src/Train.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.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 time
import random
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

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

# custom functions #
import utils


# In[ ]:


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

### Single-GPU config ###
## Feel free to uncomment the below 4 lines and comment out all the multi-gpu config code to simplify things for single-gpu
# from accelerate import Accelerator
# num_devices = torch.cuda.device_count()
# if num_devices==0: num_devices = 1
# accelerator = Accelerator(split_batches=False)
# global_batch_size = 128
                          
### Multi-GPU config ###
from accelerate import Accelerator, DeepSpeedPlugin
num_devices = torch.cuda.device_count()
if num_devices==0: num_devices = 1
if num_devices <= 1 and utils.is_interactive():
    # 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"] = "128" # set this to your batch size!
    global_batch_size = os.environ["GLOBAL_BATCH_SIZE"]

# 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'] = int(os.environ["GLOBAL_BATCH_SIZE"]) // num_devices
    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.json")
accelerator = Accelerator(split_batches=False, deepspeed_plugin=deepspeed_plugin)


# In[ ]:


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


# # Configurations

# In[3]:


# if running this interactively, can specify jupyter_args here for argparser to use
if utils.is_interactive():
    # Example use
    jupyter_args = f"--data_path=/fsx/proj-fmri/shared/mindeyev2_dataset \
                    --model_name=test \
                    --subj=1 --batch_size={global_batch_size} --n_samples_save=0 \
                    --max_lr=3e-4 --mixup_pct=.66 --num_epochs=12 --ckpt_interval=999 --no-use_image_aug"

    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[4]:


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 prior",
)
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(
    "--use_image_aug",action=argparse.BooleanOptionalAction,default=True,
    help="whether to use image augmentation",
)
parser.add_argument(
    "--num_epochs",type=int,default=240,
    help="number of epochs of training",
)
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,
)
parser.add_argument(
    "--n_samples_save",type=int,default=0,choices=[0,1],
    help="Number of reconstructions for monitoring progress, 0 will speed up training",
)

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)

print("global batch_size", batch_size)
batch_size = int(batch_size / num_devices)
print("batch_size", batch_size)


# In[5]:


outdir = os.path.abspath(f'../train_logs/{model_name}')
if not os.path.exists(outdir):
    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[6]:


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"
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=False, 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=False, pin_memory=True)


# ### check dataloaders are working

# In[7]:


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

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


# ## Load voxel betas, K-means clustering model, and images

# In[8]:


# load betas
f = h5py.File(f'{data_path}/betas_all_subj0{subj}.hdf5', 'r')
voxels = f['betas'][:]
print(f"subj0{subj} betas loaded into memory")
voxels = torch.Tensor(voxels).to("cpu").half()
if subj==1:
    voxels = torch.hstack((voxels, torch.zeros((len(voxels), 5))))
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").half()
print("images", images.shape)


# In[9]:


from models import Clipper
eva02_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
hidden_dim = 4096


# In[10]:


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


# In[11]:


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.linear = torch.nn.Linear(input_size, out_features)
    def forward(self, x):
        return self.linear(x)
        
model.ridge = RidgeRegression(voxels.shape[1], out_features=hidden_dim)
utils.count_params(model.ridge)
utils.count_params(model)

b = torch.randn((2,voxels.shape[1]))
print(b.shape, model.ridge(b).shape)


# In[12]:


from functools import partial
class BrainNetwork(nn.Module):
    def __init__(self, out_dim=768, in_dim=15724, clip_size=768, h=4096, n_blocks=4, norm_type='ln', act_first=False, use_projector=True, drop1=.5, drop2=.15):
        super().__init__()
        norm_func = partial(nn.BatchNorm1d, num_features=h) if norm_type == 'bn' else partial(nn.LayerNorm, normalized_shape=h)
        act_fn = partial(nn.ReLU, inplace=True) if norm_type == 'bn' else nn.GELU
        act_and_norm = (act_fn, norm_func) if act_first else (norm_func, act_fn)
        self.mlp = nn.ModuleList([
            nn.Sequential(
                nn.Linear(h, h),
                *[item() for item in act_and_norm],
                nn.Dropout(drop2)
            ) for _ in range(n_blocks)
        ])
        self.lin1 = nn.Linear(h, out_dim, bias=True)
        self.n_blocks = n_blocks
        self.clip_size = clip_size
        self.use_projector = use_projector
        if use_projector:
            self.projector = 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)
            )
        
    def forward(self, x):
        residual = x
        for res_block in range(self.n_blocks):
            x = self.mlp[res_block](x)
            x += residual
            residual = x
        x = x.reshape(len(x), -1)
        x = self.lin1(x)
        if self.use_projector:
            return self.projector(x.reshape(len(x), -1, self.clip_size))
        return x

model.backbone = BrainNetwork(in_dim=hidden_dim, clip_size=clip_emb_dim, out_dim=clip_seq_dim*clip_emb_dim, use_projector=True)
utils.count_params(model.backbone)
utils.count_params(model)

b = torch.randn((2,hidden_dim))
print(b.shape, model.backbone(b).shape)


# In[13]:


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(num_epochs*(num_train*num_devices//batch_size)),
        last_epoch=-1
    )
elif lr_scheduler_type == 'cycle':
    total_steps=int(num_epochs*(num_train*num_devices//batch_size))
    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!")


# # Weights and Biases

# In[14]:


# params for wandb
if local_rank==0 and wandb_log: # only use main process for wandb logging
    import wandb
    
    wandb_project = 'stability'
    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,
      "clip_variant": clip_variant,
      "batch_size": batch_size,
      "num_epochs": num_epochs,
      "use_image_aug": use_image_aug,
      "max_lr": max_lr,
      "lr_scheduler_type": lr_scheduler_type,
      "mixup_pct": mixup_pct,
      "num_train": num_train,
      "num_test": num_test,
      "seed": seed,
      "distributed": distributed,
      "num_devices": num_devices,
      "world_size": world_size,
    }
    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[15]:


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'])
    diffusion_prior.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'])
        diffusion_prior.load_state_dict(checkpoint['model_state_dict'])
        del checkpoint
torch.cuda.empty_cache()


# In[16]:


model, optimizer, train_dl, test_dl, lr_scheduler = accelerator.prepare(
model, optimizer, train_dl, test_dl, lr_scheduler
)


# In[17]:


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()
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.

    for train_i, (behav, past_behav, future_behav, old_behav) in enumerate(train_dl):
        with torch.cuda.amp.autocast():
            optimizer.zero_grad()

            voxel = voxels[behav[:,0,5].cpu().long()].to(device)
            image = images[behav[:,0,0].cpu().long()].to(device)
            
            if use_image_aug: image = img_augment(image)
            
            clip_target = eva02_model.embed_image(image.float())
            assert not torch.any(torch.isnan(clip_target))

            if epoch < int(mixup_pct * num_epochs):
                voxel, perm, betas, select = utils.mixco(voxel)

            voxel_ridge = model.ridge(voxel)
            
            clip_voxels = 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 = loss_clip
            
            utils.check_loss(loss)

            accelerator.backward(loss)
            optimizer.step()
    
            losses.append(loss.item())
            lrs.append(optimizer.param_groups[0]['lr'])
    
            # forward and backward top 1 accuracy        
            labels = torch.arange(len(clip_target_norm)).to(clip_voxels_norm.device) 
            fwd_percent_correct += utils.topk(utils.batchwise_cosine_similarity(clip_voxels_norm, clip_target_norm), labels, k=1)
            bwd_percent_correct += utils.topk(utils.batchwise_cosine_similarity(clip_target_norm, clip_voxels_norm), labels, k=1)

            if lr_scheduler_type is not None:
                lr_scheduler.step()

    model.eval()
    for test_i, (behav, past_behav, future_behav, old_behav) in enumerate(test_dl):
        with torch.no_grad():
            with torch.cuda.amp.autocast():    
                # all test samples should be loaded per batch such that test_i should never exceed 0
                if len(behav) != num_test: print("!",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.randperm(len(test_voxel))[:300]
                voxel = test_voxel[random_indices].to(device)
                image = test_image[random_indices].to(device)
                assert len(image) == 300
        
                clip_target = eva02_model.embed_image(image.float())
                
                voxel_ridge = model.ridge(voxel)
                
                clip_voxels = 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)
                    
                loss = loss_clip
                
                utils.check_loss(loss)
        
                test_losses.append(loss.item())
        
                # 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)
                test_bwd_percent_correct += utils.topk(utils.batchwise_cosine_similarity(clip_target_norm, clip_voxels_norm), labels, k=1)

    if local_rank==0:      
        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.item() / (train_i + 1),
            "train/bwd_pct_correct": bwd_percent_correct.item() / (train_i + 1),
            "test/test_fwd_pct_correct": test_fwd_percent_correct.item() / (test_i + 1),
            "test/test_bwd_pct_correct": test_bwd_percent_correct.item() / (test_i + 1),
            }
        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)