Upload folder using huggingface_hub

config/base.py

@@ -0,0 +1,112 @@
+import ml_collections
+
+
+def get_config():
+    config = ml_collections.ConfigDict()
+
+    ###### General ######
+    # run name for wandb logging and checkpoint saving -- if not provided, will be auto-generated based on the datetime.
+    config.run_name = ""
+    config.debug = False
+
+    # random seed for reproducibility.
+    config.seed = 42
+    # top-level logging directory for checkpoint saving.
+    config.logdir = "logs"
+    # number of epochs to train for. each epoch is one round of sampling from the model followed by training on those
+    # samples.
+    config.num_epochs = 100000
+    # number of epochs between saving model checkpoints.
+    config.save_freq = 10
+    config.eval_freq = 10
+    # mixed precision training. options are "fp16", "bf16", and "no". half-precision speeds up training significantly.
+    config.mixed_precision = "bf16"
+    # allow tf32 on Ampere GPUs, which can speed up training.
+    config.allow_tf32 = True
+    # resume training from a checkpoint. either an exact checkpoint directory (e.g. checkpoint_50), or a directory
+    # containing checkpoints, in which case the latest one will be used. `config.use_lora` must be set to the same value
+    # as the run that generated the saved checkpoint.
+    config.resume_from = ""
+    # whether or not to use LoRA.
+    config.use_lora = True
+    config.dataset = ""
+    config.resolution = 768
+
+    ###### Pretrained Model ######
+    config.pretrained = pretrained = ml_collections.ConfigDict()
+    # base model to load. either a path to a local directory, or a model name from the HuggingFace model hub.
+    pretrained.model = ""
+    # revision of the model to load.
+    pretrained.revision = ""
+
+    ###### Sampling ######
+    config.sample = sample = ml_collections.ConfigDict()
+    # number of sampler inference steps.
+    sample.num_steps = 40
+    sample.eval_num_steps = 40
+    # classifier-free guidance weight. 1.0 is no guidance.
+    sample.guidance_scale = 4.5
+    # batch size (per GPU!) to use for sampling.
+    sample.train_batch_size = 1
+    sample.num_image_per_prompt = 1
+    sample.test_batch_size = 1
+    # number of batches to sample per epoch. the total number of samples per epoch is `num_batches_per_epoch *
+    # batch_size * num_gpus`.
+    sample.num_batches_per_epoch = 2
+    # Whether use all samples in a batch to compute std
+    sample.global_std = True
+    # noise level
+    sample.noise_level = 1.0
+
+    ###### Training ######
+    config.train = train = ml_collections.ConfigDict()
+    # batch size (per GPU!) to use for training.
+    train.batch_size = 1
+    # learning rate.
+    train.learning_rate = 3e-4
+    # Adam beta1.
+    train.adam_beta1 = 0.9
+    # Adam beta2.
+    train.adam_beta2 = 0.999
+    # Adam weight decay.
+    train.adam_weight_decay = 1e-4
+    # Adam epsilon.
+    train.adam_epsilon = 1e-8
+    # number of gradient accumulation steps. the effective batch size is `batch_size * num_gpus *
+    # gradient_accumulation_steps`.
+    train.gradient_accumulation_steps = 1
+    # maximum gradient norm for gradient clipping.
+    train.max_grad_norm = 1.0
+    # number of inner epochs per outer epoch. each inner epoch is one iteration through the data collected during one
+    # outer epoch's round of sampling.
+    train.num_inner_epochs = 1
+    # clip advantages to the range [-adv_clip_max, adv_clip_max].
+    train.adv_clip_max = 5
+    # the fraction of timesteps to train on. if set to less than 1.0, the model will be trained on a subset of the
+    # timesteps for each sample. this will speed up training but reduce the accuracy of policy gradient estimates.
+    train.timestep_fraction = 0.99
+    # kl ratio
+    train.beta = 0.0001
+    # pretrained lora path
+    train.lora_path = None
+    train.ema = True
+
+    ###### Prompt Function ######
+    # prompt function to use. see `prompts.py` for available prompt functions.
+    config.prompt_fn = ""
+    # kwargs to pass to the prompt function.
+    config.prompt_fn_kwargs = {}
+
+    ###### Reward Function ######
+    # reward function to use. see `rewards.py` for available reward functions.
+    config.reward_fn = ml_collections.ConfigDict()
+    config.save_dir = ""
+    # whether to save all generated images during periodic eval.
+    config.save_eval_images = True
+    # optional custom directory for eval images; empty means {save_dir}/eval_images.
+    config.eval_image_dir = ""
+
+    ###### Per-Prompt Stat Tracking ######
+    config.per_prompt_stat_tracking = True
+
+    return config

config/kontext_nft.py

@@ -0,0 +1,97 @@
+import imp
+import os
+
+base = imp.load_source("base", os.path.join(os.path.dirname(__file__), "base.py"))
+
+
+def get_config(name):
+    return globals()[name]()
+
+def _get_config(base_model="kontext", n_gpus=1, gradient_step_per_epoch=1, reward_fn={}, name=""):
+    config = base.get_config()
+
+    config.base_model = base_model
+    config.dataset = "../edit-r1-dataset"
+    
+    config.pretrained.model = "black-forest-labs/FLUX.1-Kontext-dev"
+    config.sample.num_steps = 6
+    config.sample.eval_num_steps = 15
+    config.sample.guidance_scale = 2.5
+    config.resolution = 512
+    config.train.beta = 0.0001
+    config.sample.noise_level = 0.7
+    bsz = 3
+
+    config.sample.num_image_per_prompt = 12
+
+    config.sample.ban_std_thres = 0.05
+    config.sample.ban_mean_thres = 0.9
+    config.sample.ban_prompt = False
+
+    num_groups = 24
+
+    while True:
+        if bsz < 1:
+            assert False, "Cannot find a proper batch size."
+        if (
+            num_groups * config.sample.num_image_per_prompt % (n_gpus * bsz) == 0
+            and bsz * n_gpus % config.sample.num_image_per_prompt == 0
+        ):
+            n_batch_per_epoch = num_groups * config.sample.num_image_per_prompt // (n_gpus * bsz)
+            if n_batch_per_epoch % gradient_step_per_epoch == 0:
+                config.sample.train_batch_size = bsz
+                config.sample.num_batches_per_epoch = n_batch_per_epoch
+                config.train.batch_size = config.sample.train_batch_size
+                config.train.gradient_accumulation_steps = (
+                    config.sample.num_batches_per_epoch // gradient_step_per_epoch
+                )
+                break
+        bsz -= 1
+
+    # special design, the test set has a total of 1018/2212/2048 for ocr/geneval/pickscore, to make gpu_num*bs*n as close as possible to it, because when the number of samples cannot be divided evenly by the number of cards, multi-card will fill the last batch to ensure each card has the same number of samples, affecting gradient synchronization.
+    config.sample.test_batch_size = bsz
+    if n_gpus > 32:
+        config.sample.test_batch_size = config.sample.test_batch_size // 2
+
+    config.prompt_fn = "geneval"
+
+    config.run_name = f"nft_{base_model}_{name}"
+    config.save_dir = f"logs/nft/{base_model}/{name}"
+    config.reward_fn = reward_fn
+
+    config.decay_type = 1
+    config.beta = 1.0
+    config.train.adv_mode = "all"
+
+    # config.sample.guidance_scale = 1.0
+    config.sample.deterministic = True
+    config.sample.solver = "dpm2"
+    return config
+
+def kontext_mllm_reward():
+    reward_fn = {
+        "mllm_score_continue": 1.0,
+    }
+    config = _get_config(
+        base_model="kontext",
+        n_gpus=24,
+        gradient_step_per_epoch=1,
+        reward_fn=reward_fn,
+        name="mllm_score_continue",
+    )
+    return config
+
+def kontext_mllm_reward_ban_prompt():
+    reward_fn = {
+        "mllm_score_continue": 1.0,
+    }
+    config = _get_config(
+        base_model="kontext",
+        n_gpus=24,
+        gradient_step_per_epoch=1,
+        reward_fn=reward_fn,
+        name="mllm_score_continue_ban_prompt",
+    )
+    config.sample.ban_prompt = True
+    config.sample.ban_std_thres = 0.05
+    return config

config/qwen_image_edit_nft.py

@@ -0,0 +1,175 @@
+import imp
+import os
+
+base = imp.load_source("base", os.path.join(os.path.dirname(__file__), "base.py"))
+
+
+def get_config(name):
+    return globals()[name]()
+
+def _get_config(
+    base_model="qwen_image_edit",
+    n_gpus=1,
+    gradient_step_per_epoch=1,
+    reward_fn={},
+    name="",
+    sample_batch_size_start=3,
+    num_groups=24,
+    num_image_per_prompt=12,
+):
+    config = base.get_config()
+
+    config.base_model = base_model
+    config.transformer_path = None
+    config.dataset = "../edit-r1-dataset"
+    
+    config.pretrained.model = "Qwen/Qwen-Image-Edit-2509"
+    config.sample.num_steps = 6
+    config.sample.eval_num_steps = 15
+    config.sample.guidance_scale = 1.0
+    config.resolution = 512
+    config.train.beta = 0.0001
+    config.sample.noise_level = 0.7
+    bsz = sample_batch_size_start
+    config.sample.num_image_per_prompt = num_image_per_prompt
+
+    config.sample.ban_std_thres = 0.05
+    config.sample.ban_mean_thres = 0.9
+    config.sample.ban_prompt = False
+    while True:
+        if bsz < 1:
+            assert False, "Cannot find a proper batch size."
+        if (
+            num_groups * config.sample.num_image_per_prompt % (n_gpus * bsz) == 0
+            and bsz * n_gpus % config.sample.num_image_per_prompt == 0
+        ):
+            n_batch_per_epoch = num_groups * config.sample.num_image_per_prompt // (n_gpus * bsz)
+            if n_batch_per_epoch % gradient_step_per_epoch == 0:
+                config.sample.train_batch_size = bsz
+                config.sample.num_batches_per_epoch = n_batch_per_epoch
+                config.train.batch_size = config.sample.train_batch_size
+                config.train.gradient_accumulation_steps = (
+                    config.sample.num_batches_per_epoch // gradient_step_per_epoch
+                )
+                break
+        bsz -= 1
+
+    # special design, the test set has a total of 1018/2212/2048 for ocr/geneval/pickscore, to make gpu_num*bs*n as close as possible to it, because when the number of samples cannot be divided evenly by the number of cards, multi-card will fill the last batch to ensure each card has the same number of samples, affecting gradient synchronization.
+    config.sample.test_batch_size = bsz
+    if n_gpus > 32:
+        config.sample.test_batch_size = config.sample.test_batch_size // 2
+
+    config.prompt_fn = "geneval"
+
+    config.run_name = f"nft_{base_model}_{name}"
+    config.save_dir = f"logs/nft/{base_model}/{name}"
+    config.reward_fn = reward_fn
+
+    config.decay_type = 1
+    config.beta = 1.0
+    config.train.adv_mode = "all"
+
+    # config.sample.guidance_scale = 1.0
+    config.sample.deterministic = True
+    config.sample.solver = "dpm2"
+    return config
+
+def qwen_mllm_reward():
+    reward_fn = {
+        "mllm_score_continue": 1.0,
+    }
+    config = _get_config(
+        base_model="qwen_image_edit",
+        n_gpus=48,
+        gradient_step_per_epoch=1,
+        reward_fn=reward_fn,
+        name="mllm_score_continue",
+    )
+    config.sample.ban_prompt = True
+    config.sample.ban_std_thres = 0.05
+    return config
+
+
+def qwen_mllm_reward_h200_single():
+    reward_fn = {
+        "mllm_score_continue": 1.0,
+    }
+    config = _get_config(
+        base_model="qwen_image_edit",
+        n_gpus=1,
+        gradient_step_per_epoch=4,
+        reward_fn=reward_fn,
+        name="mllm_score_continue_h200_single",
+        sample_batch_size_start=12,
+        num_groups=24,
+        num_image_per_prompt=12,
+    )
+    config.sample.ban_prompt = True
+    config.sample.ban_std_thres = 0.05
+    config.pretrained.model = "/cpfs01/VideoGen/user_data/shenzhebei/Qwen-Image-Edit-2511"
+    config.sample.test_batch_size = 4
+    return config
+
+
+def qwen_mllm_reward_h200_dual():
+    reward_fn = {
+        "mllm_score_continue": 1.0,
+    }
+    config = _get_config(
+        base_model="qwen_image_edit",
+        n_gpus=2,
+        gradient_step_per_epoch=4,
+        reward_fn=reward_fn,
+        name="mllm_score_continue_h200_dual",
+        sample_batch_size_start=6,
+        num_groups=24,
+        num_image_per_prompt=12,
+    )
+    config.sample.ban_prompt = True
+    config.sample.ban_std_thres = 0.05
+    config.pretrained.model = "/cpfs01/VideoGen/user_data/shenzhebei/Qwen-Image-Edit-2511"
+    config.sample.test_batch_size = 4
+    return config
+
+
+def qwen_mllm_reward_h200_six():
+    reward_fn = {
+        "mllm_score_continue": 1.0,
+    }
+    config = _get_config(
+        base_model="qwen_image_edit",
+        n_gpus=6,
+        gradient_step_per_epoch=4,
+        reward_fn=reward_fn,
+        name="mllm_score_continue_h200_six",
+        sample_batch_size_start=12,
+        num_groups=24,
+        num_image_per_prompt=12,
+    )
+    config.sample.ban_prompt = True
+    config.sample.ban_std_thres = 0.05
+    config.pretrained.model = "/cpfs01/VideoGen/user_data/shenzhebei/Qwen-Image-Edit-2511"
+    config.sample.test_batch_size = 4
+    return config
+
+
+def qwen_mllm_reward_h200_three():
+    reward_fn = {
+        "mllm_score_continue": 1.0,
+    }
+    config = _get_config(
+        base_model="qwen_image_edit",
+        n_gpus=3,
+        gradient_step_per_epoch=4,
+        reward_fn=reward_fn,
+        name="mllm_score_continue_h200_three",
+        sample_batch_size_start=8,
+        num_groups=24,
+        num_image_per_prompt=12,
+    )
+    config.sample.ban_prompt = True
+    config.sample.ban_std_thres = 0.05
+    config.pretrained.model = "/cpfs01/VideoGen/user_data/shenzhebei/Qwen-Image-Edit-2511"
+    config.sample.test_batch_size = 4
+    return config
+