You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
pytorch-image-models/docs/models/noisy-student.md

571 lines
16 KiB

# Noisy Student (EfficientNet)
**Noisy Student Training** is a semi-supervised learning approach. It extends the idea of self-training
and distillation with the use of equal-or-larger student models and noise added to the student during learning. It has three main steps:
1. train a teacher model on labeled images
2. use the teacher to generate pseudo labels on unlabeled images
3. train a student model on the combination of labeled images and pseudo labeled images.
The algorithm is iterated a few times by treating the student as a teacher to relabel the unlabeled data and training a new student.
Noisy Student Training seeks to improve on self-training and distillation in two ways. First, it makes the student larger than, or at least equal to, the teacher so the student can better learn from a larger dataset. Second, it adds noise to the student so the noised student is forced to learn harder from the pseudo labels. To noise the student, it uses input noise such as RandAugment data augmentation, and model noise such as dropout and stochastic depth during training.
## How do I use this model on an image?
To load a pretrained model:
```python
import timm
model = timm.create_model('tf_efficientnet_b0_ns', pretrained=True)
model.eval()
```
To load and preprocess the image:
```python
import urllib
from PIL import Image
from timm.data import resolve_data_config
from timm.data.transforms_factory import create_transform
config = resolve_data_config({}, model=model)
transform = create_transform(**config)
url, filename = ("https://github.com/pytorch/hub/raw/master/images/dog.jpg", "dog.jpg")
urllib.request.urlretrieve(url, filename)
img = Image.open(filename).convert('RGB')
tensor = transform(img).unsqueeze(0) # transform and add batch dimension
```
To get the model predictions:
```python
import torch
with torch.no_grad():
out = model(tensor)
probabilities = torch.nn.functional.softmax(out[0], dim=0)
print(probabilities.shape)
# prints: torch.Size([1000])
```
To get the top-5 predictions class names:
```python
# Get imagenet class mappings
url, filename = ("https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt", "imagenet_classes.txt")
urllib.request.urlretrieve(url, filename)
with open("imagenet_classes.txt", "r") as f:
categories = [s.strip() for s in f.readlines()]
# Print top categories per image
top5_prob, top5_catid = torch.topk(probabilities, 5)
for i in range(top5_prob.size(0)):
print(categories[top5_catid[i]], top5_prob[i].item())
# prints class names and probabilities like:
# [('Samoyed', 0.6425196528434753), ('Pomeranian', 0.04062102362513542), ('keeshond', 0.03186424449086189), ('white wolf', 0.01739676296710968), ('Eskimo dog', 0.011717947199940681)]
```
Replace the model name with the variant you want to use, e.g. `tf_efficientnet_b0_ns`. You can find the IDs in the model summaries at the top of this page.
To extract image features with this model, follow the [timm feature extraction examples](https://rwightman.github.io/pytorch-image-models/feature_extraction/), just change the name of the model you want to use.
## How do I finetune this model?
You can finetune any of the pre-trained models just by changing the classifier (the last layer).
```python
model = timm.create_model('tf_efficientnet_b0_ns', pretrained=True, num_classes=NUM_FINETUNE_CLASSES)
```
To finetune on your own dataset, you have to write a training loop or adapt [timm's training
script](https://github.com/rwightman/pytorch-image-models/blob/master/train.py) to use your dataset.
## How do I train this model?
You can follow the [timm recipe scripts](https://rwightman.github.io/pytorch-image-models/scripts/) for training a new model afresh.
## Citation
```BibTeX
@misc{xie2020selftraining,
title={Self-training with Noisy Student improves ImageNet classification},
author={Qizhe Xie and Minh-Thang Luong and Eduard Hovy and Quoc V. Le},
year={2020},
eprint={1911.04252},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
```
<!--
Type: model-index
Collections:
- Name: Noisy Student
Paper:
Title: Self-training with Noisy Student improves ImageNet classification
URL: https://paperswithcode.com/paper/self-training-with-noisy-student-improves
Models:
- Name: tf_efficientnet_b0_ns
In Collection: Noisy Student
Metadata:
FLOPs: 488688572
Parameters: 5290000
File Size: 21386709
Architecture:
- 1x1 Convolution
- Average Pooling
- Batch Normalization
- Convolution
- Dense Connections
- Dropout
- Inverted Residual Block
- Squeeze-and-Excitation Block
- Swish
Tasks:
- Image Classification
Training Techniques:
- AutoAugment
- FixRes
- Label Smoothing
- Noisy Student
- RMSProp
- RandAugment
- Weight Decay
Training Data:
- ImageNet
- JFT-300M
Training Resources: Cloud TPU v3 Pod
ID: tf_efficientnet_b0_ns
LR: 0.128
Epochs: 700
Dropout: 0.5
Crop Pct: '0.875'
Momentum: 0.9
Batch Size: 2048
Image Size: '224'
Weight Decay: 1.0e-05
Interpolation: bicubic
RMSProp Decay: 0.9
Label Smoothing: 0.1
BatchNorm Momentum: 0.99
Stochastic Depth Survival: 0.8
Code: https://github.com/rwightman/pytorch-image-models/blob/9a25fdf3ad0414b4d66da443fe60ae0aa14edc84/timm/models/efficientnet.py#L1427
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_efficientnet_b0_ns-c0e6a31c.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 78.66%
Top 5 Accuracy: 94.37%
- Name: tf_efficientnet_b1_ns
In Collection: Noisy Student
Metadata:
FLOPs: 883633200
Parameters: 7790000
File Size: 31516408
Architecture:
- 1x1 Convolution
- Average Pooling
- Batch Normalization
- Convolution
- Dense Connections
- Dropout
- Inverted Residual Block
- Squeeze-and-Excitation Block
- Swish
Tasks:
- Image Classification
Training Techniques:
- AutoAugment
- FixRes
- Label Smoothing
- Noisy Student
- RMSProp
- RandAugment
- Weight Decay
Training Data:
- ImageNet
- JFT-300M
Training Resources: Cloud TPU v3 Pod
ID: tf_efficientnet_b1_ns
LR: 0.128
Epochs: 700
Dropout: 0.5
Crop Pct: '0.882'
Momentum: 0.9
Batch Size: 2048
Image Size: '240'
Weight Decay: 1.0e-05
Interpolation: bicubic
RMSProp Decay: 0.9
Label Smoothing: 0.1
BatchNorm Momentum: 0.99
Stochastic Depth Survival: 0.8
Code: https://github.com/rwightman/pytorch-image-models/blob/9a25fdf3ad0414b4d66da443fe60ae0aa14edc84/timm/models/efficientnet.py#L1437
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_efficientnet_b1_ns-99dd0c41.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 81.39%
Top 5 Accuracy: 95.74%
- Name: tf_efficientnet_b2_ns
In Collection: Noisy Student
Metadata:
FLOPs: 1234321170
Parameters: 9110000
File Size: 36801803
Architecture:
- 1x1 Convolution
- Average Pooling
- Batch Normalization
- Convolution
- Dense Connections
- Dropout
- Inverted Residual Block
- Squeeze-and-Excitation Block
- Swish
Tasks:
- Image Classification
Training Techniques:
- AutoAugment
- FixRes
- Label Smoothing
- Noisy Student
- RMSProp
- RandAugment
- Weight Decay
Training Data:
- ImageNet
- JFT-300M
Training Resources: Cloud TPU v3 Pod
ID: tf_efficientnet_b2_ns
LR: 0.128
Epochs: 700
Dropout: 0.5
Crop Pct: '0.89'
Momentum: 0.9
Batch Size: 2048
Image Size: '260'
Weight Decay: 1.0e-05
Interpolation: bicubic
RMSProp Decay: 0.9
Label Smoothing: 0.1
BatchNorm Momentum: 0.99
Stochastic Depth Survival: 0.8
Code: https://github.com/rwightman/pytorch-image-models/blob/9a25fdf3ad0414b4d66da443fe60ae0aa14edc84/timm/models/efficientnet.py#L1447
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_efficientnet_b2_ns-00306e48.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 82.39%
Top 5 Accuracy: 96.24%
- Name: tf_efficientnet_b3_ns
In Collection: Noisy Student
Metadata:
FLOPs: 2275247568
Parameters: 12230000
File Size: 49385734
Architecture:
- 1x1 Convolution
- Average Pooling
- Batch Normalization
- Convolution
- Dense Connections
- Dropout
- Inverted Residual Block
- Squeeze-and-Excitation Block
- Swish
Tasks:
- Image Classification
Training Techniques:
- AutoAugment
- FixRes
- Label Smoothing
- Noisy Student
- RMSProp
- RandAugment
- Weight Decay
Training Data:
- ImageNet
- JFT-300M
Training Resources: Cloud TPU v3 Pod
ID: tf_efficientnet_b3_ns
LR: 0.128
Epochs: 700
Dropout: 0.5
Crop Pct: '0.904'
Momentum: 0.9
Batch Size: 2048
Image Size: '300'
Weight Decay: 1.0e-05
Interpolation: bicubic
RMSProp Decay: 0.9
Label Smoothing: 0.1
BatchNorm Momentum: 0.99
Stochastic Depth Survival: 0.8
Code: https://github.com/rwightman/pytorch-image-models/blob/9a25fdf3ad0414b4d66da443fe60ae0aa14edc84/timm/models/efficientnet.py#L1457
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_efficientnet_b3_ns-9d44bf68.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 84.04%
Top 5 Accuracy: 96.91%
- Name: tf_efficientnet_b4_ns
In Collection: Noisy Student
Metadata:
FLOPs: 5749638672
Parameters: 19340000
File Size: 77995057
Architecture:
- 1x1 Convolution
- Average Pooling
- Batch Normalization
- Convolution
- Dense Connections
- Dropout
- Inverted Residual Block
- Squeeze-and-Excitation Block
- Swish
Tasks:
- Image Classification
Training Techniques:
- AutoAugment
- FixRes
- Label Smoothing
- Noisy Student
- RMSProp
- RandAugment
- Weight Decay
Training Data:
- ImageNet
- JFT-300M
Training Resources: Cloud TPU v3 Pod
ID: tf_efficientnet_b4_ns
LR: 0.128
Epochs: 700
Dropout: 0.5
Crop Pct: '0.922'
Momentum: 0.9
Batch Size: 2048
Image Size: '380'
Weight Decay: 1.0e-05
Interpolation: bicubic
RMSProp Decay: 0.9
Label Smoothing: 0.1
BatchNorm Momentum: 0.99
Stochastic Depth Survival: 0.8
Code: https://github.com/rwightman/pytorch-image-models/blob/9a25fdf3ad0414b4d66da443fe60ae0aa14edc84/timm/models/efficientnet.py#L1467
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_efficientnet_b4_ns-d6313a46.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 85.15%
Top 5 Accuracy: 97.47%
- Name: tf_efficientnet_b5_ns
In Collection: Noisy Student
Metadata:
FLOPs: 13176501888
Parameters: 30390000
File Size: 122404944
Architecture:
- 1x1 Convolution
- Average Pooling
- Batch Normalization
- Convolution
- Dense Connections
- Dropout
- Inverted Residual Block
- Squeeze-and-Excitation Block
- Swish
Tasks:
- Image Classification
Training Techniques:
- AutoAugment
- FixRes
- Label Smoothing
- Noisy Student
- RMSProp
- RandAugment
- Weight Decay
Training Data:
- ImageNet
- JFT-300M
Training Resources: Cloud TPU v3 Pod
ID: tf_efficientnet_b5_ns
LR: 0.128
Epochs: 350
Dropout: 0.5
Crop Pct: '0.934'
Momentum: 0.9
Batch Size: 2048
Image Size: '456'
Weight Decay: 1.0e-05
Interpolation: bicubic
RMSProp Decay: 0.9
Label Smoothing: 0.1
BatchNorm Momentum: 0.99
Stochastic Depth Survival: 0.8
Code: https://github.com/rwightman/pytorch-image-models/blob/9a25fdf3ad0414b4d66da443fe60ae0aa14edc84/timm/models/efficientnet.py#L1477
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_efficientnet_b5_ns-6f26d0cf.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 86.08%
Top 5 Accuracy: 97.75%
- Name: tf_efficientnet_b6_ns
In Collection: Noisy Student
Metadata:
FLOPs: 24180518488
Parameters: 43040000
File Size: 173239537
Architecture:
- 1x1 Convolution
- Average Pooling
- Batch Normalization
- Convolution
- Dense Connections
- Dropout
- Inverted Residual Block
- Squeeze-and-Excitation Block
- Swish
Tasks:
- Image Classification
Training Techniques:
- AutoAugment
- FixRes
- Label Smoothing
- Noisy Student
- RMSProp
- RandAugment
- Weight Decay
Training Data:
- ImageNet
- JFT-300M
Training Resources: Cloud TPU v3 Pod
ID: tf_efficientnet_b6_ns
LR: 0.128
Epochs: 350
Dropout: 0.5
Crop Pct: '0.942'
Momentum: 0.9
Batch Size: 2048
Image Size: '528'
Weight Decay: 1.0e-05
Interpolation: bicubic
RMSProp Decay: 0.9
Label Smoothing: 0.1
BatchNorm Momentum: 0.99
Stochastic Depth Survival: 0.8
Code: https://github.com/rwightman/pytorch-image-models/blob/9a25fdf3ad0414b4d66da443fe60ae0aa14edc84/timm/models/efficientnet.py#L1487
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_efficientnet_b6_ns-51548356.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 86.45%
Top 5 Accuracy: 97.88%
- Name: tf_efficientnet_b7_ns
In Collection: Noisy Student
Metadata:
FLOPs: 48205304880
Parameters: 66349999
File Size: 266853140
Architecture:
- 1x1 Convolution
- Average Pooling
- Batch Normalization
- Convolution
- Dense Connections
- Dropout
- Inverted Residual Block
- Squeeze-and-Excitation Block
- Swish
Tasks:
- Image Classification
Training Techniques:
- AutoAugment
- FixRes
- Label Smoothing
- Noisy Student
- RMSProp
- RandAugment
- Weight Decay
Training Data:
- ImageNet
- JFT-300M
Training Resources: Cloud TPU v3 Pod
ID: tf_efficientnet_b7_ns
LR: 0.128
Epochs: 350
Dropout: 0.5
Crop Pct: '0.949'
Momentum: 0.9
Batch Size: 2048
Image Size: '600'
Weight Decay: 1.0e-05
Interpolation: bicubic
RMSProp Decay: 0.9
Label Smoothing: 0.1
BatchNorm Momentum: 0.99
Stochastic Depth Survival: 0.8
Code: https://github.com/rwightman/pytorch-image-models/blob/9a25fdf3ad0414b4d66da443fe60ae0aa14edc84/timm/models/efficientnet.py#L1498
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_efficientnet_b7_ns-1dbc32de.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 86.83%
Top 5 Accuracy: 98.08%
- Name: tf_efficientnet_l2_ns
In Collection: Noisy Student
Metadata:
FLOPs: 611646113804
Parameters: 480310000
File Size: 1925950424
Architecture:
- 1x1 Convolution
- Average Pooling
- Batch Normalization
- Convolution
- Dense Connections
- Dropout
- Inverted Residual Block
- Squeeze-and-Excitation Block
- Swish
Tasks:
- Image Classification
Training Techniques:
- AutoAugment
- FixRes
- Label Smoothing
- Noisy Student
- RMSProp
- RandAugment
- Weight Decay
Training Data:
- ImageNet
- JFT-300M
Training Resources: Cloud TPU v3 Pod
Training Time: 6 days
ID: tf_efficientnet_l2_ns
LR: 0.128
Epochs: 350
Dropout: 0.5
Crop Pct: '0.96'
Momentum: 0.9
Batch Size: 2048
Image Size: '800'
Weight Decay: 1.0e-05
Interpolation: bicubic
RMSProp Decay: 0.9
Label Smoothing: 0.1
BatchNorm Momentum: 0.99
Stochastic Depth Survival: 0.8
Code: https://github.com/rwightman/pytorch-image-models/blob/9a25fdf3ad0414b4d66da443fe60ae0aa14edc84/timm/models/efficientnet.py#L1520
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_efficientnet_l2_ns-df73bb44.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 88.35%
Top 5 Accuracy: 98.66%
-->