a-next_token.py

import math
import random
from functools import partial
from typing import Optional, Tuple, Union
from itertools import chain

import torch
import torch.nn as nn
from torch.nn import functional as F

from models.basic_var import AdaLNBeforeHead, AdaLNSelfAttn
from models.helpers import gumbel_softmax_with_rng, sample_with_top_k_top_p_
from models.vqvae import VQVAE, VectorQuantizer2


patch_nums=(1, 2, 3)
L = sum(pn ** 2 for pn in patch_nums)
first_l = patch_nums[0] ** 2
num_stages_minus_1 = len(patch_nums) - 1 


B=8
C=1024
Cvae=32
init_std = math.sqrt(1 / C / 3)
num_classes=1000
class_emb = nn.Embedding(num_classes + 1, C)

# input
vae_local = VQVAE()
quant: VectorQuantizer2 = vae_local.quantize
vae_proxy: Tuple[VQVAE] = (vae_local,)
vae_quant_proxy: Tuple[VectorQuantizer2] = (quant,)
word_embed = nn.Linear(Cvae, C)
uniform_prob = torch.full((1, num_classes), fill_value=1.0 / num_classes, dtype=torch.float32) 

# Initialize random number generator
rng = torch.Generator()
rng.manual_seed(42)  # Set a fixed seed for reproducibility

label_B = torch.multinomial(uniform_prob, num_samples=B, replacement=True, generator=rng).reshape(B)    

sos = cond_BD = class_emb(torch.cat((label_B, torch.full_like(label_B, fill_value=num_classes)), dim=0))

# absolute position embedding
pos_1LC = []
for i, pn in enumerate(patch_nums):
    pe = torch.empty(1, pn*pn, C)
    nn.init.trunc_normal_(pe, mean=0, std=init_std)
    pos_1LC.append(pe)
pos_1LC = torch.cat(pos_1LC, dim=1)     # 1, L, C
assert tuple(pos_1LC.shape) == (1, L, C)
pos_1LC = nn.Parameter(pos_1LC)
# level embedding (similar to GPT's segment embedding, used to distinguish different levels of token pyramid)
lvl_embed = nn.Embedding(len(patch_nums), C)
nn.init.trunc_normal_(lvl_embed.weight.data, mean=0, std=init_std)

pos_start = nn.Parameter(torch.empty(1, first_l, C))

# attention mask
d: torch.Tensor = torch.cat([torch.full((pn*pn,), i) for i, pn in enumerate(patch_nums)]).view(1, L, 1)
dT = d.transpose(1, 2)    # dT: 11L
lvl_1L = dT[:, 0].contiguous()
attn_bias_for_masking = torch.where(d >= dT, 0., -torch.inf).reshape(1, 1, L, L)
       

lvl_pos = lvl_embed(lvl_1L) + pos_1LC
next_token_map = sos.unsqueeze(1).expand(2 * B, first_l, -1) + pos_start.expand(2 * B, first_l, -1) + lvl_pos[:, :first_l]

cur_L = 0
f_hat = sos.new_zeros(B, Cvae, patch_nums[-1], patch_nums[-1])


for si, pn in enumerate(patch_nums):   # si: i-th segment
    print("si pn")
    print(si, pn)
    print()
    ratio = si / num_stages_minus_1
    # last_L = cur_L
    print("cur_L")
    cur_L += pn*pn
    print(cur_L)
    print()
    

    h_BChw = torch.randn(B, L, C)
    h_BChw = h_BChw.transpose_(1, 2).reshape(B, Cvae, pn, pn)
    f_hat, next_token_map = vae_quant_proxy[0].get_next_autoregressive_input(si, len(patch_nums), f_hat, h_BChw)
            
    if si != num_stages_minus_1:   # prepare for next stage
        next_token_map = next_token_map.view(B, Cvae, -1).transpose(1, 2)
        print(next_token_map)
        print()
        next_token_map = word_embed(next_token_map) + lvl_pos[:, cur_L:cur_L + patch_nums[si+1] ** 2]
        print(next_token_map)
        print()
        next_token_map = next_token_map.repeat(2, 1, 1)   # double the batch sizes due to CFG
        print(next_token_map)
        print()

#print(f_hat)