style: modify code comments

VideoVerses · Jan 16, 2025 · 14a4128 · 14a4128
1 parent 653ca9e
commit 14a4128
Show file tree

Hide file tree

Showing 3 changed files with 45 additions and 41 deletions.
diff --git a/configs/001_videocrafter2/vc2_t2v_lora.yaml b/configs/001_videocrafter2/vc2_t2v_lora.yaml
@@ -1,11 +1,11 @@
 model:
   base_learning_rate: 6.0e-06 # 1.5e-04
   scale_lr: False
-  # empty_params_only: True # disable this means finetuning all parameters
+  # empty_params_only: True # comment this means finetuning all parameters
   target: src.base.ddpm3d.LVDMFlow
   params:
     lora_args:
-      # lora_ckpt: "/path/to/lora.ckpt" # first time training, no need; this is for resume training. 
+      # lora_ckpt: "/path/to/lora.ckpt" # no need for the first-time training, only used for resume training. 
       target_modules:  ["to_q", "to_k", "to_v"]
       lora_rank: 4
       lora_alpha: 1
@@ -127,7 +127,7 @@ lightning:
     image_logger:
       target: src.utils.callbacks.ImageLogger
       params:
-        batch_frequency: 100000
+        batch_frequency: 1000
         max_images: 2
         to_local: True # save videos into files
         log_images_kwargs:

diff --git a/configs/004_cogvideox/cogvideo2b.yaml b/configs/004_cogvideox/cogvideo2b.yaml
@@ -1,48 +1,50 @@
 model:
-  # there might be differet to load from hf and resume from pl 
-  # pretrained_checkpoint: "THUDM/CogVideoX-2b"
   base_learning_rate: 6e-6
   target: src.cogvideo_hf.cogvideo_pl.CogVideoXWorkFlow
   params: 
-    # first stage model; cond stage model ; denoising model ; scheduler 
+    # VAE of CogVideoX
     first_stage_config:
       target: diffusers.AutoencoderKLCogVideoX
       params:
         pretrained_model_name_or_path: checkpoints/cogvideo/CogVideoX-2b 
         subfolder: "vae"
+
+    # Text encoder (T5) of CogVideoX
     cond_stage_config:
       target: src.lvdm.modules.encoders.condition.FrozenT5Embedder
       params:
         version: "DeepFloyd/t5-v1_1-xxl"
         device: "cuda"
         max_length: 226 
         freeze: True
-    # denosier config equal to unet config in vc 
+
+    # Denosier model
     denoiser_config:
       target: diffusers.CogVideoXTransformer3DModel
       params:
         pretrained_model_name_or_path: checkpoints/cogvideo/CogVideoX-2b 
         subfolder: "transformer"
-        load_dtype: fp16 # bf16 5b fp16 2B 
+        load_dtype: fp16 # bf16 5b / fp16 2B 
         # revision: null 
         # variant: null
-    adapter_config: # the whole dict is remoable 
-        target: peft.LoraConfig
-        params:
-            r: 4
-            lora_alpha: 1.0 
-            init_lora_weights: True
-            target_modules: ["to_k", "to_q", "to_v", "to_out.0"]
+
+    # Lora module 
+    adapter_config:
+      target: peft.LoraConfig
+      params:
+        r: 4
+        lora_alpha: 1.0 
+        init_lora_weights: True
+        target_modules: ["to_k", "to_q", "to_v", "to_out.0"]
 
-    # sampler config. Wrap it.       
+    # Diffusion sampling scheduler
     scheduler_config:
       target: diffusers.CogVideoXDPMScheduler
       params:
         pretrained_model_name_or_path: checkpoints/cogvideo/CogVideoX-2b 
         subfolder: scheduler
 
-## training config 
-### data , can a toy dataset given 
+# data configs
 data:
   target: src.data.lightning_data.DataModuleFromConfig
   params:
@@ -66,23 +68,23 @@ data:
         cache_dir: ~/.cache
         id_token: null
 
-### training_step in cogvideoxft
+# training configs
 lightning:
   trainer:
     benchmark: True
     num_nodes: 1
     accumulate_grad_batches: 2
     max_epochs: 2000
-    precision: 32 # training precision
+    precision: 32
   callbacks:
     image_logger:
       target: src.utils.callbacks.ImageLogger
       params:
         batch_frequency: 100000
         max_images: 2
-        to_local: True # save videos into files
+        to_local: True # save videos into local files
         log_images_kwargs:
-          unconditional_guidance_scale: 12 # need this, otherwise it is grey
+          unconditional_guidance_scale: 6
     metrics_over_trainsteps_checkpoint:
       target: pytorch_lightning.callbacks.ModelCheckpoint
       params:

diff --git a/configs/004_cogvideox/cogvideo5b.yaml b/configs/004_cogvideox/cogvideo5b.yaml
@@ -1,48 +1,50 @@
 model:
-  # there might be differet to load from hf and resume from pl 
-  # pretrained_checkpoint: "THUDM/CogVideoX-2b"
   base_learning_rate: 6e-6
   target: src.cogvideo_hf.cogvideo_pl.CogVideoXWorkFlow
   params: 
-    # first stage model; cond stage model ; denoising model ; scheduler 
+    # VAE of CogVideoX
     first_stage_config:
       target: diffusers.AutoencoderKLCogVideoX
       params:
         pretrained_model_name_or_path: checkpoints/cogvideo/CogVideoX-5b 
         subfolder: "vae"
+
+    # Text encoder (T5) of CogVideoX
     cond_stage_config:
       target: src.lvdm.modules.encoders.condition.FrozenT5Embedder
       params:
         version: "DeepFloyd/t5-v1_1-xxl"
         device: "cuda"
         max_length: 226 
         freeze: True
-    # denosier config equal to unet config in vc 
+
+    # Denosier model
     denoiser_config:
       target: diffusers.CogVideoXTransformer3DModel
       params:
         pretrained_model_name_or_path: checkpoints/cogvideo/CogVideoX-5b 
         subfolder: "transformer"
-        load_dtype: fp16 # bf16 5b fp16 2B 
+        load_dtype: fp16 # bf16 5b / fp16 2B 
         # revision: null 
         # variant: null
-    adapter_config: # the whole dict is remoable 
-        target: peft.LoraConfig
-        params:
-            r: 4
-            lora_alpha: 1.0 
-            init_lora_weights: True
-            target_modules: ["to_k", "to_q", "to_v", "to_out.0"]
+
+    # Lora module 
+    adapter_config: 
+      target: peft.LoraConfig
+      params:
+        r: 4
+        lora_alpha: 1.0 
+        init_lora_weights: True
+        target_modules: ["to_k", "to_q", "to_v", "to_out.0"]
 
-    # sampler config. Wrap it.       
+    # Diffusion sampling scheduler
     scheduler_config:
       target: diffusers.CogVideoXDPMScheduler
       params:
         pretrained_model_name_or_path: checkpoints/cogvideo/CogVideoX-5b
         subfolder: scheduler
 
-## training config 
-### data , can a toy dataset given 
+# data configs
 data:
   target: src.data.lightning_data.DataModuleFromConfig
   params:
@@ -66,23 +68,23 @@ data:
         cache_dir: ~/.cache
         id_token: null
 
-### training_step in cogvideoxft
+# training configs
 lightning:
   trainer:
     benchmark: True
     num_nodes: 1
     accumulate_grad_batches: 2
     max_epochs: 2000
-    precision: 32 # training precision
+    precision: 32
   callbacks:
     image_logger:
       target: src.utils.callbacks.ImageLogger
       params:
         batch_frequency: 100000
         max_images: 2
-        to_local: True # save videos into files
+        to_local: True # save videos into local files
         log_images_kwargs:
-          unconditional_guidance_scale: 12 # need this, otherwise it is grey
+          unconditional_guidance_scale: 6
     metrics_over_trainsteps_checkpoint:
       target: pytorch_lightning.callbacks.ModelCheckpoint
       params: