Update configs

config/base.py

@@ -42,7 +42,7 @@ def get_config():
     ###### Sampling ######
     config.sample = sample = ml_collections.ConfigDict()
     # number of sampler inference steps.
-    sample.num_steps = 10
+    sample.num_steps = 50
     # eta parameter for the DDIM sampler. this controls the amount of noise injected into the sampling process, with 0.0
     # being fully deterministic and 1.0 being equivalent to the DDPM sampler.
     sample.eta = 1.0
@@ -61,7 +61,7 @@ def get_config():
     # whether to use the 8bit Adam optimizer from bitsandbytes.
     train.use_8bit_adam = False
     # learning rate.
-    train.learning_rate = 1e-4
+    train.learning_rate = 3e-4
     # Adam beta1.
     train.adam_beta1 = 0.9
     # Adam beta2.
@@ -82,7 +82,7 @@ def get_config():
     # sampling will be used during training.
     train.cfg = True
     # clip advantages to the range [-adv_clip_max, adv_clip_max].
-    train.adv_clip_max = 10
+    train.adv_clip_max = 5
     # the PPO clip range.
     train.clip_range = 1e-4
     # the fraction of timesteps to train on. if set to less than 1.0, the model will be trained on a subset of the

config/dgx.py

@@ -5,28 +5,91 @@ import os
 base = imp.load_source("base", os.path.join(os.path.dirname(__file__), "base.py"))
 
 
-def get_config():
+def compressibility():
     config = base.get_config()
 
-    config.pretrained.model = "runwayml/stable-diffusion-v1-5"
+    config.pretrained.model = "CompVis/stable-diffusion-v1-4"
 
-    config.mixed_precision = "fp16"
-    config.allow_tf32 = True
-    config.use_lora = False
+    config.num_epochs = 100
+    config.use_lora = True
+    config.save_freq = 1
+    config.num_checkpoint_limit = 100000000
 
-    config.train.batch_size = 4
-    config.train.gradient_accumulation_steps = 2
-    config.train.learning_rate = 3e-5
-    config.train.clip_range = 1e-4
-
-    # sampling
-    config.sample.num_steps = 50
+    # the DGX machine I used had 8 GPUs, so this corresponds to 8 * 8 * 4 = 256 samples per epoch.
     config.sample.batch_size = 8
     config.sample.num_batches_per_epoch = 4
 
+    # this corresponds to (8 * 4) / (4 * 2) = 4 gradient updates per epoch.
+    config.train.batch_size = 4
+    config.train.gradient_accumulation_steps = 2
+
+    # prompting
+    config.prompt_fn = "imagenet_animals"
+    config.prompt_fn_kwargs = {}
+
+    # rewards
+    config.reward_fn = "jpeg_compressibility"
+
     config.per_prompt_stat_tracking = {
         "buffer_size": 16,
         "min_count": 16,
     }
 
     return config
+
+
+def incompressibility():
+    config = compressibility()
+    config.reward_fn = "jpeg_incompressibility"
+    return config
+
+
+def aesthetic():
+    config = compressibility()
+    config.num_epochs = 200
+    config.reward_fn = "aesthetic_score"
+
+    # this reward is a bit harder to optimize, so I used 2 gradient updates per epoch.
+    config.train.gradient_accumulation_steps = 4
+
+    config.prompt_fn = "simple_animals"
+    config.per_prompt_stat_tracking = {
+        "buffer_size": 32,
+        "min_count": 16,
+    }
+    return config
+
+
+def prompt_image_alignment():
+    config = compressibility()
+
+    config.num_epochs = 200
+    # for this experiment, I reserved 2 GPUs for LLaVA inference so only 6 could be used for DDPO. the total number of
+    # samples per epoch is 8 * 6 * 6 = 288.
+    config.sample.batch_size = 8
+    config.sample.num_batches_per_epoch = 6
+
+    # again, this one is harder to optimize, so I used (8 * 6) / (4 * 6) = 2 gradient updates per epoch.
+    config.train.batch_size = 4
+    config.train.gradient_accumulation_steps = 6
+
+    # prompting
+    config.prompt_fn = "nouns_activities"
+    config.prompt_fn_kwargs = {
+        "nouns_file": "simple_animals.txt",
+        "activities_file": "activities.txt",
+    }
+
+    # rewards
+    config.reward_fn = "llava_bertscore"
+
+    config.per_prompt_stat_tracking = {
+        "buffer_size": 32,
+        "min_count": 16,
+    }
+
+    return config
+
+
+def get_config(name):
+    return globals()[name]()