Using Paddlepaddle to recognize flowers

PaddlePaddle is a Chinese based DL framework developped by China’s Google Baidu. If you are interested in exploring this DL framework, here is an example of how to use paddlepaddle to train a flower classification model.

Install paddlepaddle

Following the official instruction, which is pretty clear.

Prepare the code

git clone https://github.com/PaddlePaddle/PaddleClas.git

Dataset

Download the data

# Go to the root path of the PaddleClas repo
cd dataset/
wget https://paddle-imagenet-models-name.bj.bcebos.com/data/flowers102.zip
unzip flowers102.zip

After downlading the dataset, you will have

flowers102
├── flowers102_label_list.txt
├── jpg
│   ├── image_00001.jpg
│   ├── image_00002.jpg
│   ├── image_00003.jpg
│   ├── image_00004.jpg
...
│   ├── image_08185.jpg
│   ├── image_08186.jpg
│   ├── image_08187.jpg
│   ├── image_08188.jpg
│   └── image_08189.jpg
├── train_extra_list.txt
├── train_list.txt
└── val_list.txt

Which means you have 8189 images in total. Some samples like the following:

How many labels (categories) do we have?

If you check the flowers102_label_list.txt, as the name suggested, we have 102 different types.

0 pink primrose
1 hard-leaved pocket orchid
2 canterbury bells
3 sweet pea
4 english marigold
5 tiger lily
6 moon orchid
7 bird of paradise
8 monkshood
9 globe thistle
...
96 mallow
97 mexican petunia
98 bromelia
99 blanket flower
100 trumpet creeper
101 blackberry lily

Train and val split

We have the file of

train_extra_list.txt train_list.txt val_list.txt

Here is the samples in train_list.txt which indicate the train dataset’s

file name and label id.

jpg/image_06765.jpg 0
jpg/image_06755.jpg 0
jpg/image_06768.jpg 0
jpg/image_06736.jpg 0
jpg/image_06744.jpg 0
jpg/image_06766.jpg 0
jpg/image_06771.jpg 0
jpg/image_06750.jpg 0
jpg/image_06741.jpg 0
jpg/image_06762.jpg 0
jpg/image_05145.jpg 1
jpg/image_05137.jpg 1
jpg/image_05142.jpg 1
jpg/image_05115.jpg 1
jpg/image_05091.jpg 1
jpg/image_05106.jpg 1

Train the model

In order to run the code, you need install several packages except the paddlepaddle

pip install tqdm
conda install -c anaconda scikit-learn
pip install pyyaml
conda install -c conda-forge opencv
pip install visualdl

Using this command to train from a pretrained model

python tools/train.py -c ./ppcls/configs/quick_start/new_user/ShuffleNetV2_x0_25.yaml  -o Arch.pretrained=True

If the training process goes well, you will see something like this

[2022/01/24 23:15:44] root INFO:             channel_first : False
[2022/01/24 23:15:44] root INFO:             to_rgb : True
[2022/01/24 23:15:44] root INFO:         ResizeImage :
[2022/01/24 23:15:44] root INFO:             resize_short : 256
[2022/01/24 23:15:44] root INFO:         CropImage :
[2022/01/24 23:15:44] root INFO:             size : 224
[2022/01/24 23:15:44] root INFO:         NormalizeImage :
[2022/01/24 23:15:44] root INFO:             mean : [0.485, 0.456, 0.406]
[2022/01/24 23:15:44] root INFO:             order :
[2022/01/24 23:15:44] root INFO:             scale : 1.0/255.0
[2022/01/24 23:15:44] root INFO:             std : [0.229, 0.224, 0.225]
[2022/01/24 23:15:44] root INFO:         ToCHWImage : None
[2022/01/24 23:15:44] root INFO: Loss :
[2022/01/24 23:15:44] root INFO:     Eval :
[2022/01/24 23:15:44] root INFO:         CELoss :
[2022/01/24 23:15:44] root INFO:             weight : 1.0
[2022/01/24 23:15:44] root INFO:     Train :
[2022/01/24 23:15:44] root INFO:         CELoss :
[2022/01/24 23:15:44] root INFO:             weight : 1.0
[2022/01/24 23:15:44] root INFO: Metric :
[2022/01/24 23:15:44] root INFO:     Eval :
[2022/01/24 23:15:44] root INFO:         TopkAcc :
[2022/01/24 23:15:44] root INFO:             topk : [1, 5]
[2022/01/24 23:15:44] root INFO:     Train :
[2022/01/24 23:15:44] root INFO:         TopkAcc :
[2022/01/24 23:15:44] root INFO:             topk : [1, 5]
[2022/01/24 23:15:44] root INFO: Optimizer :
[2022/01/24 23:15:44] root INFO:     lr :
[2022/01/24 23:15:44] root INFO:         learning_rate : 0.0125
[2022/01/24 23:15:44] root INFO:         name : Cosine
[2022/01/24 23:15:44] root INFO:         warmup_epoch : 5
[2022/01/24 23:15:44] root INFO:     momentum : 0.9
[2022/01/24 23:15:44] root INFO:     name : Momentum
[2022/01/24 23:15:44] root INFO:     regularizer :
[2022/01/24 23:15:44] root INFO:         coeff : 1e-05
[2022/01/24 23:15:44] root INFO:         name : L2
[2022/01/24 23:15:44] root INFO: profiler_options : None
[2022/01/24 23:15:44] root INFO: train with paddle 2.2.1 and device CUDAPlace(0)
W0124 23:15:44.645875 39134 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.0, Runtime API Version: 11.0
W0124 23:15:45.481689 39134 device_context.cc:465] device: 0, cuDNN Version: 8.0.
[2022/01/24 23:17:25] root INFO: unique_endpoints {''}
[2022/01/24 23:17:25] root INFO: Downloading ResNet50_vd_pretrained.pdparams from https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/legendary_models/ResNet50_vd_pretrained.pdparams
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 151674/151674 [00:44<00:00, 3438.65it/s]
/home/ubuntu/anaconda3/envs/paddle/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py:1436: UserWarning: Skip loading for fc.weight. fc.weight receives a shape [2048, 1000], but the expected shape is [2048, 102].
warnings.warn(("Skip loading for {}. ".format(key) + str(err)))
/home/ubuntu/anaconda3/envs/paddle/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py:1436: UserWarning: Skip loading for fc.bias. fc.bias receives a shape [1000], but the expected shape is [102].
warnings.warn(("Skip loading for {}. ".format(key) + str(err)))
...
 0.33236, batch_cost: 0.20485s, reader_cost: 0.00014, ips: 156.21518 images/sec, eta: 0:01:08
[2022/01/24 23:23:39] root INFO: [Train][Epoch 10/20][Iter: 30/32]lr: 0.00941, top1: 0.91230, top5: 0.97681, CELoss: 0.37659, loss: 0.37659, batch_cost: 0.20435s, reader_cost: 0.00012, ips: 156.59497 images/sec, eta: 0:01:05
[2022/01/24 23:23:39] root INFO: [Train][Epoch 10/20][Avg]top1: 0.91373, top5: 0.97745, CELoss: 0.37335, loss: 0.37335
[2022/01/24 23:23:40] root INFO: [Eval][Epoch 10][Iter: 0/16]CELoss: 0.75603, loss: 0.75603, top1: 0.82812, top5: 0.93750, batch_cost: 1.02958s, reader_cost: 0.91494, ips: 62.16100 images/sec
[2022/01/24 23:23:42] root INFO: [Eval][Epoch 10][Iter: 10/16]CELoss: 0.24584, loss: 0.24584, top1: 0.92188, top5: 1.00000, batch_cost: 0.17600s, reader_cost: 0.06913, ips: 363.64605 images/sec
[2022/01/24 23:23:43] root INFO: [Eval][Epoch 10][Avg]CELoss: 0.37863, loss: 0.37863, top1: 0.90000, top5: 0.97843
[2022/01/24 23:23:49] root INFO: Already save model in ./output/ResNet50_vd/best_model
[2022/01/24 23:23:49] root INFO: [Eval][Epoch 10][best metric: 0.8999999983637941]
[2022/01/24 23:23:50] root INFO: Already save model in ./output/ResNet50_vd/epoch_10
[2022/01/24 23:23:54] root INFO: Already save model in ./output/ResNet50_vd/latest
[2022/01/24 23:23:55] root INFO: [Train][Epoch 11/20][Iter: 0/32]lr: 0.00934, top1: 0.93750, top5: 1.00000, CELoss: 0.15017, loss: 0.15017, batch_cost: 0.21860s, reader_cost: 0.01467, ips: 146.38531 images/sec, eta: 0:01:09
[2022/01/24 23:23:57] root INFO: [Train][Epoch 11/20][Iter: 10/32]lr: 0.00898, top1: 0.93466, top5: 0.99148, CELoss: 0.24013, loss: 0.24013, batch_cost: 0.20525s, reader_cost: 0
...
[2022/01/24 23:26:21] root INFO: [Train][Epoch 20/20][Avg]top1: 0.98235, top5: 0.99608, CELoss: 0.07628, loss: 0.07628
[2022/01/24 23:26:22] root INFO: [Eval][Epoch 20][Iter: 0/16]CELoss: 0.59098, loss: 0.59098, top1: 0.87500, top5: 0.95312, batch_cost: 1.02575s, reader_cost: 0.90770, ips: 62.39326 images/sec
[2022/01/24 23:26:24] root INFO: [Eval][Epoch 20][Iter: 10/16]CELoss: 0.28884, loss: 0.28884, top1: 0.92188, top5: 0.98438, batch_cost: 0.16941s, reader_cost: 0.06267, ips: 377.79246 images/sec
[2022/01/24 23:26:25] root INFO: [Eval][Epoch 20][Avg]CELoss: 0.23456, loss: 0.23456, top1: 0.94706, top5: 0.98627
[2022/01/24 23:26:25] root INFO: [Eval][Epoch 20][best metric: 0.9470588244643866]
[2022/01/24 23:26:25] root INFO: Already save model in ./output/ResNet50_vd/epoch_20
[2022/01/24 23:26:26] root INFO: Already save model in ./output/ResNet50_vd/latest

If you check you GPU usage

nvidia-smi

You can see

Inference

python tools/infer.py -c ./ppcls/configs/quick_start/ResNet50_vd.yaml -o Infer.infer_imgs=dataset/flowers102/jpg/image_00001.jpg -o Global.pretrained_model=output/ResNet50_vd/best_model

output

Pretty good, the image tested has the label of 76. The model’s prediction matches.

[{'class_ids': [76, 11, 9, 70, 12], 'scores': [0.99713, 0.00109, 0.00042, 0.00016, 0.00012], 'file_name': 'dataset/flowers102/jpg/image_00001.jpg', 'label_names': ['passion flower', "colt's foot", 'globe thistle', 'gazania', 'king protea']}]

Reference

[1]https://github.com/PaddlePaddle/PaddleClas/blob/release/2.3/docs/zh_CN/quick_start/quick_start_classification_new_user.md