Function get_train_accuracy

cnn_class2/siamese.py:291–352

(threshold=0.85)

# calculate accuracy before training
# since the dataset is imbalanced, we'll report tp, tn, fp, fn
def get_train_accuracy(threshold=0.85):
  positive_distances = []
  negative_distances = []

  tp = 0
  tn = 0
  fp = 0
  fn = 0

  batch_size = 64
  x_batch_1 = np.zeros([batch_size] + list(img.shape))
  x_batch_2 = np.zeros([batch_size] + list(img.shape))
  n_batches = int(np.ceil(len(train_positives) / batch_size))
  for i in range(n_batches):
    print(f"pos batch: {i+1}/{n_batches}")
    pos_batch_indices = train_positives[i * batch_size: (i + 1) * batch_size]

    # fill up x_batch and y_batch
    j = 0
    for idx1, idx2 in pos_batch_indices:
      x_batch_1[j] = train_images[idx1]
      x_batch_2[j] = train_images[idx2]
      j += 1

    x1 = x_batch_1[:j]
    x2 = x_batch_2[:j]
    distances = model.predict([x1, x2]).flatten()
    positive_distances += distances.tolist()

    # update tp, tn, fp, fn
    tp += (distances < threshold).sum()
    fn += (distances > threshold).sum()

  n_batches = int(np.ceil(len(train_negatives) / batch_size))
  for i in range(n_batches):
    print(f"neg batch: {i+1}/{n_batches}")
    neg_batch_indices = train_negatives[i * batch_size: (i + 1) * batch_size]

    # fill up x_batch and y_batch
    j = 0
    for idx1, idx2 in neg_batch_indices:
      x_batch_1[j] = train_images[idx1]
      x_batch_2[j] = train_images[idx2]
      j += 1

    x1 = x_batch_1[:j]
    x2 = x_batch_2[:j]
    distances = model.predict([x1, x2]).flatten()
    negative_distances += distances.tolist()

    # update tp, tn, fp, fn
    fp += (distances < threshold).sum()
    tn += (distances > threshold).sum()

  tpr = tp / (tp + fn)
  tnr = tn / (tn + fp)
  print(f"sensitivity (tpr): {tpr}, specificity (tnr): {tnr}")