Rs/docs update #1398

robertgshaw2-redhat · 2023-02-27T14:46:08Z

No description provided.

review-notebook-app · 2023-02-27T14:46:13Z

Check out this pull request on

See visual diffs & provide feedback on Jupyter Notebooks.

Powered by ReviewNB

mwitiderrick · 2023-03-03T05:46:59Z

integrations/ultralytics-yolov5/README.md

-Recipes encode the instructions and hyperparameters for sparsifying a model using modifiers to the training process.
-The modifiers can range from pruning and quantization to learning rate and weight decay.
-When appropriately combined, it becomes possible to create highly sparse and accurate models.
+Neural Magic has pre-sparsified each version of YOLOv5. These models can be deployed directly or can be fine-tuned onto custom dataset via sparse transfer learning. This


These models can be deployed directly or fine-tuned onto a custom datasets via sparse transfer learning.

mwitiderrick · 2023-03-03T05:47:17Z

integrations/torchvision/README.md

+
+### SparseZoo
+
+Neural Magic has pre-sparsified versions of common Torchvision models such as ResNet-50. These models can be deployed directly or can be fine-tuned onto custom dataset via sparse transfer learning. This makes it easy to create a sparse image classification model trained on your dataset.


These models can be deployed directly or fine-tuned onto a custom datasets via sparse transfer learning

mwitiderrick · 2023-03-03T05:51:44Z

integrations/torchvision/README.md

+
+Just like the CLI, the Python API uses YAML-based recipes to encode the parameters of the sparsification process, allowing youto add SparseML with just a few lines of code.
+
+The `ScheduleModifierManager` class is responsible for parsing the YAML recipes and overriding the standard PyTorch model and optimizer objects, 


The ScheduleModifierManager class parses the YAML recipes and overrides the standard PyTorch model and optimizer objects, encoding the logic of the sparsity algorithms from the recipe

mwitiderrick · 2023-03-03T05:52:27Z

integrations/torchvision/README.md

+
+### Overview
+
+Sparse Transfer is quite similiar to the typical transfer learing process used to train NLP models, where we fine-tune a pretrained checkpoint onto a smaller downstream dataset. With Sparse Transfer Learning, we simply start the fine-tuning process from a pre-sparsified checkpoint and maintain sparsity while the training process occurs.


Sparse Transfer is similar to the typical transfer learning process used to train NLP models, where we fine-tune a pre-trained checkpoint onto a smaller downstream dataset. With Sparse Transfer Learning, we start fine-tuning from a pre-sparsified checkpoint and maintain sparsity during training.

mwitiderrick · 2023-03-03T05:53:39Z

integrations/torchvision/README.md

+
+### Kick off Training
+
+We can start Sparse Transfer Learning by passing a starting checkpoint and recipe to the training script. For Sparse Transfer, we will use a recipe that instructs SparseML to maintain sparsity during training and to quantize the model. The starting checkpoitn and transfer recipe are specified by the following SparseZoo stub:


checkpoitn -> checkpoint

mwitiderrick · 2023-03-03T05:54:20Z

integrations/torchvision/README.md

+
+The key `Modifiers` for sparse transfer learning are the following:
+- `ConstantPruningModifier` instructs SparseML to maintain the sparsity structure of the network during the fine-tuning process
+- `QuantizationModifier` instructs SparseML to apply quantization aware training to quantize the weights over the final epochs


quantization aware -> quantization-aware

mwitiderrick · 2023-03-03T05:55:06Z

integrations/ultralytics-yolov5/README.md


-To benchmark the model, run
+To transfer learn this sparsified model to other datasets you may have to adjust certain hyperparameters in this recipe and/or training script such as the optimizer type, the number of epochs, and the learning rates.


comma before such and other

mwitiderrick · 2023-03-03T06:00:55Z

integrations/ultralytics-yolov5/tutorials/sparse-transfer-learning.md

+limitations under the License.
+-->
+
+# Sparse Transfer Learning


Sparse Transfer Learning with YOLOv5 and SparseML

mwitiderrick · 2023-03-03T06:02:16Z

integrations/ultralytics-yolov5/tutorials/sparse-transfer-learning.md

+
+## Sparse Transfer Learning with VOC
+
+Let's try a simple example of Sparse Transfer Learning YOLOv5s onto the VOC dataset.


Let's try -> Let's implement

Let's implement a simple Sparse Transfer Learning example with YOLOv5s and the VOC dataset.

mwitiderrick · 2023-03-03T06:15:17Z

integrations/ultralytics-yolov5/tutorials/sparse-transfer-learning.md

+
+Run the following to download the dataset from Google Drive:
+
+```python


I don't think this code will work for everyone reading this tutorial. Can you upload the data to G Drive, make it public, and use something like gdown or googledrivedownloader to download it?

mwitiderrick · 2023-03-03T06:18:15Z

integrations/ultralytics-yolov5/tutorials/sparse-transfer-learning.md

+
+## Wrapping Up
+
+Checkout the DeepSparse repository for more information on deploying your sparse YOLOv5 models with DeepSparse for GPU class performance on CPUs.


Check out the DeepSparse repository for more information on deploying your sparse YOLOv5 models with DeepSparse for GPU class performance on CPUs.

Can we link to the repository?

mwitiderrick · 2023-03-03T06:27:54Z

integrations/ultralytics-yolov5/tutorials/sparse-transfer-learning.md

+
+## Overview
+
+Sparse Transfer is quite similiar to the typical YOLOv5 training, where we fine-tune a checkpoint pretrained on COCO onto a smaller downstream dataset. However, with Sparse Transfer Learning, we simply start the fine-tuning process from a pre-sparsified YOLOv5 and maintain sparsity while the training process occurs.


Sparse Transfer is similar to the typical YOLOv5 training, where we fine-tune a checkpoint pre-trained on COCO onto a smaller downstream dataset. However, with Sparse Transfer Learning, we start fine-tuning from a pre-sparsified YOLOv5 and maintain sparsity during training.

mwitiderrick · 2023-03-03T06:29:14Z

integrations/ultralytics-yolov5/tutorials/sparse-transfer-learning.md

+- `--weights` specifies the starting checkpoint for the training process. Here, we passed a SparseZoo stub, which
+identifies the 75% pruned-quantized YOLOv5s model in the SparseZoo. The script downloads the PyTorch model to begin training. In addition to SparseZoo stubs, you can also pass a local path to a PyTorch checkpoint.
+
+- `--recipe` specifies the transfer learning recipe. Recipes are YAML files that declare the sparsity related algorithms


sparsity related -> sparsity-related

mwitiderrick · 2023-03-03T06:29:57Z

integrations/ultralytics-yolov5/tutorials/sparse-transfer-learning.md

+
+- `--recipe` specifies the transfer learning recipe. Recipes are YAML files that declare the sparsity related algorithms
+ that SparseML should apply. For transfer learning, the recipe instructs SparseML to maintain sparsity during training
+ and to apply quantization over the final epochs. In this case, we passed a SparseZoo stub, which instructs SparseML


In this case, we passed a SparseZoo stub instructing SparseML to download a premade YOLOv5s transfer learning recipe.

mwitiderrick · 2023-03-03T06:30:29Z

integrations/ultralytics-yolov5/tutorials/sparse-transfer-learning.md

+
+The "Modifiers" encode how SparseML should modify the training process for Sparse Transfer Learning.
+- `ConstantPruningModifier` tells SparseML to pin weights at 0 over all epochs, maintaining the sparsity structure of the network
+- `QuantizationModifier` tells SparseML to quanitze the weights with quantization aware training over the last 5 epochs


quantization aware -> quantization-aware
quanitze -> quantize

mwitiderrick · 2023-03-03T06:32:02Z

integrations/ultralytics-yolov5/tutorials/sparse-transfer-learning.md

+
+SparseZoo contains mutliple variants of each version of YOLOv5 at various levels of sparsity, which can be fine-tuned onto your dataset. 
+
+[Checkout the full list!](https://sparsezoo.neuralmagic.com/?page=1&domain=cv&sub_domain=detection) 


Checkout -> Check out

mwitiderrick · 2023-03-03T06:32:40Z

integrations/ultralytics-yolov5/tutorials/sparse-transfer-learning.md

+#### 2. Create Labels
+
+After using a tool like [Roboflow Annotate](https://roboflow.com/annotate?ref=ultralytics) to label your data, export your labels to the YOLO Format, with one
+`*.txt` file per image (if not objects are in the image, no `*.txt` file is require).


if not objects -> if no objects
is require -> is required

mwitiderrick · 2023-03-03T06:33:49Z

integrations/ultralytics-yolov5/tutorials/sparse-transfer-learning.md

+../datasets/coco128/labels/im0.txt  # label
+```
+
+For more details, checkout the [custom dataset set tutorial](https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data) in the Ultralytics repository.


checkout -> check out

mwitiderrick · 2023-03-03T07:06:15Z