Skip to content

Commit

Permalink
Tidy up manual eval
Browse files Browse the repository at this point in the history
  • Loading branch information
@SylviaWhittle
SylviaWhittle committed Jun 26, 2024
1 parent 28e1d7d commit bd7f655
Showing 1 changed file with 23 additions and 121 deletions.
144 changes: 23 additions & 121 deletions src/manual_eval.ipynb
View file
Original file line number Diff line number Diff line change
Expand Up @@ -53,7 +53,6 @@
"test_size = int(0.2 * len(image_indexes))\n",
"print(f\"Test size: {test_size}\")\n",
"\n",
"# Randomly select test_size number of indexes\n",
"test_indexes = np.random.choice(image_indexes, test_size, replace=False)\n",
"print(f\"Test indexes: {test_indexes}\")"
]
Expand All @@ -76,18 +75,15 @@
" # false_neg = tf.reduce_sum(y_true) - true_pos\n",
" # return 1 - (true_pos + smooth) / (true_pos + false_pos + false_neg + smooth)\n",
"\n",
" intersection = tf.reduce_sum(y_true * y_pred)\n",
" print(f\"intersection: {intersection}\")\n",
" union = tf.reduce_sum(y_true) + tf.reduce_sum(y_pred) - intersection\n",
" print(f\"union: {union}\")\n",
" # return 1 - ((intersection + smooth) / (union + smooth))\n",
" intersection_plus_smooth = intersection + smooth\n",
" print(f\"intersection_plus_smooth: {intersection_plus_smooth}\")\n",
" union_plus_smooth = union + smooth\n",
" print(f\"union_plus_smooth: {union_plus_smooth}\")\n",
" intersection_over_union = intersection_plus_smooth / union_plus_smooth\n",
" print(f\"intersection_over_union: {intersection_over_union}\")\n",
" return 1 - intersection_over_union\n",
" assert y_true.shape == y_pred.shape\n",
" assert y_true.dtype == bool\n",
"\n",
" # convert to float\n",
" y_true_float = tf.cast(y_true, tf.float32)\n",
"\n",
" intersection = tf.reduce_sum(y_true_float * y_pred)\n",
" union = tf.reduce_sum(y_true_float) + tf.reduce_sum(y_pred) - intersection\n",
" return 1 - (intersection + smooth) / (union + smooth)\n",
"\n",
"\n",
"def dice_loss_simple(y_true, y_pred, smooth=1e-7):\n",
Expand All @@ -102,92 +98,15 @@
" # false_neg = tf.reduce_sum(y_true) - true_pos\n",
" # return 1 - (2 * true_pos + smooth) / (2 * true_pos + false_pos + false_neg + smooth)\n",
"\n",
" intersection = tf.reduce_sum(y_true * y_pred)\n",
" union_double_count = tf.reduce_sum(y_true) + tf.reduce_sum(y_pred)\n",
" return 1 - (2 * intersection + smooth) / (union_double_count + smooth)\n",
"\n",
"\n",
"# test it\n",
" assert y_true.shape == y_pred.shape\n",
" assert y_true.dtype == bool\n",
"\n",
"ground_truth = np.zeros((10, 10))\n",
"# ground_truth[2:8, 2:8] = 1\n",
"ground_truth[5, 5:7] = 1\n",
"plt.imshow(ground_truth, cmap=\"gray\")\n",
"plt.show()\n",
" # convert to float\n",
" y_true_float = tf.cast(y_true, tf.float32)\n",
"\n",
"prediction = np.zeros((10, 10))\n",
"# prediction[3:7, 3:7] = 1\n",
"prediction[5, 5] = 1\n",
"prediction[5, 6] = 0.8\n",
"prediction[5, 4] = 0.8\n",
"prediction[5, 3] = 0.8\n",
"plt.imshow(prediction, cmap=\"gray\")\n",
"plt.show()\n",
"\n",
"print(\n",
" f\"iou loss: {iou_loss_simple(ground_truth, prediction):>5.2f} iou score: {(1 - iou_loss_simple(ground_truth, prediction)):>5.2f}\"\n",
")\n",
"print(\n",
" f\"dice loss: {dice_loss_simple(ground_truth, prediction):>5.2f} dice score: {(1 - dice_loss_simple(ground_truth, prediction)):>5.2f}\"\n",
")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"def iou_loss(y_true, y_pred, smooth=1e-5):\n",
" \"\"\"Intersection over Union loss function.\n",
"\n",
" Parameters\n",
" ----------\n",
" y_true : tf.Tensor\n",
" True values.\n",
" y_pred : tf.Tensor\n",
" Predicted values.\n",
" smooth : float\n",
" Smoothing factor to prevent division by zero.\n",
"\n",
" Returns\n",
" -------\n",
" iou : tf.Tensor\n",
" The IoU loss.\n",
" \"\"\"\n",
" # Ensure the tensors are of the same shape\n",
" y_true = tf.squeeze(y_true, axis=-1) if y_true.shape[-1] == 1 else y_true\n",
" y_pred = tf.squeeze(y_pred, axis=-1) if y_pred.shape[-1] == 1 else y_pred\n",
" y_true = tf.cast(y_true, tf.float32)\n",
" y_pred = tf.cast(y_pred, tf.float32)\n",
"\n",
" intersection = tf.reduce_sum(y_true * y_pred, axis=(1, 2))\n",
" sum_of_squares_pred = tf.reduce_sum(tf.square(y_pred), axis=(1, 2))\n",
" sum_of_squares_true = tf.reduce_sum(tf.square(y_true), axis=(1, 2))\n",
" # print(f\"intersection: {intersection} union: {sum_of_squares_pred + sum_of_squares_true - intersection}\")\n",
" iou = 1 - (intersection + smooth) / (sum_of_squares_pred + sum_of_squares_true - intersection + smooth)\n",
" # return the mean over all samples in the batch, as a single value\n",
" return float(tf.reduce_mean(iou))\n",
"\n",
"\n",
"# test it\n",
"\n",
"ground_truth = np.zeros((10, 10))\n",
"# ground_truth[2:8, 2:8] = 1\n",
"ground_truth[5, 5:7] = 1\n",
"plt.imshow(ground_truth, cmap=\"gray\")\n",
"plt.show()\n",
"\n",
"prediction = np.zeros((10, 10))\n",
"# prediction[3:7, 3:7] = 1\n",
"prediction[5, 5] = 1\n",
"plt.imshow(prediction, cmap=\"gray\")\n",
"plt.show()\n",
"\n",
"ground_truth = np.expand_dims(ground_truth, axis=0)\n",
"prediction = np.expand_dims(prediction, axis=0)\n",
"\n",
"print(iou_loss(ground_truth, prediction))"
" intersection = tf.reduce_sum(y_true_float * y_pred)\n",
" union_double_count = tf.reduce_sum(y_true_float) + tf.reduce_sum(y_pred)\n",
" return 1 - (2 * intersection + smooth) / (union_double_count + smooth)"
]
},
{
Expand All @@ -212,7 +131,7 @@
" image = np.array(pil_image).astype(np.float32)\n",
" pil_mask = Image.fromarray(mask)\n",
" pil_mask = pil_mask.resize((512, 512))\n",
" mask = np.array(pil_mask).astype(bool).astype(np.float32)\n",
" mask = np.array(pil_mask).astype(bool)\n",
"\n",
" # Add batch dimension\n",
" image = np.expand_dims(image, axis=0)\n",
Expand All @@ -231,26 +150,16 @@
" predicted_mask = np.squeeze(predicted_mask, axis=-1)\n",
" mask = np.squeeze(mask, axis=-1)\n",
"\n",
" print(predicted_mask.shape)\n",
" print(mask.shape)\n",
"\n",
" plt.imshow(predicted_mask, cmap=\"gray\")\n",
" plt.show()\n",
" plt.imshow(mask, cmap=\"gray\")\n",
" plt.show()\n",
"\n",
" # Calculate the dice loss\n",
" dice_loss_value = dice_loss_simple(predicted_mask.astype(np.float64), mask.astype(np.float64))\n",
" # print(f\"dice loss: {dice_loss_value}\")\n",
" dice_loss_value = dice_loss_simple(mask, predicted_mask)\n",
" dice_totals += dice_loss_value\n",
"\n",
" # Calculate the iou loss\n",
" iou_loss_value = iou_loss_simple(predicted_mask, mask)\n",
" print(f\"iou loss: {iou_loss_value}\")\n",
" iou_loss_value = iou_loss_simple(mask, predicted_mask)\n",
" iou_totals += iou_loss_value\n",
"\n",
" # Calculate the binary crossentropy loss\n",
" binary_crossentropy_loss_value = tf.reduce_mean(\n",
" binary_crossentropy_loss_value = 1 - tf.reduce_mean(\n",
" tf.keras.losses.binary_crossentropy(\n",
" mask.astype(np.float32), predicted_mask.astype(np.float32), from_logits=True\n",
" )\n",
Expand All @@ -261,7 +170,6 @@
" f\"Index {test_index} Losses: dice: {dice_loss_value}, iou: {iou_loss_value}, binary_crossentropy: {binary_crossentropy_loss_value}\"\n",
" )\n",
"\n",
" break\n",
"\n",
"# Calculate the average losses\n",
"average_dice_loss = dice_totals / test_size\n",
Expand All @@ -270,17 +178,11 @@
"\n",
"print(\n",
" f\"Average Losses: dice: {average_dice_loss}, iou: {average_iou_loss}, binary_crossentropy: {average_binary_crossentropy_loss}\"\n",
")\n",
"print(\n",
" f\"Average Scores: dice: {1 - average_dice_loss}, iou: {1 - average_iou_loss}, binary_crossentropy: {1 - average_binary_crossentropy_loss}\"\n",
")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"model.summary()"
]
}
],
"metadata": {
Expand Down

0 comments on commit bd7f655

Please sign in to comment.