chapter 10 11 12 代码接口更新

chapter_10_attention_mechanisms/3_multihead-attention.ipynb

@@ -24,7 +24,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
+   "execution_count": null,
    "metadata": {
     "origin_pos": 2,
     "tab": [
@@ -35,7 +35,8 @@
    "source": [
     "from d2l import mindspore as d2l\n",
     "import mindspore\n",
-    "from mindspore import nn"
+    "from mindspore import nn\n",
+    "import mindspore.mint as mint"
    ]
   },
   {
@@ -71,24 +72,28 @@
     "        self.W_k = nn.Dense(key_size, num_hiddens, has_bias=has_bias)\n",
     "        self.W_v = nn.Dense(value_size, num_hiddens, has_bias=has_bias)\n",
     "        self.W_o = nn.Dense(num_hiddens, num_hiddens, has_bias=has_bias)\n",
+    "        # mint 中没有 expand_dims，保留 tile 和 reshape\n",
+    "        self.tile = mint.tile \n",
+    "        self.reshape = mint.reshape\n",
+    "        # 移除 self.expand_dims = mint.expand_dims，将在 construct 中使用 Tensor 方法\n",
     "\n",
     "    def construct(self, queries, keys, values, valid_lens):\n",
     "        # queries，keys，values的形状:\n",
     "        # (batch_size，查询或者“键－值”对的个数，num_hiddens)\n",
-    "        # valid_lens 的形状:\n",
-    "        # (batch_size，)或(batch_size，查询的个数)\n",
-    "        # 经过变换后，输出的queries，keys，values　的形状:\n",
-    "        # (batch_size*num_heads，查询或者“键－值”对的个数，\n",
-    "        # num_hiddens/num_heads)\n",
     "        queries = transpose_qkv(self.W_q(queries), self.num_heads)\n",
     "        keys = transpose_qkv(self.W_k(keys), self.num_heads)\n",
     "        values = transpose_qkv(self.W_v(values), self.num_heads)\n",
     "\n",
     "        if valid_lens is not None:\n",
-    "            # 在轴0，将第一项（标量或者矢量）复制num_heads次，\n",
-    "            # 然后如此复制第二项，然后诸如此类。\n",
-    "            valid_lens = d2l.repeat(\n",
-    "                valid_lens, repeats=self.num_heads, axis=0)\n",
+    "            \n",
+    "            # 1. 扩展 valid_lens 形状至 (batch_size, 1)\n",
+    "            valid_lens = valid_lens.expand_dims(1)\n",
+    "            \n",
+    "            # 2. 沿着轴 1 (新增维度) 重复 num_heads 次，形状变为 (batch_size, num_heads)\n",
+    "            valid_lens = self.tile(valid_lens, (1, self.num_heads))\n",
+    "            \n",
+    "            # 3. 展平为 (batch_size * num_heads,)\n",
+    "            valid_lens = self.reshape(valid_lens, (-1,))\n",
     "\n",
     "        # output的形状:(batch_size*num_heads，查询的个数，\n",
     "        # num_hiddens/num_heads)\n",
@@ -167,18 +172,26 @@
     ]
    },
    "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "[WARNING] ME(42683:281472939304512,MainProcess):2025-12-08-17:59:06.972.000 [mindspore/nn/layer/basic.py:174] For Dropout, this parameter `keep_prob` will be deprecated, please use `p` instead.\n",
+      "[WARNING] ME(42683:281472939304512,MainProcess):2025-12-08-17:59:06.984.000 [mindspore/nn/layer/basic.py:200] For Dropout, this parameter `keep_prob` will be deprecated, please use `p` instead.\n"
+     ]
+    },
     {
      "data": {
       "text/plain": [
-       "MultiHeadAttention<\n",
-       "  (attention): DotProductAttention<\n",
-       "    (dropout): Dropout<keep_prob=0.5>\n",
-       "    >\n",
-       "  (W_q): Dense<input_channels=100, output_channels=100>\n",
-       "  (W_k): Dense<input_channels=100, output_channels=100>\n",
-       "  (W_v): Dense<input_channels=100, output_channels=100>\n",
-       "  (W_o): Dense<input_channels=100, output_channels=100>\n",
-       "  >"
+       "MultiHeadAttention(\n",
+       "  (attention): DotProductAttention(\n",
+       "    (dropout): Dropout(keep_prob=0.5)\n",
+       "  )\n",
+       "  (W_q): Dense(input_channels=100, output_channels=100)\n",
+       "  (W_k): Dense(input_channels=100, output_channels=100)\n",
+       "  (W_v): Dense(input_channels=100, output_channels=100)\n",
+       "  (W_o): Dense(input_channels=100, output_channels=100)\n",
+       ")"
       ]
      },
      "execution_count": 5,
@@ -195,7 +208,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 6,
+   "execution_count": null,
    "metadata": {
     "origin_pos": 16,
     "tab": [
@@ -217,18 +230,18 @@
    "source": [
     "batch_size, num_queries = 2, 4\n",
     "num_kvpairs, valid_lens = 6, d2l.tensor([3, 2], mindspore.int32)\n",
-    "X = d2l.ones((batch_size, num_queries, num_hiddens))\n",
-    "Y = d2l.ones((batch_size, num_kvpairs, num_hiddens))\n",
+    "X = mint.ones((batch_size, num_queries, num_hiddens))\n",
+    "Y = mint.ones((batch_size, num_kvpairs, num_hiddens))\n",
     "attention(X, Y, Y, valid_lens).shape"
    ]
   }
  ],
  "metadata": {
   "celltoolbar": "Slideshow",
   "kernelspec": {
-   "display_name": "Python 3 (ipykernel)",
+   "display_name": "Python 3.10",
    "language": "python",
-   "name": "python3"
+   "name": "py310"
   },
   "language_info": {
    "codemirror_mode": {
@@ -240,7 +253,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.7.5"
+   "version": "3.10.14"
   },
   "rise": {
    "autolaunch": true,
@@ -251,4 +264,4 @@
  },
  "nbformat": 4,
  "nbformat_minor": 4
-}
+}