final version

Author: raynardj

Diff for: 3.6_convolution_and_image_problem.ipynb

@@ -578,21 +578,6 @@
     "conv_layers = conv_model.features"
    ]
   },
-  {
-   "cell_type": "code",
-   "execution_count": 24,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "\n",
-    "# dense_conv1 = nn.Sequential(*[getattr(conv_model.features,nn_name) for nn_name in [\"conv0\",\"norm0\",\"relu0\",\"pool0\",\"denseblock1\",\"transition1\",\n",
-    "#                                                                                    \"denseblock2\",\"transition2\",\"denseblock3\",\"transition3\",]])\n",
-    "\n",
-    "# dense_conv2 = nn.Sequential(*[getattr(conv_model.features,nn_name) for nn_name in [\"denseblock4\",\"norm5\"]])\n",
-    "\n",
-    "# conv_model"
-   ]
-  },
   {
    "cell_type": "code",
    "execution_count": 25,
@@ -645,6 +630,11 @@
    "outputs": [],
    "source": [
     "def action(*args,**kwargs):\n",
+    "    \"\"\"\n",
+    "    single training step, \n",
+    "    take in data, spit out loss/ metric\n",
+    "    and \n",
+    "    \"\"\"\n",
     "    x,y = args[0]\n",
     "    y = torch.LongTensor(np.array(y).astype(int))\n",
     "    if CUDA:\n",
@@ -710,7 +700,8 @@
      "name": "stderr",
      "output_type": "stream",
      "text": [
-      "⭐[ep_0_i_1914]\tacc\t0.556✨\tloss\t1.793:  64%|██████▍   | 1916/3001 [30:49<17:27,  1.04it/s]"
+      "⭐[ep_0_i_2999]\tacc\t0.506✨\tloss\t1.678: 100%|██████████| 3001/3001 [48:23<00:00,  1.03it/s]\n",
+      "😎[val_ep_0_i_33]\tacc\t0.548😂\tloss\t1.832:  22%|██▏       | 34/156 [00:21<01:16,  1.60it/s]"
      ]
     }
    ],
@@ -728,6 +719,45 @@
     "save_model(top_half_,\"food_top.0.0.1.npy\")"
    ]
   },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Excercise"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Please work on at least 2 of the following challenge\n",
+    "\n",
+    "1. Optimize all the layers instead of only linear classifier\n",
+    "2. Optimize the last convblock(the conv block close the linear layer) and linear classifier\n",
+    "3. Try other image classify datasets, like [monekey image set](https://www.kaggle.com/slothkong/10-monkey-species) or [flower classifying problem](https://www.kaggle.com/alxmamaev/flowers-recognition) or [blood cell images](https://www.kaggle.com/paultimothymooney/blood-cells). You'll pretty soon find out convolutional neural network is a universal tool for this kind of problem\n",
+    "4. Try keras to work out a better accuaracy, keras import pretrained model by ```from keras.application import ... ```"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "how we break down a pytorch model to several pytorch models:\n",
+    "\n",
+    "```python\n",
+    "dense_conv1 = nn.Sequential(*[getattr(conv_model.features,nn_name) for nn_name in [\"conv0\",\"norm0\",\"relu0\",\"pool0\",\"denseblock1\",\"transition1\",\n",
+    "                                                                                   \"denseblock2\",\"transition2\",\"denseblock3\",\"transition3\",]])\n",
+    "\n",
+    "dense_conv2 = nn.Sequential(*[getattr(conv_model.features,nn_name) for nn_name in [\"denseblock4\",\"norm5\"]])\n",
+    "```\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": []
+  },
   {
    "cell_type": "code",
    "execution_count": null,