feat: add usf to paddle

ppdiffusers/examples/class_conditional_image_generation/DiT/evo_search.yml

@@ -0,0 +1,22 @@
+baseline:
+    timestep_type: ["linspace", "leading", "trailing"]
+    orders: [3, 2, 1]
+parents:
+    rank_prob: 1.0
+    rank_bar: 2
+    absolute_bar: 1
+crossover:
+    prob: 0.15
+    better_prob: 0.6
+mutate:
+    order:
+        prob: 0.2
+        dist:
+          1: 0.3
+          2: 0.5
+          3: 0.2
+    timestep:
+        prob: 0.2
+        scale: 3
+metric:
+    indicator: -1

ppdiffusers/examples/class_conditional_image_generation/DiT/gen_fixed_noise.py

@@ -0,0 +1,8 @@
+import paddle
+import numpy as np
+
+paddle.seed(1234)  # 固定随机种子
+noise = paddle.randn([5000, 4, 32, 32], dtype="float32")  # 举例：形状按你模型需要调整
+# np.save("/path/to/dit_fixed_noise_B5000.npy", noise.numpy())  # 保存为 .npy
+# # 或保存为 .pdparams
+paddle.save({"fixed_noise": noise}, "/share/chenqian-local/PaddleMIX/ppdiffusers/examples/class_conditional_image_generation/dit_fixed_noise_B5000.pdparams")

ppdiffusers/examples/class_conditional_image_generation/DiT/infer_demo_dit copy.py

@@ -0,0 +1,40 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+from paddlenlp.trainer import set_seed
+
+from ppdiffusers import DDIMScheduler, DiTPipeline, DPMSolverMultistepScheduler
+
+dtype = paddle.float32
+pipe = DiTPipeline.from_pretrained("facebook/DiT-XL-2-256", paddle_dtype=dtype)
+# import ipdb; ipdb.set_trace()
+# use DDIMScheduler for inference
+# pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+pipe.scheduler.config.algorithm_type = "dpmsolver"
+pipe.scheduler.config.solver_order = 3
+words = ["golden retriever"]  # class_ids [207]
+class_ids = pipe.get_label_ids(words)
+class_ids = [206,207]
+import ipdb; ipdb.set_trace()
+# import ipdb; ipdb.set_trace()
+timesteps_list = [999, 899, 799, 699, 599, 499, 399, 299, 199, 99]
+order_list = [1, 2, 3, 1, 1, 2, 2, 2, 2, 1]
+# generate image
+set_seed(42)
+generator = paddle.Generator().manual_seed(0)
+image = pipe(class_labels=class_ids, num_inference_steps=10, generator=generator, timesteps_list = timesteps_list, order_list = order_list).images[0]
+import ipdb; ipdb.set_trace()
+image.save("result_DiT_golden_retriever_dpm_10_2.png")

ppdiffusers/examples/class_conditional_image_generation/DiT/infer_with_result.py

@@ -0,0 +1,40 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+from paddlenlp.trainer import set_seed
+
+from ppdiffusers import DDIMScheduler, DiTPipeline, DPMSolverMultistepScheduler
+
+dtype = paddle.float32
+pipe = DiTPipeline.from_pretrained("facebook/DiT-XL-2-256", paddle_dtype=dtype)
+# import ipdb; ipdb.set_trace()
+# use DDIMScheduler for inference
+# pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+pipe.scheduler.config.algorithm_type = "dpmsolver"
+pipe.scheduler.config.solver_order = 3
+words = ["golden retriever"]  # class_ids [207]
+class_ids = pipe.get_label_ids(words)
+search_result = paddle.load("/share/chenqian-local/PaddleMIX/ppdiffusers/examples/class_conditional_image_generation/DiT/search_result_debug/0/0.pdparams")
+
+timesteps_list = search_result[0]["timesteps"]
+order_list = search_result[0]["orders"]
+# import ipdb; ipdb.set_trace()
+# generate image
+set_seed(42)
+generator = paddle.Generator().manual_seed(0)
+image = pipe(class_labels=class_ids, num_inference_steps=10, generator=generator, timesteps_list = timesteps_list, order_list = order_list).images[0]
+# import ipdb; ipdb.set_trace()
+image.save("search_result_DiT_golden_retriever_dpm_10.png")

ppdiffusers/examples/class_conditional_image_generation/DiT/search.py

@@ -0,0 +1,168 @@
+import paddle
+from paddlenlp.trainer import set_seed
+import yaml
+import argparse
+
+from ppdiffusers import DDIMScheduler, DiTPipeline, DPMSolverMultistepScheduler
+import os
+import shutil
+import random
+from search_utils import *
+
+def load_prompts(prompt_path):
+    with open(prompt_path, "r") as f:
+        prompts = [line.strip() for line in f.readlines()]
+    return prompts
+
+def eval_coeff(coeff, cfg, generator ,prompts, pipe, save_dir, fixed_noise):
+    print(f"Begin to evaluate {coeff}")
+    sample_idx = 0
+
+    # Clean and create save directory
+    if os.path.exists(save_dir):
+        shutil.rmtree(save_dir)
+    os.makedirs(save_dir, exist_ok=True)
+
+    # Generate samples
+    for i in range(0, len(prompts), cfg.batch_size):
+        # import ipdb; ipdb.set_trace()
+        batch_prompts = prompts[i : i + cfg.batch_size]
+
+        # Sample using DiT with custom orders and timesteps
+        # samples = scheduler.sample(
+        #     model,
+        #     text_encoder,
+        #     z_size=(vae.out_channels, *latent_size),
+        #     prompts=batch_prompts,
+        #     device=device,
+        #     additional_args=model_args,
+        #     orders=coeff["orders"],
+        #     timesteps=coeff["timesteps"],
+        #     input_noise=fixed_noise[sample_idx:sample_idx+cfg.batch_size],
+        # )
+        # import ipdb; ipdb.set_trace()
+        samples = pipe(class_labels=batch_prompts, num_inference_steps=10, generator=generator, timesteps_list = coeff["timesteps"], order_list = coeff["orders"], fixed_noise = fixed_noise[sample_idx:sample_idx+cfg.batch_size]).images
+        # import ipdb; ipdb.set_trace()
+
+        # Decode VAE latents to images
+
+        for idx, sample in enumerate(samples):
+            print(f"Prompt: {batch_prompts[idx]}")
+            save_path = os.path.join(save_dir, f"sample_{sample_idx}.png")
+            sample.save(save_path)
+            sample_idx += 1
+    from paddle_fid.fid_score import calculate_fid_given_paths
+    fid = calculate_fid_given_paths(
+        [save_dir,"/share/public-nfs/chenqian/var/dataset/imagenet256/VIRTUAL_imagenet256_labeled.npz"], 
+        batch_size=256,
+        dims=2048,
+        num_workers=8,
+    )
+
+    return -fid
+
+
+
+def main(args):
+    dtype = paddle.float32
+    cfg = args
+    # import ipdb; ipdb.set_trace()
+
+    fixed_noise = paddle.load("/share/chenqian-local/PaddleMIX/ppdiffusers/examples/class_conditional_image_generation/dit_fixed_noise_B5000.pdparams")
+
+    fixed_noise = fixed_noise["fixed_noise"]
+    prompts = load_prompts(cfg.prompt_path)
+    prompts = list(map(int, prompts))
+
+    pipe = DiTPipeline.from_pretrained("facebook/DiT-XL-2-256", paddle_dtype=dtype)
+    pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config) 
+    pipe.scheduler.config.algorithm_type = "dpmsolver"
+    pipe.scheduler.config.solver_order = 3 
+    generator = paddle.Generator().manual_seed(0)  
+    # set data save path
+    os.makedirs(os.path.join(cfg.data_path, cfg.split), exist_ok=True)
+    os.makedirs(os.path.join(cfg.data_path, "baselines"), exist_ok=True)
+
+    # search
+    with open(cfg.search_config, "r") as f:
+        search_cfg = yaml.safe_load(f)
+    baseline_done = False
+    data_num = len([d for d in os.listdir(os.path.join(cfg.data_path, cfg.split)) if ".pdparams" in str(d)])
+    image_save_dir = os.path.join(cfg.data_path, cfg.split, "images")
+
+    # Main search loop
+    while(1):
+        print(f"Random check: {random.uniform(0, 1)}")
+
+        # Baseline evaluation phase
+        if not baseline_done:
+            baseline, data_path = get_baseline(search_cfg, cfg.budget, cfg.data_path)
+            if baseline == -1:
+                baseline_done = True
+                print("All baselines evaluated, starting evolutionary search...")
+            else:
+                score = eval_coeff(baseline, cfg, generator, prompts, pipe, image_save_dir, fixed_noise)
+                paddle.save([baseline, score], data_path)
+
+                print(f"Save baseline {[baseline, score]} to {data_path}")
+
+                # delete occ file
+                str_data_path = str(data_path)
+                assert ".pdparams" in str_data_path
+                occ_path = str(data_path).replace(".pdparams", ".occ")
+                if os.path.exists(occ_path):
+                    os.remove(occ_path)
+
+                continue
+
+        # Evolutionary search phase
+        population = get_population(cfg.data_path, search_cfg)
+
+        # Decide between crossover and mutation
+        if random.uniform(0, 1) < search_cfg["crossover"]["prob"]:
+            # Crossover operation
+            parents_1, parents_2 = select_parents(population, search_cfg, num=2)
+            print(f"Choose {parents_1} and {parents_2} as the crossover parents")
+            new_coeff = crossover(parents_1[0], parents_2[0], search_cfg)
+        else:
+            new_coeff = None
+
+        if new_coeff is not None:
+            parent = new_coeff
+        else:
+            # Mutation operation
+            parent = select_parents(population, search_cfg, num=1)[0]
+            print(f"Choose {parent} as the mutation parent")
+
+        new_coeff = mutate(parent, search_cfg)
+
+        # Evaluate new coefficient
+        score = eval_coeff(new_coeff, cfg, generator, prompts, pipe, image_save_dir, fixed_noise)
+
+        # Save result
+        result_path = os.path.join(cfg.data_path, cfg.split, f"{data_num}.pdparams")
+        paddle.save([new_coeff, score], result_path)
+        print(f"Save {[new_coeff, score]} to {result_path}")
+        data_num += 1
+
+        # Update the search config (in case it was modified externally)
+        try:
+            with open(cfg.search_config, "r") as f:
+                search_cfg = yaml.safe_load(f)
+        except Exception as e:
+            print(f"Warning: Could not reload search config: {e}")
+
+if __name__ == "__main__":
+
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--search_config", type=str, default="/share/chenqian-local/PaddleMIX/ppdiffusers/examples/class_conditional_image_generation/DiT/evo_search.yml")
+    parser.add_argument("--budget", type=int, default=10)
+    parser.add_argument("--data_path", type=str, default="/share/chenqian-local/PaddleMIX/ppdiffusers/examples/class_conditional_image_generation/DiT/search_result_debug")
+    parser.add_argument("--split", type=str, default="0")
+    parser.add_argument("--prompt_path", type=str, default="/share/chenqian-local/PaddleMIX/ppdiffusers/examples/class_conditional_image_generation/DiT/dit_5k.txt")  
+    parser.add_argument("--batch_size", type=int, default=40) 
+    args = parser.parse_args()
+
+    # import ipdb; ipdb.set_trace()
+    main(args)