Focal loss optimisation

segmentation_models_pytorch/losses/focal.py

@@ -70,20 +70,21 @@ def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:
 
         elif self.mode == MULTICLASS_MODE:
             num_classes = y_pred.size(1)
-            loss = 0
 
-            # Filter anchors with -1 label from loss computation
+            # If ignore_index parameter is passed, treat it as an extra class during one-hot encoding and remove it later
+            # One-hot encoding the labels allows us to vectorise the focal loss computation
             if self.ignore_index is not None:
-                not_ignored = y_true != self.ignore_index
+                y_true[y_true == self.ignore_index] = num_classes
+                y_true_one_hot = torch.nn.functional.one_hot(y_true,num_classes = num_classes + 1)
+                y_true_one_hot = y_true_one_hot[ ... , : -1]
 
-            for cls in range(num_classes):
-                cls_y_true = (y_true == cls).long()
-                cls_y_pred = y_pred[:, cls, ...]
+            else:     
+                y_true_one_hot = torch.nn.functional.one_hot(y_true,num_classes = num_classes)
 
-                if self.ignore_index is not None:
-                    cls_y_true = cls_y_true[not_ignored]
-                    cls_y_pred = cls_y_pred[not_ignored]
+            y_true_one_hot = torch.permute(y_true_one_hot,(0,3,1,2))
 
-                loss += self.focal_loss_fn(cls_y_pred, cls_y_true)
+            # Multiplying the loss by num_classes in order to stay consistent with the earlier loss computation which did not
+            # take a classwise average of the loss
+            loss = num_classes * self.focal_loss_fn(y_pred, y_true_one_hot)
 
         return loss

segmentation_models_pytorch/losses/focal_loss_optimisation_benchmarking.ipynb

@@ -0,0 +1,195 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "245a88c9",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from segmentation_models_pytorch.losses import BINARY_MODE, MULTICLASS_MODE, MULTILABEL_MODE\n",
+    "from time import time\n",
+    "from typing import Optional\n",
+    "import torch\n",
+    "import segmentation_models_pytorch\n",
+    "\n",
+    "class FocalLossVectorised(segmentation_models_pytorch.losses.FocalLoss):\n",
+    "    def __init__(\n",
+    "        self,\n",
+    "        mode: str,\n",
+    "        alpha: Optional[float] = None,\n",
+    "        gamma: Optional[float] = 2.0,\n",
+    "        ignore_index: Optional[int] = None,\n",
+    "        reduction: Optional[str] = \"mean\",\n",
+    "        normalized: bool = False,\n",
+    "        reduced_threshold: Optional[float] = None,\n",
+    "    ):\n",
+    "        \n",
+    "        super().__init__(mode = mode,alpha = alpha,gamma = gamma, ignore_index = ignore_index,reduction = reduction,\n",
+    "                         normalized = normalized,reduced_threshold = reduced_threshold)\n",
+    "    \n",
+    "    def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:\n",
+    "        if self.mode in {BINARY_MODE, MULTILABEL_MODE}:\n",
+    "            y_true = y_true.view(-1)\n",
+    "            y_pred = y_pred.view(-1)\n",
+    "\n",
+    "            if self.ignore_index is not None:\n",
+    "                # Filter predictions with ignore label from loss computation\n",
+    "                not_ignored = y_true != self.ignore_index\n",
+    "                y_pred = y_pred[not_ignored]\n",
+    "                y_true = y_true[not_ignored]\n",
+    "\n",
+    "            loss = self.focal_loss_fn(y_pred, y_true)\n",
+    "\n",
+    "        elif self.mode == MULTICLASS_MODE:\n",
+    "            num_classes = y_pred.size(1)\n",
+    "\n",
+    "            if self.ignore_index is not None:\n",
+    "                y_true[y_true == self.ignore_index] = num_classes\n",
+    "                y_true_one_hot = torch.nn.functional.one_hot(y_true,num_classes = num_classes + 1)\n",
+    "                y_true_one_hot = y_true_one_hot[ : , : , : , : -1]\n",
+    "\n",
+    "            else:     \n",
+    "                y_true_one_hot = torch.nn.functional.one_hot(y_true,num_classes = num_classes)\n",
+    "\n",
+    "            y_true_one_hot = torch.permute(y_true_one_hot,(0,3,1,2))\n",
+    "            loss = num_classes * self.focal_loss_fn(y_pred, y_true_one_hot)\n",
+    "\n",
+    "        return loss"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "9c64c3ea",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "num_classes = 20\n",
+    "batch_size = 128\n",
+    "resolution = 512\n",
+    "device = 'cuda:1' if torch.cuda.is_available() else 'cpu'"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "id": "d4d3a5f5",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "vectorised_loss_fn = FocalLossVectorised(mode = 'multiclass',ignore_index = num_classes)\n",
+    "loss_fn = segmentation_models_pytorch.losses.FocalLoss(mode = 'multiclass',ignore_index = num_classes)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "id": "5e49a5b8",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "predictions = torch.randn((batch_size,num_classes,resolution,resolution)).to(device = device)\n",
+    "labels = torch.randint(low = 0,high = num_classes+1,size = (batch_size,resolution,resolution)).to(device = device)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "id": "36a0f89b",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def benchmark(function,predictions,labels,benchmark_iterations = 100):\n",
+    "    start_time = time()\n",
+    "\n",
+    "    for _ in range(benchmark_iterations):\n",
+    "        loss = function(predictions,labels)\n",
+    "\n",
+    "    end_time = time()\n",
+    "\n",
+    "    average_time_taken = (end_time - start_time) / (benchmark_iterations)\n",
+    "\n",
+    "    print(f\"Average time taken by function {function} is {average_time_taken} seconds\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "id": "8de20667",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Average time taken by function FocalLoss() is 0.3390256547927856 seconds\n"
+     ]
+    }
+   ],
+   "source": [
+    "benchmark(loss_fn,predictions,labels)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "id": "6da16fc2",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Average time taken by function FocalLossVectorised() is 0.11771584510803222 seconds\n"
+     ]
+    }
+   ],
+   "source": [
+    "benchmark(vectorised_loss_fn,predictions,labels)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "a1083cd4",
+   "metadata": {},
+   "source": [
+    "##### CHECKING THAT OUTPUT OF NEW CLASS IS CONSISTENT WITH THE OLD ONE"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "id": "8fba3182",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "output_from_vectorised_fn = vectorised_loss_fn(predictions,labels)\n",
+    "output_from_old_fn = loss_fn(predictions,labels)\n",
+    "\n",
+    "assert torch.allclose(output_from_vectorised_fn,output_from_old_fn)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "torch",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.9.19"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}