Back | Next | Contents
Transfer Learning
Next, we'll train a model capable of classifying 20 different varieties of plants and trees from the PlantCLEF dataset.
Provided below is a 1.5GB subset that includes 10,475 training images, 1,155 validation images, and 30 test images across 20 classes of plants and trees. The classes were selected from PlantCLEF 2017 from categories that had at least 500 training images in the original dataset:
• ash
• beech
• cat-tail
• cedar
• clover
• cyprus
• daisy
• dandelion
• dogwood
• elm
• fern
• fig
• fir
• juniper
• maple
• poison ivy
• sweetgum
• sycamore
• trout lily
• tulip tree
To get started, first make sure that you have PyTorch installed on your Jetson, then download the dataset below and kick off the training script. After that, we'll test the re-trained model in TensorRT on some static images and a live camera feed.
Run these commands to download and extract the prepared PlantCLEF dataset:
$ cd ~/datasets
$ wget https://nvidia.box.com/shared/static/vbsywpw5iqy7r38j78xs0ctalg7jrg79.gz -O PlantCLEF_Subset.tar.gz
$ tar xvzf PlantCLEF_Subset.tar.gzMirrors of the dataset are available here:
- https://drive.google.com/file/d/14pUv-ZLHtRR-zCYjznr78mytFcnuR_1D/view?usp=sharing
- https://nvidia.box.com/s/vbsywpw5iqy7r38j78xs0ctalg7jrg79
We'll use the same training script that we did from the previous example, located under python/training/imagenet/. By default it's set to train a ResNet-18 model, but you can change that with the --arch flag.
To launch the training, run the following commands:
$ cd jetson-inference/python/training/imagenet
$ python train.py --model-dir=plants ~/datasets/PlantCLEF_SubsetAs training begins, you should see text from the console like the following:
Use GPU: 0 for training
=> dataset classes: 20 ['ash', 'beech', 'cattail', 'cedar', 'clover', 'cyprus', 'daisy', 'dandelion', 'dogwood', 'elm', 'fern', 'fig', 'fir', 'juniper', 'maple', 'poison_ivy', 'sweetgum', 'sycamore', 'trout_lily', 'tulip_tree']
=> using pre-trained model 'resnet18'
=> reshaped ResNet fully-connected layer with: Linear(in_features=512, out_features=20, bias=True)
Epoch: [0][ 0/1307] Time 49.345 (49.345) Data 0.561 ( 0.561) Loss 3.2172e+00 (3.2172e+00) Acc@1 0.00 ( 0.00) Acc@5 25.00 ( 25.00)
Epoch: [0][ 10/1307] Time 0.779 ( 5.211) Data 0.000 ( 0.060) Loss 2.3915e+01 (1.5221e+01) Acc@1 0.00 ( 5.68) Acc@5 12.50 ( 27.27)
Epoch: [0][ 20/1307] Time 0.765 ( 3.096) Data 0.000 ( 0.053) Loss 3.6293e+01 (2.1256e+01) Acc@1 0.00 ( 5.95) Acc@5 37.50 ( 27.38)
Epoch: [0][ 30/1307] Time 0.773 ( 2.346) Data 0.000 ( 0.051) Loss 2.8803e+00 (1.9256e+01) Acc@1 37.50 ( 6.85) Acc@5 62.50 ( 27.42)
Epoch: [0][ 40/1307] Time 0.774 ( 1.962) Data 0.000 ( 0.050) Loss 3.7734e+00 (1.5865e+01) Acc@1 12.50 ( 8.84) Acc@5 37.50 ( 29.88)
Epoch: [0][ 50/1307] Time 0.772 ( 1.731) Data 0.000 ( 0.049) Loss 3.0311e+00 (1.3756e+01) Acc@1 25.00 ( 10.29) Acc@5 37.50 ( 32.35)
Epoch: [0][ 60/1307] Time 0.773 ( 1.574) Data 0.000 ( 0.048) Loss 3.2433e+00 (1.2093e+01) Acc@1 0.00 ( 9.84) Acc@5 25.00 ( 32.79)
Epoch: [0][ 70/1307] Time 0.806 ( 1.462) Data 0.000 ( 0.048) Loss 2.9213e+00 (1.0843e+01) Acc@1 12.50 ( 8.98) Acc@5 37.50 ( 33.27)
Epoch: [0][ 80/1307] Time 0.792 ( 1.379) Data 0.000 ( 0.048) Loss 3.2370e+00 (9.8715e+00) Acc@1 0.00 ( 9.26) Acc@5 25.00 ( 34.41)
Epoch: [0][ 90/1307] Time 0.770 ( 1.314) Data 0.000 ( 0.048) Loss 2.4494e+00 (9.0905e+00) Acc@1 25.00 ( 9.75) Acc@5 75.00 ( 36.26)
Epoch: [0][ 100/1307] Time 0.801 ( 1.261) Data 0.001 ( 0.048) Loss 2.6449e+00 (8.4769e+00) Acc@1 25.00 ( 10.40) Acc@5 62.50 ( 37.00)See the Training Metrics from the previous page for a description of the statistics from the output above.
On the PlantCLEF dataset of 10,475 images, training ResNet-18 takes approximately ~15 minutes per epoch on Jetson Nano, or around 8 hours to train the model for 35 epochs. Below is a graph for analyzing the training progression of epochs versus model accuracy:
At around epoch 30, the ResNet-18 model reaches 75% Top-5 accuracy, and at epoch 65 it converges on 85% Top-5 accuracy. Interestingly these points of stability and convergence for the model occur at similiar times for ResNet-18 that they did for the previous Cat/Dog model. The model's Top-1 accuracy is 55%, which we'll find to be quite effective in practice, given the diversity and challenging content from the PlantCLEF dataset (i.e. multiple overlapping varieties of plants per image and many pictures of leaves and tree trunks that are virtually indistinguishable from one another).
By default the training script is set to run for 35 epochs, but if you don't wish to wait that long to test out your model, you can exit training early and proceed to the next step (optionally re-starting the training again later from where you left off). You can also download this completed model that was trained for a full 100 epochs from here:
Note that the models are saved under jetson-inference/python/training/imagenet/plants/, including a checkpoint from the latest epoch and the best-performing model that has the highest classification accuracy. You can change the directory that the models are saved to by altering the --model-dir flag.
Just like with the Cat/Dog example, next we need to convert our trained model from PyTorch to ONNX, so that we can load it with TensorRT:
python onnx_export.py --model-dir=plantsThis will create a model called resnet18.onnx under jetson-inference/python/training/imagenet/plants/
To classify some static test images, like before we'll use the extended command-line parameters to imagenet-console to load our customized ResNet-18 model that we re-trained above. To run these commands, the working directory of your terminal should still be located in: jetson-inference/python/training/imagenet/
DATASET=~/datasets/PlantCLEF_Subset
# C++
imagenet-console --model=plants/resnet18.onnx --input_blob=input_0 --output_blob=output_0 --labels=$DATASET/labels.txt $DATASET/test/cattail.jpg cattail.jpg
# Python
imagenet-console --model=plants/resnet18.onnx --input_blob=input_0 --output_blob=output_0 --labels=$DATASET/labels.txt $DATASET/test/cattail.jpg cattail.jpg
# C++
imagenet-console --model=plants/resnet18.onnx --input_blob=input_0 --output_blob=output_0 --labels=$DATASET/labels.txt $DATASET/test/elm.jpg elm.jpg
# Python
imagenet-console --model=plants/resnet18.onnx --input_blob=input_0 --output_blob=output_0 --labels=$DATASET/labels.txt $DATASET/test/elm.jpg elm.jpg
# C++
imagenet-console --model=plants/resnet18.onnx --input_blob=input_0 --output_blob=output_0 --labels=$DATASET/labels.txt $DATASET/test/juniper.jpg juniper.jpg
# Python
imagenet-console --model=plants/resnet18.onnx --input_blob=input_0 --output_blob=output_0 --labels=$DATASET/labels.txt $DATASET/test/juniper.jpg juniper.jpg
There are a bunch of test images included with the dataset, or you can download your own pictures to try.
If you want to classify all of the test images without having to do them individually, you can create a simple script like below that loops over them and outputs to the test_output directory under the dataset:
#!/bin/bash
NET="~/jetson-inference/python/training/imagenet/plants"
DATASET="~/datasets/PlantCLEF_Subset"
cd $DATASET
cp -r test test_output
FILES="$DATASET/test_output/*.jpg"
cd $DATA
for f in $FILES
do
echo "Processing $f"
imagenet-console --model=$NET/resnet18.onnx --input_blob=input_0 --output_blob=output_0 --labels=$DATASET/../labels.txt $f $f
doneYou can also try running your re-trained plant model on a live camera stream like below:
DATASET=~/datasets/PlantCLEF_Subset
# C++
imagenet-camera --model=plants/resnet18.onnx --input_blob=input_0 --output_blob=output_0 --labels=$DATASET/labels.txt
# Python
imagenet-camera.py --model=plants/resnet18.onnx --input_blob=input_0 --output_blob=output_0 --labels=$DATASET/labels.txt
Looks like I should be watching out for poison ivy!
Next, we're going to cover a camera-based tool for collecting and labelling your own datasets captured from live video.
Next | Collecting your own Datasets
Back | Re-training on the Cat/Dog Dataset
© 2016-2019 NVIDIA | Table of Contents