Créé avec Colaboratory

Author: ttaburet

P100.ipynb

@@ -67,6 +67,110 @@
           "name": "stdout"
         }
       ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "CPUS4nwiAlfd"
+      },
+      "source": [
+        "import keras\n",
+        "import numpy as np\n",
+        "import tensorflow as tf\n",
+        "\n",
+        "from keras.layers import Input, Dense, Lambda\n",
+        "from keras.models import Model\n",
+        "from keras import backend as K\n",
+        "\n",
+        "from tensorflow.keras.applications.resnet50 import ResNet50\n",
+        "from tensorflow.keras.preprocessing import image\n",
+        "from tensorflow.keras.applications.resnet50 import preprocess_input, decode_predictions"
+      ],
+      "execution_count": 1,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "nvu-hhKlCcsg"
+      },
+      "source": [
+        "def triplet_loss(inputs, dist='sqeuclidean', margin='maxplus'):\n",
+        "    anchor, positive, negative = inputs\n",
+        "    positive_distance = K.square(anchor - positive)\n",
+        "    negative_distance = K.square(anchor - negative)\n",
+        "    if dist == 'euclidean':\n",
+        "        positive_distance = K.sqrt(K.sum(positive_distance, axis=-1, keepdims=True))\n",
+        "        negative_distance = K.sqrt(K.sum(negative_distance, axis=-1, keepdims=True))\n",
+        "    elif dist == 'sqeuclidean':\n",
+        "        positive_distance = K.sum(positive_distance, axis=-1, keepdims=True)\n",
+        "        negative_distance = K.sum(negative_distance, axis=-1, keepdims=True)\n",
+        "    loss = positive_distance - negative_distance\n",
+        "    if margin == 'maxplus':\n",
+        "        loss = K.maximum(0.0, 1 + loss)\n",
+        "    elif margin == 'softplus':\n",
+        "        loss = K.log(1 + K.exp(loss))\n",
+        "    return K.mean(loss)\n",
+        "\n",
+        "def triplet_loss_np(inputs, dist='sqeuclidean', margin='maxplus'):\n",
+        "    anchor, positive, negative = inputs\n",
+        "    positive_distance = np.square(anchor - positive)\n",
+        "    negative_distance = np.square(anchor - negative)\n",
+        "    if dist == 'euclidean':\n",
+        "        positive_distance = np.sqrt(np.sum(positive_distance, axis=-1, keepdims=True))\n",
+        "        negative_distance = np.sqrt(np.sum(negative_distance, axis=-1, keepdims=True))\n",
+        "    elif dist == 'sqeuclidean':\n",
+        "        positive_distance = np.sum(positive_distance, axis=-1, keepdims=True)\n",
+        "        negative_distance = np.sum(negative_distance, axis=-1, keepdims=True)\n",
+        "    loss = positive_distance - negative_distance\n",
+        "    if margin == 'maxplus':\n",
+        "        loss = np.maximum(0.0, 1 + loss)\n",
+        "    elif margin == 'softplus':\n",
+        "        loss = np.log(1 + np.exp(loss))\n",
+        "    return np.mean(loss)\n",
+        "\n",
+        "def check_loss():\n",
+        "    batch_size = 10\n",
+        "    shape = (batch_size, 4096)\n",
+        "\n",
+        "    p1 = normalize(np.random.random(shape))\n",
+        "    n = normalize(np.random.random(shape))\n",
+        "    p2 = normalize(np.random.random(shape))\n",
+        "    \n",
+        "    input_tensor = [K.variable(p1), K.variable(n), K.variable(p2)]\n",
+        "    out1 = K.eval(triplet_loss(input_tensor))\n",
+        "    input_np = [p1, n, p2]\n",
+        "    out2 = triplet_loss_np(input_np)\n",
+        "\n",
+        "    assert out1.shape == out2.shape\n",
+        "    print(np.linalg.norm(out1))\n",
+        "    print(np.linalg.norm(out2))\n",
+        "    print(np.linalg.norm(out1-out2))"
+      ],
+      "execution_count": 2,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "JM-tEj92Chs_"
+      },
+      "source": [
+        "tf.random.set_seed(1234)"
+      ],
+      "execution_count": 3,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "9nqywM79C8kt"
+      },
+      "source": [
+        ""
+      ],
+      "execution_count": null,
+      "outputs": []
     }
   ]
 }