src/Train_MLPMixer-Copy1.py

#!/usr/bin/env python
# coding: utf-8

# In[1]:


# # Code to convert this notebook to .py if you want to run it via command line or with Slurm
#from subprocess import call
#command = "jupyter nbconvert Train_MLPMixer-Copy1.ipynb --to python"
#call(command,shell=True)


# # Import packages & functions

# In[2]:


import os
import sys
import json
import argparse
import numpy as np
import math
from einops import rearrange
import time
import random
import string
import h5py
from tqdm import tqdm

import webdataset as wds
import gc

import matplotlib.pyplot as plt
import torch
import torch.nn as nn
from torchvision import transforms

from accelerate import Accelerator, DeepSpeedPlugin

# tf32 data type is faster than standard float32
torch.backends.cuda.matmul.allow_tf32 = True

# custom functions #
import utils


# In[3]:


### Multi-GPU config ###
local_rank = os.getenv('RANK')
if local_rank is None: 
    local_rank = 0
else:
    local_rank = int(local_rank)
print("LOCAL RANK ", local_rank)  

num_devices = torch.cuda.device_count()
if num_devices==0: num_devices = 1

# ## UNCOMMENT BELOW SECTION AND COMMENT OUT DEEPSPEED SECTION TO AVOID USING DEEPSPEED ###
# accelerator = Accelerator(split_batches=False, mixed_precision="fp16")
# global_batch_size = batch_size = 128
# data_type = torch.float16 # change depending on your mixed_precision

### DEEPSPEED INITIALIZATION ###
if num_devices <= 1 and utils.is_interactive():
    global_batch_size = batch_size = 128
    print(f"Setting batch_size to {batch_size}")
    # can emulate a distributed environment for deepspeed to work in jupyter notebook
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = str(np.random.randint(10000)+9000)
    os.environ["RANK"] = "0"
    os.environ["LOCAL_RANK"] = "0"
    os.environ["WORLD_SIZE"] = "1"
    os.environ["GLOBAL_BATCH_SIZE"] = str(global_batch_size) # set this to your batch size!
else:
    global_batch_size = os.environ["GLOBAL_BATCH_SIZE"]    
    batch_size = int(os.environ["GLOBAL_BATCH_SIZE"]) // num_devices

# alter the deepspeed config according to your global and local batch size
if local_rank == 0:
    with open('deepspeed_config_stage2.json', 'r') as file:
        config = json.load(file)
    config['train_batch_size'] = int(os.environ["GLOBAL_BATCH_SIZE"])
    config['train_micro_batch_size_per_gpu'] = batch_size
    config['bf16'] = {'enabled': False}
    config['fp16'] = {'enabled': True}
    with open('deepspeed_config_stage2.json', 'w') as file:
        json.dump(config, file)
else:
    # give some time for the local_rank=0 gpu to prep new deepspeed config file
    time.sleep(10)
deepspeed_plugin = DeepSpeedPlugin("deepspeed_config_stage2_cpuoffload.json")
accelerator = Accelerator(split_batches=False, deepspeed_plugin=deepspeed_plugin)


# In[4]:


print("PID of this process =",os.getpid())
device = accelerator.device
print("device:",device)
num_workers = num_devices
print(accelerator.state)
world_size = accelerator.state.num_processes
distributed = not accelerator.state.distributed_type == 'NO'

# set data_type to match your mixed precision (automatically set based on deepspeed config)
if accelerator.mixed_precision == "bf16":
    data_type = torch.bfloat16
elif accelerator.mixed_precision == "fp16":
    data_type = torch.float16
else:
    data_type = torch.float32

print("distributed =",distributed, "num_devices =", num_devices, "local rank =", local_rank, "world size =", world_size, "data_type =", data_type)
print = accelerator.print # only print if local_rank=0


# # Configurations

# In[5]:


# if running this interactively, can specify jupyter_args here for argparser to use
if utils.is_interactive():
    # create random model_name
    model_name = ''.join(random.choices(string.ascii_letters + string.digits, k=10))
    model_name = model_name + "_interactive"
    print("model_name:", model_name)

    # global_batch_size and batch_size should already be defined in the above cells
    # other variables can be specified in the following string:
    jupyter_args = f"--data_path=/fsx/proj-fmri/shared/mindeyev2_dataset \
                    --model_name={model_name} \
                    --subj=1 --batch_size={batch_size} --no-blurry_recon --no-depth_recon --hidden_dim=4096 \
                    --clip_scale=1. --blur_scale=100. --depth_scale=100. \
                    --max_lr=3e-4 --mixup_pct=.66 --num_epochs=12 --ckpt_interval=999 --no-use_image_aug --no-ckpt_saving"

    jupyter_args = jupyter_args.split()
    print(jupyter_args)
    
    from IPython.display import clear_output # function to clear print outputs in cell
    get_ipython().run_line_magic('load_ext', 'autoreload')
    # this allows you to change functions in models.py or utils.py and have this notebook automatically update with your revisions
    get_ipython().run_line_magic('autoreload', '2')


# In[6]:


parser = argparse.ArgumentParser(description="Model Training Configuration")
parser.add_argument(
    "--model_name", type=str, default="testing",
    help="name of model, used for ckpt saving and wandb logging (if enabled)",
)
parser.add_argument(
    "--data_path", type=str, default="/fsx/proj-fmri/shared/natural-scenes-dataset",
    help="Path to where NSD data is stored / where to download it to",
)
parser.add_argument(
    "--subj",type=int, default=1, choices=[1,2,5,7],
)
parser.add_argument(
    "--batch_size", type=int, default=32,
    help="Batch size can be increased by 10x if only training v2c and not diffusion diffuser",
)
parser.add_argument(
    "--wandb_log",action=argparse.BooleanOptionalAction,default=False,
    help="whether to log to wandb",
)
parser.add_argument(
    "--resume_from_ckpt",action=argparse.BooleanOptionalAction,default=False,
    help="if not using wandb and want to resume from a ckpt",
)
parser.add_argument(
    "--wandb_project",type=str,default="stability",
    help="wandb project name",
)
parser.add_argument(
    "--mixup_pct",type=float,default=.33,
    help="proportion of way through training when to switch from BiMixCo to SoftCLIP",
)
parser.add_argument(
    "--blurry_recon",action=argparse.BooleanOptionalAction,default=True,
    help="whether to output blurry reconstructions",
)
parser.add_argument(
    "--depth_recon",action=argparse.BooleanOptionalAction,default=True,
    help="whether to output depth reconstructions",
)
parser.add_argument(
    "--blur_scale",type=float,default=100.,
    help="multiply loss from blurry recons by this number",
)
parser.add_argument(
    "--depth_scale",type=float,default=100.,
    help="multiply loss from depth recons by this number",
)
parser.add_argument(
    "--clip_scale",type=float,default=1.,
    help="multiply contrastive loss by this number",
)
parser.add_argument(
    "--use_image_aug",action=argparse.BooleanOptionalAction,default=True,
    help="whether to use image augmentation",
)
parser.add_argument(
    "--num_epochs",type=int,default=120,
    help="number of epochs of training",
)
parser.add_argument(
    "--hidden_dim",type=int,default=4096,
)
parser.add_argument(
    "--lr_scheduler_type",type=str,default='cycle',choices=['cycle','linear'],
)
parser.add_argument(
    "--ckpt_saving",action=argparse.BooleanOptionalAction,default=True,
)
parser.add_argument(
    "--ckpt_interval",type=int,default=5,
    help="save backup ckpt and reconstruct every x epochs",
)
parser.add_argument(
    "--seed",type=int,default=42,
)
parser.add_argument(
    "--max_lr",type=float,default=3e-4,
)

if utils.is_interactive():
    args = parser.parse_args(jupyter_args)
else:
    args = parser.parse_args()

# create global variables without the args prefix
for attribute_name in vars(args).keys():
    globals()[attribute_name] = getattr(args, attribute_name)


# In[7]:


outdir = os.path.abspath(f'../train_logs/{model_name}')
if not os.path.exists(outdir) and ckpt_saving:
    os.makedirs(outdir,exist_ok=True)
if use_image_aug:
    import kornia
    from kornia.augmentation.container import AugmentationSequential
    img_augment = AugmentationSequential(
        kornia.augmentation.RandomResizedCrop((224,224), (0.6,1), p=0.3),
        kornia.augmentation.Resize((224, 224)),
        kornia.augmentation.RandomHorizontalFlip(p=0.3),
        kornia.augmentation.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.2, hue=0.1, p=0.3),
        kornia.augmentation.RandomGrayscale(p=0.3),
        same_on_batch=False,
        data_keys=["input"],
    )


# # Prep data, models, and dataloaders

# ## Dataloader

# In[8]:


if subj==1:
    num_train = 24958
    num_test = 2770
test_batch_size = num_test

def my_split_by_node(urls): return urls
    
train_url = f"{data_path}/wds/subj0{subj}/train/" + "{0..36}.tar"
# train_url = f"{data_path}/wds/subj0{subj}/train/" + "{0..1}.tar"
print(train_url)

train_data = wds.WebDataset(train_url,resampled=False,nodesplitter=my_split_by_node)\
                    .shuffle(750, initial=1500, rng=random.Random(42))\
                    .decode("torch")\
                    .rename(behav="behav.npy", past_behav="past_behav.npy", future_behav="future_behav.npy", olds_behav="olds_behav.npy")\
                    .to_tuple(*["behav", "past_behav", "future_behav", "olds_behav"])
train_dl = torch.utils.data.DataLoader(train_data, batch_size=batch_size, shuffle=False, drop_last=True, pin_memory=True)

test_url = f"{data_path}/wds/subj0{subj}/test/" + "0.tar"
print(test_url)

test_data = wds.WebDataset(test_url,resampled=False,nodesplitter=my_split_by_node)\
                    .shuffle(750, initial=1500, rng=random.Random(42))\
                    .decode("torch")\
                    .rename(behav="behav.npy", past_behav="past_behav.npy", future_behav="future_behav.npy", olds_behav="olds_behav.npy")\
                    .to_tuple(*["behav", "past_behav", "future_behav", "olds_behav"])
test_dl = torch.utils.data.DataLoader(test_data, batch_size=test_batch_size, shuffle=False, drop_last=True, pin_memory=True)


# ### check dataloaders are working

# In[9]:


test_vox_indices = []
test_73k_images = []
for test_i, (behav, past_behav, future_behav, old_behav) in enumerate(test_dl):
    test_vox_indices = np.append(test_vox_indices, behav[:,0,5].cpu().numpy())
    test_73k_images = np.append(test_73k_images, behav[:,0,0].cpu().numpy())
test_vox_indices = test_vox_indices.astype(np.int16)
print(test_i, (test_i+1) * test_batch_size, len(test_vox_indices))
print("---\n")

train_vox_indices = []
train_73k_images = []
for train_i, (behav, past_behav, future_behav, old_behav) in enumerate(train_dl):
    train_vox_indices = np.append(train_vox_indices, behav[:,0,5].long().cpu().numpy())
    train_73k_images = np.append(train_73k_images, behav[:,0,0].cpu().numpy())
train_vox_indices = train_vox_indices.astype(np.int16)
print(train_i, (train_i+1) * batch_size, len(train_vox_indices))


# ## Load data and images

# In[10]:


# load betas
f = h5py.File(f'{data_path}/betas_all_subj0{subj}.hdf5', 'r')
# f = h5py.File(f'{data_path}/betas_subj0{subj}_thresholded_wholebrain.hdf5', 'r')

voxels = f['betas'][:]
print(f"subj0{subj} betas loaded into memory")
voxels = torch.Tensor(voxels).to("cpu").to(data_type)
print("voxels", voxels.shape)
num_voxels = voxels.shape[-1]

# load orig images
f = h5py.File(f'{data_path}/coco_images_224_float16.hdf5', 'r')
images = f['images'][:]
images = torch.Tensor(images).to("cpu").to(data_type)
print("images", images.shape)


# ## Load models

# ### CLIP image embeddings  model

# In[11]:


from models import Clipper
clip_model = Clipper("ViT-L/14", device=torch.device(f"cuda:{local_rank}"), hidden_state=True, norm_embs=True)
clip_seq_dim = 257
clip_emb_dim = 768 #1024
# hidden_dim = 4096
seq_len = 1 #2 #32 


# ### SD VAE

# In[12]:


# if blurry_recon:
#     from diffusers import AutoencoderKL
#     autoenc = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16, cache_dir="/fsx/proj-fmri/shared/cache")
#     # autoenc.load_state_dict(torch.load('../train_logs/sdxl_vae_normed/best.pth')["model_state_dict"])
#     autoenc.eval()
#     autoenc.requires_grad_(False)
#     autoenc.to(device)
#     utils.count_params(autoenc)

if blurry_recon:# or depth_recon:
    from diffusers import VQModel
    autoenc = VQModel.from_pretrained("/fsx/proj-fmri/shared/cache/models--microsoft--vq-diffusion-ithq/snapshots/3f796fb49ee559370dc638dea1d8116af131d993/vqvae", torch_dtype=data_type)
    autoenc.eval()
    autoenc.requires_grad_(False)
    autoenc.to(device)
    utils.count_params(autoenc)


# #### downsampled images

# In[13]:


if blurry_recon:
    if utils.is_interactive(): display(utils.torch_to_Image(images[[30]]))

    input_batch = images[[30]].to(device)
    print(input_batch.shape)

    downsampled_image = nn.functional.interpolate(input_batch, size=(8, 8), mode='bilinear', align_corners=False)
    re_upsampled_image = nn.functional.interpolate(downsampled_image, size=(128, 128), mode='nearest')
    re_upsampled_enc = autoenc.encode(2*re_upsampled_image-1).latents * 0.18215
    print(re_upsampled_enc.shape)
    
    if utils.is_interactive(): display(utils.torch_to_Image((autoenc.decode(re_upsampled_enc/0.18215).sample / 2 + 0.5).clamp(0,1)))


# #### MiDaS depth

# In[14]:


if depth_recon:
    from controlnet_aux.midas import MidasDetector
    
    midas_depth = MidasDetector.from_pretrained(
      "valhalla/t2iadapter-aux-models", filename="dpt_large_384.pt", model_type="dpt_large", cache_dir="/fsx/proj-fmri/shared/cache").to(device)
    midas_depth.model.eval()
    midas_depth.model.requires_grad_(False)
    midas_depth.model.to(device)
    pass


# In[15]:


if depth_recon:
    if utils.is_interactive(): display(utils.torch_to_Image(images[[30]]))

    input_batch = images[[30,31]].float().to(device)
    print(input_batch.shape)
    
    midas_emb = midas_depth.model(input_batch).unsqueeze(1)
    print(midas_emb.shape)

    prediction = utils.resize(midas_emb, 32) #/30).clamp(0,1).half() # 30 is roughly prediction.max()
    print(prediction.shape)
    
    prediction = (prediction / prediction.view(prediction.shape[0], -1).max(dim=1)[0].view(-1, 1, 1, 1).expand_as(prediction)).half()
    midas_emb_size = prediction.flatten(1).shape[1]
    print("midas_emb", prediction.shape, prediction.min(), prediction.max())
    print("midas_emb_size", midas_emb_size)
    
    if utils.is_interactive(): display(utils.torch_to_Image(utils.resize(prediction, 224))) 

    if blurry_recon:
        prediction = utils.resize(midas_emb, 128).half().repeat(1,3,1,1)
        prediction = (prediction / prediction.view(prediction.shape[0], -1).max(dim=1)[0].view(-1, 1, 1, 1).expand_as(prediction)).half()
        prediction_enc = autoenc.encode(2*prediction-1).latents * 0.18215
        print("vae midas_emb", prediction_enc.shape, prediction_enc.min(), prediction_enc.max())
    
        if utils.is_interactive(): display(utils.torch_to_Image((autoenc.decode(prediction_enc/0.18215).sample / 2 + 0.5).clamp(0,1)))


# ### MindEye modules

# In[16]:


class MindEyeModule(nn.Module):
    def __init__(self):
        super(MindEyeModule, self).__init__()
    def forward(self, x):
        return x
        
model = MindEyeModule()
model


# In[17]:


time_embedding_dim = 512

class RidgeRegression(torch.nn.Module):
    # make sure to add weight_decay when initializing optimizer
    def __init__(self, input_size, out_features): 
        super(RidgeRegression, self).__init__()
        self.out_features = out_features
        self.linear = torch.nn.Linear(input_size, out_features)
    def forward(self, x):
        return self.linear(x)
        
model.ridge = RidgeRegression(voxels.shape[1] + time_embedding_dim, out_features=hidden_dim)
utils.count_params(model.ridge)
utils.count_params(model)

b = torch.randn((2,1,voxels.shape[1]))
time_emb_test = torch.randn((2,1,time_embedding_dim))
print(b.shape, model.ridge(torch.cat((b,time_emb_test),dim=-1)).shape)


# In[59]:


num_past_voxels = 15
seq_len = 1 + 1


# In[73]:


class BrainNetwork(nn.Module):
    def __init__(self, out_dim=768, in_dim=15724, seq_len=2, h=4096, n_blocks=4, drop=.15, clip_size=768):
        super().__init__()
        self.seq_len = seq_len
        self.h = h
        self.clip_size = clip_size
        
        # Initial linear layer to match the input dimensions to hidden dimensions
        # self.lin0 = nn.Linear(in_dim, seq_len * h)
        
        # Mixer Blocks
        self.mixer_blocks1 = nn.ModuleList([
            self.mixer_block1(h, drop) for _ in range(n_blocks)
        ])
        self.mixer_blocks2 = nn.ModuleList([
            self.mixer_block2(seq_len, drop) for _ in range(n_blocks)
        ])
        
        # Output linear layer
        self.clin1 = nn.Linear(h * seq_len, out_dim, bias=True)

        # low-rank matrices
        # self.rank = 500
        # self.U = nn.Parameter(torch.randn(self.rank, out_dim))
        # self.V = nn.Parameter(torch.randn(h * seq_len, self.rank))
        # self.S = nn.Parameter(torch.randn(out_dim))

        self.clip_proj = nn.Sequential(
            nn.LayerNorm(clip_size),
            nn.GELU(),
            nn.Linear(clip_size, 2048),
            nn.LayerNorm(2048),
            nn.GELU(),
            nn.Linear(2048, 2048),
            nn.LayerNorm(2048),
            nn.GELU(),
            nn.Linear(2048, clip_size)
        )

        if blurry_recon:
            # self.blin1 = nn.Sequential(
            #     nn.Linear(out_dim, 4096, bias=True),
            #     nn.LayerNorm(4096),
            #     nn.GELU(),
            #     nn.Linear(4096, 4096))
            self.blin1 = nn.Linear(h*seq_len, 4096)
            self.bgroupnorm = nn.GroupNorm(1, 256)
            self.bupsampler = Decoder(
                in_channels=256,
                out_channels=128,
                up_block_types=["UpDecoderBlock2D","UpDecoderBlock2D","UpDecoderBlock2D"],
                block_out_channels=[32, 64, 128],
                layers_per_block=1,
            )

        if depth_recon:
            # self.dlin1 = nn.Sequential(
            #         nn.Linear(h, midas_emb_size),
            #         nn.Sigmoid(),
            #     )
            self.dlin1 = nn.Linear(h*seq_len, 4096)
            self.dgroupnorm = nn.GroupNorm(1, 256)
            self.dupsampler = Decoder(
                in_channels=256,
                out_channels=1,#128,
                up_block_types=["UpDecoderBlock2D","UpDecoderBlock2D","UpDecoderBlock2D","UpDecoderBlock2D"],
                block_out_channels=[32, 64, 128, 256],
                layers_per_block=1,
            )
        
    def mixer_block1(self, h, drop):
        return nn.Sequential(
            nn.LayerNorm(h),
            self.mlp(h, h, drop),  # Token mixing
        )

    def mixer_block2(self, seq_len, drop):
        return nn.Sequential(
            nn.LayerNorm(seq_len),
            self.mlp(seq_len, seq_len, drop)  # Channel mixing
        )
    
    def mlp(self, in_dim, out_dim, drop):
        return nn.Sequential(
            nn.Linear(in_dim, out_dim),
            nn.GELU(),
            nn.Dropout(drop),
            nn.Linear(out_dim, out_dim),
        )
        
    def forward(self, x):
        # make empty tensors for blur and depth outputs
        b,d = torch.Tensor([0.]), torch.Tensor([0.])
        
        # Initial linear layer
        # x = self.lin0(x)
        
        # Reshape to seq_len by dim
        # x = x.reshape(-1, self.seq_len, self.h)
        
        # Mixer blocks
        residual1 = x
        residual2 = x.permute(0,2,1)
        for block1, block2 in zip(self.mixer_blocks1,self.mixer_blocks2):
            x = block1(x) + residual1
            residual1 = x
            x = x.permute(0,2,1)
            
            x = block2(x) + residual2
            residual2 = x
            x = x.permute(0,2,1)
        
        # Flatten
        x = x.reshape(x.size(0), -1)
        
        c = self.clin1(x)

        # low rank linear to out dim cuts # params by nearly half compared to full linear mapping
        # c = (x @ (self.V/100) @ (self.U/100)) + self.S
        
        c = self.clip_proj(c.reshape(len(c), -1, self.clip_size))

        if blurry_recon:
            b = self.blin1(x)
            b = b.reshape(len(b), 256, 4, 4)
            b = self.bgroupnorm(b)
            b = self.bupsampler(b)
            
        if depth_recon:
            d = self.dlin1(x)#.reshape(len(x), 1, 32, 32)
            d = d.reshape(len(d), 256, 4, 4)
            d = self.dgroupnorm(d)
            d = self.dupsampler(d)
        
        return c, b, d


class TimeEmbedding(nn.Module):
    def __init__(self, embedding_time_dim=512, num_past_voxels=15):
        super().__init__()
        self.embedding_time = nn.Embedding(num_past_voxels, embedding_time_dim)
        self.num_past_voxels = num_past_voxels
        self.embedding_time_dim = embedding_time_dim

    def forward(self, time):
        # time is (batch_size,)
        time = time.long()
        time = self.embedding_time(time)
        return time # (batch_size, embedding_time_dim)
    

#model.memory_encoder = MemoryEncoder(in_dim=voxels.shape[1], out_dim=4096, num_past_voxels=15, embedding_time_dim=512)
model.time_embedding = TimeEmbedding(embedding_time_dim=512, num_past_voxels=15)

model.backbone = BrainNetwork(h=1024, in_dim=1024, seq_len=4, clip_size=clip_emb_dim, out_dim=clip_emb_dim*clip_seq_dim) 
utils.count_params(model.backbone)
utils.count_params(model)


# test that the model works on some fake data
b = torch.randn((256,4,1024))
print("b.shape",b.shape)
with torch.no_grad():
    clip_, blur_, depth_ = model.backbone(b)
print(clip_.shape, blur_.shape, depth_.shape)


# In[70]:


voxel_ridge = torch.randn(512,4096)
voxel_ridge = voxel_ridge.view(int(voxel_ridge.shape[0]/seq_len), seq_len, hidden_dim)
print("b.shape",voxel_ridge.shape)
with torch.no_grad():
    clip_, blur_, depth_ = model.backbone(voxel_ridge)
print(clip_.shape, blur_.shape, depth_.shape)


# In[64]:


no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
opt_grouped_parameters = [
    {'params': [p for n, p in model.ridge.named_parameters()], 'weight_decay': 1e-2},
    {'params': [p for n, p in model.backbone.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': 1e-2},
    {'params': [p for n, p in model.backbone.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0},
]

optimizer = torch.optim.AdamW(opt_grouped_parameters, lr=max_lr)

if lr_scheduler_type == 'linear':
    lr_scheduler = torch.optim.lr_scheduler.LinearLR(
        optimizer,
        total_iters=int(np.floor(num_epochs*(num_train/num_devices/batch_size))),
        last_epoch=-1
    )
elif lr_scheduler_type == 'cycle':
    total_steps=int(np.floor(num_epochs*(num_train/num_devices/batch_size)))
    print("total_steps", total_steps)
    lr_scheduler = torch.optim.lr_scheduler.OneCycleLR(
        optimizer, 
        max_lr=max_lr,
        total_steps=total_steps,
        final_div_factor=1000,
        last_epoch=-1, pct_start=2/num_epochs
    )
    
def save_ckpt(tag):    
    ckpt_path = outdir+f'/{tag}.pth'
    print(f'saving {ckpt_path}',flush=True)
    unwrapped_model = accelerator.unwrap_model(model)
    try:
        torch.save({
            'epoch': epoch,
            'model_state_dict': unwrapped_model.state_dict(),
            'optimizer_state_dict': optimizer.state_dict(),
            'lr_scheduler': lr_scheduler.state_dict(),
            'train_losses': losses,
            'test_losses': test_losses,
            'lrs': lrs,
            }, ckpt_path)
    except:
        print("Couldn't save... moving on to prevent crashing.")
    del unwrapped_model
        
print("\nDone with model preparations!")
utils.count_params(model)


# In[49]:


seq_len = 4


# In[57]:


voxel_ridge = torch.randn(512,4096)
voxel_ridge = voxel_ridge.view(int(voxel_ridge.shape[0]/seq_len), seq_len, hidden_dim)


# In[58]:


voxel_ridge.shape


# In[55]:


pp = None
for train_i, (behav, past_behav, future_behav, old_behav) in enumerate(train_dl):
    with torch.cuda.amp.autocast(dtype=data_type):
        #optimizer.zero_grad()

        voxel = voxels[behav[:,0,5].cpu().long()]#.to(device)
        image = images[behav[:,0,0].cpu().long()].float()#.to(device).float()

        past_15_voxels = voxels[past_behav[:,:seq_len-1,5].cpu().long()]#.to(device) # batch_size, 15, 15279
        past_15_times = torch.Tensor([i for i in range(seq_len)])#.to(device) # 15
    
        print(past_15_times)
        #for past in range(1):
        #    past_voxel = voxels[past_behav[:,past,5].cpu().long()].to(device)
        """
        #if blurry_recon:
        #    blurry_image_enc = autoenc.encode(2*utils.resize(image,128)-1).latent_dist.mode() * 0.18215
        blurry_image_enc = autoenc.encode(2*utils.resize(add_saturation(image),128)-1).latents * 0.18215

        if depth_recon:
            # depth_images = utils.resize(midas_depth.model(image).unsqueeze(1).repeat(1,3,1,1), 128)
            depth_images = utils.resize(midas_depth.model(image).unsqueeze(1), 32)
            depth_images = (depth_images / depth_images.view(depth_images.shape[0], -1).max(dim=1)[0].view(-1, 1, 1, 1).expand_as(depth_images)).half()
            depth_image_enc = depth_images # autoenc.encode(2*depth_images-1).latents * 0.18215

        if use_image_aug: 
            image = img_augment(image)

        clip_target = clip_model.embed_image(image)
        assert not torch.any(torch.isnan(clip_target))

        if epoch < int(mixup_pct * num_epochs):
            voxel, perm, betas, select = utils.mixco(voxel)
            past_voxel, _, _, _ = utils.mixco(voxel, perm=perm, betas=betas, select=select)
        """
        for p in range(seq_len-1):
            print(past_behav.shape) #128, 15, 17
            print(past_behav[:,p,-1])
            print(past_15_voxels.shape) # 128, 1, 15724
            mask = past_behav[:,p,-1] == torch.ones_like(past_behav[:,p,-1])
            print(mask) # 128
            past_15_voxels[mask, p, :] = torch.zeros_like(past_15_voxels[0, p, :])
            print(past_15_voxels)
            pp = past_15_voxels
            
        break


# In[54]:


pp[20, 0, :]


# # Weights and Biases

# In[66]:


if local_rank==0 and wandb_log: # only use main process for wandb logging
    import wandb
    wandb_project = 'mindeyev2'
    wandb_run = model_name
    wandb_notes = ''
    
    print(f"wandb {wandb_project} run {wandb_run}")
    wandb.login(host='https://stability.wandb.io')#, relogin=True)
    wandb_config = {
      "model_name": model_name,
      "global_batch_size": global_batch_size,
      "batch_size": batch_size,
      "num_epochs": num_epochs,
      "clip_scale": clip_scale,
      "blur_scale": blur_scale,
      "use_image_aug": use_image_aug,
      "max_lr": max_lr,
      "mixup_pct": mixup_pct,
      "num_train": num_train,
      "num_test": num_test,
      "ckpt_interval": ckpt_interval,
      "ckpt_saving": ckpt_saving,
      "seed": seed,
      "distributed": distributed,
      "num_devices": num_devices,
      "world_size": world_size,
      "train_url": train_url,
      "test_url": test_url,
    }
    print("wandb_config:\n",wandb_config)
    if True: # wandb_auto_resume
        print("wandb_id:",model_name)
        wandb.init(
            id = model_name,
            project=wandb_project,
            name=wandb_run,
            config=wandb_config,
            notes=wandb_notes,
            resume="allow",
        )
    else:
        wandb.init(
            project=wandb_project,
            name=wandb_run,
            config=wandb_config,
            notes=wandb_notes,
        )
else:
    wandb_log = False


# # Main

# In[67]:


epoch = 0
losses, test_losses, lrs = [], [], []
best_test_loss = 1e9
soft_loss_temps = utils.cosine_anneal(0.004, 0.0075, num_epochs - int(mixup_pct * num_epochs))

# Optionally resume from checkpoint #
if resume_from_ckpt:
    print("\n---resuming from last.pth ckpt---\n")
    try:
        checkpoint = torch.load(outdir+'/last.pth', map_location='cpu')
    except:
        print('last.pth failed... trying last_backup.pth')
        checkpoint = torch.load(outdir+'/last_backup.pth', map_location='cpu')
    epoch = checkpoint['epoch']
    print("Epoch",epoch)
    optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
    lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
    model.load_state_dict(checkpoint['model_state_dict'])
    del checkpoint
elif wandb_log:
    if wandb.run.resumed:
        print("\n---resuming from last.pth ckpt---\n")
        try:
            checkpoint = torch.load(outdir+'/last.pth', map_location='cpu')
        except:
            print('last.pth failed... trying last_backup.pth')
            checkpoint = torch.load(outdir+'/last_backup.pth', map_location='cpu')
        epoch = checkpoint['epoch']
        print("Epoch",epoch)
        optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
        lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
        model.load_state_dict(checkpoint['model_state_dict'])
        del checkpoint
torch.cuda.empty_cache()


# In[68]:


model, optimizer, train_dl, lr_scheduler = accelerator.prepare(
model, optimizer, train_dl, lr_scheduler
)
# leaving out test_dl since we will only have local_rank 0 device do evals


# In[ ]:


def add_saturation(image, alpha=2):
    gray_image = 0.2989 * image[:, 0, :, :] + 0.5870 * image[:, 1, :, :] + 0.1140 * image[:, 2, :, :]
    gray_image = gray_image.unsqueeze(1).expand_as(image)
    saturated_image = alpha * image + (1 - alpha) * gray_image
    return torch.clamp(saturated_image, 0, 1)


# In[65]:


print(f"{model_name} starting with epoch {epoch} / {num_epochs}")
progress_bar = tqdm(range(epoch,num_epochs), ncols=1200, disable=(local_rank!=0))
test_image, test_voxel = None, None
mse = nn.MSELoss()
l1 = nn.L1Loss()

for epoch in progress_bar:
    model.train()
    
    fwd_percent_correct = 0.
    bwd_percent_correct = 0.
    test_fwd_percent_correct = 0.
    test_bwd_percent_correct = 0.

    loss_clip_total = 0.
    loss_blurry_total = 0.
    loss_depth_total = 0.
    test_loss_clip_total = 0.
    test_loss_blurry_total = 0.
    test_loss_depth_total = 0.

    blurry_pixcorr = 0.
    test_blurry_pixcorr = 0. # needs >.456 to beat low-level subj01 results in mindeye v1
    
    for train_i, (behav, past_behav, future_behav, old_behav) in enumerate(train_dl):
        with torch.cuda.amp.autocast(dtype=data_type):
            optimizer.zero_grad()
    
            #voxel = voxels[behav[:,0,5].cpu().long()].to(device)
            #image = images[behav[:,0,0].cpu().long()].to(device).float()
            
            #past_15_voxels = voxels[past_behav[:,:,5].cpu().long()].to(device) # batch_size, 15, 15279
            #past_15_times = torch.Tensor([i for i in range(seq_len - 1)]).to(device) # 15
            
            voxel = voxels[behav[:,0,5].cpu().long()].to(device)
            image = images[behav[:,0,0].cpu().long()].to(device).float()

            past_15_voxels = voxels[past_behav[:,:seq_len-1,5].cpu().long()].to(device) # batch_size, 15, 15279
            past_15_times = torch.Tensor([i for i in range(seq_len-1)]).to(device) # 15
            #for past in range(1):
            #    past_voxel = voxels[past_behav[:,past,5].cpu().long()].to(device)
    
            if blurry_recon:
                # blurry_image_enc = autoenc.encode(2*utils.resize(image,128)-1).latent_dist.mode() * 0.18215
                blurry_image_enc = autoenc.encode(2*utils.resize(add_saturation(image),128)-1).latents * 0.18215

            if depth_recon:
                # depth_images = utils.resize(midas_depth.model(image).unsqueeze(1).repeat(1,3,1,1), 128)
                depth_images = utils.resize(midas_depth.model(image).unsqueeze(1), 32)
                depth_images = (depth_images / depth_images.view(depth_images.shape[0], -1).max(dim=1)[0].view(-1, 1, 1, 1).expand_as(depth_images)).half()
                depth_image_enc = depth_images # autoenc.encode(2*depth_images-1).latents * 0.18215
            
            if use_image_aug: 
                image = img_augment(image)
    
            clip_target = clip_model.embed_image(image)
            assert not torch.any(torch.isnan(clip_target))
    
            if epoch < int(mixup_pct * num_epochs):
                voxel, perm, betas, select = utils.mixco(voxel)
                past_voxel, _, _, _ = utils.mixco(voxel, perm=perm, betas=betas, select=select)
                
            for p in range(seq_len-1):
                #print(past_behav.shape) #128, 15, 17
                #print(past_behav[:,p,-1])
                #print(past_15_voxels.shape) # 128, 1, 15724
                mask = past_behav[:,p,-1] == torch.ones_like(past_behav[:,p,-1])
                #print(mask) # 128
                past_15_voxels[mask, p, :] = torch.zeros_like(past_15_voxels[0, p, :])
                #print(past_15_voxels)
                
            past_15_voxels = past_15_voxels.reshape(-1, past_15_voxels.shape[-1])
            past_15_times = past_15_times.repeat(voxel.shape[0], 1)
            past_15_times = past_15_times.reshape(-1)
            time_embeddings = model.time_embedding(past_15_times)
            past_info_full = torch.cat((past_15_voxels, time_embeddings), dim=-1)
            
            positional_current_voxel = torch.zeros((voxel.shape[0], time_embeddings.shape[-1])).to(voxel.device)
            voxel = torch.cat((voxel, positional_current_voxel), dim=-1)
            voxel_ridge = model.ridge(torch.cat((voxel, past_info_full), dim=-2))
            voxel_ridge = voxel_ridge.view(int(voxel_ridge.shape[0]/seq_len), seq_len, hidden_dim)
            #unsqueeze(1) # bz * 2, 1, 4096
            
            # past_voxel_ridge = model.ridge(past_voxel)
            # voxel_ridge = torch.cat((voxel_ridge.unsqueeze(1), past_voxel_ridge.unsqueeze(1)), axis=1)
            print(voxel_ridge.shape)
    
            clip_voxels, blurry_image_enc_, depth_image_enc_ = model.backbone(voxel_ridge)
            
            clip_voxels_norm = nn.functional.normalize(clip_voxels.flatten(1), dim=-1)
            clip_target_norm = nn.functional.normalize(clip_target.flatten(1), dim=-1)
    
            if epoch < int(mixup_pct * num_epochs):                
                loss_clip = utils.mixco_nce(
                    clip_voxels_norm,
                    clip_target_norm,
                    temp=.006, 
                    perm=perm, betas=betas, select=select)
            else:
                epoch_temp = soft_loss_temps[epoch-int(mixup_pct*num_epochs)]
                loss_clip = utils.soft_clip_loss(
                    clip_voxels_norm,
                    clip_target_norm,
                    temp=epoch_temp)

            loss_clip_total += loss_clip.item()
            loss_clip *= clip_scale
            loss = loss_clip
    
            if blurry_recon:
                downsampled_image = nn.functional.interpolate(image, size=(8, 8), mode='bilinear', align_corners=False)
                re_upsampled_image = add_saturation(nn.functional.interpolate(downsampled_image, size=(128, 128), mode='nearest'))
                re_upsampled_enc = autoenc.encode(2*re_upsampled_image-1).latents * 0.18215
                
                loss_blurry = (l1(blurry_image_enc_, blurry_image_enc) + l1(blurry_image_enc_, re_upsampled_enc))
                loss_blurry += l1(torch.var(blurry_image_enc), torch.var(blurry_image_enc_))
                loss_blurry_total += loss_blurry.item()
                loss_blurry *= blur_scale
                loss += loss_blurry

            if depth_recon:
                loss_depth = l1(depth_image_enc_, depth_image_enc)
                # loss_depth += l1(torch.var(depth_image_enc_), torch.var(depth_image_enc))
                loss_depth_total += loss_depth.item()
                loss_depth *= depth_scale
                loss += loss_depth
    
            # forward and backward top 1 accuracy        
            labels = torch.arange(len(clip_target_norm)).to(clip_voxels_norm.device) 
            fwd_percent_correct += utils.topk(torch.abs(utils.batchwise_cosine_similarity(clip_voxels_norm, clip_target_norm)), labels, k=1).item()
            bwd_percent_correct += utils.topk(torch.abs(utils.batchwise_cosine_similarity(clip_target_norm, clip_voxels_norm)), labels, k=1).item()
    
            if blurry_recon:
                with torch.no_grad():
                    # only doing pixcorr eval on a subset of the samples per batch because its costly & slow to compute autoenc.decode()
                    random_samps = np.random.choice(np.arange(len(voxel)), size=batch_size//5, replace=False)
                    # random_samps = np.arange(batch_size//5)
                    blurry_recon_images = (autoenc.decode(blurry_image_enc_[random_samps]/0.18215).sample/ 2 + 0.5).clamp(0,1)
                    # pixcorr_origsize_nanmean is computationally less intense than utils.pixcorr and uses nanmean instead of mean
                    pixcorr = utils.pixcorr_origsize_nanmean(image[random_samps], blurry_recon_images)
                    # pixcorr = utils.pixcorr(image[random_samps], blurry_recon_images)
                    # loss += (1 - pixcorr)
                    blurry_pixcorr += pixcorr.item()
                    # utils.check_loss(pixcorr)

            utils.check_loss(loss)
            accelerator.backward(loss)
            optimizer.step()
    
            losses.append(loss.item())
            lrs.append(optimizer.param_groups[0]['lr'])
    
            if lr_scheduler_type is not None:
                lr_scheduler.step()

    model.eval()
    if local_rank==0:
        with torch.no_grad(), torch.cuda.amp.autocast(dtype=data_type): 
            for test_i, (behav, past_behav, future_behav, old_behav) in enumerate(test_dl):  
                # all test samples should be loaded per batch such that test_i should never exceed 0
                assert len(behav) == num_test
                
                ## Average same-image repeats ##
                if test_image is None:
                    voxel = voxels[behav[:,0,5].cpu().long()]
                    image = behav[:,0,0].cpu().long()
                    
                    unique_image, sort_indices = torch.unique(image, return_inverse=True)
                    for im in unique_image:
                        locs = torch.where(im == image)[0]
                        if test_image is None:
                            test_image = images[im][None]
                            test_voxel = torch.mean(voxel[locs],axis=0)[None]
                        else:
                            test_image = torch.vstack((test_image, images[im][None]))
                            test_voxel = torch.vstack((test_voxel, torch.mean(voxel[locs],axis=0)[None]))
    
                # random sample of 300
                random_indices = torch.arange(len(test_voxel))[:300]
                voxel = test_voxel[random_indices].to(device)
                image = test_image[random_indices].to(device)
                assert len(image) == 300
                
                current_past_behav = past_behav[random_indices]

                past_15_voxels = voxels[current_past_behav[:,:seq_len-1,5].cpu().long()].to(device) # batch_size, 15, 15279
                past_15_times = torch.Tensor([i for i in range(seq_len-1)]).to(device) # 15

                if blurry_recon:
                    # blurry_image_enc = autoenc.encode(2*utils.resize(image,128)-1).latent_dist.mode() * 0.18215
                    blurry_image_enc = autoenc.encode(2*utils.resize(add_saturation(image),128)-1).latents * 0.18215

                if depth_recon:
                    # depth_images = utils.resize(midas_depth.model(image).unsqueeze(1).repeat(1,3,1,1), 128)
                    depth_images = utils.resize(midas_depth.model(image).unsqueeze(1), 32)
                    depth_images = (depth_images / depth_images.view(depth_images.shape[0], -1).max(dim=1)[0].view(-1, 1, 1, 1).expand_as(depth_images)).half()
                    depth_image_enc = depth_images # autoenc.encode(2*depth_images-1).latents * 0.18215
            
                clip_target = clip_model.embed_image(image.float())
                

                past_15_voxels = past_15_voxels.reshape(-1, past_15_voxels.shape[-1])
                past_15_times = past_15_times.repeat(voxel.shape[0], 1)
                past_15_times = past_15_times.reshape(-1)
                time_embeddings = model.time_embedding(past_15_times)
                past_info_full = torch.cat((past_15_voxels, time_embeddings), dim=-1)

                positional_current_voxel = torch.zeros((voxel.shape[0], time_embeddings.shape[-1])).to(voxel.device)
                voxel = torch.cat((voxel, positional_current_voxel), dim=-1)
                voxel_ridge = model.ridge(torch.cat((voxel, past_info_full), dim=-2)).unsqueeze(1)
                
                #voxel_ridge = model.ridge(voxel).unsqueeze(1)

                # voxel_ridge = torch.cat((voxel_ridge.unsqueeze(1),voxel_ridge.unsqueeze(1)),axis=1)
                
                clip_voxels, blurry_image_enc_, depth_image_enc_ = model.backbone(voxel_ridge)
                
                clip_voxels_norm = nn.functional.normalize(clip_voxels.flatten(1), dim=-1)
                clip_target_norm = nn.functional.normalize(clip_target.flatten(1), dim=-1)
        
                loss_clip = utils.soft_clip_loss(
                    clip_voxels_norm,
                    clip_target_norm,
                    temp=.006)
                test_loss_clip_total += loss_clip.item()
                loss_clip = loss_clip * clip_scale
                loss = loss_clip

                if blurry_recon:
                    downsampled_image = nn.functional.interpolate(image, size=(8, 8), mode='bilinear', align_corners=False)
                    re_upsampled_image = add_saturation(nn.functional.interpolate(downsampled_image, size=(128, 128), mode='nearest'))
                    re_upsampled_enc = autoenc.encode(2*re_upsampled_image-1).latents * 0.18215
                    
                    loss_blurry = (l1(blurry_image_enc_, blurry_image_enc) + l1(blurry_image_enc_, re_upsampled_enc))
                    loss_blurry += l1(torch.var(blurry_image_enc), torch.var(blurry_image_enc_))
                    test_loss_blurry_total += loss_blurry.item()
                    loss_blurry *= blur_scale
                    loss += loss_blurry
    
                    # halving the batch size because the decoder is computationally heavy
                    blurry_recon_images = (autoenc.decode(blurry_image_enc_[:len(voxel)//2]/0.18215).sample / 2 + 0.5).clamp(0,1)
                    blurry_recon_images = torch.vstack((blurry_recon_images, (autoenc.decode(blurry_image_enc_[len(voxel)//2:]/0.18215).sample / 2 + 0.5).clamp(0,1)))
                    pixcorr = utils.pixcorr(image, blurry_recon_images)
                    loss += (1 - pixcorr)
                    test_blurry_pixcorr += pixcorr.item()

                if depth_recon:
                    loss_depth = l1(depth_image_enc_, depth_image_enc)
                    # loss_depth += l1(torch.var(depth_image_enc_), torch.var(depth_image_enc))
                    test_loss_depth_total += loss_depth.item()
                    loss_depth *= depth_scale
                    loss += loss_depth
        
                # forward and backward top 1 accuracy        
                labels = torch.arange(len(clip_target_norm)).to(clip_voxels_norm.device) 
                test_fwd_percent_correct += utils.topk(utils.batchwise_cosine_similarity(clip_voxels_norm, clip_target_norm), labels, k=1).item()
                test_bwd_percent_correct += utils.topk(utils.batchwise_cosine_similarity(clip_target_norm, clip_voxels_norm), labels, k=1).item()

                utils.check_loss(loss)                
                test_losses.append(loss.item())

            # if utils.is_interactive(): clear_output(wait=True)
            print("---")
            
            assert (test_i+1) == 1
            logs = {"train/loss": np.mean(losses[-(train_i+1):]),
                "test/loss": np.mean(test_losses[-(test_i+1):]),
                "train/lr": lrs[-1],
                "train/num_steps": len(losses),
                "test/num_steps": len(test_losses),
                "train/fwd_pct_correct": fwd_percent_correct / (train_i + 1),
                "train/bwd_pct_correct": bwd_percent_correct / (train_i + 1),
                "test/test_fwd_pct_correct": test_fwd_percent_correct / (test_i + 1),
                "test/test_bwd_pct_correct": test_bwd_percent_correct / (test_i + 1),
                "train/loss_clip_total": loss_clip_total / (train_i + 1),
                "train/loss_blurry_total": loss_blurry_total / (train_i + 1),
                "test/loss_clip_total": test_loss_clip_total / (test_i + 1),
                "test/loss_blurry_total": test_loss_blurry_total / (test_i + 1),
                "train/blurry_pixcorr": blurry_pixcorr / (train_i + 1),
                "test/blurry_pixcorr": test_blurry_pixcorr / (test_i + 1),
                "train/loss_depth_total": loss_depth_total / (train_i + 1),
                "test/loss_depth_total": test_loss_depth_total / (test_i + 1),
                }
    
            if blurry_recon:    
                # transform blurry recon latents to images and plot it
                fig, axes = plt.subplots(1, 8, figsize=(10, 4))
                jj=-1
                for j in [0,1,2,3]:
                    jj+=1
                    axes[jj].imshow(utils.torch_to_Image((autoenc.decode(blurry_image_enc[[j]]/0.18215).sample / 2 + 0.5).clamp(0,1)))
                    axes[jj].axis('off')
                    jj+=1
                    axes[jj].imshow(utils.torch_to_Image((autoenc.decode(blurry_image_enc_[[j]]/0.18215).sample / 2 + 0.5).clamp(0,1)))
                    axes[jj].axis('off')
                
                if wandb_log:
                    logs[f"test/recons"] = wandb.Image(fig, caption=f"epoch{epoch:03d}")
                    plt.close()
                else:
                    plt.show()

            if depth_recon:
                # transform blurry recon latents to images and plot it
                fig, axes = plt.subplots(1, 8, figsize=(10, 4))
                # axes[0].imshow(utils.torch_to_Image((autoenc.decode(depth_image_enc[[0]]/0.18215).sample / 2 + 0.5).clamp(0,1)))
                # axes[1].imshow(utils.torch_to_Image((autoenc.decode(depth_image_enc_[[0]]/0.18215).sample / 2 + 0.5).clamp(0,1)))
                jj=-1
                for j in [0,1,2,3]:
                    jj+=1
                    axes[jj].imshow(utils.torch_to_Image(utils.resize(depth_image_enc[[j]].view(1,1,32,32).clamp(0,1), 224)))
                    axes[jj].axis('off')
                    jj+=1
                    axes[jj].imshow(utils.torch_to_Image(utils.resize(depth_image_enc_[[j]].view(1,1,32,32).clamp(0,1), 224)))
                    axes[jj].axis('off')
                if wandb_log:
                    logs[f"test/depth_recons"] = wandb.Image(fig, caption=f"epoch{epoch:03d}")
                    plt.close()
                else:
                    plt.show()
            
            progress_bar.set_postfix(**logs)
    
            # Save model checkpoint and reconstruct
            if epoch % ckpt_interval == 0:
                if not utils.is_interactive():
                    save_ckpt(f'last')
                    
            if wandb_log: wandb.log(logs)

    # wait for other GPUs to catch up if needed
    accelerator.wait_for_everyone()
    torch.cuda.empty_cache()
    gc.collect()

print("\n===Finished!===\n")
if ckpt_saving:
    save_ckpt(f'last')
if not utils.is_interactive():
    sys.exit(0)


# In[ ]:


plt.plot(losses)
plt.show()
plt.plot(test_losses)
plt.show()


# # Retrieve nearest neighbor in the training set using test set data

# In[ ]:


annots = np.load("/fsx/proj-fmri/shared/mindeyev2_dataset/COCO_73k_annots_curated.npy")


# In[ ]:


ii=2
all_indices = np.unique(train_73k_images) #np.hstack((test_vox_indices[ii],train_vox_indices))
with torch.no_grad(), torch.cuda.amp.autocast():
    for batch in tqdm(range(0,len(all_indices),512)):
        if batch==0:
            clip_target = clip_model.embed_image(images[all_indices[batch:batch+512]]).cpu()
        else:
            target = clip_model.embed_image(images[all_indices[batch:batch+512]]).cpu()
            clip_target = torch.vstack((clip_target,target))
    clip_target_norm = nn.functional.normalize(clip_target.flatten(1), dim=-1)

    voxel = test_voxel[[ii]].to(device)
    image = test_image[[ii]].to(device)

    print("Original Image (test set)")
    display(utils.torch_to_Image(image))
    
    clip_target = clip_model.embed_image(image).cpu()
    # clip_target_norm = torch.vstack((clip_target_norm, nn.functional.normalize(clip_target.flatten(1), dim=-1)))
    
    voxel_ridge = model.ridge(voxel).unsqueeze(1)
    clip_voxels, _, _ = model.backbone(voxel_ridge)    
    clip_voxels_norm = nn.functional.normalize(clip_voxels.flatten(1), dim=-1)
    clip_voxels_norm = nn.functional.normalize(clip_target.flatten(1), dim=-1)

    print("clip_voxels_norm", clip_voxels_norm.shape)
    print("clip_target_norm", clip_target_norm.shape)
    
    sortt = torch.argsort(utils.batchwise_cosine_similarity(clip_voxels_norm.cpu(), 
                                                            clip_target_norm).flatten()).flip(0)
    picks = all_indices[sortt[:5]]

    print("\nNearest neighbors in training set")
    for ip,p in enumerate(picks):
        display(utils.torch_to_Image(images[[p]]))
        # print(utils.select_annotations([annots[int(p)]]))
        if ip==0: predicted_caption = utils.select_annotations([annots[int(p)]])[0]

print("\n=====\npredicted_caption:\n", predicted_caption)


# # Feed into Stable Diffusion XL for reconstructions

# In[ ]:


from diffusers import StableDiffusionXLPipeline
pipe = StableDiffusionXLPipeline.from_pretrained(
    "/fsx/proj-fmri/shared/cache/models--stabilityai--stable-diffusion-xl-base-1.0/snapshots/f898a3e026e802f68796b95e9702464bac78d76f", torch_dtype=torch.float16, variant="fp16", use_safetensors=True
)
pipe.to("cuda")
pass


# In[ ]:


prompt = predicted_caption
recon = pipe(prompt=prompt).images[0]


# In[ ]:


print("Seen image")
display(utils.torch_to_Image(image))

print("Reconstruction")
utils.torch_to_Image(utils.resize(transforms.ToTensor()(recon),224))