124M["context_length"]:
    print("Not enough tokens for the validation loader. "
          "Try to lower the `GPT_CONFIG_124M['context_length']` or "
          "decrease the `training_ratio`")

print("Train loader:")
for x, y in train_loader:
    print(x.shape, y.shape)

print("\nValidation loader:")
for x, y in val_loader:
    print(x.shape, y.shape)

train_tokens = 0
for input_batch, target_batch in train_loader:
    train_tokens += input_batch.numel()

val_tokens = 0
for input_batch, target_batch in val_loader:
    val_tokens += input_batch.numel()

print("Training tokens:", train_tokens)
print("Validation tokens:", val_tokens)
print("All tokens:", train_tokens + val_tokens)
```

### Select device for training & pre calculations

The following code just select the device to use and calculates a training loss and validation loss (without having trained anything yet) as a starting point.

```python
# Indicate the device to use

if torch.cuda.is_available():
    device = torch.device("cuda")
elif torch.backends.mps.is_available():
    device = torch.device("mps")
else:
    device = torch.device("cpu")

print(f"Using {device} device.")

model.to(device) # no assignment model = model.to(device) necessary for nn.Module classes

# Pre-calculate losses without starting yet
torch.manual_seed(123) # For reproducibility due to the shuffling in the data loader

with torch.no_grad(): # Disable gradient tracking for efficiency because we are not training, yet
    train_loss = calc_loss_loader(train_loader, model, device)
    val_loss = calc_loss_loader(val_loader, model, device)

print("Training loss:", train_loss)
print("Validation loss:", val_loss)
```

### Training functions

The function `generate_and_print_sample` will just get a context and generate some tokens in order to get a feeling about how good is the model at that point. This is called by `train_model_simple` on each step.

The function `evaluate_model` is called as frequently as indicate to the training function and it's used to measure the train loss and the validation loss at that point in the model training.

Then the big function `train_model_simple` is the one that actually train the model. It expects:

* The train data loader (with the data already separated and prepared for training)
* The validator loader
* The **optimizer** to use during training: This is the function that will use the gradients and will update the parameters to reduce the loss. In this case, as you will see, `AdamW` is used, but there are many more.
  * `optimizer.zero_grad()` is called to reset the gradients on each round to not accumulate them.
  * The **`lr`** param is the **learning rate** which determines the **size of the steps** taken during the optimization process when updating the model's parameters. A **smaller** learning rate means the optimizer **makes smaller updates** to the weights, which can lead to more **precise** convergence but might **slow down** training. A **larger** learning rate can speed up training but **risks overshooting** the minimum of the loss function (**jump over** the point where the loss function is minimized).
  * **Weight Decay** modifies the **Loss Calculation** step by adding an extra term that penalizes large weights. This encourages the optimizer to find solutions with smaller weights, balancing between fitting the data well and keeping the model simple preventing overfitting in machine learning models by discouraging the model from assigning too much importance to any single feature.
    * Traditional optimizers like SGD with L2 regularization couple weight decay with the gradient of the loss function. However, **AdamW** (a variant of Adam optimizer) decouples weight decay from the gradient update, leading to more effective regularization.
* The device to use for training
* The number of epochs: Number of times to go over the training data
* The evaluation frequency: The frequency to call `evaluate_model`
* The evaluation iteration: The number of batches to use when evaluating the current state of the model when calling `generate_and_print_sample`
* The start context: Which the starting sentence to use when calling `generate_and_print_sample`
* The tokenizer

```python
# Functions to train the data
def train_model_simple(model, train_loader, val_loader, optimizer, device, num_epochs,
                       eval_freq, eval_iter, start_context, tokenizer):
    # Initialize lists to track losses and tokens seen
    train_losses, val_losses, track_tokens_seen = [], [], []
    tokens_seen, global_step = 0, -1

    # Main training loop
    for epoch in range(num_epochs):
        model.train()  # Set model to training mode
        
        for input_batch, target_batch in train_loader:
            optimizer.zero_grad() # Reset loss gradients from previous batch iteration
            loss = calc_loss_batch(input_batch, target_batch, model, device)
            loss.backward() # Calculate loss gradients
            optimizer.step() # Update model weights using loss gradients
            tokens_seen += input_batch.numel()
            global_step += 1

            # Optional evaluation step
            if global_step % eval_freq == 0:
                train_loss, val_loss = evaluate_model(
                    model, train_loader, val_loader, device, eval_iter)
                train_losses.append(train_loss)
                val_losses.append(val_loss)
                track_tokens_seen.append(tokens_seen)
                print(f"Ep {epoch+1} (Step {global_step:06d}): "
                      f"Train loss {train_loss:.3f}, Val loss {val_loss:.3f}")

        # Print a sample text after each epoch
        generate_and_print_sample(
            model, tokenizer, device, start_context
        )

    return train_losses, val_losses, track_tokens_seen


def evaluate_model(model, train_loader, val_loader, device, eval_iter):
    model.eval() # Set in eval mode to avoid dropout
    with torch.no_grad():
        train_loss = calc_loss_loader(train_loader, model, device, num_batches=eval_iter)
        val_loss = calc_loss_loader(val_loader, model, device, num_batches=eval_iter)
    model.train() # Back to training model applying all the configurations
    return train_loss, val_loss


def generate_and_print_sample(model, tokenizer, device, start_context):
    model.eval() # Set in eval mode to avoid dropout
    context_size = model.pos_emb.weight.shape[0]
    encoded = text_to_token_ids(start_context, tokenizer).to(device)
    with torch.no_grad():
        token_ids = generate_text(
            model=model, idx=encoded,
            max_new_tokens=50, context_size=context_size
        )
    decoded_text = token_ids_to_text(token_ids, tokenizer)
    print(decoded_text.replace("\n", " "))  # Compact print format
    model.train() # Back to training model applying all the configurations
```

{% hint style="info" %}
To improve the learning rate there are a couple relevant techniques called **linear warmup** and **cosine decay.**

**Linear warmup** consist on define an initial learning rate and a maximum one and consistently update it after each epoch. This is because starting the training with smaller weight updates decreases the risk of the model encountering large, destabilizing updates during its training phase.\
**Cosine decay** is a technique that **gradually reduces the learning rate** following a half-cosine curve **after the warmup** phase, slowing weight updates to **minimize the risk of overshooting** the loss minima and ensure training stability in later phases.

_Note that these improvements aren't included in the previous code._
{% endhint %}

### Start training

```python
import time
start_time = time.time()

torch.manual_seed(123)
model = GPTModel(GPT_CONFIG_124M)
model.to(device)
optimizer = torch.optim.AdamW(model.parameters(), lr=0.0004, weight_decay=0.1)

num_epochs = 10
train_losses, val_losses, tokens_seen = train_model_simple(
    model, train_loader, val_loader, optimizer, device,
    num_epochs=num_epochs, eval_freq=5, eval_iter=5,
    start_context="Every effort moves you", tokenizer=tokenizer
)

end_time = time.time()
execution_time_minutes = (end_time - start_time) / 60
print(f"Training completed in {execution_time_minutes:.2f} minutes.")
```

### Print training evolution

With the following function it's possible to print the evolution of the model while it was being trained.

```python
import matplotlib.pyplot as plt
from matplotlib.ticker import MaxNLocator
import math
def plot_losses(epochs_seen, tokens_seen, train_losses, val_losses):
    fig, ax1 = plt.subplots(figsize=(5, 3))
    ax1.plot(epochs_seen, train_losses, label="Training loss")
    ax1.plot(
        epochs_seen, val_losses, linestyle="-.", label="Validation loss"
    )
    ax1.set_xlabel("Epochs")
    ax1.set_ylabel("Loss")
    ax1.legend(loc="upper right")
    ax1.xaxis.set_major_locator(MaxNLocator(integer=True))
    ax2 = ax1.twiny()
    ax2.plot(tokens_seen, train_losses, alpha=0)
    ax2.set_xlabel("Tokens seen")
    fig.tight_layout()
    plt.show()

    # Compute perplexity from the loss values
    train_ppls = [math.exp(loss) for loss in train_losses]
    val_ppls = [math.exp(loss) for loss in val_losses]
    # Plot perplexity over tokens seen
    plt.figure()
    plt.plot(tokens_seen, train_ppls, label='Training Perplexity')
    plt.plot(tokens_seen, val_ppls, label='Validation Perplexity')
    plt.xlabel('Tokens Seen')
    plt.ylabel('Perplexity')
    plt.title('Perplexity over Training')
    plt.legend()
    plt.show()

epochs_tensor = torch.linspace(0, num_epochs, len(train_losses))
plot_losses(epochs_tensor, tokens_seen, train_losses, val_losses)
```

### Save the model

It's possible to save the model + optimizer if you want to continue training later:

```python
# Save the model and the optimizer for later training
torch.save({
    "model_state_dict": model.state_dict(),
    "optimizer_state_dict": optimizer.state_dict(),
    },
"/tmp/model_and_optimizer.pth"
)
# Note that this model with the optimizer occupied close to 2GB

# Restore model and optimizer for training
checkpoint = torch.load("/tmp/model_and_optimizer.pth", map_location=device)

model = GPTModel(GPT_CONFIG_124M)
model.load_state_dict(checkpoint["model_state_dict"])
optimizer = torch.optim.AdamW(model.parameters(), lr=5e-4, weight_decay=0.1)
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
model.train(); # Put in training mode
```

Or just the model if you are planing just on using it:

```python
# Save the model
torch.save(model.state_dict(), "model.pth")

# Load it
model = GPTModel(GPT_CONFIG_124M)

model.load_state_dict(torch.load("model.pth", map_location=device))

model.eval() # Put in eval mode
```

## Loading GPT2 weights

There 2 quick scripts to load the GPT2 weights locally. For both you can clone the repository [https://github.com/rasbt/LLMs-from-scratch](https://github.com/rasbt/LLMs-from-scratch) locally, then:

* The script [https://github.com/rasbt/LLMs-from-scratch/blob/main/ch05/01\_main-chapter-code/gpt\_generate.py](https://github.com/rasbt/LLMs-from-scratch/blob/main/ch05/01\_main-chapter-code/gpt\_generate.py) will download all the weights and transform the formats from OpenAI to the ones expected by our LLM. The script is also prepared with the needed configuration and with the prompt: "Every effort moves you"
* The script [https://github.com/rasbt/LLMs-from-scratch/blob/main/ch05/02\_alternative\_weight\_loading/weight-loading-hf-transformers.ipynb](https://github.com/rasbt/LLMs-from-scratch/blob/main/ch05/02\_alternative\_weight\_loading/weight-loading-hf-transformers.ipynb) allows you to load any of the GPT2 weights locally (just change the `CHOOSE_MODEL` var) and predict text from some prompts.

## References

* [https://www.manning.com/books/build-a-large-language-model-from-scratch](https://www.manning.com/books/build-a-large-language-model-from-scratch)