In this post, we build a **Decoder-Only Transformer** from scratch, using **only numpy**.
I wanted to put this together to see if I can find an easier way to build this very popular architecture, while at the same time, seeing if it helps someone else.
As you go through, if you find I missed anything or have some suggestions for improvement, please do not hesitate to drop me a line.
As we go through, we build a decoder-only transformer that can generate baby names.
The original paper for transformer **Attention is all you need**: https://arxiv.org/pdf/1706.03762
For this problem, we will use character level tokenization.
Text for training: https:/raw.githubusercontent.com/karpathy/makemore/refs/heads/master/names.txt
Start by importing our libraries.
# We will keep it simple as stated above using numpy # We will also use matplotlib for visualization import numpy as np import matplotlib.pyplot as plt
Preparing our data for the model
We setup a configuration class that holds our hyperparameters
# Let us config a data class class Config: d_model = 16 # The embedding dimensions n_heads = 4 # When we get to multi-head attention, we will need this d_head = 4 # We could calculate this manually by doing d_model // n_heads n_layers = 2 # We are going to stack two layers, that is two decoder blocks. batch_size = 1 # Batch size of 1. For simplicity and easier visualization text = 'Welcome to the world of AI' # The test our untrained model should generate # instantiate the config cfg = Config() cfg ----------- <__main__.Config at 0x77ecd644c050>
Let's build a function to create our vocab. This is overkill but hey, we should learn to write dry code as much as possible 😀
def build_vocab(text): ''' text: The full text return: chars: The chars in vocabulary stoi: maps/encodes characters to numbers itos: unmaps/decode numbers back to characters ''' chars = sorted(list(set(text))) # get a list of unique characters in the input text # Convert the text to numbers stoi = { ch:i for i,ch in enumerate(chars, start=1)} # Go back from numbers to text itos = { i:ch for ch,i in stoi.items()} return chars, stoi, itos # Test the function chars, stoi, itos = build_vocab(cfg.text) print(f'[*] Here are the characters: {chars}') print(f'[*] Here are the characters: {"".join(chars)}') print(f'[*] Here is the stoi mapping/encoding: {stoi}') print(f'[*] Here is the itos un-mapping/decoding: {itos}') # Setup the vocab size vocab_size = len(chars) print(f'Vocab size / unique tokens: {vocab_size}') -----------
[*] Here are the characters: [' ', 'A', 'I', 'W', 'c', 'd', 'e', 'f', 'h', 'l', 'm', 'o', 'r', 't', 'w']
[*] Here are the characters: AIWcdefhlmortw
[*] Here is the stoi mapping/encoding: {' ': 1, 'A': 2, 'I': 3, 'W': 4, 'c': 5, 'd': 6, 'e': 7, 'f': 8, 'h': 9, 'l': 10, 'm': 11, 'o': 12, 'r': 13, 't': 14, 'w': 15}
[*] Here is the itos un-mapping/decoding: {1: ' ', 2: 'A', 3: 'I', 4: 'W', 5: 'c', 6: 'd', 7: 'e', 8: 'f', 9: 'h', 10: 'l', 11: 'm', 12: 'o', 13: 'r', 14: 't', 15: 'w'}
Vocab size / unique tokens: 15Let us take a different view of this mapping by using pandas.
# Import pandas as pd import pandas as pd df = pd.DataFrame(stoi.items(), columns=['char', 'num']) df.style.hide(axis='index')
We do the same thing for the number to strings
df = pd.DataFrame(itos.items(), columns=['num', 'char']) df.style.hide(axis='index')
With above in place, we now have a clear understanding, of one way to map text to numbers and back from numbers to text.
Let's build on this to setup an encoder function. This function is what will be called on future text, using the vocabulary we defined above. Remember, our vocab is the unique characters we have within the string "Welcome to the world of AI".
encode = lambda text, stoi: [ stoi.get(ch) for ch in text ] # Test the encoder encode(text='Welcome', stoi=stoi) --------------- [4, 7, 10, 5, 12, 11, 7]
As we said earlier, if we encode from text to numbers, we have to be able to revert that process. While the computer needs numbers to train on, we cannot provide back those numbers to humans. We need to give humans something that is understandable. Hence the need for the decoder to revert the mapping.
# This maps us back from numbers to chars decode = lambda indices, itos: ''.join([ itos.get(i) for i in indices ]) # Test the encoder decode(encode(text='Welcome', stoi=stoi), itos=itos)
------------'Welcome'Now that we know the encoder and decoder works, let us get all our tokens from the text "Welcome to the world of AI" . At the same time, we make a 1-dimension NumPy. We also add a new (batch) dimension also, moving the input form a list to a 2-dimension NumPy array.
tokens = np.array(encode(text=cfg.text, stoi=stoi), dtype=np.int32)[None, :] print(f'Here are the tokens: \n{tokens} | tokens dtype: {tokens.dtype} | shape: {tokens.shape} | Dims: {tokens.ndim}') # Extract the batch and time dimensions and put them into separate variables B, T = tokens.shape # (batch, timestep) ------------- Here are the tokens: [[ 4 7 10 5 12 11 7 1 14 12 1 14 9 7 1 15 12 13 10 6 1 12 8 1 2 3]] | tokens dtype: int32 | shape: (1, 26) | Dims: 2
We are making progress, let us setup our X from the tokens. We are using a batch size of 1 for simplicity.
We use batch size of one as it is easy for us to visualize as we go along.
I like visuals and you should too ;-)
# This also means we will feed the entire sequence into the model X = tokens[:, :-1] # (We are predicting the next token) Y = tokens[:, 1:] # the 1 is the next token # Peek into the data print(f'Here is the X: {X}') print(f'Here is the Y: {Y}') ------------- Here is the X: [[ 4 7 10 5 12 11 7 1 14 12 1 14 9 7 1 15 12 13 10 6 1 12 8 1 2]] Here is the Y: [[ 7 10 5 12 11 7 1 14 12 1 14 9 7 1 15 12 13 10 6 1 12 8 1 2 3]]
What do we take away from above? When the model sees 4, we would like it to predict 7. When it sees the sequence of 4, 7, we would like it to predict 10. When it sees, 4, 7, 10, we would like it to predict 5. That pattern continues ...
Let's prepare to visualize our tokens. Setup a function for this even though we don't need to.
# Let us visualize above def plot_token_indices(tokens, title='Token Indices over time'): ''' tokens: np.array of shape (B, T) ''' assert tokens.shape[0] == 1, f'We are working with 1 full row' t = np.arange(tokens.shape[1]) plt.figure(figsize=(15,4)) plt.title(title) plt.bar(x=t, height=tokens[0]) plt.xticks(ticks=range(0, len(cfg.text),1), labels=cfg.text) plt.yticks(ticks=range(0,15,1)) plt.ylabel('Token Index') plt.xlabel('Sequence') plt.grid(axis='y') plt.show() # Test the function plot_token_indices(tokens=tokens)
Above shows our sequence and the index positions for each token. For example, we see that w has a value of 4, e has a value of 6, space has a value of 0, etc.
With this in place, let's work on our core numerical primitives.
Stable Softmax / Cross-entropy from logits / LayerNorm / Dropout / GELU
First up
Softmax
Softmax is a core activation function used in machine learning tasks. It is used to convert the outputs - usually the raw logits - into a probability distribution. We setup our Softmax via a function. We also consider numerical stability as we build this out.
# Setup a numerically stable implementation of softmax def softmax_stable(logits, axis=-1): ''' Numerically stale softmax implementation logits: np.array(..., D) D Is vocab size ''' # First up find the max value in the logits max_logits = np.max(logits, axis=axis, keepdims=True) # Shift the logits by the max shifted = logits - max_logits exp_shifted = np.exp(shifted) probs = exp_shifted / np.sum(exp_shifted, axis=axis, keepdims=True) return probs # Suppress scientific notation np.set_printoptions(suppress=True) # Test the function ----------------- array([0.00078972, 0.11720525, 0.01586201, 0.86603615, 0.00010688])
Cool we seem to have a stable Softmax. Lets plot Softmax and also see the impact temperature can have on the probabilities. We learned a lot about temperature, top_p and top_k in the first post in this series: Welcome to the world of AI - Understanding temperature, top_p and top_k
# Create a 100 evenly spaced points between -5 and +5 x = np.linspace(-5, 5, 100) for temp in [0.5, 1, 2.9, 0.1, 3]: probs = softmax_stable(x/temp) plt.plot(x, probs, label=f'Temp-{temp}') plt.legend() plt.title('Softmax sensitivity to temperature');
What we see above, is that a lower temperature results in sharper probabilities. Larger temperature, results in flatter probability distributions. As mentioned, we learned alot about temperature, top_p and top_k in the first post in this series: **Welcome to the world of AI - Understanding temperature, top_p and top_k**
We need to be very careful here as even though we went through the process to make this numerically stable, we still have a situation where if these values are too "large" this Softmax output can - or should I say will - converge to a one-hot vector.
As you see below, once the values are "large" Softmax converges to a one-hot vector. Here is an example of that situation:
softmax_stable(np.array([-20., 30, 100, 50, -4])) ---------------- array([0., 0., 1., 0., 0.])
Well Softmax converging to a one-hot vector is not the only problem we have here. The other problem is if we take the naive Softmax. We can already see large values causes overflow. Hence we see the *inf* below
a = np.array([-20., 30, 1000, 50, -4]) np.exp(a) ----------------- /tmp/ipykernel_157535/1527753011.py:2: RuntimeWarning: overflow encountered in exp np.exp(a) array([2.06115362e-09, 1.06864746e+13, inf, 5.18470553e+21, 1.83156389e-02])
When we try to compute the Softmax using the naive method. We see that we have additional overflows and nan values.Hence the reason why we need to ensure we are using the stable method.
# Overflow and nans np.exp(a) / np.sum(np.exp(a), axis=-1, keepdims=True) --------------- /tmp/ipykernel_157535/844943855.py:2: RuntimeWarning: overflow encountered in exp np.exp(a) / np.sum(np.exp(a), axis=-1, keepdims=True) /tmp/ipykernel_157535/844943855.py:2: RuntimeWarning: invalid value encountered in divide np.exp(a) / np.sum(np.exp(a), axis=-1, keepdims=True) array([ 0., 0., nan, 0., 0.])
Let us now jump to the
Cross-entropy loss
If you are doing anything with classification in neural networks, you are more than likely using cross-entropy loss. If you are doing binary classification, you are more than likely using Binary Cross-entropy. For a multi-class problem, you may be using Categorical Cross-entropy or maybe Sparse Categorical Cross-entropy. All different flavours of Cross-entropy loss.
Let us build a Cross entropy function.
# Cross entropy loss def cross_entropy_loss(logits, targets): ''' logits: (B, T, vocab_size) targets: (B, T) Returns scalar loss. Single value ''' B, T, V = logits.shape probs = softmax_stable(logits=logits, axis=-1) # Now let us get the log probability at those index positions log_probs = np.log(probs[np.arange(B)[:, None], np.arange(T)[None, :], targets ]) loss = -np.mean(log_probs) return loss # The function targets = np.array([0,1,1,0,1]) logits = np.array([-2., 3, 1, 5, -4]) cross_entropy_loss(logits=logits.reshape(1, 1, -1), targets=targets) ------------- np.float64(4.143828630781675)
The result of Cross-entropy is a single (scalar) value that tells us how well the model is learning. The closer this loss is to 0, the higher the model accuracy. So the objective is to minimize the loss.
LayerNorm
https://arxiv.org/pdf/1607.06450
While this might seem as only being used for normalization, LayerNorm is also used to condition the residual update scale.
Normalization also helps with speeding up the training process. Layer normalization is done on a per record - single training example - case. This normalization method uses the same technique at training time and test time.
# With the loss calculated, let us setup LayerNorm class LayerNorm: def __init__(self, d_model, eps=1e-5): self.d_model = d_model self.eps = eps # The scale and bias will be learned self.gamma = np.ones((d_model,), dtype=np.float32) self.beta = np.zeros((d_model,), dtype=np.float32) def __call__(self, x): ''' x: (B, T, d_model) ''' mean = np.mean(x, axis=-1, keepdims=True) var = np.var(x, axis=-1, keepdims=True) # Perform standardization x_hat = (x - mean) / np.sqrt(var + self.eps) # Do the scaling and shifting out = self.gamma * x_hat + self.beta return out
Let us now visualize this.
# Set the seed for repeatability np.random.seed(10) B, T, D = 1, vocab_size, cfg.d_model x = np.random.randn(B, T, D).astype(np.float32) * 3.0 + 5.0 # Just shift and scale a bit ln = LayerNorm(d_model=D) y = ln(x) # Flatten x x_flat = x.reshape(-1, D) y_flat = y.reshape(-1, D) plt.figure(figsize=(10,5)) plt.subplot(1,2,1) plt.title(f'Pre-LayerNormalization: \nmean:{x_flat.flatten().mean():.4f} \nstd:{x_flat.flatten().std():.4f}') plt.hist(x=x_flat.flatten(), bins=50) plt.vlines(x=x_flat.flatten().mean(), ymin=0, ymax=20, label='mean', color='r') plt.vlines(x=x_flat.flatten().mean() + x_flat.flatten().std() * 1, ymin=0, ymax=20, label='+1 std', color='k') plt.vlines(x=x_flat.flatten().mean() + x_flat.flatten().std() * -1, ymin=0, ymax=20, label='-1 std', color='k') plt.legend() plt.subplot(1,2,2) plt.title(f'Post-LayerNormalization: \nmean:{y_flat.flatten().mean():.4f} \nstd:{y_flat.flatten().std():.4f} ') plt.hist(x=y_flat.flatten(), bins=50) plt.tight_layout() plt.vlines(x=y_flat.flatten().mean(), ymin=0, ymax=20, label='mean', color='r') plt.vlines(x=y_flat.flatten().mean() + (1 * y_flat.flatten().std()), ymin=0, ymax=20, label='+ 1 std', color='k') plt.vlines(x=y_flat.flatten().mean() - (1 * y_flat.flatten().std()), ymin=0, ymax=20, label='-1 std', color='k') plt.legend() plt.show()
Without LayerNorm, we have a mean of 5.1 on the left and a standard deviation of 2.9. On the right we have a mean of 0 and a standard deviation of 1. This is what we typically want when training our models.
With LayerNorm in place and its visualization, let's see what Dropout is about
Dropout
Dropout is a regularization strategy that is used to address overfitting. Dropout - disable - neurons during training of the neural network.
Overfitting is a term you will hear alot about in machine learning. It is where the model has learned not only the patterns in the data but potentially the noise also. Thus while the model may train and have an accuracy of 100% and a loss of 0, during inference time, the model is quite inconsistent. That is to say the model will have high variance and low bias.
Dropout is one mechanism used to address overfitting. It is what is called a regularization strategy.
# Setup our dropout class class Dropout: def __init__(self, p=0.1): self.p = p self.training = True def __call__(self, x): if not self.training or self.p == 0: return x mask = ( np.random.rand(*x.shape) > self.p).astype(x.dtype) # Implement invert dropout: scale by 1/(1-p) at train time only return mask * x / (1.0 - self.p)
Test the Dropout
B, T, D = (1, 5, 4) x = np.ones((B, T, D), dtype=np.float32) print(x) # Setup dropout do = Dropout(p=0.5) # Set training to True do.training = True print(f'0.5 dropout:\n{do(x)}') # Disable dropout do.training = False do(x) ------------------ [[[1. 1. 1. 1.] [1. 1. 1. 1.] [1. 1. 1. 1.] [1. 1. 1. 1.] [1. 1. 1. 1.]]] 0.5 dropout: [[[2. 0. 2. 0.] [2. 2. 0. 0.] [2. 0. 0. 0.] [0. 2. 0. 2.] [2. 2. 2. 2.]]]
Now that we have an understanding of dropout, let's go ahead and wrap this up with th Gaussian Error Linear Unit (GELU) activation function
GELU - Gaussian Error Linear Unit
GELU is considered to be a high performance activation function. Activation functions are what introduces the non-linearity in neural networks. It weights inputs by their values. GELU also includes property from dropout and ReLU.
# Define GELU def gelu(x): ''' This is the approximate version using Tanh x: np.array ''' return 0.5 * x * ( 1.0 + np.tanh( np.sqrt(2.0 / np.pi) * (x + 0.044715 * (x**3) ) ) ) # Test the functio x = np.linspace(-4, 4, 400) # Implement ReLU so we can compare y_relu = np.maximum(0, x) y_gelu = gelu(x)
Let's now visualize the effect GELU has on our data.
# plot GELU plt.figure(figsize=(8, 4)) plt.subplot(121) plt.plot(x, y_relu, label='ReLU') plt.legend() plt.subplot(122) plt.plot(x, y_gelu, label='GELU') plt.legend() plt.show()
We can see above that while ReLU puts everything below 0 to exactly 0, this is not the case with GELU. With GELU, small negative values are possible while large negative values are clipped at 0.
We have most of the tools we need so far to move ahead with building our model. Let's move on to Token Embeddings and Learned Positional Encodings.
the positional embeddings will have shape (max_seq_len, d_model).
Each position (time step) will have a trainable vector.
In our case, our token embeddings will be (vocab_size, d_model)
Our initial residual stream will be
residual = token_embed + pos_embed
Token Embeddings and Learned Positional Encodings
# Setup an embedding class class Embeddings: def __init__(self, vocab_size, d_model, max_len): self.vocab_size = vocab_size self.d_model = d_model self.max_len = max_len # Our token embeddings will be: (vocab_size, d_model) # We will also use this for weight tying strategy later when setting up our Language Model (LM) Head self.W_tok = (np.random.randn(vocab_size+1, d_model) / np.sqrt(d_model) ).astype(np.float32) # Learned positional embeddings: (max_len, d_model) self.W_pos = (np.random.randn(max_len, d_model) / np.sqrt(d_model) ).astype(np.float32) def __call__(self, x): ''' x: (B, T) our integer token indices Returns: residual stream (B, T, d_model) ''' B, T = x.shape assert T <= max_len, f'Sequence length: {T} is greater than max len: {self.max_len} ' # Setup the token embeddings tok_emb = self.W_tok[x] # (B, T, d_model) # Setup the positional embeddings pos_emb = self.W_pos[None, :T, :] # (1, T, d_model) - This is for broadcasting residual = tok_emb + pos_emb return residual, tok_emb, pos_emb # Just something to start with max_len = 64 # Set a manual seed so our results are the same np.random.seed(10) emb = Embeddings(vocab_size=vocab_size, d_model=cfg.d_model, max_len=max_len) # Time to build the initial residual stream from x residual, tok_emb, pos_emb = emb(X) # All shapes or now (1, T-1, d_model) residual.shape, tok_emb.shape, pos_emb.shape
--------------((1, 25, 16), (1, 25, 16), (1, 25, 16))Cool, we setup our residual, we got our token and positional embeddings.
# Visualize the untrained positional embeddings def plot_positional_embeddings_heatmap(W_pos, num_positions=16): num_positions = min(num_positions, W_pos.shape[0]) plt.figure(dpi=150) plt.title(f'Learned positional embeddings: First: {num_positions}') plt.imshow(W_pos[:num_positions], aspect='auto', cmap='coolwarm') plt.colorbar() plt.xlabel('d_model') plt.ylabel('Position') plt.yticks(ticks=range(0, len(cfg.text),1), labels=cfg.text) plt.xticks(ticks=range(0, cfg.d_model, 1)) plt.show() plot_positional_embeddings_heatmap(emb.W_pos, num_positions=32)
At this point, we have no structure above as no learning has been done as yet.
- Each row is a position.
- Each column is one of our 16 embedding dimensions.
- Notice that these are not smooth.
- We also see roughly same variance across
- It also looks like no two positions look identical.
Think about this as our first view as the positions embedding into the tokens
def plot_token_vs_pos_norms(tok_emb, pos_emb): ''' tok_emb, pos_emb: (B, T, d_model) ''' assert tok_emb.shape == pos_emb.shape B, T, D = tok_emb.shape tok_norms = np.linalg.norm(tok_emb, axis=-1)[0] # (T,) pos_norms = np.linalg.norm(pos_emb, axis=-1)[0] # (T,) plt.figure(figsize=(8,3)) t = np.arange(T) plt.plot(t, tok_norms, label=f'Token embedding norms - mean: {tok_norms.mean():.4f}') plt.plot(t, pos_norms, label=f'Positional embedding norms - mean: {pos_norms.mean():.4f}') plt.xlabel('Position {t}') plt.ylabel('L2 norm') plt.legend() plt.show() # Test the function plot_token_vs_pos_norms(tok_emb, pos_emb)
Our data has d_model = 16 dimensions at this time. We cannot visualize this, so let's leverage PCA to bring this data down.
We see the average mean norm is about the same. This means they are about the same scale
If the positional norms are too small, the model may struggle to learn
At the same time, we don't want the positional embeddings to be too large. We do not wish to overwhelm the token identity
What we want is a balanced representation. This looks somewhat balanced when we look at the mean
We could leverage sklearn's PCA but let's build our own just for the fun of it.
# Setup def pca_2d(x): ''' x: (n_rows, d_dimensions) Returns: (N, 2) ''' x_mean = x.mean(axis=0, keepdims=True) x_centered = x - x_mean cov = x_centered.T @ x_centered / (x_centered.shape[0] - 1) eigvals, eigvecs = np.linalg.eigh(cov) idx = np.argsort(eigvals)[::-1] eigvecs = eigvecs[:, idx[:2]] # (D, 2) return x_centered @ eigvecs # (N, 2) # Test the function pca_2d(tok_emb.reshape(-1, 16))[:5] ---------------- array([[ 0.255776 , 0.23570058], [-0.9714201 , 0.5525704 ], [-0.19787998, 0.3391831 ], [-0.63930357, 0.12790056], [-0.07901763, -0.9264408 ]], dtype=float32)
Visualization time ...
# Let's visualize this now def plot_pca_token_vs_token_plus_pos(tok_emb, pos_emb): ''' Compare geometry of token embeddings vs token + pos ''' B, T, D = tok_emb.shape # Reshape the embeddings for PCA # We have three dimensions but only need 2 tok_flat = tok_emb.reshape(B*T, D) pos_flat = pos_emb.reshape(B*T, D) tok_pos_flat = (tok_emb + pos_emb).reshape(B*T, D) # Leverage PCA tok_pca = pca_2d(tok_flat) pos_pca = pca_2d(pos_flat) tok_pos_pca = pca_2d(tok_pos_flat) plt.figure(figsize=(12,4)) plt.subplot(131) plt.title('Token embeddings PCA') plt.scatter(tok_pca[:, 0], tok_pca[:, 1], c=np.arange(T).repeat(B), cmap='viridis') for idx, ch in enumerate(chars): plt.text(tok_pca[idx, 0], tok_pca[idx, 1], s=ch, fontsize=15) plt.subplot(132) plt.title('POS embeddings PCA') plt.scatter(pos_pca[:, 0], pos_pca[:, 1], c=np.arange(T).repeat(B), cmap='viridis') for idx, ch in enumerate(chars): plt.text(pos_pca[idx, 0], pos_pca[idx, 1], s=ch, fontsize=15) plt.subplot(133) plt.title('Token + position embeddings PCA') plt.scatter(tok_pos_pca[:, 0], tok_pos_pca[:, 1], c=np.arange(T).repeat(B), cmap='viridis') for idx, ch in enumerate(chars): plt.text(tok_pos_pca[idx, 0], tok_pos_pca[idx, 1], s=ch, fontsize=15) plt.tight_layout() plt.show() plot_pca_token_vs_token_plus_pos(tok_emb, pos_emb)
What should we take away from these images above. Here are a few things:
1. Transformer encodes some structure, even before we interact with attention or the feed forward network.
2. We want to know how adding the position embeddings change the geometry of the token embeddings
Let us move on to a masked single head attention. We will do the mask single head before moving to multi-head attention.
Single Head Masked self-attention mechanism
We have our residual (token_embeddings + positional_embeddings) with shape (1,25, 16)
At this point we have (B, T, d_model) in the end this will be (B, T, d_head). Remember we will only have one head to start, so d_head will equal to d_model.
# Define a he single head attention def single_head_attention(x, W_q, W_k, W_v): ''' x: (B, T, d_model) W_q: (d_model, d_model) W_k: (d_model, d_model) W_v: (d_model, d_model) Returns: attn_out: (B, T, d_model) attn_weights: (B, T, T) scores_raw: (B, T, T) scores_masked: (B, T, T) ''' # Get the shape B, T, D = x.shape # perform the projections to Q, K, V Q = x @ W_q # (B, T, d_model) K = x @ W_k # (B, T, d_model) V = x @ W_v # (B, T, d_model) # With the projections in place, # let get scaled dot-product attention scores scores_raw = (Q @ K.transpose(0,2,1)) / np.sqrt(cfg.d_model) # (B, T, T) # Setup the causal mask mask = np.triu(np.ones((T, T), dtype=bool), k=1) scores_masked = scores_raw.copy() scores_masked[:, mask] = -1e9 # (B, T, T) # Softmax attn_weights = softmax_stable(scores_masked, axis=-1) # (B, T, T) # Get the weighted values attn_out = attn_weights @ V # (B, T, d_model) return attn_out, attn_weights, scores_raw, scores_masked
# disable scientific notation np.set_printoptions(suppress=True) # Setup the weight matricies # We scale the initial weights here by 0.02, just to make them a bit smaller to help the training # We are basically scaling the standard deviation here so it is closer to 0 with ~0.02 std W_q = np.random.randn(cfg.d_model, cfg.d_model).astype(np.float32) * 0.02 W_k = np.random.randn(cfg.d_model, cfg.d_model).astype(np.float32) * 0.02 W_v = np.random.randn(cfg.d_model, cfg.d_model).astype(np.float32) * 0.02 # test the function attn_out, attn_weights, scores_raw, scores_masked = single_head_attention(residual, W_q , W_k, W_v) # Confirm the shapes print(f'Residua shape: {residual.shape} -> (B, T, d_model)') print(f'Attn out shape: {attn_out.shape} -> (B, T, d_model)') print(f'Attn weights shape: {attn_weights.shape} -> (B, T, T)') print(f'Scores raw shape: {scores_raw.shape} -> (B, T, T) ') print(f'Scores masked shape: {scores_masked.shape} -> (B, T, T)') print(f'W_q mean: {W_q.mean():.4f} | W_q std: {W_q.std():.4f}')
-------------Residua shape: (1, 25, 16) -> (B, T, d_model)
Attn out shape: (1, 25, 16) -> (B, T, d_model)
Attn weights shape: (1, 25, 25) -> (B, T, T)
Scores raw shape: (1, 25, 25) -> (B, T, T)
Scores masked shape: (1, 25, 25) -> (B, T, T)
W_q mean: -0.0002 | W_q std: 0.0189plt.figure(figsize=(15,4)) plt.subplot(141) plt.imshow(scores_raw[0], aspect='auto', cmap='viridis') plt.title('Scores pre-masking') plt.xlabel('Key Position') plt.ylabel('Query position') plt.subplot(142) plt.imshow(scores_masked[0], aspect='auto') plt.title('Scores post-masking') plt.xlabel('Key Position') #plt.ylabel('Query position') plt.subplot(143) plt.imshow(attn_weights[0], aspect='auto', cmap='viridis') plt.title('Attention Weights') plt.xlabel('Key Position') #plt.ylabel('Query position') plt.subplot(144) plt.imshow(attn_out[0], aspect='auto', cmap='viridis') plt.title('Attention Output') plt.xlabel('Key Position') #plt.ylabel('Query position') plt.colorbar() plt.tight_layout() plt.show()
**Pre-masking**
For the pre-masking, the query is the row and the column is the key
We still do not have any structure as yet in the pre-masking plot
right now, each token is most likely similar to itself
This represents the unrestricted attention landscape
We can conclude this is how the model would attend if there was no autoregressive behaviour
This is the raw nature of the residual stream
**Post-masking**
What do we take away from the post-masking.
Keep in mind, this is the same matrix as the pre-masking. Only difference now is the upper triangle has been replaced with -inf
The mask prevents the model from looking to the future
Every token can only attend to itself and the tokens preceding it.
This is the core idea behind autoregressive generation
**Attention weights**
This now says where each token looks
Position 0 can only attend to itself
Earlier positions distribute the attention across earlier tokens
Think about this as a routing mechanism where the attention flows across the sequences
This is the model communicating with itself
**Attention output**
Finally, we have the attention output
# Plot the per attention weights attn_weights.shape # Get the shape data B, T, d_model = attn_weights.shape # Get the bar plot plt.figure(figsize=(15,10)) for i in range(28): plt.subplot(7, 4,i+1) plt.bar(np.arange(T), attn_weights[0, i]) plt.title(f'attn distribution for pos: {i}:{cfg.text[i]}') plt.xlabel('key position') plt.ylabel('attn weights') plt.xticks(ticks=range(0,25,1)) if i == attn_weights.shape[1] - 1: break plt.tight_layout() #plt.bar(np.arange(T), attn_weights[0, 10])
Visualize ...
What do you take away from above.
The one bar in the first plot, means that the model can only attend to the first token. Basically itself.
For position 5 for example, the model can only attend to positions 0-4. These values sum to 1 for the probabilities
Some positions may strongly prefer one earlier token
Overall, we can look at attention as a probability distribution. From a local perspective the model is attending to nearby tokens. From the global perspective the model attends broadly. It is self-focused when the model attends mostly to itself.
Plot the update norm to the residual. Visualize once again.
# We have a larger residual norm than the update norm # This is what we want upd_norms = np.linalg.norm(attn_out[0], axis=-1) res_norms = np.linalg.norm(residual[0], axis=-1) plt.plot(np.arange(T), upd_norms, label=f'Attention update norm mean: {upd_norms.mean():.4f}') plt.plot(np.arange(T), res_norms, label=f'residual norm mean {res_norms.mean():.4f}') plt.xlabel('Positions T') plt.ylabel('L2 Norm') plt.title("Attention update vs residual norm (single head)") plt.legend() plt.show()
Think of attention as an additive update not a replacement.
The update is usually small relative to the residual stream
Now that we understand how a single attention head works, let us move on to multi-head attention.
Multi-head attention
We build our own multi-head attention mechanism.
# Let us do this via a class class MultiHeadSelfAttention: def __init__(self, d_model, n_heads, dropout_p=0.0): # Let us ensure that the d_model is divisible by n_heads assert d_model % n_heads == 0, f'd_model: {d_model} not divisible by n_heads: {n_heads}' self.d_model = d_model # Each head shares the same input # but will see different subspaces hence difference perspectives self.n_heads = n_heads self.d_head = d_model // n_heads # Using one QKV projection: (d_model, 3*d_model) # This approach is also more efficient self.W_qkv = (np.random.randn(d_model, 3 * d_model) * 0.02).astype(np.float32) # Also setup our input projection # We need this to fuse the heads back together # Fuse the information from the different heads together self.W_o = (np.random.randn(d_model, d_model) * 0.002).astype(np.float32) # Setup dropout self.dropout = Dropout(p=dropout_p) def __call__(self, x): ''' x> (B, T, d_model) returns: out: (B, T, d_model) attn_weights: (B, n_heads, T, T) ''' # Capture the shape information B, T, D = x.shape # do our first linear projection to QKV qkv = x @ self.W_qkv # (B, T, 3*d_model) # 3 is included below for the each of the QKV qkv = qkv.reshape(B, T, 3, self.n_heads, self.d_head) # (B, T, 3, n_heads, d_head) # Transpose the dimensions qkv = np.transpose(qkv, axes=(2, 0, 3, 1, 4)) # (3, B, n_heads, T, d_head) # Extract the Q, K, V # Each of these now have a shape of (B, n_heads, T, d_head ) Q, K, V = qkv[0], qkv[1], qkv[2] # Scaled dot-product attention per head # shape (n_heads, T, T) scores = (Q @ K.transpose(0, 1, 3, 2)) / np.sqrt(self.d_head) # Setup the causal mask mask = np.triu(np.ones((T, T), dtype=bool), k=1) # (T,T) scores_masked = scores.copy() scores_masked[:, :, mask] = -1e9 # Apply softmax attn_weights = softmax_stable(scores_masked, axis=-1) # (B, n_heads, T, T) # Get the weighted sum of values attn_out = attn_weights @ V # (B, n_heads, T, d_head) # Let us put these heads back together attn_out = attn_out.transpose(0, 2, 1, 3).reshape(B, T, self.d_model) # (B, T, d_model) # Final output projection from the attention mechanism out = attn_out @ self.W_o # (B, T, d_model) # Let's add a dropout if needed out = self.dropout(out) return out, attn_weights, scores, scores_masked
Testing the class
# Test the class mha = MultiHeadSelfAttention(d_model=cfg.d_model, n_heads=cfg.n_heads) mha_out, mha_attn_weights, mha_scores_raw, mha_scores_masked = mha(x=residual) # confirming the shapes we saw above mha_out.shape, mha_attn_weights.shape, mha_scores_raw.shape, mha_scores_masked.shape --------------- ((1, 25, 16), (1, 4, 25, 25), (1, 4, 25, 25), (1, 4, 25, 25))
With the output from the multi-head self-attention, let us visualize some of the items we retrieved
plt.figure(figsize=(15,4)) plt.suptitle('Plots of masked heads') for i in range(cfg.n_heads): plt.subplot(1,4,i+1) plt.imshow(mha_attn_weights[0, i], cmap='viridis') plt.title(f'head: {i}') plt.xlabel('Key Position') if i == 0: plt.ylabel('Query') plt.tight_layout()
Remember this model has not been trained as yet, hence as we move from the first row (query) down to the last row, the probabilities are more diffused. This is why the colours move from bright above to seemingly the same as you go further down.
Plot the output variance
We see the residual var is larger than the MHA out variance.
# Plot the output variance # We see the residual var is larger than the MHA out variance residual_var = residual[0].var(axis=-1) mha_out_var = mha_out[0].var(axis=-1) B, T, D = residual.shape t = np.arange(T) plt.plot(t, residual_var, label=f'residual (input) variance. Mean: {residual_var.mean():.4f}') plt.plot(t, mha_out_var, label=f'MHA output variance. Mean: {mha_out_var.mean():.4f}') plt.xlabel('Positions ') plt.ylabel('Variance across d_model') plt.title('Variance of Residual vs MHA Out') plt.legend() plt.show()
We have the multi-head self-attention mechanism in place. Let's move on to the feed forward network (FFN).
Feed Forward
The FFN is where the computation occurs
This FFN is position wise. There is no mixing across the time dimension.
class FeedForward: def __init__(self, d_model, ffn_expansion=4, dropout_p=0.0): ''' ffn_expansion=4: 4 * d_model ''' self.d_model = d_model self.d_hidden = ffn_expansion * d_model # Our first linear projection self.W1 = (np.random.randn(d_model, self.d_hidden)*0.02).astype(np.float32) self.b1 = np.zeros((self.d_hidden,), dtype=np.float32) # Our second linear projection self.W2 = (np.random.randn(self.d_hidden, self.d_model)*0.02).astype(np.float32) self.b2 = np.zeros((self.d_model,), dtype=np.float32) # Setup dropout self.dropout = Dropout(p=dropout_p) def __call__(self, x): ''' x: (B, T, d_model) returns: (B, T, d_model) ''' # Apply the first linear layer h = x @ self.W1 + self.b1 # (B, T, d_hidden) # Apply the activation function h = gelu(h) # Final linear projection and get the output of the ffn out = h @ self.W2 + self.b2 # (B, T, d_model) # Apply dropout if available out = self.dropout(out) return out # Test the function ffn = FeedForward(d_model=cfg.d_model) # Realistically, we should test this on the output of the MHA # ffn(mha_out).shape, # Let's test it on our residual, the original input ffn(residual).shape # Visualization of the effects of the FFN on the input ffn_pre_activation = residual@ ffn.W1 + ffn.b1 ffn_post_activation = gelu(ffn_pre_activation) plt.figure(figsize=(10,4)) plt.subplot(121) plt.title('FFN pre-GELU activations') plt.hist(ffn_pre_activation.flatten()) plt.subplot(122) plt.title('FFN post-GELU activations') plt.hist(ffn_post_activation.flatten()) plt.tight_layout() plt.show()
Above, we see the impact the activation function has on the input.
Let's take a scatter plot
x_flat = residual.reshape(-1, cfg.d_model) y_flat = ffn(residual).reshape(-1, cfg.d_model) plt.figure(figsize=(15,15)) for i in range(16): plt.subplot(4,4,i+1) plt.scatter(x_flat[:, i ], y_flat[:, i]) plt.title(f'ffn input vs out for dim: {i}') plt.xlabel(f'input at dim: {i}') plt.ylabel(f'output at dim: {i}') plt.tight_layout()
One take away from here is how the FFN reshapes the input vectors
# Remember when we called GELU some neurons will become 0 # Let's calculate how many of those neurons are 0s sparsity = np.mean(ffn_post_activation > 0, axis=(0,1)) plt.title(f'FFN neuron activation sparsity') plt.plot(sparsity) plt.xlabel('Hidden Neuron Index') plt.ylabel('Fraction Active');
# Get the norms residual_norms = np.linalg.norm(residual[0], axis=-1) out_norm = np.linalg.norm(mha_out[0], axis=-1) plt.plot(t, out_norm, label=f'MHA out norm mean: {out_norm.mean():.4f}') plt.plot(t, residual_norms, label=f'Residual norm: {residual_norms.mean():.4f}') plt.title(f'Residual norms vs MHA out norm') plt.xlabel('Position') plt.ylabel('L2 norm') plt.legend()
With all of this in place let's go ahead and put it all together.
We will use our pre LayerNorm
residual connection
MHA + FFN combined
Basically, let us put together a decoder block
Decoder Block
Like we have done before, let us build a Class
class DecoderBlock: def __init__(self, d_model, n_heads, ffn_expansion=4, attn_dropout=0.0, ffn_dropout=0.0): # Setup the layer norms self.ln1 = LayerNorm(d_model=d_model) self.ln2 = LayerNorm(d_model=d_model) # Setup MHA self.mha = MultiHeadSelfAttention(d_model=d_model, n_heads=cfg.n_heads, dropout_p=attn_dropout) # Setup the FFN self.ffn = FeedForward(d_model=d_model, ffn_expansion=ffn_expansion, dropout_p=ffn_dropout) def __call__(self, x): ''' residual: (B, T, d_model) returns: residual_out: (B, T, d_model) cache: dict of intermediates for visualizations ''' cache = {} # MHA Block x_norm1 = self.ln1(x) mha_out, attn_weights, scores_raw, scores_masked = self.mha(x) # Get the residual after the MHA residual_mha = x + mha_out # Residual updates # Cache some results cache['x_norm1'] = x_norm1 cache['mha_out'] = mha_out cache['attn_weights'] = attn_weights cache['scores_masked'] = scores_masked cache['residual_after_mha'] = residual_mha # FFN Block x_norm2 = self.ln2(residual_mha) ffn_out = self.ffn(x_norm2) residual_mha_ffn_out = residual_mha + ffn_out cache['x_norm2'] = x_norm2 cache['ffn_out'] = ffn_out cache['residual_mha_ffn_out'] = residual_mha_ffn_out return residual_mha_ffn_out, cache
# Test the class decoder_block = DecoderBlock(d_model=cfg.d_model, n_heads=cfg.n_heads) decoder_out, decoder_cache = decoder_block(x=residual) decoder_out.shape, decoder_cache.keys()
---------------((1, 25, 16),
dict_keys(['x_norm1', 'mha_out', 'attn_weights', 'scores_masked', 'residual_after_mha', 'x_norm2', 'ffn_out', 'residual_mha_ffn_out']))Now that we have put together one decoder block, let's put together a stack.
Stack of Decoders
class DecoderStack: def __init__(self, n_layers, d_model, n_heads, ffn_expansion=4, attn_dropout_p=0.0, ffn_dropout_p=0.0): self.n_layers = n_layers self.blocks = [ DecoderBlock(d_model=cfg.d_model, n_heads=cfg.n_heads) for _ in range(n_layers)] def __call__(self, x): ''' x: residual (B, T, d_model) returns: residual: (B, T, d_model) all_caches: list[dict] per layer ''' all_caches = [] for layer_idx, block in enumerate(self.blocks): x, cache = block(x) cache['layer_idx'] = layer_idx all_caches.append(cache) return x, all_caches
Test the stack
decoder_stack = DecoderStack(n_layers=cfg.n_layers, d_model=cfg.d_model, n_heads=cfg.n_heads) residual_final, caches = decoder_stack(residual) residual_final.shape, [ i.keys() for i in caches ] -------------- ((1, 25, 16), [dict_keys(['x_norm1', 'mha_out', 'attn_weights', 'scores_masked', 'residual_after_mha', 'x_norm2', 'ffn_out', 'residual_mha_ffn_out', 'layer_idx']), dict_keys(['x_norm1', 'mha_out', 'attn_weights', 'scores_masked', 'residual_after_mha', 'x_norm2', 'ffn_out', 'residual_mha_ffn_out', 'layer_idx'])])
# Let's visualize what we just built layer_indices = [] norms_before = [] norms_after = [] x = residual for cache in caches: layer_idx = cache['layer_idx'] res_after = cache['residual_mha_ffn_out'] norms_before.append(np.linalg.norm(residual[0], axis=-1).mean()) norms_after.append(np.linalg.norm(res_after[0], axis=-1).mean()) layer_indices.append(layer_idx) x = res_after plt.plot(layer_indices, norms_before, label='Residual norm (before layer)') plt.plot(layer_indices, norms_after, label='Residual Norm afterr') plt.xlabel('Layer') plt.ylabel('Mean L2 norm over positions') plt.title("Residual norm evolution across layers") plt.legend();
print(f'Norms before: {norms_before}') print(f'Norms after: {norms_after}') ---------------- Norms before: [np.float32(1.4254605), np.float32(1.4254605)] Norms after: [np.float64(1.426371919459707), np.float64(1.4268161008242064)]
We can see above the residual stream barely changes across layers
Residual connections add small updates
norms_before# Let's visualize what we just built layers = [] mha_updates = [] ffn_updates = [] x = residual for cache in caches: layer_idx = cache['layer_idx'] res_before = x res_after_mha = cache['residual_after_mha'] res_after = cache['residual_mha_ffn_out'] mha_update = res_after_mha - res_before ffn_update = res_after - res_after_mha mha_updates.append(np.linalg.norm(mha_update[0], axis=-1).mean()) ffn_updates.append(np.linalg.norm(ffn_update[0], axis=-1).mean()) layers.append(layer_idx) x = res_after plt.plot(layers, mha_updates, label='MHA update norm') plt.plot(layers, ffn_updates, label='FFN update norm') plt.xlabel('Layer') plt.ylabel('Mean L2 norm over positions') plt.title("Update magnitude per layer") plt.legend();
Let's prepare to wrap this up by visualizing the residual stream.
# Create a heatmap of the residual stream B, T, D = residual.shape activations = [residual[0]] x = residual.copy() for cache in caches: x =cache['residual_mha_ffn_out'] activations.append(x[0]) activations = np.stack(activations, axis=0) Lp1 = activations.shape[0] plt.figure(figsize=(15,15)) plt.imshow(activations.reshape(Lp1, T * D ), aspect='auto', cmap='coolwarm') plt.xlabel('Position x d_model') plt.ylabel('Layer (0 - input)') plt.title("Residual stream evolution across layers") plt.colorbar();
Above, we are visualizing the entire residual stream for all layers
We have a model at 16 dimensions. We have 0-25 positions 16*26. Hence the reason for the 400 points above.
Think about this as seeing how the model's internal representation evolves as it goes deeper.
We see similar patters across rows
Let us move ahead with setting up the language head, Softmax and loss ..
LM Head
We will use the weight tying approach.
class LMHead: def __init__(self, W_tok): ''' W_tok: (vocab_size, d_model) We reuse tok embeddings as output weights (weight tying) ''' self.W_out = W_tok # (vocab_size, d_model) def __call__(self, x): ''' residual: (B, T, d_model) returns: logits: (B, T, vocab_size) ''' B, T, D = x.shape V, D2 = self.W_out.shape #(B, T, D) @ (D, vocab_size) -> (B, T, vocab_size) logits = x @ self.W_out.T return logits # Test the function lmh = LMHead(W_tok=emb.W_tok) logits = lmh(residual_final) logits.shape --------- (1, 25, 16)
With the logits in place, let us now grab the probabilities.
out_preds = np.argmax(logits, axis=-1)[0] # Here is our prediction for our untrained model ''.join([ itos[i] for i in out_preds])
Well at this point, we could calculate the loss and backpropagate, etc. However, the objective was to build a transformer, not train a transformer. I think we have achieved this objective.
To train a transformer, we will take an easier route in the final part of this series: "Building and training fully functional Decoder-Only transformer."
# Put it all together class DecoderOnlyTransformer: def __init__(self, d_model, n_heads, n_layers, dropout_p, W_tok): self.decoder_stack = DecoderStack(n_layers=n_layers, d_model=d_model, n_heads=n_heads) self.lm_head = LMHead(W_tok=W_tok) def __call__(self, x): x, _ = self.decoder_stack(x) x = self.lm_head(x) return x # Setup the full Decoder only transformer transformer = DecoderOnlyTransformer(d_model=cfg.d_model, n_heads=cfg.n_heads, n_layers=cfg.n_layers, dropout_p=0.0, W_tok=emb.W_tok) # Get the logits logits = transformer(residual) logits.shape
With the logits, we can get the probabilities again if we wish. Maybe you wanted to plot the probability distribution or something.
out_preds = np.argmax(logits, axis=-1)[0]
And finally, we generate from our untrained model.
# Here is our generation for our untrained model ''.join([ itos[i] for i in out_preds]) ----------- 'Welco e fo ohe world of A'
Well that's it for this second post in this series. If you find something I could have or should have done differently, let me know.
Let us do this with Torch now and leverage Andrej Karpathy's makemore series to generate new baby names.
Posts in this series:
- Git Notebook:
2: Welcome to the world of AI - Learning about the Decoder-Only Transformer - From scratch with NumPy
No comments:
Post a Comment