Replace softmax with stable version + small corrections

lib/ex_vision/classification/efficientnet_v2_l.ex

@@ -15,8 +15,8 @@ defmodule ExVision.Classification.EfficientNet_V2_L do
     image
     |> ExVision.Utils.resize({480, 480})
     |> NxImage.normalize(
-      Nx.tensor([0.5, 0.5, 0.5]),
-      Nx.tensor([0.5, 0.5, 0.5]),
+      Nx.f32([0.5, 0.5, 0.5]),
+      Nx.f32([0.5, 0.5, 0.5]),
       channels: :first
     )
   end

lib/ex_vision/classification/efficientnet_v2_m.ex

@@ -15,8 +15,8 @@ defmodule ExVision.Classification.EfficientNet_V2_M do
     image
     |> ExVision.Utils.resize({480, 480})
     |> NxImage.normalize(
-      Nx.tensor([0.485, 0.456, 0.406]),
-      Nx.tensor([0.229, 0.224, 0.225]),
+      Nx.f32([0.485, 0.456, 0.406]),
+      Nx.f32([0.229, 0.224, 0.225]),
       channels: :first
     )
   end

lib/ex_vision/classification/efficientnet_v2_s.ex

@@ -15,8 +15,8 @@ defmodule ExVision.Classification.EfficientNet_V2_S do
     image
     |> ExVision.Utils.resize({384, 384})
     |> NxImage.normalize(
-      Nx.tensor([0.485, 0.456, 0.406]),
-      Nx.tensor([0.229, 0.224, 0.225]),
+      Nx.f32([0.485, 0.456, 0.406]),
+      Nx.f32([0.229, 0.224, 0.225]),
       channels: :first
     )
   end

lib/ex_vision/classification/generic_classifier.ex

@@ -4,8 +4,6 @@ defmodule ExVision.Classification.GenericClassifier do
   # Contains a default implementation of post processing for TorchVision classifiers
   # To use: `use ExVision.Classification.GenericClassifier`
 
-  alias ExVision.Utils
-
   alias ExVision.Types.ImageMetadata
 
   @typep output_t() :: %{atom() => number()}
@@ -15,7 +13,7 @@ defmodule ExVision.Classification.GenericClassifier do
     scores
     |> Nx.backend_transfer()
     |> Nx.flatten()
-    |> Utils.softmax()
+    |> Axon.Activations.softmax(axis: [0])
     |> Nx.to_flat_list()
     |> then(&Enum.zip(categories, &1))
     |> Map.new()

lib/ex_vision/classification/mobilenet_v3_small.ex

@@ -15,8 +15,8 @@ defmodule ExVision.Classification.MobileNetV3Small do
     image
     |> ExVision.Utils.resize({224, 224})
     |> NxImage.normalize(
-      Nx.tensor([0.485, 0.456, 0.406]),
-      Nx.tensor([0.229, 0.224, 0.225]),
+      Nx.f32([0.485, 0.456, 0.406]),
+      Nx.f32([0.229, 0.224, 0.225]),
       channels: :first
     )
   end

lib/ex_vision/classification/squeezenet1_1.ex

@@ -15,8 +15,8 @@ defmodule ExVision.Classification.SqueezeNet1_1 do
     image
     |> ExVision.Utils.resize({224, 224})
     |> NxImage.normalize(
-      Nx.tensor([0.485, 0.456, 0.406]),
-      Nx.tensor([0.229, 0.224, 0.225]),
+      Nx.f32([0.485, 0.456, 0.406]),
+      Nx.f32([0.229, 0.224, 0.225]),
       channels: :first
     )
   end

lib/ex_vision/instance_segmentation/maskrcnn_resnet50_fpn_v2.ex

@@ -6,6 +6,8 @@ defmodule ExVision.InstanceSegmentation.MaskRCNN_ResNet50_FPN_V2 do
     model: "maskrcnn_resnet50_fpn_v2_instance_segmentation.onnx",
     categories: "priv/categories/coco_categories.json"
 
+  import ExVision.Utils
+
   require Logger
 
   alias ExVision.Types.BBoxWithMask
@@ -46,16 +48,10 @@ defmodule ExVision.InstanceSegmentation.MaskRCNN_ResNet50_FPN_V2 do
     scale_x = w / 224
     scale_y = h / 224
 
-    bboxes =
-      bboxes
-      |> Nx.squeeze(axes: [0])
-      |> Nx.multiply(Nx.tensor([scale_x, scale_y, scale_x, scale_y]))
-      |> Nx.round()
-      |> Nx.as_type(:s64)
-      |> Nx.to_list()
+    bboxes = process_bbox(bboxes, Nx.tensor([scale_x, scale_y, scale_x, scale_y]))
 
-    scores = scores |> Nx.squeeze(axes: [0]) |> Nx.to_list()
-    labels = labels |> Nx.squeeze(axes: [0]) |> Nx.to_list()
+    scores = unbatch(scores)
+    labels = unbatch(labels)
 
     masks =
       masks

lib/ex_vision/keypoint_detection/keypointrcnn_resnet50_fpn.ex

@@ -6,6 +6,8 @@ defmodule ExVision.KeypointDetection.KeypointRCNN_ResNet50_FPN do
     model: "keypointrcnn_resnet50_fpn_keypoint_detector.onnx",
     categories: "priv/categories/no_person_or_person.json"
 
+  import ExVision.Utils
+
   require Logger
 
   alias ExVision.Types.BBoxWithKeypoints
@@ -67,26 +69,14 @@ defmodule ExVision.KeypointDetection.KeypointRCNN_ResNet50_FPN do
     scale_x = w / 224
     scale_y = h / 224
 
-    bboxes =
-      bboxes
-      |> Nx.squeeze(axes: [0])
-      |> Nx.multiply(Nx.tensor([scale_x, scale_y, scale_x, scale_y]))
-      |> Nx.round()
-      |> Nx.as_type(:s64)
-      |> Nx.to_list()
-
-    scores = scores |> Nx.squeeze(axes: [0]) |> Nx.to_list()
-    labels = labels |> Nx.squeeze(axes: [0]) |> Nx.to_list()
-
-    keypoints_list =
-      keypoints_list
-      |> Nx.squeeze(axes: [0])
-      |> Nx.multiply(Nx.tensor([scale_x, scale_y, 1]))
-      |> Nx.round()
-      |> Nx.as_type(:s64)
-      |> Nx.to_list()
-
-    keypoints_scores_list = keypoints_scores_list |> Nx.squeeze(axes: [0]) |> Nx.to_list()
+    bboxes = process_bbox(bboxes, Nx.tensor([scale_x, scale_y, scale_x, scale_y]))
+
+    scores = unbatch(scores)
+    labels = unbatch(labels)
+
+    keypoints_list = process_bbox(keypoints_list, Nx.tensor([scale_x, scale_y, 1]))
+
+    keypoints_scores_list = unbatch(keypoints_scores_list)
 
     [bboxes, scores, labels, keypoints_list, keypoints_scores_list]
     |> Enum.zip()

lib/ex_vision/object_detection/generic_detector.ex

@@ -4,6 +4,8 @@ defmodule ExVision.ObjectDetection.GenericDetector do
   # Contains a default implementation of pre and post processing for TorchVision detectors
   # To use: `use ExVision.ObjectDetection.GenericDetector`
 
+  import ExVision.Utils
+
   require Logger
 
   alias ExVision.Types.{BBox, ImageMetadata}
@@ -29,16 +31,10 @@ defmodule ExVision.ObjectDetection.GenericDetector do
     scale_x = w / 224
     scale_y = h / 224
 
-    bboxes =
-      bboxes
-      |> Nx.squeeze(axes: [0])
-      |> Nx.multiply(Nx.tensor([scale_x, scale_y, scale_x, scale_y]))
-      |> Nx.round()
-      |> Nx.as_type(:s64)
-      |> Nx.to_list()
+    bboxes = process_bbox(bboxes, Nx.tensor([scale_x, scale_y, scale_x, scale_y]))
 
-    scores = scores |> Nx.squeeze(axes: [0]) |> Nx.to_list()
-    labels = labels |> Nx.squeeze(axes: [0]) |> Nx.to_list()
+    scores = unbatch(scores)
+    labels = unbatch(labels)
 
     [bboxes, scores, labels]
     |> Enum.zip()

lib/ex_vision/utils.ex

@@ -1,7 +1,6 @@
 defmodule ExVision.Utils do
   @moduledoc false
 
-  import Nx.Defn
   require Nx
   require Image
   alias ExVision.Types
@@ -86,8 +85,7 @@ defmodule ExVision.Utils do
 
   defp ensure_grad_3(tensor) do
     tensor
-    |> Nx.shape()
-    |> tuple_size()
+    |> Nx.rank()
     |> case do
       3 -> [tensor]
       4 -> tensor |> Nx.to_batched(1) |> Stream.map(&Nx.squeeze(&1, axes: [0])) |> Enum.to_list()
@@ -149,10 +147,6 @@ defmodule ExVision.Utils do
     Enum.map(outputs, fn {name, _type, _shape} -> name end)
   end
 
-  defn softmax(x) do
-    Nx.divide(Nx.exp(x), Nx.sum(Nx.exp(x)))
-  end
-
   @spec batched_run(atom(), ExVision.Model.input_t()) :: ExVision.Model.output_t()
   def batched_run(process_name, input) when is_list(input) do
     Nx.Serving.batched_run(process_name, input)
@@ -161,4 +155,19 @@ defmodule ExVision.Utils do
   def batched_run(process_name, input) do
     process_name |> batched_run([input]) |> hd()
   end
+
+  @spec process_bbox(Nx.Tensor.t(), Nx.Tensor.t()) :: [integer()]
+  def process_bbox(bbox, scales) do
+    bbox
+    |> Nx.squeeze(axes: [0])
+    |> Nx.multiply(scales)
+    |> Nx.round()
+    |> Nx.as_type(:s64)
+    |> Nx.to_list()
+  end
+
+  @spec unbatch(Nx.Tensor.t()) :: [number()]
+  def unbatch(batched_value) do
+    batched_value |> Nx.squeeze(axes: [0]) |> Nx.to_list()
+  end
 end