text_seq_len=TEXTSEQLEN, # text sequence length
depth=12, # should aim to be 64
heads=16, # attention heads
dim_head=64, # attention head dimension
attn_dropout=0.1, # attention dropout
ff_dropout=0.1 # feedforward dropout
).cuda()
dalle.load_state_dict(torch.load("dalle-small.pth"))
"""
text = torch.randint(0, NUM_TOKENS, (BATCH_SIZE, TEXTSEQLEN))
images = torch.randn(BATCH_SIZE, 3, IMAGE_SIZE, IMAGE_SIZE)
mask = torch.ones_like(text).bool()
"""
tokenDset = token_dataset('./coco/merged-smallsample.txt')
# do the above for a long time with a lot of data ... then
num_pics = 30
def denorm(img: torch.Tensor):
mean = torch.mean(img)
min_maxrange = (torch.max(img) - torch.min(img))
return (((img - mean) / (min_maxrange) + 0.5) * 255)
for i in range(30):
test_text = "犬が地面に寝そべっている写真"
images = torch.randn(BATCH_SIZE, 3, IMAGE_SIZE, IMAGE_SIZE)
mask = torch.ones_like(text).bool()
"""
cap = dset.CocoCaptions(
root='./coco/images',
annFile='./coco/annotations/captions_val2014.json',
transform=transforms.Compose([
#transforms.RandomCrop((IMAGE_SIZE,IMAGE_SIZE),pad_if_needed=True),
#transforms.Grayscale(),
transforms.Resize((IMAGE_SIZE, IMAGE_SIZE), Image.BILINEAR),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
]))
tokenDset = token_dataset('./coco/merged-1000.txt')
VAEloss = []
for epoch in range(EPOCHS):
for i in range(DATASET_SIZE):
#print(i,":",tokenDset.getRand(i),img.size())
optimizerVAE.zero_grad()
img, _ = cap[i]
img = img.unsqueeze(0).cuda()
#print(img.size())
if i % 10 == 0:
print("VAE epoch {} / {}".format(i + epoch * DATASET_SIZE,
EPOCHS * DATASET_SIZE))
loss = vae(img, return_recon_loss=True)
VAEloss.append(loss.cpu().detach().numpy())