flaxdiff 0.1.36.3__py3-none-any.whl → 0.1.36.5__py3-none-any.whl

flaxdiff/data/sources/tfds.py

@@ -4,6 +4,8 @@ import grain.python as pygrain
 from flaxdiff.utils import AutoTextTokenizer
 from typing import Dict
 import random
+import augmax
+import jax
 
 # -----------------------------------------------------------------------------------------------#
 # Oxford flowers and other TFDS datasources -----------------------------------------------------#
@@ -47,6 +49,15 @@ def tfds_augmenters(image_scale, method):
         interpolation = cv2.INTER_CUBIC
     else:
         interpolation = cv2.INTER_AREA
+        
+    augments = augmax.Chain(
+        augmax.HorizontalFlip(0.5),
+        augmax.RandomContrast((-0.05, 0.05), 1.),
+        augmax.RandomBrightness((-0.2, 0.2), 1.)
+    )
+
+    augments = jax.jit(augments, backend="cpu")
+        
     class augmenters(pygrain.MapTransform):
         def __init__(self, *args, **kwargs):
             super().__init__(*args, **kwargs)
@@ -56,6 +67,7 @@ def tfds_augmenters(image_scale, method):
             image = element['image']
             image = cv2.resize(image, (image_scale, image_scale),
                             interpolation=interpolation)
+            # image = augments(image)
             # image = (image - 127.5) / 127.5
             caption = labelizer(element)
             results = self.tokenize(caption)

flaxdiff/schedulers/__init__.py

@@ -1,6 +1,6 @@
 from .discrete import DiscreteNoiseScheduler
 from .common import NoiseScheduler, GeneralizedNoiseScheduler
-from .cosine import CosineNoiseSchedule, ContinuousNoiseScheduler, CosineGeneralNoiseScheduler
+from .cosine import CosineNoiseScheduler, ContinuousNoiseScheduler, CosineGeneralNoiseScheduler
 from .linear import LinearNoiseSchedule
 from .sqrt import SqrtContinuousNoiseScheduler
 from .karras import KarrasVENoiseScheduler, SimpleExpNoiseScheduler, EDMNoiseScheduler

flaxdiff/schedulers/cosine.py

@@ -12,7 +12,7 @@ def cosine_beta_schedule(timesteps, start_angle=0.008, end_angle=0.999):
     betas = 1 - (alphas_bar[1:] / alphas_bar[:-1])
     return np.clip(betas, 0, end_angle)
 
-class CosineNoiseSchedule(DiscreteNoiseScheduler):
+class CosineNoiseScheduler(DiscreteNoiseScheduler):
     def __init__(self, timesteps, beta_start=0.008, beta_end=0.999, *args, **kwargs):
         super().__init__(timesteps, beta_start, beta_end, schedule_fn=cosine_beta_schedule, *args, **kwargs)
 

flaxdiff/trainer/diffusion_trainer.py

@@ -14,6 +14,7 @@ from typing import Dict, Callable, Sequence, Any, Union, Tuple, Type
 from ..schedulers import NoiseScheduler
 from ..predictors import DiffusionPredictionTransform, EpsilonPredictionTransform
 from ..samplers.common import DiffusionSampler
+from ..samplers.ddim import DDIMSampler
 
 from flaxdiff.utils import RandomMarkovState
 
@@ -179,9 +180,6 @@ class DiffusionTrainer(SimpleTrainer):
                 nloss = loss_fn(preds, expected_output)
                 # Ignore the loss contribution of images with zero standard deviation
                 nloss *= noise_schedule.get_weights(noise_level)
-                # nloss = jnp.mean(nloss, axis=(1,2,3))
-                # nloss = jnp.where(is_non_zero, nloss, 0)
-                # nloss = jnp.mean(nloss, where=nloss != 0)
                 nloss = jnp.mean(nloss)
                 loss = nloss
                 return loss
@@ -224,11 +222,11 @@ class DiffusionTrainer(SimpleTrainer):
         if distributed_training:
             train_step = shard_map(train_step, mesh=self.mesh, in_specs=(P(), P(), P('data'), P('data')), 
                                    out_specs=(P(), P(), P()))
-            train_step = jax.jit(train_step)
+        train_step = jax.jit(train_step)
             
         return train_step
 
-    def _define_vaidation_step(self, sampler_class: Type[DiffusionSampler]):
+    def _define_vaidation_step(self, sampler_class: Type[DiffusionSampler]=DDIMSampler, sampling_noise_schedule: NoiseScheduler=None):
         model = self.model
         encoder = self.encoder
         autoencoder = self.autoencoder
@@ -241,7 +239,7 @@ class DiffusionTrainer(SimpleTrainer):
             sampler = sampler_class(
                 model=model,
                 params=state.ema_params,
-                noise_schedule=self.noise_schedule,
+                noise_schedule=self.noise_schedule if sampling_noise_schedule is None else sampling_noise_schedule,
                 model_output_transform=self.model_output_transform,
                 image_size=self.input_shapes['x'][0],
                 null_labels_seq=null_labels_full,
@@ -311,10 +309,11 @@ class DiffusionTrainer(SimpleTrainer):
             print("Error logging images to wandb", e)
             traceback.print_exc()
     
-    def fit(self, data, training_steps_per_epoch, epochs, val_steps_per_epoch=8, sampler_class=None):
+    def fit(self, data, training_steps_per_epoch, epochs, val_steps_per_epoch=8, sampler_class: Type[DiffusionSampler]=DDIMSampler, sampling_noise_schedule: NoiseScheduler=None):
         local_batch_size = data['local_batch_size']
         validation_step_args = {
             "sampler_class": sampler_class,
+            "sampling_noise_schedule": sampling_noise_schedule,
         }
         super().fit(
             data, 

{flaxdiff-0.1.36.3.dist-info → flaxdiff-0.1.36.5.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: flaxdiff
-Version: 0.1.36.3
+Version: 0.1.36.5
 Summary: A versatile and easy to understand Diffusion library
 Author-email: Ashish Kumar Singh <ashishkmr472@gmail.com>
 License-Expression: MIT
@@ -96,7 +96,7 @@ Also, few of the text may be generated with help of github copilot, so please ex
 ### Schedulers
 Implemented in `flaxdiff.schedulers`:
 - **LinearNoiseSchedule** (`flaxdiff.schedulers.LinearNoiseSchedule`): A beta-parameterized discrete scheduler.
-- **CosineNoiseSchedule** (`flaxdiff.schedulers.CosineNoiseSchedule`): A beta-parameterized discrete scheduler.
+- **CosineNoiseScheduler** (`flaxdiff.schedulers.CosineNoiseScheduler`): A beta-parameterized discrete scheduler.
 - **ExpNoiseSchedule** (`flaxdiff.schedulers.ExpNoiseSchedule`): A beta-parameterized discrete scheduler.
 - **CosineContinuousNoiseScheduler** (`flaxdiff.schedulers.CosineContinuousNoiseScheduler`): A continuous scheduler.
 - **CosineGeneralNoiseScheduler** (`flaxdiff.schedulers.CosineGeneralNoiseScheduler`): A continuous sigma parameterized cosine scheduler.
@@ -147,43 +147,81 @@ sticking to the versions mentioned in the requirements.txt
 Here is a simplified example to get you started with training a diffusion model using FlaxDiff:
 
 ```python
-from flaxdiff.schedulers import EDMNoiseScheduler
+from flaxdiff.schedulers import EDMNoiseScheduler, KarrasVENoiseScheduler
 from flaxdiff.predictors import KarrasPredictionTransform
-from flaxdiff.models.simple_unet import SimpleUNet as UNet
+from flaxdiff.models.simple_unet import Unet
 from flaxdiff.trainer import DiffusionTrainer
+from flaxdiff.data.datasets import get_dataset_grain
+from flaxdiff.utils import defaultTextEncodeModel
+from flaxdiff.samplers.euler import EulerAncestralSampler
 import jax
+import jax.numpy as jnp
 import optax
 from datetime import datetime
 
 BATCH_SIZE = 16
-IMAGE_SIZE = 64
+IMAGE_SIZE = 128
 
 # Define noise scheduler
 edm_schedule = EDMNoiseScheduler(1, sigma_max=80, rho=7, sigma_data=0.5)
-
+karas_ve_schedule = KarrasVENoiseScheduler(1, sigma_max=80, rho=7, sigma_data=0.5)
 # Define model
-unet = UNet(emb_features=256, 
-            feature_depths=[64, 128, 256, 512],
-            attention_configs=[{"heads":4}, {"heads":4}, {"heads":4}, {"heads":4}, {"heads":4}],
+unet = Unet(emb_features=256, 
+            feature_depths=[64, 64, 128, 256, 512],
+            attention_configs=[
+                None,
+                {"heads":8, "dtype":jnp.float16, "flash_attention":False, "use_projection":True, "use_self_and_cross":True}, 
+                {"heads":8, "dtype":jnp.float16, "flash_attention":False, "use_projection":True, "use_self_and_cross":True}, 
+                {"heads":8, "dtype":jnp.float16, "flash_attention":False, "use_projection":True, "use_self_and_cross":True}, 
+                {"heads":8, "dtype":jnp.float16, "flash_attention":False, "use_projection":False, "use_self_and_cross":False}
+                ],
             num_res_blocks=2,
-            num_middle_res_blocks=1)
-
+            num_middle_res_blocks=1
+)
 # Load dataset
-data, datalen = get_dataset("oxford_flowers102", batch_size=BATCH_SIZE, image_scale=IMAGE_SIZE)
+data = get_dataset_grain("oxford_flowers102", batch_size=BATCH_SIZE, image_scale=IMAGE_SIZE)
+datalen = data['train_len']
 batches = datalen // BATCH_SIZE
 
+input_shapes = {
+    "x": (IMAGE_SIZE, IMAGE_SIZE, 3),
+    "temb": (),
+    "textcontext": (77, 768)
+}
+text_encoder = defaultTextEncodeModel()
+
+# Construct a validation set by the prompts
+val_prompts = ['water tulip', ' a water lily', ' a water lily', ' a photo of a rose', ' a photo of a rose', ' a water lily', ' a water lily', ' a photo of a marigold', ' a photo of a marigold']
+
+def get_val_dataset(batch_size=8):
+    for i in range(0, len(val_prompts), batch_size):
+        prompts = val_prompts[i:i + batch_size]
+        tokens = text_encoder.tokenize(prompts)
+        yield tokens
+
+data['test'] = get_val_dataset
+data['test_len'] = len(val_prompts)
+
 # Define optimizer
 solver = optax.adam(2e-4)
 
 # Create trainer
-trainer = DiffusionTrainer(unet, optimizer=solver, 
-                           noise_schedule=edm_schedule,
-                           rngs=jax.random.PRNGKey(4), 
-                           name="Diffusion_SDE_VE_" + datetime.now().strftime("%Y-%m-%d_%H:%M:%S"),
-                           model_output_transform=KarrasPredictionTransform(sigma_data=edm_schedule.sigma_data))
+trainer = DiffusionTrainer(
+    unet, optimizer=solver, 
+    input_shapes=input_shapes,
+    noise_schedule=edm_schedule,
+    rngs=jax.random.PRNGKey(4), 
+    name="Diffusion_SDE_VE_" + datetime.now().strftime("%Y-%m-%d_%H:%M:%S"),
+    model_output_transform=KarrasPredictionTransform(sigma_data=edm_schedule.sigma_data),
+    encoder=text_encoder,
+    distributed_training=True,
+    wandb_config = {
+        "project": 'mlops-msml605-project',
+        "name": f"prototype-{datetime.now().strftime('%Y-%m-%d_%H:%M:%S')}",
+})
 
 # Train the model
-final_state = trainer.fit(data, batches, epochs=2000)
+final_state = trainer.fit(data, batches, epochs=2000, sampler_class=EulerAncestralSampler, sampling_noise_schedule=karas_ve_schedule)
 ```
 
 ### Inference Example
@@ -301,8 +339,8 @@ Images generated by the following prompts using classifier free guidance with gu
 `Training Epochs: 1000`
 `Steps per epoch: 511`
 
-`Training Noise Schedule: CosineNoiseSchedule`
-`Inference Noise Schedule: CosineNoiseSchedule`
+`Training Noise Schedule: CosineNoiseScheduler`
+`Inference Noise Schedule: CosineNoiseScheduler`
 
 `Model: UNet(emb_features=256, 
             feature_depths=[64, 128, 256, 512],
@@ -321,8 +359,8 @@ Images generated by the following prompts using classifier free guidance with gu
 `Training Epochs: 1000`
 `Steps per epoch: 511`
 
-`Training Noise Schedule: CosineNoiseSchedule`
-`Inference Noise Schedule: CosineNoiseSchedule`
+`Training Noise Schedule: CosineNoiseScheduler`
+`Inference Noise Schedule: CosineNoiseScheduler`
 
 `Model: UNet(emb_features=256, 
             feature_depths=[64, 128, 256, 512],

{flaxdiff-0.1.36.3.dist-info → flaxdiff-0.1.36.5.dist-info}/RECORD

@@ -5,7 +5,7 @@ flaxdiff/data/dataset_map.py,sha256=hcHaoR2IbNQmfyPUhYd6_8xinurxxCqawQijAsDI0Ek,
 flaxdiff/data/datasets.py,sha256=YUMoSvF2yAyikRvRofZVlHwfEOU3zXSSG4KkLnVfpoA,5626
 flaxdiff/data/online_loader.py,sha256=1Fi_QRixxRzbt602nORINcDeHEccvCrBpagrz4PURYg,12499
 flaxdiff/data/sources/gcs.py,sha256=11ZuQhvMyJRLg21DgVdzO5qEuae7zgzTXGNOskF-cbs,3380
-flaxdiff/data/sources/tfds.py,sha256=WA3h9lyR4yotCNEmJON2noIN-2HNcqhf6zigx1XXsMI,2481
+flaxdiff/data/sources/tfds.py,sha256=7n-uobG_UvkD5mU_1ovPd9kb6xJrbEKFFXdVEHDunts,2781
 flaxdiff/metrics/inception.py,sha256=a5kjMCPMT9gB88c_HCKiek-2vsAyoE35K7nDt4h4pVI,31843
 flaxdiff/metrics/utils.py,sha256=YuuOfqvqgIjsceupwNeJ59vQ2TnGeNMIyKdkIqOmoNg,1702
 flaxdiff/models/__init__.py,sha256=FAivVYXxM2JrCFIXf-C3374RB2Hth25dBrzOeNFhH1U,26
@@ -27,10 +27,10 @@ flaxdiff/samplers/euler.py,sha256=Htb-IJeu7jSgY6mvgYr9yl9pUnos49vijlVk5IQsRps,27
 flaxdiff/samplers/heun_sampler.py,sha256=UyI-hSlyWvt-7VEUJj27zjgyzKkGVl8fDUHV-YpSOCc,1421
 flaxdiff/samplers/multistep_dpm.py,sha256=3Wu3MrMLYaBb1ObraTbWrJmtEtU0adl1dDbz5fPJ4Gs,2735
 flaxdiff/samplers/rk4_sampler.py,sha256=1j1pES_Q2QiaURvEWeedbbT1LHmkc3jsu0GgH83qBL0,1926
-flaxdiff/schedulers/__init__.py,sha256=3id390WEfdf-MN-oLSPAhlRFIXrFWr6ioAHPAwURJyE,375
+flaxdiff/schedulers/__init__.py,sha256=EcABJ5UqsfeFXD9ypbgjVSYb6IKm7bcmsUbqEwVpHUc,376
 flaxdiff/schedulers/common.py,sha256=b-W4iI-aqScpVE8VZbBpiYvAVI6rqDkUP-C_hEVBwCI,4151
 flaxdiff/schedulers/continuous.py,sha256=5c_niOA20fxJ5oJDi09FfayIRogBGwtfG0XThW2IUZk,334
-flaxdiff/schedulers/cosine.py,sha256=9ban0dFHLMm35wQvaBT4nCQwPGmzNsXwQ1xI0oppmJI,2005
+flaxdiff/schedulers/cosine.py,sha256=EtU3SjJaP9R9ULHNiYrX9jBLSsAGKPGteHiwOzWNzYo,2006
 flaxdiff/schedulers/discrete.py,sha256=O54wH2HVu3olJA71NxgAXFW9cr6B6Gl-DR_uZeytpds,3319
 flaxdiff/schedulers/exp.py,sha256=cPTnUJpYdzJRRZqMLYQz0rRUCpEmaP2tXhRumLx94jA,605
 flaxdiff/schedulers/karras.py,sha256=4GN120kGwdxxU-h2mVdhBVy9IORkUMm_vvz3XjthBcI,3355
@@ -38,10 +38,10 @@ flaxdiff/schedulers/linear.py,sha256=6003F5ISq1Wc0h6UAzY95MJgsDIKGMhBzbiVALpea0k
 flaxdiff/schedulers/sqrt.py,sha256=1F84ZgQPuoNMhe6yxGTR2G0h7dPOZtm4UDQOakbSsEU,445
 flaxdiff/trainer/__init__.py,sha256=T-vUVq4zHcMK6kpCsG4Gu8vn71q6lZD-lg-Ul7yKfEk,128
 flaxdiff/trainer/autoencoder_trainer.py,sha256=hxihkRL9WCIQVGOP-pc1jjjIUaRXDLcNo3_erTKsuWM,7049
-flaxdiff/trainer/diffusion_trainer.py,sha256=ajOWBgFFwXP_VQScUjcuPoaB4Gk02aF0Ls5LNlA8wqA,12691
+flaxdiff/trainer/diffusion_trainer.py,sha256=zde_nRzsC2GD5KNCn5Qjw9ldHi7L_-teJhcUNUDCdcQ,12815
 flaxdiff/trainer/simple_trainer.py,sha256=lmRo8N0bMupIyS3ejPvPtxoskY_3GLC8iyJE6u4TIWc,21990
 flaxdiff/trainer/video_diffusion_trainer.py,sha256=gMkKpnKNTo8QhTx5ptEEkc7W5-7rzXIr9queU53hXyQ,2197
-flaxdiff-0.1.36.3.dist-info/METADATA,sha256=9XaZMJ6SMFP7OUn2tp9v5FQveMGoxvuiyxdJ8SmMd8w,22310
-flaxdiff-0.1.36.3.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
-flaxdiff-0.1.36.3.dist-info/top_level.txt,sha256=-2-nXnfkJgSfkki1tjm5Faw6Dso7vhtdn2szwCdX5CQ,9
-flaxdiff-0.1.36.3.dist-info/RECORD,,
+flaxdiff-0.1.36.5.dist-info/METADATA,sha256=Bk4FoPnJ0DlpficfpLQ9t0SaE13xe26xs2gEb2BYdfI,23985
+flaxdiff-0.1.36.5.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
+flaxdiff-0.1.36.5.dist-info/top_level.txt,sha256=-2-nXnfkJgSfkki1tjm5Faw6Dso7vhtdn2szwCdX5CQ,9
+flaxdiff-0.1.36.5.dist-info/RECORD,,