removed answers :(

notebook/labs/03 - SVM vs LogReg.ipynb

@@ -7,6 +7,8 @@
     "# CSCI-UA 0473 - Introduction to Machine Learning\n",
     "## Wednesday, February 07, 2018\n",
     "\n",
+    "**Please download\n",
+    "\n",
     "This lab will discuss the difference between SVM and Logistic Regression, and go through the implementation of both using naive `numpy` (not really, because we are using `autograd` to skip manual computation of the gradients).\n",
     "\n",
     "We will also go through the usage of `autograd` and `scipy.optimize.minimize()` in greater details, and take a look at numerical stability issues.\n",
@@ -324,20 +326,17 @@
     "    # (1)\n",
     "    # Get the bias term from the last element of @params, and get the weight vector from the rest.\n",
     "    # YOUR CODE HERE\n",
-    "    w, b = params[:-1], params[-1]\n",
     "    \n",
     "    # (2)\n",
     "    # Compute the score.  Do not forget the bias term!\n",
     "    # Your score should be a 1D array with num_samples elements for a 2D @x, or a scalar for a 1D @x.\n",
     "    # YOUR CODE HERE\n",
-    "    s = x @ w + b\n",
     "    \n",
     "    # (3) If @pre is True, return the score itself.  Otherwise, return the sign of the score.\n",
     "    # YOUR CODE HERE\n",
-    "    return s if pre else numpy.sign(s)\n",
     "    \n",
     "    # (4) Delete the \"pass\" statement if you have completed everything above.\n",
-    "    #pass\n",
+    "    pass\n",
     "\n",
     "# Empirical cost function.\n",
     "# y: 1D array or a scalar containing the ground truth label(s).  Each element is either +1 or -1.\n",
@@ -348,27 +347,24 @@
     "    # (1)\n",
     "    # Call svm() to get the pre-sign score.\n",
     "    # YOUR CODE HERE\n",
-    "    s = svm(x, params, True)\n",
     "    \n",
     "    # (2)\n",
     "    # Compute hinge loss, which is the \"max\" term within the summation, for each example.\n",
     "    # YOUR CODE HERE\n",
-    "    l = numpy.maximum(0, 1 - y * s)\n",
     "    \n",
     "    # (3)\n",
     "    # If avg is False, return the per-example hinge loss.  Otherwise, return the empirical cost function.\n",
     "    # Recall that the cost function is the sum of an average and a norm.\n",
     "    # YOUR CODE HERE\n",
-    "    return l if not avg else numpy.mean(l) + lambda_ * numpy.sum(params ** 2)\n",
     "    \n",
     "    # (4) Delete the \"pass\" statement if you have completed everything above.\n",
-    "    #pass\n",
+    "    pass\n",
     "\n",
     "def svm_cost(w, x, y, lambda_=1):\n",
     "    return svm_dist(y, x, w, avg=True, lambda_=lambda_)\n",
     "\n",
     "# Compute the gradient.\n",
-    "svm_rule = grad(svm_cost, 0)   # YOUR CODE HERE: replace None with the correct function call."
+    "svm_rule = None   # YOUR CODE HERE: replace None with the correct function call."
    ]
   },
   {
@@ -479,7 +475,7 @@
     "params = 1e-4 * numpy.random.randn(n_dim+1); params[-1] = 0.\n",
     "\n",
     "# (1) Call scipy's minimize method on SVM.\n",
-    "svm_model = scipy.optimize.minimize(svm_cost, params, (x_train, y_train_svm), jac=svm_rule) # YOUR CODE HERE: replace None with the function call.\n",
+    "svm_model = None # YOUR CODE HERE: replace None with the function call.\n",
     "\n",
     "params_svm = svm_model.x\n",
     "\n",