41 KiB
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:
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 some times slightly modify
Previous code used here but already explained in previous sections
"""
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
# 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:
# 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-kmeans that we will start reducing to-infall 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
temperaturemeans that every probability will be divided by the temperature value. A value of0.1will improve the highest probability compared with the lowest one, while a temperature of5for 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
softmaxfunction 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
multinomialis 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%.
# 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, 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
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.
# 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 %}
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.
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:
# 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.
# 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,
AdamWis used, but there are many more.optimizer.zero_grad()is called to reset the gradients on each round to not accumulate them.- The
lrparam 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
# 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
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.
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:
# 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:
# 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 locally, then:
- The script 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 allows you to load any of the GPT2 weights locally (just change the
CHOOSE_MODELvar) and predict text from some prompts.