Add MovieGen - TAE (#778)

* add tae

* linting

* rm redundancy

* fix inflate

* add dev plan

* Update README.md

* update psnr w/ opl

* add report

* Update report.md

* Update report.md

* Update report.md

* Update report.md

* Update report.md

* fix linting

* add transformer to tech report

* Update report.md

* linting

* rename to moviegen

* rm irrelevant files

* add report link

* use mint.pad

* linting

examples/animatediff/ad/models/diffusion/ddpm.py

@@ -365,7 +365,10 @@ def construct(self, x: ms.Tensor, text_tokens: ms.Tensor, control=None, **kwargs
         """
 
         # 1. get image/video latents z using vae
-        z = self.get_latents(x)
+        if self.emb_cache:
+            z = x
+        else:
+            z = self.get_latents(x)
 
         # 2. sample timestep and add noise to latents
         t = self.uniform_int(

examples/moviegen/README.md

@@ -0,0 +1,151 @@
+# Movie Gen Video based on MindSpore
+
+This project is built on the [Movie Gen](https://arxiv.org/abs/2410.13720) paper by Meta for video generation, personalization, and editing. We aim to explore an efficient implementation  based on MindSpore and Ascend NPUs. See our [report](docs/report.md) for more details.
+
+## 📑 Development Plan
+
+This project is in an early stage and under active development. We welcome the open-source community to contribute to this project!
+
+- Temporal Autoencoder (TAE)
+  - [x] Inference
+  - [x] Training
+- MovieGenVideo-5B (T2I/V)
+  - [x] Inference
+  - [x] Training stage 1: T2I 256px
+  - [x] Training stage 2: T2I/V 256px 256frames
+  - [ ] Training stage 3: T2I/V 768px 256frames (under training)
+  - [x] Web Demo (Gradio)
+- MovieGenVideo-30B (T2I/V)
+  - [x] Inference
+  - [ ] Mixed parallelism training  (support DP+SP+CP+TP+MP+Zero3, under training)
+- Personalized-MovieGenVideo (PT2V)
+  - [ ] Inference
+  - [ ] Training
+- MovieGen-Edit
+  - [ ] Inference
+  - [ ] Training
+
+
+## Temporal Autoencoder (TAE)
+
+
+### Requirements
+
+| mindspore  | ascend driver  |  firmware   |cann toolkit/kernel |
+|:----------:|:--------------:|:-----------:|:------------------:|
+|   2.3.1    |    24.1.RC2    | 7.3.0.1.231 |   8.0.RC2.beta1    |
+
+### Prepare weights
+
+We use SD3.5 VAE to initialize the spatial layers of TAE, considering they have the same number of latent channels, i.e. 16.
+
+1. Download SD3.5 VAE from [huggingface](https://huggingface.co/stabilityai/stable-diffusion-3.5-large/tree/main/vae)
+
+2. Inflate VAE checkpoint for TAE initialization by
+
+```shell
+python inflate_vae_to_tae.py --src /path/to/sd3.5_vae/diffusion_pytorch_model.safetensors --target models/tae_vae2d.ckpt
+```
+
+### Prepare datasets
+
+We need to prepare a csv annotation file listing the path to each input video related to the root folder, indicated by the `video_folder` argument. An example is
+```
+video
+dance/vid001.mp4
+dance/vid002.mp4
+dance/vid003.mp4
+...
+```
+
+Taking UCF-101 for example, please download the [UCF-101](https://www.crcv.ucf.edu/data/UCF101.php) dataset and extract it to `datasets/UCF-101` folder.
+
+
+### Training
+
+TAE is trained to optimize the reconstruction loss, perceptual loss, and the outlier penalty loss (OPL) proposed in the MovieGen paper.
+
+To launch training, please run
+
+```shell
+python scripts/train_tae.py \
+--config configs/tae/train/mixed_256x256x32.yaml \
+--output_path /path/to/save_ckpt_and_log \
+--csv_path /path/to/video_train.csv  \
+--video_folder /path/to/video_root_folder  \
+```
+
+Unlike the paper, we found that OPL loss doesn't benefit the training outcome in our ablation study (w/ OPl PSNR is 31.17). Thus we disable OPL loss by default. You may enable it by appending `--use_outlier_penalty_loss True`
+
+For more details on the arguments, please run `python scripts/train_tae.py --help`
+
+
+### Evaluation
+
+To run video reconstruction with the trained TAE model and evaluate the PSNR and SSIM on the test set, please run
+
+```shell
+python scripts/inference_tae.py \
+--ckpt_path /path/to/tae.ckpt \
+--batch_size 2 \
+--num_frames 32  \
+--image_size 256 \
+--csv_path  /path/to/video_test.csv  \
+--video_folder /path/to/video_root_folder  \
+```
+
+The reconstructed videos will be saved in `samples/recons`.
+
+#### Performance
+
+Here, we report the training performance and evaluation results on the UCF-101 dataset.
+
+Experiments are tested on ascend 910* with mindspore 2.3.1 graph mode.
+
+| model name      |  cards | batch size | resolution |  precision | jit level |   graph compile | s/step     | PSNR | SSIM | recipe |
+| :--:         | :---:   | :--:       | :--:       |  :--:       | :--:       | :--:      |:--:    | :--:   |:--:   |:--:   |
+| TAE  |  1     | 1      | 256x256x32   |  bf16    |   O0  | 2 mins |   2.18     | 31.35     |   0.92       |  [config](configs/tae/train/mixed_256x256x32.yaml) |
+
+
+### Usages for Latent Diffusion Models
+
+<details>
+<summary>View more</summary>
+
+#### Encoding video
+
+```python
+from mg.models.tae.tae import TemporalAutoencoder, TAE_CONFIG
+
+# may set use_tile=True to save memory
+tae = TemporalAutoencoder(
+    pretrained='/path/to/tae.ckpt',
+    use_tile=False,
+    )
+
+# x - a batch of videos, shape (b c t h w)
+z, _, _ = tae.encode(x)
+
+
+# you may scale z by:
+# z = TAE_CONFIG['scaling_factor'] * (z - TAE_CONFIG['shift_factor'])
+
+```
+
+For detailed arguments, please refer to the docstring in [tae.py](mg/models/tae/tae.py)
+
+### Decoding video latent
+
+```python
+
+# if z is scaled, you should unscale at first:
+# z = z / TAE_CONFIG['scaling_factor'] + TAE_CONFIG['shift_factor']
+
+# z - a batch of video latent, shape (b c t h w)
+x = tae.decode(z)
+
+# for image decoding, set num_target_frames to discard the spurious frames
+x = tae.decode(z, num_target_frames=1)
+```
+
+</details>

examples/moviegen/configs/tae/train/mixed_256x256x16.yaml

@@ -0,0 +1,46 @@
+# model
+pretrained_model_path: "models/tae_vae2d.ckpt"
+
+# loss
+perceptual_loss_weight: 1.0
+kl_loss_weight: 1.e-6
+use_outlier_penalty_loss: False  # OPL bring no benefit in our experiments
+mixed_strategy: "mixed_video_image"
+mixed_image_ratio: 0.2
+
+# data
+dataset_name: "video"
+csv_path: "../videocomposer/datasets/webvid5_copy.csv"
+video_folder: "../videocomposer/datasets/webvid5"
+frame_stride: 1
+num_frames: 16
+image_size: 256
+crop_size: 256
+# flip: True
+
+# training recipe
+seed: 42
+use_discriminator: False
+batch_size: 1
+clip_grad: True
+max_grad_norm: 1.0
+start_learning_rate: 1.e-5
+scale_lr: False
+weight_decay: 0.
+
+dtype: "fp32"
+amp_level: "O0"
+use_recompute: False
+
+epochs: 2000
+ckpt_save_interval: 50
+init_loss_scale: 1024.
+loss_scaler_type: dynamic
+
+scheduler: "constant"
+use_ema: False
+
+output_path: "outputs/tae_train"
+
+# ms settting
+jit_level: O0

examples/moviegen/configs/tae/train/mixed_256x256x32.yaml

@@ -0,0 +1,46 @@
+# model
+pretrained_model_path: "models/tae_vae2d.ckpt"
+
+# loss
+perceptual_loss_weight: 1.0
+kl_loss_weight: 1.e-6
+use_outlier_penalty_loss: False  # OPL bring no benefit in our experiments
+mixed_strategy: "mixed_video_image"
+mixed_image_ratio: 0.2
+
+# data
+dataset_name: "video"
+csv_path: "../videocomposer/datasets/webvid5_copy.csv"
+video_folder: "../videocomposer/datasets/webvid5"
+frame_stride: 1
+num_frames: 32
+image_size: 256
+crop_size: 256
+# flip: True
+
+# training recipe
+seed: 42
+use_discriminator: False
+batch_size: 1
+clip_grad: True
+max_grad_norm: 1.0
+start_learning_rate: 1.e-5
+scale_lr: False
+weight_decay: 0.
+
+dtype: "bf16"
+amp_level: "O2"   # reduce memory cost
+use_recompute: True
+
+epochs: 2000
+ckpt_save_interval: 50
+init_loss_scale: 1024.
+loss_scaler_type: dynamic
+
+scheduler: "constant"
+use_ema: False
+
+output_path: "outputs/tae_train"
+
+# ms settting
+jit_level: O0