discojs/src/models/gpt/layers: fix weight initializations

discojs/src/models/gpt/layers.ts

@@ -67,13 +67,14 @@ tf.serialization.registerClass(LogLayer)
 
 type CausalSelfAttentionConfig =
     ConstructorParameters<typeof tf.layers.Layer>[0]
-    & Record<'blockSize' | 'nHead' | 'nEmbd' | 'dropout', number>
+    & Record<'blockSize' | 'nHead' | 'nEmbd' | 'dropout' | 'nLayer', number>
 
 class CausalSelfAttention extends tf.layers.Layer {
   static readonly className = 'CausalSelfAttention'
 
   private readonly nHead: number
   private readonly nEmbd: number
+  private readonly nLayer: number
   private readonly dropout: number
   private readonly mask: tf.Tensor2D
   cAttnKernel?: tf.LayerVariable
@@ -88,6 +89,7 @@ class CausalSelfAttention extends tf.layers.Layer {
 
     this.nEmbd = config.nEmbd
     this.nHead = config.nHead
+    this.nLayer = config.nLayer
     this.dropout = config.dropout
 
     // mask is a lower triangular matrix filled with 1
@@ -102,7 +104,7 @@ class CausalSelfAttention extends tf.layers.Layer {
       'c_attn.weight',
       [this.nEmbd, 3 * this.nEmbd],
       'float32',
-      tf.initializers.glorotNormal({})
+      tf.initializers.randomNormal({ mean:0, stddev:0.02 }) // use same init as GPT2
     )
     this.cAttnBias = this.addWeight(
       'c_attn.bias',
@@ -115,7 +117,12 @@ class CausalSelfAttention extends tf.layers.Layer {
       'c_proj.kernel',
       [this.nEmbd, this.nEmbd],
       'float32',
-      tf.initializers.glorotNormal({})
+      // the input keeps accumulating through the residual stream so we
+      // scale the initialization with the nb of layers to keep a unit std
+      // Sources:
+      // https://github.com/karpathy/build-nanogpt/blob/6104ab1b53920f6e2159749676073ff7d815c1fa/train_gpt2.py#L103
+      // https://youtu.be/l8pRSuU81PU?si=5GcKfi_kPgLgvtg2&t=4640
+      tf.initializers.randomNormal({ mean:0, stddev: 0.02 * Math.sqrt(2 * this.nLayer) })
     )
     this.cProjBias = this.addWeight(
       'c_proj.bias',
@@ -255,27 +262,31 @@ class GELU extends tf.layers.Layer {
 tf.serialization.registerClass(GELU)
 
 type MLPConfig = ConstructorParameters<typeof tf.layers.Layer>[0] &
-  Required<ModelSize> & Record<'blockSize' | 'residDrop', number>
+  Required<ModelSize> & Record<'blockSize' | 'residDrop' | 'nLayer', number>
 
 function MLP(config: MLPConfig): tf.LayersModel {
   return tf.sequential({ layers: [
     tf.layers.dense({
       name: config.name + `.mlp.c_fc`,
       units: 4 * config.nEmbd,
       inputDim: config.nEmbd,
-      inputShape: [config.blockSize, config.nEmbd]
+      inputShape: [config.blockSize, config.nEmbd],
+      kernelInitializer: tf.initializers.randomNormal({ mean: 0, stddev: 0.02 }),
     }),
     new GELU(),
     tf.layers.dense({
-        name: config.name + '.mlp.c_proj',
-        units: config.nEmbd,
-        inputDim: 4 * config.nEmbd,
-      inputShape: [config.blockSize, 4 * config.nEmbd]
-      }),
-      tf.layers.dropout({
-        name: config.name + '.mlp.drop',
-        rate: config.residDrop
+      name: config.name + '.mlp.c_proj',
+      units: config.nEmbd,
+      inputDim: 4 * config.nEmbd,
+      inputShape: [config.blockSize, 4 * config.nEmbd],
+      kernelInitializer: tf.initializers.randomNormal({
+        mean: 0, stddev: 0.02 * Math.sqrt(2 * config.nLayer)
       }),
+    }),
+    tf.layers.dropout({
+      name: config.name + '.mlp.drop',
+      rate: config.residDrop
+    }),
   ]})
 }
 
@@ -362,7 +373,7 @@ class LMEmbedding extends tf.layers.Layer {
       'wte', //use same name as GPT2
       [this.vocabSize, this.nEmbd],
       'float32',
-      tf.initializers.randomNormal({})
+      tf.initializers.randomNormal({ mean:0, stddev:0.02 })
     )
   }
 
@@ -452,7 +463,7 @@ export function GPTArchitecture(config: Required<GPTConfig>): tf.LayersModel {
     name: config.name + '.wpe',
     inputDim: config.blockSize,
     outputDim: config.nEmbd,
-    embeddingsInitializer: 'zeros'
+    embeddingsInitializer: tf.initializers.randomNormal({ mean: 0, stddev: 0.02 }),
   }).apply(range) as tf.SymbolicTensor
 
   if (config.debug) {
@@ -474,7 +485,12 @@ export function GPTArchitecture(config: Required<GPTConfig>): tf.LayersModel {
     ).apply(x)
   }
   // Normalization
-  x = tf.layers.layerNormalization({ name: config.name + '.ln_f', epsilon: 1e-5 })
+  x = tf.layers.layerNormalization({
+    name: config.name + '.ln_f',
+    epsilon: 1e-5,
+    gammaInitializer: 'ones',
+    betaInitializer: 'zeros',
+  })
     .apply(x)
   if (config.debug) {
     x = new LogLayer({ name: 'ln_f_log' }).apply(x)