notebooks/explaining/saliency-gradient-backpropagation.ipynb

{"nbformat":4,"nbformat_minor":0,"metadata":{"colab":{"name":"1-saliency-grad-backprop.ipynb","provenance":[{"file_id":"1SQ9Vtu5a-YWR0urOdJBadcW8C4gUDy6d","timestamp":1613482411118}],"collapsed_sections":[],"toc_visible":true},"kernelspec":{"display_name":"Python 3","language":"python","name":"python3"},"language_info":{"codemirror_mode":{"name":"ipython","version":3},"file_extension":".py","mimetype":"text/x-python","name":"python","nbconvert_exporter":"python","pygments_lexer":"ipython3","version":"3.8.5"}},"cells":[{"cell_type":"markdown","metadata":{"id":"_AbalcTKgGgw"},"source":["# Saliency using Gradient Backpropagation\n","\n","Implement the saliency algorithm described in https://arxiv.org/pdf/1312.6034.pdf."]},{"cell_type":"code","metadata":{"colab":{"base_uri":"https://localhost:8080/"},"id":"_psmrM7LN_WG","executionInfo":{"status":"ok","timestamp":1614081801023,"user_tz":180,"elapsed":47266,"user":{"displayName":"Lucas","photoUrl":"https://lh3.googleusercontent.com/a-/AOh14GiQTUp6domdv59RQ9rxuJo49Q7CXVKQp2RuAgkgeQ=s64","userId":"01670254553878112810"}},"outputId":"2720bb0e-5ec7-44dd-d352-ca2ef6c05596"},"source":["from google.colab import drive\n","drive.mount('/content/drive')\n","\n","! cp -r '/content/drive/MyDrive/Colab Notebooks/cs-no/6.grads/config' ."],"execution_count":1,"outputs":[{"output_type":"stream","text":["Mounted at /content/drive\n","\u001b[K     |████████████████████████████████| 706kB 9.0MB/s \n","\u001b[?25h"],"name":"stdout"}]},{"cell_type":"code","metadata":{"scrolled":true,"id":"ctkvP4pGN2VA","executionInfo":{"status":"ok","timestamp":1614082716320,"user_tz":180,"elapsed":521,"user":{"displayName":"Lucas","photoUrl":"https://lh3.googleusercontent.com/a-/AOh14GiQTUp6domdv59RQ9rxuJo49Q7CXVKQp2RuAgkgeQ=s64","userId":"01670254553878112810"}}},"source":["import tensorflow as tf\n","from config import config, data, model, utils\n","\n","from config.utils import to_image, normalize\n","\n","utils.setup_clean_image_plotting()"],"execution_count":22,"outputs":[]},{"cell_type":"code","metadata":{"id":"ysC8003bxIE7","executionInfo":{"status":"ok","timestamp":1614081885102,"user_tz":180,"elapsed":36791,"user":{"displayName":"Lucas","photoUrl":"https://lh3.googleusercontent.com/a-/AOh14GiQTUp6domdv59RQ9rxuJo49Q7CXVKQp2RuAgkgeQ=s64","userId":"01670254553878112810"}}},"source":["images = data.load_images()\n","x = model.preprocess(images)"],"execution_count":5,"outputs":[]},{"cell_type":"code","metadata":{"id":"2jz1ZKnmi3AX","colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"status":"ok","timestamp":1614081892814,"user_tz":180,"elapsed":41525,"user":{"displayName":"Lucas","photoUrl":"https://lh3.googleusercontent.com/a-/AOh14GiQTUp6domdv59RQ9rxuJo49Q7CXVKQp2RuAgkgeQ=s64","userId":"01670254553878112810"}},"outputId":"5c7dcf0c-9695-417c-8b4c-7985e78925a9"},"source":["nn = model.build()\n","\n","logits = nn(x, training=False)\n","preds = tf.argmax(logits, axis=1)\n","probs = tf.nn.softmax(logits)\n","decoded = model.decode(probs.numpy(), top=2)"],"execution_count":6,"outputs":[{"output_type":"stream","text":["Downloading data from https://storage.googleapis.com/keras-applications/efficientnetb2.h5\n","37437440/37432240 [==============================] - 1s 0us/step\n","Downloading data from https://storage.googleapis.com/download.tensorflow.org/data/imagenet_class_index.json\n","40960/35363 [==================================] - 0s 0us/step\n"],"name":"stdout"}]},{"cell_type":"code","metadata":{"colab":{"base_uri":"https://localhost:8080/","height":1000,"output_embedded_package_id":"1VPwxPmxeiXJ1N1cRhlqOAa8b6fPMR8q5"},"id":"Ck9aq1YXxoKw","executionInfo":{"status":"ok","timestamp":1614081899347,"user_tz":180,"elapsed":45356,"user":{"displayName":"Lucas","photoUrl":"https://lh3.googleusercontent.com/a-/AOh14GiQTUp6domdv59RQ9rxuJo49Q7CXVKQp2RuAgkgeQ=s64","userId":"01670254553878112810"}},"outputId":"117df111-37d3-45c5-e006-f93c3b0098b8"},"source":["utils.plot(utils.to_image(images),\n","           titles=[f'{f}\\n{p[0][1]} {p[0][2]:.0%}'\n","                  for f, p in zip(data.class_names, decoded)],\n","           rows=4)"],"execution_count":7,"outputs":[{"output_type":"display_data","data":{"text/plain":"Output hidden; open in https://colab.research.google.com to view."},"metadata":{}}]},{"cell_type":"code","metadata":{"id":"byntNcS7wn9G","executionInfo":{"status":"ok","timestamp":1614081924507,"user_tz":180,"elapsed":10935,"user":{"displayName":"Lucas","photoUrl":"https://lh3.googleusercontent.com/a-/AOh14GiQTUp6domdv59RQ9rxuJo49Q7CXVKQp2RuAgkgeQ=s64","userId":"01670254553878112810"}}},"source":["def activation_loss(y, units):\n","    return tf.gather(y, units, axis=1, batch_dims=1)\n","\n","@tf.function\n","def gradients(inputs, units):\n","    with tf.GradientTape() as tape:\n","        tape.watch(inputs)\n","        y = nn(inputs)\n","        loss = activation_loss(y, units)\n","    \n","    grads = tape.gradient(loss, inputs)\n","\n","    return loss, grads"],"execution_count":8,"outputs":[]},{"cell_type":"code","metadata":{"id":"oCUeAz9ExTF8","executionInfo":{"status":"ok","timestamp":1614082744078,"user_tz":180,"elapsed":6403,"user":{"displayName":"Lucas","photoUrl":"https://lh3.googleusercontent.com/a-/AOh14GiQTUp6domdv59RQ9rxuJo49Q7CXVKQp2RuAgkgeQ=s64","userId":"01670254553878112810"}}},"source":["_, g = gradients(x, preds)\n","\n","sa = normalize(tf.reduce_sum(tf.abs(g), axis=-1))\n","sp = normalize(tf.reduce_sum(tf.nn.relu(g), axis=-1))\n","sn = normalize(tf.reduce_sum(tf.nn.relu(-g), axis=-1))"],"execution_count":23,"outputs":[]},{"cell_type":"code","metadata":{"colab":{"base_uri":"https://localhost:8080/","height":1000,"output_embedded_package_id":"160DcWPAIghQjA7nUFxECk6EjwlrNt980"},"id":"9hJUPz-WxYWa","executionInfo":{"status":"ok","timestamp":1614083026765,"user_tz":180,"elapsed":39210,"user":{"displayName":"Lucas","photoUrl":"https://lh3.googleusercontent.com/a-/AOh14GiQTUp6domdv59RQ9rxuJo49Q7CXVKQp2RuAgkgeQ=s64","userId":"01670254553878112810"}},"outputId":"35e7319d-95d6-4250-eb39-42387b9a5254"},"source":["columns = [\n","    to_image(images),\n","    sa.numpy(),\n","    sp.numpy(),\n","    sn.numpy(),\n","    to_image(images * sa[..., tf.newaxis])\n","]\n","titles = ['original',\n","          'absolute gradients',\n","          'positive gradients',\n","          'negative gradients',\n","          'original * abs(gradients)']\n","\n","utils.plot(sum(zip(*columns), ()),\n","           titles=titles,\n","           rows=len(images))"],"execution_count":30,"outputs":[{"output_type":"display_data","data":{"text/plain":"Output hidden; open in https://colab.research.google.com to view."},"metadata":{}}]}]}