Merge pull request #1992 from rockerBOO/flux-ip-noise-gamma

Add IP noise gamma for Flux

Author: kohya-ss

library/flux_train_utils.py

@@ -366,8 +366,6 @@ def get_sigmas(noise_scheduler, timesteps, device, n_dim=4, dtype=torch.float32)
     step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
 
     sigma = sigmas[step_indices].flatten()
-    while len(sigma.shape) < n_dim:
-        sigma = sigma.unsqueeze(-1)
     return sigma
 
 
@@ -410,42 +408,34 @@ def compute_loss_weighting_for_sd3(weighting_scheme: str, sigmas=None):
 
 
 def get_noisy_model_input_and_timesteps(
-    args, noise_scheduler, latents, noise, device, dtype
+    args, noise_scheduler, latents: torch.Tensor, noise: torch.Tensor, device, dtype
 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
     bsz, _, h, w = latents.shape
-    sigmas = None
-
+    assert bsz > 0, "Batch size not large enough"
+    num_timesteps = noise_scheduler.config.num_train_timesteps
     if args.timestep_sampling == "uniform" or args.timestep_sampling == "sigmoid":
-        # Simple random t-based noise sampling
+        # Simple random sigma-based noise sampling
         if args.timestep_sampling == "sigmoid":
             # https://github.com/XLabs-AI/x-flux/tree/main
-            t = torch.sigmoid(args.sigmoid_scale * torch.randn((bsz,), device=device))
+            sigmas = torch.sigmoid(args.sigmoid_scale * torch.randn((bsz,), device=device))
         else:
-            t = torch.rand((bsz,), device=device)
+            sigmas = torch.rand((bsz,), device=device)
 
-        timesteps = t * 1000.0
-        t = t.view(-1, 1, 1, 1)
-        noisy_model_input = (1 - t) * latents + t * noise
+        timesteps = sigmas * num_timesteps
     elif args.timestep_sampling == "shift":
         shift = args.discrete_flow_shift
-        logits_norm = torch.randn(bsz, device=device)
-        logits_norm = logits_norm * args.sigmoid_scale  # larger scale for more uniform sampling
-        timesteps = logits_norm.sigmoid()
-        timesteps = (timesteps * shift) / (1 + (shift - 1) * timesteps)
-
-        t = timesteps.view(-1, 1, 1, 1)
-        timesteps = timesteps * 1000.0
-        noisy_model_input = (1 - t) * latents + t * noise
+        sigmas = torch.randn(bsz, device=device)
+        sigmas = sigmas * args.sigmoid_scale  # larger scale for more uniform sampling
+        sigmas = sigmas.sigmoid()
+        sigmas = (sigmas * shift) / (1 + (shift - 1) * sigmas)
+        timesteps = sigmas * num_timesteps
     elif args.timestep_sampling == "flux_shift":
-        logits_norm = torch.randn(bsz, device=device)
-        logits_norm = logits_norm * args.sigmoid_scale  # larger scale for more uniform sampling
-        timesteps = logits_norm.sigmoid()
-        mu = get_lin_function(y1=0.5, y2=1.15)((h // 2) * (w // 2))
-        timesteps = time_shift(mu, 1.0, timesteps)
-
-        t = timesteps.view(-1, 1, 1, 1)
-        timesteps = timesteps * 1000.0
-        noisy_model_input = (1 - t) * latents + t * noise
+        sigmas = torch.randn(bsz, device=device)
+        sigmas = sigmas * args.sigmoid_scale  # larger scale for more uniform sampling
+        sigmas = sigmas.sigmoid()
+        mu = get_lin_function(y1=0.5, y2=1.15)((h // 2) * (w // 2)) # we are pre-packed so must adjust for packed size 
+        sigmas = time_shift(mu, 1.0, sigmas)
+        timesteps = sigmas * num_timesteps
     else:
         # Sample a random timestep for each image
         # for weighting schemes where we sample timesteps non-uniformly
@@ -456,12 +446,24 @@ def get_noisy_model_input_and_timesteps(
             logit_std=args.logit_std,
             mode_scale=args.mode_scale,
         )
-        indices = (u * noise_scheduler.config.num_train_timesteps).long()
+        indices = (u * num_timesteps).long()
         timesteps = noise_scheduler.timesteps[indices].to(device=device)
-
-        # Add noise according to flow matching.
         sigmas = get_sigmas(noise_scheduler, timesteps, device, n_dim=latents.ndim, dtype=dtype)
-        noisy_model_input = sigmas * noise + (1.0 - sigmas) * latents
+
+    # Broadcast sigmas to latent shape
+    sigmas = sigmas.view(-1, 1, 1, 1)
+
+    # Add noise to the latents according to the noise magnitude at each timestep
+    # (this is the forward diffusion process)
+    if args.ip_noise_gamma:
+        xi = torch.randn_like(latents, device=latents.device, dtype=dtype)
+        if args.ip_noise_gamma_random_strength:
+            ip_noise_gamma = (torch.rand(1, device=latents.device, dtype=dtype) * args.ip_noise_gamma)
+        else:
+            ip_noise_gamma = args.ip_noise_gamma
+        noisy_model_input = (1.0 - sigmas) * latents + sigmas * (noise + ip_noise_gamma * xi)
+    else:
+        noisy_model_input = (1.0 - sigmas) * latents + sigmas * noise
 
     return noisy_model_input.to(dtype), timesteps.to(dtype), sigmas
 

tests/library/test_flux_train_utils.py

@@ -0,0 +1,220 @@
+import pytest
+import torch
+from unittest.mock import MagicMock, patch
+from library.flux_train_utils import (
+    get_noisy_model_input_and_timesteps,
+)
+
+# Mock classes and functions
+class MockNoiseScheduler:
+    def __init__(self, num_train_timesteps=1000):
+        self.config = MagicMock()
+        self.config.num_train_timesteps = num_train_timesteps
+        self.timesteps = torch.arange(num_train_timesteps, dtype=torch.long)
+
+
+# Create fixtures for commonly used objects
+@pytest.fixture
+def args():
+    args = MagicMock()
+    args.timestep_sampling = "uniform"
+    args.weighting_scheme = "uniform"
+    args.logit_mean = 0.0
+    args.logit_std = 1.0
+    args.mode_scale = 1.0
+    args.sigmoid_scale = 1.0
+    args.discrete_flow_shift = 3.1582
+    args.ip_noise_gamma = None
+    args.ip_noise_gamma_random_strength = False
+    return args
+
+
+@pytest.fixture
+def noise_scheduler():
+    return MockNoiseScheduler(num_train_timesteps=1000)
+
+
+@pytest.fixture
+def latents():
+    return torch.randn(2, 4, 8, 8)
+
+
+@pytest.fixture
+def noise():
+    return torch.randn(2, 4, 8, 8)
+
+
+@pytest.fixture
+def device():
+    # return "cuda" if torch.cuda.is_available() else "cpu"
+    return "cpu"
+
+
+# Mock the required functions
+@pytest.fixture(autouse=True)
+def mock_functions():
+    with (
+        patch("torch.sigmoid", side_effect=torch.sigmoid),
+        patch("torch.rand", side_effect=torch.rand),
+        patch("torch.randn", side_effect=torch.randn),
+    ):
+        yield
+
+
+# Test different timestep sampling methods
+def test_uniform_sampling(args, noise_scheduler, latents, noise, device):
+    args.timestep_sampling = "uniform"
+    dtype = torch.float32
+
+    noisy_input, timesteps, sigmas = get_noisy_model_input_and_timesteps(args, noise_scheduler, latents, noise, device, dtype)
+
+    assert noisy_input.shape == latents.shape
+    assert timesteps.shape == (latents.shape[0],)
+    assert sigmas.shape == (latents.shape[0], 1, 1, 1)
+    assert noisy_input.dtype == dtype
+    assert timesteps.dtype == dtype
+
+
+def test_sigmoid_sampling(args, noise_scheduler, latents, noise, device):
+    args.timestep_sampling = "sigmoid"
+    args.sigmoid_scale = 1.0
+    dtype = torch.float32
+
+    noisy_input, timesteps, sigmas = get_noisy_model_input_and_timesteps(args, noise_scheduler, latents, noise, device, dtype)
+
+    assert noisy_input.shape == latents.shape
+    assert timesteps.shape == (latents.shape[0],)
+    assert sigmas.shape == (latents.shape[0], 1, 1, 1)
+
+
+def test_shift_sampling(args, noise_scheduler, latents, noise, device):
+    args.timestep_sampling = "shift"
+    args.sigmoid_scale = 1.0
+    args.discrete_flow_shift = 3.1582
+    dtype = torch.float32
+
+    noisy_input, timesteps, sigmas = get_noisy_model_input_and_timesteps(args, noise_scheduler, latents, noise, device, dtype)
+
+    assert noisy_input.shape == latents.shape
+    assert timesteps.shape == (latents.shape[0],)
+    assert sigmas.shape == (latents.shape[0], 1, 1, 1)
+
+
+def test_flux_shift_sampling(args, noise_scheduler, latents, noise, device):
+    args.timestep_sampling = "flux_shift"
+    args.sigmoid_scale = 1.0
+    dtype = torch.float32
+
+    noisy_input, timesteps, sigmas = get_noisy_model_input_and_timesteps(args, noise_scheduler, latents, noise, device, dtype)
+
+    assert noisy_input.shape == latents.shape
+    assert timesteps.shape == (latents.shape[0],)
+    assert sigmas.shape == (latents.shape[0], 1, 1, 1)
+
+
+def test_weighting_scheme(args, noise_scheduler, latents, noise, device):
+    # Mock the necessary functions for this specific test
+    with patch("library.flux_train_utils.compute_density_for_timestep_sampling", 
+               return_value=torch.tensor([0.3, 0.7], device=device)), \
+         patch("library.flux_train_utils.get_sigmas", 
+               return_value=torch.tensor([[0.3], [0.7]], device=device).view(-1, 1, 1, 1)):
+               
+        args.timestep_sampling = "other"  # Will trigger the weighting scheme path
+        args.weighting_scheme = "uniform"
+        args.logit_mean = 0.0
+        args.logit_std = 1.0
+        args.mode_scale = 1.0
+        dtype = torch.float32
+        
+        noisy_input, timesteps, sigmas = get_noisy_model_input_and_timesteps(
+            args, noise_scheduler, latents, noise, device, dtype
+        )
+        
+        assert noisy_input.shape == latents.shape
+        assert timesteps.shape == (latents.shape[0],)
+        assert sigmas.shape == (latents.shape[0], 1, 1, 1)
+
+
+# Test IP noise options
+def test_with_ip_noise(args, noise_scheduler, latents, noise, device):
+    args.ip_noise_gamma = 0.5
+    args.ip_noise_gamma_random_strength = False
+    dtype = torch.float32
+
+    noisy_input, timesteps, sigmas = get_noisy_model_input_and_timesteps(args, noise_scheduler, latents, noise, device, dtype)
+
+    assert noisy_input.shape == latents.shape
+    assert timesteps.shape == (latents.shape[0],)
+    assert sigmas.shape == (latents.shape[0], 1, 1, 1)
+
+
+def test_with_random_ip_noise(args, noise_scheduler, latents, noise, device):
+    args.ip_noise_gamma = 0.1
+    args.ip_noise_gamma_random_strength = True
+    dtype = torch.float32
+
+    noisy_input, timesteps, sigmas = get_noisy_model_input_and_timesteps(args, noise_scheduler, latents, noise, device, dtype)
+
+    assert noisy_input.shape == latents.shape
+    assert timesteps.shape == (latents.shape[0],)
+    assert sigmas.shape == (latents.shape[0], 1, 1, 1)
+
+
+# Test different data types
+def test_float16_dtype(args, noise_scheduler, latents, noise, device):
+    dtype = torch.float16
+
+    noisy_input, timesteps, sigmas = get_noisy_model_input_and_timesteps(args, noise_scheduler, latents, noise, device, dtype)
+
+    assert noisy_input.dtype == dtype
+    assert timesteps.dtype == dtype
+
+
+# Test different batch sizes
+def test_different_batch_size(args, noise_scheduler, device):
+    latents = torch.randn(5, 4, 8, 8)  # batch size of 5
+    noise = torch.randn(5, 4, 8, 8)
+    dtype = torch.float32
+
+    noisy_input, timesteps, sigmas = get_noisy_model_input_and_timesteps(args, noise_scheduler, latents, noise, device, dtype)
+
+    assert noisy_input.shape == latents.shape
+    assert timesteps.shape == (5,)
+    assert sigmas.shape == (5, 1, 1, 1)
+
+
+# Test different image sizes
+def test_different_image_size(args, noise_scheduler, device):
+    latents = torch.randn(2, 4, 16, 16)  # larger image size
+    noise = torch.randn(2, 4, 16, 16)
+    dtype = torch.float32
+
+    noisy_input, timesteps, sigmas = get_noisy_model_input_and_timesteps(args, noise_scheduler, latents, noise, device, dtype)
+
+    assert noisy_input.shape == latents.shape
+    assert timesteps.shape == (2,)
+    assert sigmas.shape == (2, 1, 1, 1)
+
+
+# Test edge cases
+def test_zero_batch_size(args, noise_scheduler, device):
+    with pytest.raises(AssertionError):  # expecting an error with zero batch size
+        latents = torch.randn(0, 4, 8, 8)
+        noise = torch.randn(0, 4, 8, 8)
+        dtype = torch.float32
+
+        get_noisy_model_input_and_timesteps(args, noise_scheduler, latents, noise, device, dtype)
+
+
+def test_different_timestep_count(args, device):
+    noise_scheduler = MockNoiseScheduler(num_train_timesteps=500)  # different timestep count
+    latents = torch.randn(2, 4, 8, 8)
+    noise = torch.randn(2, 4, 8, 8)
+    dtype = torch.float32
+
+    noisy_input, timesteps, sigmas = get_noisy_model_input_and_timesteps(args, noise_scheduler, latents, noise, device, dtype)
+
+    assert noisy_input.shape == latents.shape
+    assert timesteps.shape == (2,)
+    # Check that timesteps are within the proper range
+    assert torch.all(timesteps < 500)