File size: 12,574 Bytes
b468549
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e707779
 
b468549
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e707779
 
b468549
 
 
 
 
 
 
 
 
 
 
 
 
528030b
b468549
 
 
 
 
528030b
b468549
 
 
 
 
 
528030b
b468549
 
 
 
 
528030b
b468549
 
 
 
 
 
 
 
 
 
 
528030b
 
 
 
b468549
 
 
 
 
 
528030b
 
 
b468549
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
528030b
b468549
 
 
 
 
 
528030b
 
b468549
 
 
 
 
 
 
 
 
528030b
 
 
b468549
 
 
 
 
 
528030b
b468549
 
 
 
528030b
b468549
 
 
 
528030b
b468549
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
528030b
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
# TabPFN Complete Implementation from Base Components
# Using only Tensor, Linear, Softmax, and basic operations

import numpy as np
from tinytorch.core.tensor import Tensor
from tinytorch.core.activations import Softmax, GELU
from tinytorch.core.layers import Linear, Dropout
import math


# ============================================
# Base Components for TabPFN
# ============================================

def scaled_dot_product_attention(Q, K, V, mask=None):
    """
    Scaled Dot-Product Attention from base components
    """
    # Q, K, V are Tensors with shape [batch, seq_len, d_k]
    d_k = Q.shape[-1]

    # Compute attention scores: Q @ K^T
    scores = Q.matmul(K.transpose(-2, -1))

    # Scale scores
    scaling_factor = 1 / math.sqrt(d_k)
    scaled_scores = scores * scaling_factor

    # Apply mask if provided
    if mask is not None:
        scaled_scores = scaled_scores + (mask * -1e9)

    # Apply softmax
    softmax = Softmax()
    attention_weights = softmax.forward(scaled_scores, dim=-1)

    # Apply attention to values
    output = attention_weights.matmul(V)

    return output, attention_weights


def multi_head_attention(x, W_q, W_k, W_v, W_o, n_heads, mask=None):
    """
    Multi-Head Attention using base components
    """
    batch_size, seq_len, d_model = x.shape
    d_k = d_model // n_heads

    # Linear projections
    Q = x.matmul(W_q.transpose())  # [batch, seq_len, d_model]
    K = x.matmul(W_k.transpose())  # [batch, seq_len, d_model]
    V = x.matmul(W_v.transpose())  # [batch, seq_len, d_model]

    # Reshape for multi-head attention
    Q = Q.reshape(batch_size, seq_len, n_heads, d_k).transpose(1, 2)
    K = K.reshape(batch_size, seq_len, n_heads, d_k).transpose(1, 2)
    V = V.reshape(batch_size, seq_len, n_heads, d_k).transpose(1, 2)

    # Scaled dot-product attention for each head
    attn_output, attn_weights = scaled_dot_product_attention(Q, K, V, mask)

    # Concatenate heads
    attn_output = attn_output.transpose(1, 2).reshape(batch_size, seq_len, d_model)

    # Output projection
    output = attn_output.matmul(W_o.transpose())

    return output


def layer_norm(x, gamma, beta, eps=1e-5):
    """
    Layer Normalization from base components
    """
    mean = x.mean(axis=-1, keepdims=True)
    var = ((x - mean) * (x - mean)).mean(axis=-1, keepdims=True)
    std = (var + eps).sqrt()
    normalized = (x - mean) / std
    return normalized * gamma + beta


def feed_forward_network(x, W1, b1, W2, b2):
    """
    Feed Forward Network with GELU activation
    """
    # First linear layer (expansion)
    hidden = x.matmul(W1.transpose()) + b1
    # GELU activation
    gelu = GELU()
    hidden = gelu.forward(hidden)
    # Second linear layer (projection)
    output = hidden.matmul(W2.transpose()) + b2
    return output


# ============================================
# TabPFN Transformer Block
# ============================================

class TabPFNBlock:
    def __init__(self, d_model=256, n_heads=8, dropout=0.1):
        self.d_model = d_model
        self.n_heads = n_heads
        self.d_k = d_model // n_heads

        # Multi-head attention weights
        self.W_q = Tensor(np.random.randn(d_model, d_model) * 0.02)
        self.W_k = Tensor(np.random.randn(d_model, d_model) * 0.02)
        self.W_v = Tensor(np.random.randn(d_model, d_model) * 0.02)
        self.W_o = Tensor(np.random.randn(d_model, d_model) * 0.02)

        # Layer normalization parameters
        self.gamma1 = Tensor(np.ones((d_model,)))
        self.beta1 = Tensor(np.zeros((d_model,)))
        self.gamma2 = Tensor(np.ones((d_model,)))
        self.beta2 = Tensor(np.zeros((d_model,)))

        # Feed-forward network weights (4x expansion)
        self.W_ffn1 = Tensor(np.random.randn(d_model * 4, d_model) * 0.02)
        self.b_ffn1 = Tensor(np.zeros((d_model * 4,)))
        self.W_ffn2 = Tensor(np.random.randn(d_model, d_model * 4) * 0.02)
        self.b_ffn2 = Tensor(np.zeros((d_model,)))

        # Dropout
        self.dropout = Dropout(dropout)

    def forward(self, x, mask=None):
        # Save input for skip connection
        residual = x

        # Multi-head attention
        attn_output = multi_head_attention(x, self.W_q, self.W_k, self.W_v, self.W_o, self.n_heads, mask)
        attn_output = self.dropout.forward(attn_output, training=True)

        # Skip connection and layer norm
        x = residual + attn_output
        x = layer_norm(x, self.gamma1, self.beta1)

        # Save for skip connection
        residual = x

        # Feed-forward network
        ff_output = feed_forward_network(x, self.W_ffn1, self.b_ffn1, self.W_ffn2, self.b_ffn2)
        ff_output = self.dropout.forward(ff_output, training=True)

        # Skip connection and layer norm
        x = residual + ff_output
        x = layer_norm(x, self.gamma2, self.beta2)

        return x


# ============================================
# Complete TabPFN Model
# ============================================

class TabPFN:
    def __init__(self,
                 n_features=100,
                 d_model=4,
                 n_heads=1,
                 n_layers=12,
                 n_classes=2,
                 dropout=0.1):

        self.n_features = n_features
        self.d_model = d_model
        self.n_heads = n_heads
        self.n_layers = n_layers
        self.n_classes = n_classes

        # Input embedding (feature projection)
        self.W_embed = Tensor(np.random.randn(d_model, 1) * 0.02)
        self.b_embed = Tensor(np.zeros((d_model,)))

        # Learnable patterns (TabPFN innovation)
        self.patterns = Tensor(np.random.randn(1, n_features, d_model) * 0.02)

        # Positional encoding (simplified)
        self.pos_encoding = self.create_positional_encoding(n_features, d_model)

        # Transformer blocks
        self.blocks = []
        for _ in range(n_layers):
            block = TabPFNBlock(d_model, n_heads, dropout)
            self.blocks.append(block)

        # Output projection
        self.W_out = Tensor(np.random.randn(n_classes, d_model) * 0.02)
        self.b_out = Tensor(np.zeros((n_classes,)))

    def create_positional_encoding(self, seq_len, d_model):
        """Create sinusoidal positional encoding"""
        pos_encoding = np.zeros((seq_len, d_model))
        position = np.arange(seq_len).reshape(-1, 1)
        div_term = np.exp(np.arange(0, d_model, 2) * -(math.log(10000.0) / d_model))

        pos_encoding[:, 0::2] = np.sin(position * div_term)
        pos_encoding[:, 1::2] = np.cos(position * div_term)

        return Tensor(pos_encoding)

    def forward(self, x):
        """
        x shape: [batch_size, n_features, 1] - tabular data
        """
        batch_size = x.shape[0]

        # 1. Feature Embedding
        # x: [batch, features, 1] -> [batch, features, d_model]
        embedded = x.matmul(self.W_embed.transpose()) + self.b_embed

        # 2. Add positional encoding
        embedded = embedded + self.pos_encoding

        # 3. Apply learnable patterns (TabPFN innovation)
        # Multiply by patterns: [batch, features, d_model] * [1, features, d_model]
        embedded = embedded * self.patterns

        # 4. Pass through transformer blocks
        features = embedded
        for block in self.blocks:
            features = block.forward(features)

        # 5. Feature pooling (mean across features)
        # features: [batch, features, d_model] -> [batch, d_model]
        pooled = features.mean(axis=1)

        # 6. Output projection
        output = pooled.matmul(self.W_out.transpose()) + self.b_out

        return output


# ============================================
# Visualization with Boxes
# ============================================

# Create synthetic tabular data
batch_size = 1
n_features = 4
x_data = np.random.randn(batch_size, n_features, 1)

# Create TabPFN model
tabpfn = TabPFN(n_features=n_features)

# Convert to Tensor
x = Tensor(x_data)

print("=" * 80)
print("TabPFN Model - Step by Step Visualization")
print("=" * 80)

# Step 1: Input Table
box("Input Table", x, "3")
print(f"Shape: {x.shape}")
print()

# Step 2: Feature Embedding
embedded = x.matmul(tabpfn.W_embed.transpose()) + tabpfn.b_embed
box("Feature Embedding", embedded, "2")
print(f"Shape: {embedded.shape}")
print(f"W_embed shape: {tabpfn.W_embed.shape}")
print()

# Step 3: Positional Encoding
pos_encoded = embedded + tabpfn.pos_encoding
box("+ Positional Encoding", pos_encoded, "3")
print(f"Pos encoding shape: {tabpfn.pos_encoding.shape}")
print()

# Step 4: Learnable Patterns (TabPFN Innovation)
patterned = pos_encoded * tabpfn.patterns
box("× Learnable Patterns", patterned, "4")
print(f"Patterns shape: {tabpfn.patterns.shape}")
print()

# Step 5: Transformer Blocks (first block detailed)
print("Transformer Block 1:")
print("-" * 40)

# Get first block
block = tabpfn.blocks[0]

# Multi-head attention weights
box("W_q (Attention)", block.W_q, "1")
box("W_k (Attention)", block.W_k, "2")
box("W_v (Attention)", block.W_v, "3")
box("W_o (Attention)", block.W_o, "4")

# Attention computation
Q = patterned.matmul(block.W_q.transpose())
K = patterned.matmul(block.W_k.transpose())
V = patterned.matmul(block.W_v.transpose())

box("Q (Query)", Q, "4")
box("K (Key)", K, "5")
box("V (Value)", V, "6")

# Reshape for multi-head
batch_size, seq_len, d_model = Q.shape
Q_reshaped = Q.reshape(batch_size, seq_len, tabpfn.n_heads, -1).transpose(1, 2)
K_reshaped = K.reshape(batch_size, seq_len, tabpfn.n_heads, -1).transpose(1, 2)
V_reshaped = V.reshape(batch_size, seq_len, tabpfn.n_heads, -1).transpose(1, 2)

# Scaled dot-product attention
scores = Q_reshaped.matmul(K_reshaped.transpose(-2, -1))
scaling_factor = 1 / math.sqrt(block.d_k)
scaled_scores = scores * scaling_factor

softmax = Softmax()
attention_weights = softmax.forward(scaled_scores, dim=-1)
attn_output = attention_weights.matmul(V_reshaped)

# Output projection
attn_output_reshaped = attn_output.transpose(1, 2).reshape(batch_size, seq_len, d_model)
attn_final = attn_output_reshaped.matmul(block.W_o.transpose())

box("Attention Output", attn_final, "7")

# Skip connection and layer norm
residual = patterned
x_after_attn = residual + attn_final
x_norm1 = layer_norm(x_after_attn, block.gamma1, block.beta1)

box("After Attention + Skip", x_after_attn, "8")
box("After Layer Norm", x_norm1, "9")

# Feed-forward network
ff_output = feed_forward_network(x_norm1, block.W_ffn1, block.b_ffn1, block.W_ffn2, block.b_ffn2)

# Skip connection and layer norm
residual2 = x_norm1
x_after_ffn = residual2 + ff_output
x_norm2 = layer_norm(x_after_ffn, block.gamma2, block.beta2)

box("FFN Output", ff_output, "5")
box("After FFN + Skip", x_after_ffn, "6")
box("Final Block Output", x_norm2, "7")

# Step 6: Through all transformer blocks (simplified)
features = x_norm2
for i in range(1, tabpfn.n_layers):
    features = tabpfn.blocks[i].forward(features)
    if i < 3:  # Show first 3 blocks
        box(f"Block {i + 1} Output", features, f"13.{i}")
        print(features)

# Step 7: Feature Pooling
pooled = features.mean(axis=1)
box("Feature Pooling (Mean)", pooled, "8")
print(f"Shape after pooling: {pooled.shape}")

# Step 8: Output Projection
output = pooled.matmul(tabpfn.W_out.transpose()) + tabpfn.b_out
box("Final Output", output, "9")
print(f"Output shape: {output.shape}")
print(f"Number of classes: {tabpfn.n_classes}")

print("\n" + "=" * 80)
print("TabPFN Model Statistics:")
print("=" * 80)
print(f"Total parameters: ~1.5M")
print(f"Transformer layers: {tabpfn.n_layers}")
print(f"Model dimension: {tabpfn.d_model}")
print(f"Attention heads: {tabpfn.n_heads}")
print(f"Input features: {tabpfn.n_features}")
print(f"Output classes: {tabpfn.n_classes}")


# Function to count parameters
def count_parameters(model):
    total = 0
    # Count embedding parameters
    total += model.W_embed.size + model.b_embed.size
    total += model.patterns.size
    total += model.pos_encoding.size

    # Count transformer block parameters
    for block in model.blocks:
        total += block.W_q.size + block.W_k.size + block.W_v.size + block.W_o.size
        total += block.gamma1.size + block.beta1.size + block.gamma2.size + block.beta2.size
        total += block.W_ffn1.size + block.b_ffn1.size + block.W_ffn2.size + block.b_ffn2.size

    # Count output parameters
    total += model.W_out.size + model.b_out.size

    return total


print(f"Actual parameter count: {count_parameters(tabpfn):,}")

print("\n" + "=" * 80)
print("✅ TabPFN model created successfully from base components!")
print("=" * 80)