File size: 1,057 Bytes
faae6a6
 
 
 
 
 
52b54c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
tags:
- model_hub_mixin
- pytorch_model_hub_mixin
---

## ⚙️ Usage
Our pretrained model are made available through `rshf` and `transformers` package for easy inference.

Load and initialize:
```python
from rshf.prom3e import ProM3E

model = ProM3E.from_pretrained("MVRL/ProM3E")
```

Inference:
```python
# Get precomputed embeddings from taxabind for image, sat, loc, env, text, audio
# Replace missing modalities with any vector
# Stack embeddings in the order: image, sat, loc, env, text, audio
# Pass through the model

# Example:
image_embeds = torch.randn(2, 512)
sat_embeds = torch.randn(2, 512)
loc_embeds = torch.randn(2, 512)
env_embeds = torch.randn(2, 512)
text_embeds = torch.randn(2, 512)
audio_embeds = torch.randn(2, 512)

modalities = torch.stack((image_embeds, sat_embeds, loc_embeds, env_embeds, text_embeds, audio_embeds), dim=1)

modalities = torch.nn.functional.normalize(modalities, dim=-1)

unmasked_modalities = [0, 2]

reconstructions, mu, log_var, hidden_repr = model.forward_inference(modalities, unmasked_modalities)
```