ADD

Author: Animadversio

Diffusion_training_demo.py

@@ -156,8 +156,11 @@ def train_score_model(score_model, dataset, lr, n_epochs, batch_size, ckpt_name,
 #%%
 
 #%%
+from tqdm import tqdm
+from torch.utils.data import DataLoader, TensorDataset
 from torchvision.datasets import CelebA
 from torchvision.transforms import ToTensor, CenterCrop, Resize, Compose, Normalize
+
 tfm = Compose([
     Resize(32),
     CenterCrop(32),
@@ -167,9 +170,7 @@ def train_score_model(score_model, dataset, lr, n_epochs, batch_size, ckpt_name,
 dataset_rsz = CelebA("/home/binxuwang/Datasets", target_type=["attr"],
                     transform=tfm, download=False) # ,"identity"
 #%%
-from torch.utils.data import DataLoader, TensorDataset
-from tqdm import tqdm
-
+# def preprocess_dataset(dataset_rsz, ):
 dataloader = DataLoader(dataset_rsz, batch_size=64, num_workers=8, shuffle=False)
 x_col = []
 y_col = []
@@ -181,12 +182,12 @@ def train_score_model(score_model, dataset, lr, n_epochs, batch_size, ckpt_name,
 print(x_col.shape)
 print(y_col.shape)
 
-maxlen = (y_col.sum(dim=1)).max()
 nantoken = 40
-yseq_data = torch.ones(y_col.size(0), maxlen,
-                       dtype=int).fill_(nantoken)
+maxlen = (y_col.sum(dim=1)).max()
+yseq_data = torch.ones(y_col.size(0), maxlen, dtype=int).fill_(nantoken)
 
 saved_dataset = TensorDataset(x_col, yseq_data)
+# return saved_dataset
 #%%
 import matplotlib.pyplot as plt
 

StableDiffusion_exps.py

@@ -35,6 +35,8 @@ def recursive_print(module, prefix="", depth=0, deepest=3):
 ).to("cuda")
 def dummy_checker(images, **kwargs): return images, False
 pipe.safety_checker = dummy_checker
+#%%
+recursive_print(pipe.unet, deepest=2)
 #%% Text to
 # prompt = "a photo of an ballerina riding a horse on mars"
 prompt = "A ballerina riding a Harley Motorcycle, CG Art"
@@ -76,13 +78,13 @@ def save_latents(i, t, latents):
 prompt = "A ballerina chasing her cat running on the grass in the style of Monet"
 prompt = "A kitty cat dressed like Lincoln, old timey style"
 with autocast("cuda"):
-    image = pipe(prompt, callback=None)["sample"][0]  # plot_show_callback
+    image = pipe(prompt, callback=plot_show_callback)["sample"][0]  # plot_show_callback
 
 image.save("cat_Lincoln.png")
 plt_show_image(image)
 #%%
 len(latents_reservoir)
-plt_show_image(latents_reservoir[-10][0, [0, 1, 2,], :].permute(1, 2, 0).cpu().numpy() / 1.6 + 0.4)
+plt_show_image(latents_reservoir[-1][0, [0, 1, 2,], :].permute(1, 2, 0).cpu().numpy() / 1.6 + 0.4)
 #%% Visualize architecture
 
 #%% Full unets