Update README.md
Browse files
README.md
CHANGED
|
@@ -6,31 +6,40 @@ tags:
|
|
| 6 |
- diffusion
|
| 7 |
library_name: diffuse
|
| 8 |
---
|
|
|
|
| 9 |
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
| 11 |
|
| 12 |
-
|
| 13 |
|
| 14 |
## Model Details
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
-
|
| 17 |
-
- **Checkpoint Step**: 45,000
|
| 18 |
-
- **Precision**: float32
|
| 19 |
-
- **Framework**: JAX/Flax (NNX)
|
| 20 |
-
- **Format**: msgpack
|
| 21 |
|
| 22 |
## Usage
|
| 23 |
|
|
|
|
|
|
|
| 24 |
```python
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
from flax import nnx, serialization
|
| 26 |
from huggingface_hub import hf_hub_download
|
| 27 |
-
import importlib.util
|
| 28 |
|
| 29 |
# Download model weights and config
|
| 30 |
-
model_path = hf_hub_download(repo_id="jcopo/mnist", filename="model.msgpack")
|
| 31 |
-
config_path = hf_hub_download(repo_id="jcopo/mnist", filename="config.py")
|
| 32 |
|
| 33 |
# Load config to get model architecture
|
|
|
|
| 34 |
spec = importlib.util.spec_from_file_location("model_config", config_path)
|
| 35 |
config_module = importlib.util.module_from_spec(spec)
|
| 36 |
spec.loader.exec_module(config_module)
|
|
@@ -43,13 +52,10 @@ with open(model_path, "rb") as f:
|
|
| 43 |
state_dict = serialization.from_bytes(None, f.read())
|
| 44 |
|
| 45 |
# Restore weights into model
|
| 46 |
-
nnx.
|
|
|
|
|
|
|
| 47 |
model.eval() # Set to evaluation mode
|
| 48 |
|
| 49 |
-
|
| 50 |
-
# output = model(input_data)
|
| 51 |
```
|
| 52 |
-
|
| 53 |
-
## Training Configuration
|
| 54 |
-
|
| 55 |
-
This model was trained with the Triax framework using the configuration saved in the checkpoint.
|
|
|
|
| 6 |
- diffusion
|
| 7 |
library_name: diffuse
|
| 8 |
---
|
| 9 |
+
---
|
| 10 |
|
| 11 |
+
## Mnist Generation
|
| 12 |
+
Flow matching diffusion model trained for mnist generation.
|
| 13 |
+
Use with [**diffuse**](https://github.com/jcopo/diffuse), a JAX/Flax sampling library.
|
| 14 |
+
Light enough to run on CPU
|
| 15 |
|
| 16 |
+
---
|
| 17 |
|
| 18 |
## Model Details
|
| 19 |
+
* **Framework:** JAX/Flax (NNX)
|
| 20 |
+
* **Format:** msgpack
|
| 21 |
+
* **Prediction Type:** Velocity (Flow Matching)
|
| 22 |
|
| 23 |
+
---
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
|
| 25 |
## Usage
|
| 26 |
|
| 27 |
+
### Download and Load Model
|
| 28 |
+
|
| 29 |
```python
|
| 30 |
+
import os
|
| 31 |
+
|
| 32 |
+
import jax
|
| 33 |
+
import jax.numpy as jnp
|
| 34 |
from flax import nnx, serialization
|
| 35 |
from huggingface_hub import hf_hub_download
|
|
|
|
| 36 |
|
| 37 |
# Download model weights and config
|
| 38 |
+
model_path = hf_hub_download(repo_id="{jcopo/mnist}", filename="model.msgpack")
|
| 39 |
+
config_path = hf_hub_download(repo_id="{jcopo/mnist}", filename="config.py")
|
| 40 |
|
| 41 |
# Load config to get model architecture
|
| 42 |
+
import importlib.util
|
| 43 |
spec = importlib.util.spec_from_file_location("model_config", config_path)
|
| 44 |
config_module = importlib.util.module_from_spec(spec)
|
| 45 |
spec.loader.exec_module(config_module)
|
|
|
|
| 52 |
state_dict = serialization.from_bytes(None, f.read())
|
| 53 |
|
| 54 |
# Restore weights into model
|
| 55 |
+
graphdef, state = nnx.split(model)
|
| 56 |
+
state.replace_by_pure_dict(state_dict)
|
| 57 |
+
model = nnx.merge(graphdef, state)
|
| 58 |
model.eval() # Set to evaluation mode
|
| 59 |
|
| 60 |
+
print("✅ Model loaded successfully!")
|
|
|
|
| 61 |
```
|
|
|
|
|
|
|
|
|
|
|
|