diff --git a/.gitea/workflows/embed.yaml b/.gitea/workflows/embed.yaml index 22c5bda..091db58 100644 --- a/.gitea/workflows/embed.yaml +++ b/.gitea/workflows/embed.yaml @@ -34,4 +34,4 @@ jobs: VECTORDB_TOKEN: ${{ secrets.VECTORDB_TOKEN }} run: | cd VectorLoader - python -m src.run --full + python -m src.run diff --git a/setup.py b/setup.py index b6c17c3..769f912 100644 --- a/setup.py +++ b/setup.py @@ -2,7 +2,7 @@ from setuptools import setup, find_packages setup( name="aiunn", - version="0.1.2", + version="0.2.0", packages=find_packages(where="src"), package_dir={"": "src"}, install_requires=[ diff --git a/src/aiunn/__init__.py b/src/aiunn/__init__.py index 16f468f..4a57351 100644 --- a/src/aiunn/__init__.py +++ b/src/aiunn/__init__.py @@ -3,4 +3,4 @@ from .upsampler.aiunn import aiuNN from .upsampler.config import aiuNNConfig from .inference.inference import aiuNNInference -__version__ = "0.1.2" \ No newline at end of file +__version__ = "0.2.0" \ No newline at end of file diff --git a/src/aiunn/inference/inference.py b/src/aiunn/inference/inference.py index d288931..1ff5994 100644 --- a/src/aiunn/inference/inference.py +++ b/src/aiunn/inference/inference.py @@ -12,13 +12,13 @@ class aiuNNInference: Inference class for aiuNN upsampling model. Handles model loading, image upscaling, and output processing. """ - def __init__(self, model_path: str, precision: Optional[str] = None, device: Optional[str] = None): + def __init__(self, model_path: str, device: Optional[str] = None): """ Initialize the inference class by loading the aiuNN model. Args: model_path: Path to the saved model directory - precision: Optional precision setting ('fp16', 'bf16', or None for default) + device: Optional device specification ('cuda', 'cpu', or None for auto-detection) """ @@ -30,7 +30,7 @@ class aiuNNInference: self.device = device # Load the model with specified precision - self.model = aiuNN.load(model_path, precision=precision) + self.model = aiuNN.from_pretrained(model_path) self.model.to(self.device) self.model.eval() @@ -160,54 +160,11 @@ class aiuNNInference: return binary_data - def process_batch(self, - images: List[Union[str, Image.Image]], - output_dir: Optional[str] = None, - save_format: str = 'PNG', - return_binary: bool = False) -> Union[List[Image.Image], List[bytes], None]: - """ - Process multiple images in batch. - - Args: - images: List of input images (paths or PIL Images) - output_dir: Optional directory to save results - save_format: Format to use when saving images - return_binary: Whether to return binary data instead of PIL Images - - Returns: - List of processed images or binary data, or None if only saving - """ - results = [] - - for i, img in enumerate(images): - # Upscale the image - upscaled = self.upscale(img) - - # Save if output directory is provided - if output_dir: - # Extract filename if input is a path - if isinstance(img, str): - filename = os.path.basename(img) - base, _ = os.path.splitext(filename) - else: - base = f"upscaled_{i}" - - output_path = os.path.join(output_dir, f"{base}.{save_format.lower()}") - self.save(upscaled, output_path, format=save_format) - - # Add to results based on return type - if return_binary: - results.append(self.convert_to_binary(upscaled, format=save_format)) - else: - results.append(upscaled) - - return results if (not output_dir or return_binary or not save_format) else None - # Example usage (can be removed) if __name__ == "__main__": # Initialize inference with a model path - inferencer = aiuNNInference("path/to/model", precision="bf16") + inferencer = aiuNNInference("path/to/model") # Upscale a single image upscaled_image = inferencer.upscale("input_image.jpg") @@ -217,10 +174,4 @@ if __name__ == "__main__": # Convert to binary binary_data = inferencer.convert_to_binary(upscaled_image) - - # Process a batch of images - inferencer.process_batch( - ["image1.jpg", "image2.jpg"], - output_dir="output_folder", - save_format="PNG" - ) \ No newline at end of file + \ No newline at end of file diff --git a/src/aiunn/upsampler/aiunn.py b/src/aiunn/upsampler/aiunn.py index ecb3c21..71c77f3 100644 --- a/src/aiunn/upsampler/aiunn.py +++ b/src/aiunn/upsampler/aiunn.py @@ -2,19 +2,19 @@ import os import torch import torch.nn as nn import warnings -from aiia.model.Model import AIIA, AIIAConfig, AIIABase +from aiia.model.Model import AIIAConfig, AIIABase +from transformers import PreTrainedModel from .config import aiuNNConfig import warnings -class aiuNN(AIIA): - def __init__(self, base_model: AIIA, config:aiuNNConfig): - super().__init__(base_model.config) - self.base_model = base_model - +class aiuNN(PreTrainedModel): + config_class = aiuNNConfig + def __init__(self, config: aiuNNConfig): + super().__init__(config) # Pass the unified base configuration using the new parameter. self.config = config - + # Enhanced approach scale_factor = self.config.upsample_scale out_channels = self.base_model.config.num_channels * (scale_factor ** 2) @@ -26,118 +26,18 @@ class aiuNN(AIIA): ) self.pixel_shuffle = nn.PixelShuffle(scale_factor) - + def load_base_model(self, base_model: PreTrainedModel): + self.base_model = base_model + def forward(self, x): + if self.base_model is None: + raise ValueError("Base model is not loaded. Call 'load_base_model' before forwarding.") x = self.base_model(x) # Get base features x = self.pixel_shuffle_conv(x) # Expand channels for shuffling x = self.pixel_shuffle(x) # Rearrange channels into spatial dimensions return x - @classmethod - def load(cls, path, precision: str = None, **kwargs): - """ - Load a aiuNN model from disk with automatic detection of base model type. - - Args: - path (str): Directory containing the stored configuration and model parameters. - precision (str, optional): Desired precision for the model's parameters. - **kwargs: Additional keyword arguments to override configuration parameters. - - Returns: - An instance of aiuNN with loaded weights. - """ - # Load the configuration - config = aiuNNConfig.load(path) - - # Determine the device - device = 'cuda' if torch.cuda.is_available() else 'cpu' - - # Load the state dictionary - state_dict = torch.load(os.path.join(path, "model.pth"), map_location=device) - - # Import all model types - from aiia.model.Model import AIIABase, AIIABaseShared, AIIAExpert, AIIAmoe, AIIAchunked, AIIArecursive - - # Helper function to detect base class type from key patterns - def detect_base_class_type(keys_prefix): - if any(f"{keys_prefix}.shared_layer" in key for key in state_dict.keys()): - return AIIABaseShared - else: - return AIIABase - - # Detect base model type - base_model = None - - # Check for AIIAmoe with multiple experts - if any("base_model.experts" in key for key in state_dict.keys()): - # Count the number of experts - max_expert_idx = -1 - for key in state_dict.keys(): - if "base_model.experts." in key: - try: - parts = key.split("base_model.experts.")[1].split(".") - expert_idx = int(parts[0]) - max_expert_idx = max(max_expert_idx, expert_idx) - except (ValueError, IndexError): - pass - - if max_expert_idx >= 0: - # Determine the type of base_cnn each expert is using - base_class_for_experts = detect_base_class_type("base_model.experts.0.base_cnn") - - # Create AIIAmoe with the detected expert count and base class - base_model = AIIAmoe(config, num_experts=max_expert_idx+1, base_class=base_class_for_experts, **kwargs) - - # Check for AIIAchunked or AIIArecursive - elif any("base_model.chunked_cnn" in key for key in state_dict.keys()): - if any("recursion_depth" in key for key in state_dict.keys()): - # This is an AIIArecursive model - base_class = detect_base_class_type("base_model.chunked_cnn.base_cnn") - base_model = AIIArecursive(config, base_class=base_class, **kwargs) - else: - # This is an AIIAchunked model - base_class = detect_base_class_type("base_model.chunked_cnn.base_cnn") - base_model = AIIAchunked(config, base_class=base_class, **kwargs) - - # Check for AIIAExpert - elif any("base_model.base_cnn" in key for key in state_dict.keys()): - # Determine which base class the expert is using - base_class = detect_base_class_type("base_model.base_cnn") - base_model = AIIAExpert(config, base_class=base_class, **kwargs) - - # If none of the above, use AIIABase or AIIABaseShared directly - else: - base_class = detect_base_class_type("base_model") - base_model = base_class(config, **kwargs) - - # Create the aiuNN model with the detected base model - model = cls(base_model, config=base_model.config) - - # Handle precision conversion - dtype = None - if precision is not None: - if precision.lower() == 'fp16': - dtype = torch.float16 - elif precision.lower() == 'bf16': - if device == 'cuda' and not torch.cuda.is_bf16_supported(): - warnings.warn("BF16 is not supported on this GPU. Falling back to FP16.") - dtype = torch.float16 - else: - dtype = torch.bfloat16 - else: - raise ValueError("Unsupported precision. Use 'fp16', 'bf16', or leave as None.") - - if dtype is not None: - for key, param in state_dict.items(): - if torch.is_tensor(param): - state_dict[key] = param.to(dtype) - - # Load the state dict - model.load_state_dict(state_dict) - return model - - if __name__ == "__main__": from aiia import AIIABase, AIIAConfig @@ -146,11 +46,11 @@ if __name__ == "__main__": ai_config = aiuNNConfig() base_model = AIIABase(config) # Instantiate Upsampler from the base model (works correctly). - upsampler = aiuNN(base_model, config=ai_config) - + upsampler = aiuNN(config=ai_config) + upsampler.load_base_model(base_model) # Save the model (both configuration and weights). - upsampler.save("aiunn") + upsampler.save_pretrained("aiunn") # Now load using the overridden load method; this will load the complete model. - upsampler_loaded = aiuNN.load("aiunn", precision="bf16") + upsampler_loaded = aiuNN.from_pretrained("aiunn") print("Updated configuration:", upsampler_loaded.config.__dict__) diff --git a/tests/inference/test_inference.py b/tests/inference/test_inference.py index 39dea9a..5583d75 100644 --- a/tests/inference/test_inference.py +++ b/tests/inference/test_inference.py @@ -21,9 +21,8 @@ def real_model(tmp_path): base_model = AIIABase(config) # Make sure aiuNN is properly configured with all required attributes - upsampler = aiuNN(base_model, config=ai_config) - # Ensure the upsample attribute is properly set if needed - # upsampler.upsample = ... # Add any necessary initialization + upsampler = aiuNN(config=ai_config) + upsampler.load_base_model(base_model) # Save the model and config to temporary directory save_path = str(model_dir / "save") @@ -40,10 +39,10 @@ def real_model(tmp_path): json.dump(config_data, f) # Save model - upsampler.save(save_path) + upsampler.save_pretrained(save_path) # Load model in inference mode - inference_model = aiuNNInference(model_path=save_path, precision='fp16', device='cpu') + inference_model = aiuNNInference(model_path=save_path, device='cpu') return inference_model @@ -88,12 +87,3 @@ def test_convert_to_binary(inference): result = inference.convert_to_binary(test_image) assert isinstance(result, bytes) assert len(result) > 0 - -def test_process_batch(inference): - # Create test images - test_array = np.zeros((100, 100, 3), dtype=np.uint8) - test_images = [Image.fromarray(test_array) for _ in range(2)] - - results = inference.process_batch(test_images) - assert len(results) == 2 - assert all(isinstance(img, Image.Image) for img in results) \ No newline at end of file diff --git a/tests/upsampler/test_aiunn.py b/tests/upsampler/test_aiunn.py index aae0813..cdf11bc 100644 --- a/tests/upsampler/test_aiunn.py +++ b/tests/upsampler/test_aiunn.py @@ -10,39 +10,21 @@ def test_save_and_load_model(): config = AIIAConfig() ai_config = aiuNNConfig() base_model = AIIABase(config) - upsampler = aiuNN(base_model, config=ai_config) - + upsampler = aiuNN(config=ai_config) + upsampler.load_base_model(base_model) # Save the model save_path = os.path.join(tmpdirname, "model") - upsampler.save(save_path) + upsampler.save_pretrained(save_path) # Load the model - loaded_upsampler = aiuNN.load(save_path) + loaded_upsampler = aiuNN.from_pretrained(save_path) # Verify that the loaded model is the same as the original model assert isinstance(loaded_upsampler, aiuNN) - assert loaded_upsampler.config.__dict__ == upsampler.config.__dict__ + assert loaded_upsampler.config.hidden_size == upsampler.config.hidden_size + assert loaded_upsampler.config._activation_function == upsampler.config._activation_function + assert loaded_upsampler.config.architectures == upsampler.config.architectures -def test_save_and_load_model_with_precision(): - # Create a temporary directory to save the model - with tempfile.TemporaryDirectory() as tmpdirname: - # Create configurations and build a base model - config = AIIAConfig() - ai_config = aiuNNConfig() - base_model = AIIABase(config) - upsampler = aiuNN(base_model, config=ai_config) - - # Save the model - save_path = os.path.join(tmpdirname, "model") - upsampler.save(save_path) - - # Load the model with precision 'bf16' - loaded_upsampler = aiuNN.load(save_path, precision="bf16") - - # Verify that the loaded model is the same as the original model - assert isinstance(loaded_upsampler, aiuNN) - assert loaded_upsampler.config.__dict__ == upsampler.config.__dict__ if __name__ == "__main__": test_save_and_load_model() - test_save_and_load_model_with_precision() \ No newline at end of file