fixed bugs hw3-2 base

Author: Татьяна Гайнцева

homework03/homework03_part2_autoencoders_basic.ipynb

@@ -276,8 +276,10 @@
     "    model.to(device)\n",
     "    gd = optim.Adamax(model.parameters(), lr=0.002)\n",
     "    dataloader = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=True)\n",
-    "    losses = []\n",
+    "    \n",
     "    for epoch in range(num_epoch):\n",
+    "        losses = []\n",
+    "        \n",
     "        for i, (batch) in enumerate(dataloader):\n",
     "            gd.zero_grad()\n",
     "            loss = model.batch_loss(batch.cuda())\n",
@@ -613,7 +615,6 @@
    },
    "outputs": [],
    "source": [
-    "\n",
     "theoretical_std = (X[:100].std()**2 + 0.5**2)**.5\n",
     "our_std = apply_gaussian_noise(X[:100],sigma=0.5).std()\n",
     "assert abs(theoretical_std - our_std) < 0.01, \"Standard deviation does not match it's required value. Make sure you use sigma as std.\"\n",
@@ -655,35 +656,15 @@
    "outputs": [],
    "source": [
     "def train_noise(model, dataset, num_epoch=50):\n",
-    "    #<Your code: define train function for denoising autoencoder as train function above>"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "colab": {},
-    "colab_type": "code",
-    "id": "MKQK4HJL_eX2"
-   },
-   "outputs": [],
-   "source": [
-    "X_train_noise = apply_gaussian_noise(X_train)\n",
-    "X_test_noise = apply_gaussian_noise(X_test)"
+    "    #<Your code: define train function for denoising autoencoder as train function above>\n",
+    "    #<Think carefully, what should be ground-truth image for computing loss function>"
    ]
   },
   {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "colab": {},
-    "colab_type": "code",
-    "id": "8ClYTlpa_eX4"
-   },
-   "outputs": [],
+   "cell_type": "markdown",
+   "metadata": {},
    "source": [
-    "X_train_tensor_n = torch.from_numpy(X_train_noise).type(torch.DoubleTensor)\n",
-    "X_test_tensor_n = torch.Tensor(X_test_noise).type(torch.DoubleTensor)"
+    "__Note:__ You may change the way the training with noise is done, if you want. For example, you may change Dataloader or batch_loss function in model and leave train function unchanged."
    ]
   },
   {
@@ -701,7 +682,7 @@
    "outputs": [],
    "source": [
     "aenc = pca_autoencoder()\n",
-    "train(aenc, X_train_tensor_n, 50)"
+    "train_noise(aenc, X_train_tensor, 50)"
    ]
   },
   {
@@ -723,7 +704,7 @@
     "id": "HVj0NPXV3liL"
    },
    "source": [
-    "** Let's evaluate!!! **"
+    "**Let's evaluate!!!**"
    ]
   },
   {
@@ -842,7 +823,7 @@
     "def get_similar(image, n_neighbors=5):\n",
     "    assert len(image.shape)==3,\"image must be [batch,height,width,3]\"\n",
     "\n",
-    "    code = aenc.enc(image.cuda(device)).detach().cpu().numpy()\n",
+    "    code = #<Your code: encode image into latent code>\n",
     "    \n",
     "    (distances,),(idx,) = #<Your code: using lshf.kneighbors find nearest neighbors>\n",
     "    \n",
@@ -1023,7 +1004,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.6.2"
+   "version": "3.6.9"
   }
  },
  "nbformat": 4,