File size: 9,044 Bytes
9384dae dd84920 9384dae c0c50b7 9384dae dd84920 9384dae dd84920 9384dae dd84920 9384dae |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 |
---
tags:
- monai
- medical
library_name: monai
license: apache-2.0
---
# Model Overview
A pre-trained model for volumetric (3D) Brats MRI 3D Latent Diffusion Generative Model.
This model is trained on BraTS 2016 and 2017 data from [Medical Decathlon](http://medicaldecathlon.com/), using the Latent diffusion model [1].
This model is a generator for creating images like the Flair MRIs based on BraTS 2016 and 2017 data. It was trained as a 3d latent diffusion model and accepts Gaussian random noise as inputs to produce an image output. The `train_autoencoder.json` file describes the training process of the variational autoencoder with GAN loss. The `train_diffusion.json` file describes the training process of the 3D latent diffusion model.
In this bundle, the autoencoder uses perceptual loss, which is based on ResNet50 with pre-trained weights (the network is frozen and will not be trained in the bundle). In default, the `pretrained` parameter is specified as `False` in `train_autoencoder.json`. To ensure correct training, changing the default settings is necessary. There are two ways to utilize pretrained weights:
1. if set `pretrained` to `True`, ImageNet pretrained weights from [torchvision](https://pytorch.org/vision/stable/_modules/torchvision/models/resnet.html#ResNet50_Weights) will be used. However, the weights are for non-commercial use only.
2. if set `pretrained` to `True` and specifies the `perceptual_loss_model_weights_path` parameter, users are able to load weights from a local path. This is the way this bundle used to train, and the pre-trained weights are from some internal data.
Please note that each user is responsible for checking the data source of the pre-trained models, the applicable licenses, and determining if suitable for the intended use.
#### Example synthetic image
An example result from inference is shown below:
**This is a demonstration network meant to just show the training process for this sort of network with MONAI. To achieve better performance, users need to use larger dataset like [Brats 2021](https://www.synapse.org/#!Synapse:syn25829067/wiki/610865) and have GPU with memory larger than 32G to enable larger networks and attention layers.**
## MONAI Generative Model Dependencies
This bundle requires to install [MONAI generative models](https://github.com/Project-MONAI/GenerativeModels).
## Data
The training data is BraTS 2016 and 2017 from the Medical Segmentation Decathalon. Users can find more details on the dataset (`Task01_BrainTumour`) at http://medicaldecathlon.com/.
- Target: Image Generation
- Task: Synthesis
- Modality: MRI
- Size: 388 3D volumes (1 channel used)
## Training Configuration
If you have a GPU with less than 32G of memory, you may need to decrease the batch size when training. To do so, modify the `train_batch_size` parameter in the [configs/train_autoencoder.json](../configs/train_autoencoder.json) and [configs/train_diffusion.json](../configs/train_diffusion.json) configuration files.
### Training Configuration of Autoencoder
The autoencoder was trained using the following configuration:
- GPU: at least 32GB GPU memory
- Actual Model Input: 112 x 128 x 80
- AMP: False
- Optimizer: Adam
- Learning Rate: 1e-5
- Loss: L1 loss, perceptual loss, KL divergence loss, adversarial loss, GAN BCE loss
#### Input
1 channel 3D MRI Flair patches
#### Output
- 1 channel 3D MRI reconstructed patches
- 8 channel mean of latent features
- 8 channel standard deviation of latent features
### Training Configuration of Diffusion Model
The latent diffusion model was trained using the following configuration:
- GPU: at least 32GB GPU memory
- Actual Model Input: 36 x 44 x 28
- AMP: False
- Optimizer: Adam
- Learning Rate: 1e-5
- Loss: MSE loss
#### Training Input
- 8 channel noisy latent features
- a long int that indicates the time step
#### Training Output
8 channel predicted added noise
#### Inference Input
8 channel noise
#### Inference Output
8 channel denoised latent features
### Memory Consumption Warning
If you face memory issues with data loading, you can lower the caching rate `cache_rate` in the configurations within range [0, 1] to minimize the System RAM requirements.
## Performance
#### Training Loss
## MONAI Bundle Commands
In addition to the Pythonic APIs, a few command line interfaces (CLI) are provided to interact with the bundle. The CLI supports flexible use cases, such as overriding configs at runtime and predefining arguments in a file.
For more details usage instructions, visit the [MONAI Bundle Configuration Page](https://docs.monai.io/en/latest/config_syntax.html).
### Execute Autoencoder Training
#### Execute Autoencoder Training on single GPU
```
python -m monai.bundle run --config_file configs/train_autoencoder.json
```
Please note that if the default dataset path is not modified with the actual path (it should be the path that contains `Task01_BrainTumour`) in the bundle config files, you can also override it by using `--dataset_dir`:
```
python -m monai.bundle run --config_file configs/train_autoencoder.json --dataset_dir <actual dataset path>
```
#### Override the `train` config to execute multi-GPU training for Autoencoder
To train with multiple GPUs, use the following command, which requires scaling up the learning rate according to the number of GPUs.
```
torchrun --standalone --nnodes=1 --nproc_per_node=8 -m monai.bundle run --config_file "['configs/train_autoencoder.json','configs/multi_gpu_train_autoencoder.json']" --lr 8e-5
```
#### Check the Autoencoder Training result
The following code generates a reconstructed image from a random input image.
We can visualize it to see if the autoencoder is trained correctly.
```
python -m monai.bundle run --config_file configs/inference_autoencoder.json
```
An example of reconstructed image from inference is shown below. If the autoencoder is trained correctly, the reconstructed image should look similar to original image.
### Execute Latent Diffusion Training
#### Execute Latent Diffusion Model Training on single GPU
After training the autoencoder, run the following command to train the latent diffusion model. This command will print out the scale factor of the latent feature space. If your autoencoder is well trained, this value should be close to 1.0.
```
python -m monai.bundle run --config_file "['configs/train_autoencoder.json','configs/train_diffusion.json']"
```
#### Override the `train` config to execute multi-GPU training for Latent Diffusion Model
To train with multiple GPUs, use the following command, which requires scaling up the learning rate according to the number of GPUs.
```
torchrun --standalone --nnodes=1 --nproc_per_node=8 -m monai.bundle run --config_file "['configs/train_autoencoder.json','configs/train_diffusion.json','configs/multi_gpu_train_autoencoder.json','configs/multi_gpu_train_diffusion.json']" --lr 8e-5
```
#### Execute inference
The following code generates a synthetic image from a random sampled noise.
```
python -m monai.bundle run --config_file configs/inference.json
```
#### Export checkpoint to TorchScript file
The Autoencoder can be exported into a TorchScript file.
```
python -m monai.bundle ckpt_export autoencoder_def --filepath models/model_autoencoder.ts --ckpt_file models/model_autoencoder.pt --meta_file configs/metadata.json --config_file configs/inference.json
```
# References
[1] Rombach, Robin, et al. "High-resolution image synthesis with latent diffusion models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022. https://openaccess.thecvf.com/content/CVPR2022/papers/Rombach_High-Resolution_Image_Synthesis_With_Latent_Diffusion_Models_CVPR_2022_paper.pdf
# License
Copyright (c) MONAI Consortium
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
|