init

Author: Peter Featherstone

.vscode/settings.json

@@ -0,0 +1,27 @@
+{
+    "launch": {
+        "version": "0.2.0",
+        "compounds": [],
+        "configurations": [
+            {
+                "name": "TRAIN",
+                "type": "python",
+                "request": "launch",
+                "cwd": "${workspaceFolder}/src",
+                "program": "train_coco.py",
+                "args": ["--batchsize", "16",
+                         "--trainRoot", "/data/coco/images/train2017/",
+                         "--trainAnn",  "/data/coco/annotations/instances_train2017.json",
+                         "--valRoot",   "/data/coco/images/val2017/",
+                         "--valAnn",    "/data/coco/annotations/instances_val2017.json"]
+            },
+            {
+                "name": "THIS",
+                "type": "python",
+                "request": "launch",
+                "cwd": "${workspaceFolder}/src",
+                "program": "${file}"
+            }
+        ]
+    }
+}

README.md

@@ -0,0 +1,32 @@
+# TinyYolo #
+
+If you like tinygrad, Pytorch and Yolo, then you'll like TinyYolo.
+
+## What this is ##
+
+* This repo uses the new "Tiny-Oriented-Programming" paradigm invented by [TinyGrad](https://github.com/tinygrad/tinygrad) to implement a set of popular Yolo models in (almost) pure Pytorch in under 1000 lines of code.
+
+* The goals are:
+  * Repo must be under 1000 lines of code
+  * Use pure Torch (with a sprinkle of Torchvision and [Einops](https://github.com/arogozhnikov/einops)) to implement models and loss functions
+  * Cannot sacrifice readability in favour of lines of code.
+  * However, reduce lines of code where possible
+  * Be as concise as possible
+  * Code must strive to be self-documenting
+  * No config files like Darknet .cfg or YAML for model configuration. This usually makes code less readable and less hackage.
+
+## What this isn't ##
+
+* This isn't a framework. If you want a fine-tuned framework then use darknet or ultralytics
+
+* A highly optimized library (I haven't implement conv+bn fusion, or anything like that yet)
+
+## What's provided ##
+
+* `models.py`: contains all the Yolo models, which automatically calculate loss when targets are provided in forward function.
+* `test.py`: tests the models with pretrained weights from darknet, ultralytics and Yolov7.
+* `train_coco.py`: trains on COCO using [lightning](https://lightning.ai/)
+
+## Observations ##
+
+* Yolov7 uses `eps=0.001` and `momentum=0.03` in `nn.Batchnorm2d`. That's unusual. I wonder what effects that has on training.

download_weights.sh

@@ -0,0 +1,5 @@
+wget https://pjreddie.com/media/files/yolov3-tiny.weights
+wget https://pjreddie.com/media/files/yolov3.weights
+wget https://github.com/ultralytics/yolov3/releases/download/v8/yolov3-spp.weights
+wget https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4.weights
+wget https://github.com/AlexeyAB/darknet/releases/download/yolov4/yolov4-tiny.weights

images/dog.jpg

@@ -0,0 +1,10 @@
+numpy
+torch
+torchvision
+einops
+onnx
+lightning
+tensorboard
+matplotlib
+pycocotools
+pyqt5