LeNet was proposed in 1998, a typical convolutional neural network. It was used for digit recognition and got big success.
The architectural definition of network refers to the following papers:
[1] Y.Lecun, L.Bottou, Y.Bengio, P.Haffner. Gradient-Based Learning Applied to Document Recognition. Proceedings of the IEEE. 1998.
The following table lists LeNet mnist checkpoints. The model verifies the accuracy of Top-1 and Top-5.
| MindSpore | MindSpore | |||
|---|---|---|---|---|
| Model | Top-1 (%) | Top-5 (%) | Download | Config |
| LeNet | 98.768 | 1.0 | model | config |
| Parameter | Default | Description |
|---|---|---|
| device_target | GPU | Hardware device |
| data_url | Path to training dataset | |
| pretrained | False | Path to pretrained model |
| dataset_sink_mode | True | Data sinking mode |
| num_classes | 10 | Number of dataset classifications |
| batch_size | 32 | Number of batch size |
| repeat_num | 1 | Number of data repetitions |
| momentum | 0.9 | Momentum parameter |
| epoch_size | 10 | Number of epoch |
| keep_checkpoint_max | 10 | Maximum number of checkpoints saved |
| ckpt_save_dir | './lenet' | Save path of checkpoint |
| learning_rate | 0.01 | Number of learning rate |
| resize | 32 | Resize the height and weight of picture |
| download | False | Download Mnist train dataset |
-
The following configuration for training.
python lenet_mnist_train.py --data_url ./mnist --download True
output:
Epoch:[0/10], step:[1875/1875], loss:[0.248/0.248], time:7872.273, lr:0.01000. Epoch time:14200.658, per step time:7.574, avg loss:0.248 Epoch:[1/10], step:[1875/1875], loss:[0.028/0.028], time:4733.176, lr:0.01000. Epoch time:4733.927, per step time:2.525, avg loss:0.028 Epoch:[2/10], step:[1875/1875], loss:[0.200/0.200], time:4725.960, lr:0.01000. Epoch time:4726.731, per step time:2.521, avg loss:0.200 Epoch:[3/10], step:[1875/1875], loss:[0.047/0.047], time:4712.884, lr:0.01000. Epoch time:4713.662, per step time:2.514, avg loss:0.047 Epoch:[4/10], step:[1875/1875], loss:[0.002/0.002], time:4672.785, lr:0.01000. Epoch time:4673.390, per step time:2.492, avg loss:0.002 Epoch:[5/10], step:[1875/1875], loss:[0.000/0.000], time:4630.773, lr:0.01000. Epoch time:4631.456, per step time:2.470, avg loss:0.000 Epoch:[6/10], step:[1875/1875], loss:[0.000/0.000], time:4596.333, lr:0.01000. Epoch time:4597.269, per step time:2.452, avg loss:0.000 Epoch:[7/10], step:[1875/1875], loss:[0.031/0.031], time:4682.164, lr:0.01000. Epoch time:4682.911, per step time:2.498, avg loss:0.031 Epoch:[8/10], step:[1875/1875], loss:[0.004/0.004], time:4639.663, lr:0.01000. Epoch time:4640.499, per step time:2.475, avg loss:0.004 Epoch:[9/10], step:[1875/1875], loss:[0.042/0.042], time:4660.191, lr:0.01000. Epoch time:4660.966, per step time:2.486, avg loss:0.042 -
The following configuration uses yaml file for training.
python examples/classification/tools/train_config.py -c mindvision/classification/config/lenet/lenet_mnist.yaml
output:
Epoch:[0/10], step:[1875/1875], loss:[0.289/0.289], time:4501.820, lr:0.01000. Epoch time:10293.452, per step time:5.490, avg loss:0.289 Epoch:[1/10], step:[1875/1875], loss:[0.029/0.029], time:2001.829, lr:0.00976. Epoch time:2002.693, per step time:1.068, avg loss:0.029 Epoch:[2/10], step:[1875/1875], loss:[0.094/0.094], time:1905.051, lr:0.00905. Epoch time:1905.938, per step time:1.017, avg loss:0.094 Epoch:[3/10], step:[1875/1875], loss:[0.122/0.122], time:2028.579, lr:0.00794. Epoch time:2029.425, per step time:1.082, avg loss:0.122 Epoch:[4/10], step:[1875/1875], loss:[0.002/0.002], time:2000.762, lr:0.00655. Epoch time:2001.407, per step time:1.067, avg loss:0.002 Epoch:[5/10], step:[1875/1875], loss:[0.000/0.000], time:1850.426, lr:0.00500. Epoch time:1851.062, per step time:0.987, avg loss:0.000 Epoch:[6/10], step:[1875/1875], loss:[0.000/0.000], time:1875.080, lr:0.00345. Epoch time:1875.705, per step time:1.000, avg loss:0.000 Epoch:[7/10], step:[1875/1875], loss:[0.001/0.001], time:1937.989, lr:0.00206. Epoch time:1938.506, per step time:1.034, avg loss:0.001 Epoch:[8/10], step:[1875/1875], loss:[0.000/0.000], time:1964.432, lr:0.00095. Epoch time:1964.922, per step time:1.048, avg loss:0.000 Epoch:[9/10], step:[1875/1875], loss:[0.004/0.004], time:1926.403, lr:0.00024. Epoch time:1926.939, per step time:1.028, avg loss:0.004
-
The following configuration for eval.
python lenet_mnist_eval.py --data_url ./mnist --pretrained True --download True
output:
{'Top_1_Accuracy': 0.9876802884615384, 'Top_5_Accuracy': 1.0} -
The following configuration uses yaml file for eval.
python examples/classification/tools/eval_config.py -c mindvision/classification/config/lenet/lenet_mnist.yaml
output:
{'Top_1_Accuracy': 0.9876802884615384, 'Top_5_Accuracy': 1.0}
-
The following configuration for infer.
python lenet_mnist_infer.py --data_url ./infer --pretrained True
output:
{4: 'four'}