flaxdiff 0.2.3__tar.gz → 0.2.5__tar.gz

{flaxdiff-0.2.3 → flaxdiff-0.2.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: flaxdiff
-Version: 0.2.3
+Version: 0.2.5
 Summary: A versatile and easy to understand Diffusion library
 Author-email: Ashish Kumar Singh <ashishkmr472@gmail.com>
 License-Expression: MIT

{flaxdiff-0.2.3 → flaxdiff-0.2.5}/flaxdiff/data/dataloaders.py

@@ -258,7 +258,7 @@ def get_dataset_grain(
     image_scale=256,
     count=None,
     num_epochs=None,
-    method=jax.image.ResizeMethod.LANCZOS3,
+    method=None, #jax.image.ResizeMethod.LANCZOS3,
     worker_count=32,
     read_thread_count=64,
     read_buffer_size=50,

{flaxdiff-0.2.3 → flaxdiff-0.2.5}/flaxdiff/inference/pipeline.py

@@ -208,7 +208,8 @@ class DiffusionInferencePipeline(InferencePipeline):
         self,
         num_samples: int,
         resolution: int,
-        conditioning_data: Optional[List[Union[Tuple, Dict]]] = None,  # one list per modality or list of tuples
+        conditioning_data: List[Union[Tuple, Dict]] = None,
+        conditioning_data_tokens: Tuple = None,
         sequence_length: Optional[int] = None,
         diffusion_steps: int = 50,
         guidance_scale: float = 1.0,
@@ -256,5 +257,6 @@ class DiffusionInferencePipeline(InferencePipeline):
             steps_override=steps_override,
             priors=priors,
             rngstate=rngstate,
-            conditioning=conditioning_data
+            conditioning=conditioning_data,
+            model_conditioning_inputs=conditioning_data_tokens,
         )

flaxdiff-0.2.5/flaxdiff/metrics/common.py

@@ -0,0 +1,11 @@
+from typing import Callable
+from dataclasses import dataclass
+
+@dataclass
+class EvaluationMetric:
+    """
+    Evaluation metrics for the diffusion model.
+    The function is given generated samples batch [B, H, W, C] and the original batch.
+    """
+    function: Callable
+    name: str

flaxdiff-0.2.5/flaxdiff/metrics/images.py

@@ -0,0 +1,59 @@
+from .common import EvaluationMetric
+import jax
+import jax.numpy as jnp
+    
+def get_clip_metric(
+    modelname: str = "openai/clip-vit-large-patch14",
+):
+    from transformers import AutoProcessor, FlaxCLIPModel
+    model = FlaxCLIPModel.from_pretrained(modelname, dtype=jnp.float16)
+    processor = AutoProcessor.from_pretrained(modelname, use_fast=False, dtype=jnp.float16)
+    
+    @jax.jit
+    def calc(pixel_values, input_ids, attention_mask):
+        # Get the logits
+        generated_out = model(
+            pixel_values=pixel_values,
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+        )
+            
+        gen_img_emb = generated_out.image_embeds
+        txt_emb = generated_out.text_embeds
+
+        # 1. Normalize embeddings (essential for cosine similarity/distance)
+        gen_img_emb = gen_img_emb / (jnp.linalg.norm(gen_img_emb, axis=-1, keepdims=True) + 1e-6)
+        txt_emb = txt_emb / (jnp.linalg.norm(txt_emb, axis=-1, keepdims=True) + 1e-6)
+
+        # 2. Calculate cosine similarity
+        # Using einsum for batch dot product: batch (b), embedding_dim (d) -> bd,bd->b
+        # Calculate cosine similarity
+        similarity = jnp.einsum('bd,bd->b', gen_img_emb, txt_emb)
+
+        scaled_distance = (1.0 - similarity)
+        # 4. Average over the batch
+        mean_scaled_distance = jnp.mean(scaled_distance)
+
+        return mean_scaled_distance
+        
+    def clip_metric(
+        generated: jnp.ndarray,
+        batch
+    ):
+        original_conditions = batch['text']
+        
+        # Convert samples from [-1, 1] to [0, 255] and uint8
+        generated = (((generated + 1.0) / 2.0) * 255).astype(jnp.uint8)
+        
+        generated_inputs = processor(images=generated, return_tensors="jax", padding=True,)
+        
+        pixel_values = generated_inputs['pixel_values']
+        input_ids = original_conditions['input_ids']
+        attention_mask = original_conditions['attention_mask']
+        
+        return calc(pixel_values, input_ids, attention_mask)
+    
+    return EvaluationMetric(
+        function=clip_metric,
+        name='clip_similarity'
+    )

{flaxdiff-0.2.3 → flaxdiff-0.2.5}/flaxdiff/trainer/general_diffusion_trainer.py

@@ -27,6 +27,8 @@ from flax.training import dynamic_scale as dynamic_scale_lib
 from .diffusion_trainer import TrainState, DiffusionTrainer
 import shutil
 
+from flaxdiff.metrics.common import EvaluationMetric
+
 def generate_modelname(
     dataset_name: str,
     noise_schedule_name: str,
@@ -103,15 +105,6 @@ def generate_modelname(
     # model_name = f"{model_name}-{config_hash}"
     return model_name
 
-@dataclass
-class EvaluationMetric:
-    """
-    Evaluation metrics for the diffusion model.
-    The function is given generated samples batch [B, H, W, C] and the original batch.
-    """
-    function: Callable
-    name: str
-
 class GeneralDiffusionTrainer(DiffusionTrainer):
     """
     General trainer for diffusion models supporting both images and videos.

{flaxdiff-0.2.3 → flaxdiff-0.2.5}/flaxdiff.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: flaxdiff
-Version: 0.2.3
+Version: 0.2.5
 Summary: A versatile and easy to understand Diffusion library
 Author-email: Ashish Kumar Singh <ashishkmr472@gmail.com>
 License-Expression: MIT

{flaxdiff-0.2.3 → flaxdiff-0.2.5}/flaxdiff.egg-info/SOURCES.txt

@@ -25,6 +25,9 @@ flaxdiff/inference/pipeline.py
 flaxdiff/inference/utils.py
 flaxdiff/inputs/__init__.py
 flaxdiff/inputs/encoders.py
+flaxdiff/metrics/__init__.py
+flaxdiff/metrics/common.py
+flaxdiff/metrics/images.py
 flaxdiff/metrics/inception.py
 flaxdiff/metrics/psnr.py
 flaxdiff/metrics/ssim.py

{flaxdiff-0.2.3 → flaxdiff-0.2.5}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "flaxdiff"
-version = "0.2.3"
+version = "0.2.5"
 description = "A versatile and easy to understand Diffusion library"
 readme = "README.md"
 authors = [