Merge pull request #1080 from bghira/main

merge masked loss + reg image fixes

helpers/data_backend/factory.py

@@ -559,15 +559,19 @@ def configure_multi_databackend(args: dict, accelerator, text_encoders, tokenize
             or backend["id"] in StateTracker.get_data_backends()
         ):
             raise ValueError("Each dataset needs a unique 'id' field.")
+        info_log(f"Configuring data backend: {backend['id']}")
+        conditioning_type = backend.get("conditioning_type")
+        if backend.get("dataset_type") == "conditioning" or conditioning_type is not None:
+            backend['dataset_type'] = 'conditioning'
         resolution_type = backend.get("resolution_type", args.resolution_type)
         if resolution_type == "pixel_area":
-            pixel_edge_length = backend.get("resolution")
+            pixel_edge_length = backend.get("resolution", int(args.resolution))
             if pixel_edge_length is None or (
                 type(pixel_edge_length) is not int
                 or not str(pixel_edge_length).isdigit()
             ):
                 raise ValueError(
-                    f"Resolution type 'pixel_area' requires a 'resolution' field to be set in the backend config using an integer in the format: 1024"
+                    f"Resolution type 'pixel_area' requires a 'resolution' field to be set in the backend config using an integer in the format: 1024, but {pixel_edge_length} was given"
                 )
             # we'll convert pixel_area to area
             backend["resolution_type"] = "area"
@@ -596,7 +600,6 @@ def configure_multi_databackend(args: dict, accelerator, text_encoders, tokenize
                     backend["minimum_image_size"] * backend["minimum_image_size"]
                 ) / 1_000_000
 
-        info_log(f"Configuring data backend: {backend['id']}")
         # Retrieve some config file overrides for commandline arguments, eg. cropping
         init_backend = init_backend_config(backend, args, accelerator)
         StateTracker.set_data_backend_config(
@@ -748,9 +751,6 @@ def configure_multi_databackend(args: dict, accelerator, text_encoders, tokenize
             repeats=init_backend["config"].get("repeats", 0),
             **metadata_backend_args,
         )
-        conditioning_type = None
-        if backend.get("dataset_type") == "conditioning":
-            conditioning_type = backend.get("conditioning_type", "controlnet")
 
         if (
             "aspect" not in args.skip_file_discovery
@@ -1066,6 +1066,7 @@ def configure_multi_databackend(args: dict, accelerator, text_encoders, tokenize
 
         if (
             not args.vae_cache_ondemand
+            and "vaecache" in init_backend
             and "vae" not in args.skip_file_discovery
             and "vae" not in backend.get("skip_file_discovery", "")
             and "deepfloyd" not in StateTracker.get_args().model_type

helpers/image_manipulation/training_sample.py

@@ -80,7 +80,7 @@ def __init__(
         self.resolution = self.data_backend_config.get("resolution")
         self.resolution_type = self.data_backend_config.get("resolution_type")
         self.target_size_calculator = resize_helpers.get(self.resolution_type)
-        if self.target_size_calculator is None:
+        if self.target_size_calculator is None and conditioning_type not in ["mask", "controlnet"]:
             raise ValueError(f"Unknown resolution type: {self.resolution_type}")
         self._set_resolution()
         self.target_downsample_size = self.data_backend_config.get(

helpers/training/trainer.py

@@ -2412,24 +2412,30 @@ def train(self):
                     # x-prediction requires that we now subtract the noise residual from the prediction to get the target sample.
                     if (
                         hasattr(self.noise_scheduler, "config")
-                        and hasattr(self.noise_scheduler.config, "prediction_type")
-                        and self.noise_scheduler.config.prediction_type == "sample"
+                        and hasattr(
+                            self.noise_scheduler.config, "prediction_type"
+                        )
+                        and self.noise_scheduler.config.prediction_type
+                        == "sample"
                     ):
                         model_pred = model_pred - noise
 
                     parent_loss = None
+
+                    # Compute the per-pixel loss without reducing over spatial dimensions
                     if self.config.flow_matching:
-                        loss = torch.mean(
-                            ((model_pred.float() - target.float()) ** 2).reshape(
-                                target.shape[0], -1
-                            ),
-                            1,
-                        )
-                    elif self.config.snr_gamma is None or self.config.snr_gamma == 0:
+                        # For flow matching, compute the per-pixel squared differences
+                        loss = (
+                            model_pred.float() - target.float()
+                        ) ** 2  # Shape: (batch_size, C, H, W)
+                    elif (
+                        self.config.snr_gamma is None
+                        or self.config.snr_gamma == 0
+                    ):
                         training_logger.debug("Calculating loss")
                         loss = self.config.snr_weight * F.mse_loss(
                             model_pred.float(), target.float(), reduction="none"
-                        )
+                        )  # Shape: (batch_size, C, H, W)
                     else:
                         # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
                         # Since we predict the noise instead of x_0, the original formulation is slightly changed.
@@ -2442,7 +2448,8 @@ def train(self):
                             == "v_prediction"
                             or (
                                 self.config.flow_matching
-                                and self.config.flow_matching_loss == "diffusion"
+                                and self.config.flow_matching_loss
+                                == "diffusion"
                             )
                         ):
                             snr_divisor = snr + 1
@@ -2454,52 +2461,62 @@ def train(self):
                             torch.stack(
                                 [
                                     snr,
-                                    self.config.snr_gamma * torch.ones_like(timesteps),
+                                    self.config.snr_gamma
+                                    * torch.ones_like(timesteps),
                                 ],
                                 dim=1,
                             ).min(dim=1)[0]
                             / snr_divisor
-                        )
+                        )  # Shape: (batch_size,)
 
-                        # We first calculate the original loss. Then we mean over the non-batch dimensions and
-                        # rebalance the sample-wise losses with their respective loss weights.
-                        # Finally, we take the mean of the rebalanced loss.
+                        # Compute the per-pixel MSE loss without reduction
                         loss = F.mse_loss(
                             model_pred.float(), target.float(), reduction="none"
-                        )
+                        )  # Shape: (batch_size, C, H, W)
+
+                        # Reshape mse_loss_weights for broadcasting and apply to loss
+                        mse_loss_weights = mse_loss_weights.view(
+                            -1, 1, 1, 1
+                        )  # Shape: (batch_size, 1, 1, 1)
                         loss = (
-                            loss.mean(dim=list(range(1, len(loss.shape))))
-                            * mse_loss_weights
-                        )
+                            loss * mse_loss_weights
+                        )  # Shape: (batch_size, C, H, W)
 
                     # Mask the loss using any conditioning data
                     conditioning_type = batch.get("conditioning_type")
                     if conditioning_type == "mask":
-                        # adapted from: https://github.com/kohya-ss/sd-scripts/blob/main/library/custom_train_functions.py#L482
+                        # Adapted from:
+                        # https://github.com/kohya-ss/sd-scripts/blob/main/library/custom_train_functions.py#L482
                         mask_image = (
                             batch["conditioning_pixel_values"]
                             .to(dtype=loss.dtype, device=loss.device)[:, 0]
                             .unsqueeze(1)
-                        )
+                        )  # Shape: (batch_size, 1, H', W')
                         mask_image = torch.nn.functional.interpolate(
                             mask_image, size=loss.shape[2:], mode="area"
-                        )
-                        mask_image = mask_image / 2 + 0.5
-                        loss = loss * mask_image
+                        )  # Resize to match loss spatial dimensions
+                        mask_image = mask_image / 2 + 0.5  # Normalize to [0,1]
+                        loss = loss * mask_image  # Element-wise multiplication
+
+                    # Reduce the loss by averaging over channels and spatial dimensions
+                    loss = loss.mean(
+                        dim=list(range(1, len(loss.shape)))
+                    )  # Shape: (batch_size,)
+
+                    # Further reduce the loss by averaging over the batch dimension
+                    loss = loss.mean()  # Scalar value
 
-                    # reduce loss now
-                    loss = loss.mean()
                     if is_regularisation_data:
                         parent_loss = loss
 
-                    # Gather the losses across all processes for logging (if we use distributed training).
+                    # Gather the losses across all processes for logging (if using distributed training)
                     avg_loss = self.accelerator.gather(
                         loss.repeat(self.config.train_batch_size)
                     ).mean()
                     self.train_loss += (
-                        avg_loss.item() / self.config.gradient_accumulation_steps
+                        avg_loss.item()
+                        / self.config.gradient_accumulation_steps
                     )
-
                     # Backpropagate
                     grad_norm = None
                     if not self.config.disable_accelerator:

toolkit/datasets/masked_loss/generate_dataset_masks.py

@@ -1,7 +1,96 @@
 import argparse
 import os
-import shutil
-from gradio_client import Client, handle_file
+
+os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
+from typing import Union, Any, Tuple, Dict
+from unittest.mock import patch
+
+import numpy as np
+import supervision as sv
+import torch
+from PIL import Image
+from gradio_client import handle_file
+from transformers import AutoModelForCausalLM, AutoProcessor
+from transformers.dynamic_module_utils import get_imports
+
+from sam2.build_sam import build_sam2
+from sam2.sam2_image_predictor import SAM2ImagePredictor
+
+
+# Constants
+FLORENCE_CHECKPOINT = "microsoft/Florence-2-large"
+FLORENCE_OPEN_VOCABULARY_DETECTION_TASK = "<OPEN_VOCABULARY_DETECTION>"
+
+SAM_CONFIG = "sam2_hiera_l.yaml"
+SAM_CHECKPOINT = "checkpoints/sam2_hiera_large.pt"
+
+
+def load_sam_image_model(
+    device: torch.device, config: str = SAM_CONFIG, checkpoint: str = SAM_CHECKPOINT
+) -> SAM2ImagePredictor:
+    model = build_sam2(config, checkpoint, device=device)
+    return SAM2ImagePredictor(sam_model=model)
+
+
+def run_sam_inference(
+    model: Any, image: Image.Image, detections: sv.Detections
+) -> sv.Detections:
+    image_np = np.array(image.convert("RGB"))
+    model.set_image(image_np)
+    masks, scores, _ = model.predict(box=detections.xyxy, multimask_output=False)
+
+    # Ensure mask dimensions are correct
+    if len(masks.shape) == 4:
+        masks = np.squeeze(masks)
+
+    detections.mask = masks.astype(bool)
+    return detections
+
+
+def fixed_get_imports(filename: Union[str, os.PathLike]) -> list[str]:
+    """Workaround for specific import issues."""
+    if not str(filename).endswith("/modeling_florence2.py"):
+        return get_imports(filename)
+    imports = get_imports(filename)
+    if "flash_attn" in imports:
+        imports.remove("flash_attn")
+    return imports
+
+
+def load_florence_model(
+    device: torch.device, checkpoint: str = FLORENCE_CHECKPOINT
+) -> Tuple[Any, Any]:
+    with patch("transformers.dynamic_module_utils.get_imports", fixed_get_imports):
+        model = (
+            AutoModelForCausalLM.from_pretrained(checkpoint, trust_remote_code=True)
+            .to(device)
+            .eval()
+        )
+        processor = AutoProcessor.from_pretrained(checkpoint, trust_remote_code=True)
+        return model, processor
+
+
+def run_florence_inference(
+    model: Any,
+    processor: Any,
+    device: torch.device,
+    image: Image.Image,
+    task: str,
+    text: str = "",
+) -> Tuple[str, Dict]:
+    prompt = task + text
+    inputs = processor(text=prompt, images=image, return_tensors="pt").to(device)
+    generated_ids = model.generate(
+        input_ids=inputs["input_ids"],
+        pixel_values=inputs["pixel_values"],
+        max_new_tokens=1024,
+        num_beams=3,
+    )
+    generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
+    response = processor.post_process_generation(
+        generated_text, task=task, image_size=image.size
+    )
+    return generated_text, response
 
 
 def main():
@@ -27,25 +116,41 @@ def main():
         default="person",
         help='Text prompt for masking (default: "person").',
     )
-    parser.add_argument(
-        "--model",
-        type=str,
-        default="SkalskiP/florence-sam-masking",
-        help='Model name to use (default: "SkalskiP/florence-sam-masking").',
-    )
     args = parser.parse_args()
+    if args.input_dir is None or args.output_dir is None:
+        import sys
+
+        sys.exit(1)
 
     input_path = args.input_dir
     output_path = args.output_dir
     text_input = args.text_input
-    model_name = args.model
+
+    DEVICE = torch.device(
+        "cuda"
+        if torch.cuda.is_available()
+        else "mps" if torch.backends.mps.is_available() else "cpu"
+    )
+
+    # Retrieve model
+    from huggingface_hub import hf_hub_download
+
+    print(f"Downloading SAM2 to {os.getcwd()}/checkpoints.")
+    hf_hub_download(
+        "SkalskiP/florence-sam-masking",
+        repo_type="space",
+        subfolder="checkpoints",
+        local_dir="./",
+        filename="sam2_hiera_large.pt",
+    )
+
+    # Load models
+    FLORENCE_MODEL, FLORENCE_PROCESSOR = load_florence_model(device=DEVICE)
+    SAM_IMAGE_MODEL = load_sam_image_model(device=DEVICE)
 
     # Create the output directory if it doesn't exist
     os.makedirs(output_path, exist_ok=True)
 
-    # Initialize the Gradio client
-    client = Client(model_name)
-
     # Get all files in the input directory
     files = os.listdir(input_path)
 
@@ -65,16 +170,36 @@ def main():
                 continue
             # Predict the mask
             try:
-                mask_filename = client.predict(
-                    image_input=handle_file(full_path),
-                    text_input=text_input,
-                    api_name="/process_image",
+                image_input = Image.open(full_path)
+                # cast to RGB
+                image_input = image_input.convert(
+                    "RGB", dither=None, palette=Image.ADAPTIVE
+                )
+                _, result = run_florence_inference(
+                    model=FLORENCE_MODEL,
+                    processor=FLORENCE_PROCESSOR,
+                    device=DEVICE,
+                    image=image_input,
+                    task=FLORENCE_OPEN_VOCABULARY_DETECTION_TASK,
+                    text=text_input,
+                )
+                detections = sv.Detections.from_lmm(
+                    lmm=sv.LMM.FLORENCE_2, result=result, resolution_wh=image_input.size
                 )
-                # Move the generated mask to the output directory
-                shutil.move(mask_filename, mask_path)
+                if len(detections) == 0:
+                    print(f"No objects detected in {file}.")
+                    continue
+                detections = run_sam_inference(SAM_IMAGE_MODEL, image_input, detections)
+                # Combine masks if multiple detections
+                combined_mask = np.any(detections.mask, axis=0)
+                mask_image = Image.fromarray(combined_mask.astype("uint8") * 255)
+                mask_image.save(mask_path)
                 print(f"Saved mask to {mask_path}")
             except Exception as e:
                 print(f"Failed to process {file}: {e}")
+    # Clean up
+    os.remove("checkpoints/sam2_hiera_large.pt")
+    os.rmdir("checkpoints")
 
 
 if __name__ == "__main__":