src/imagenet/transfer.py

import argparse
import os
import jax
import flax
import copy
import json
import optax
import torch
import wandb
import numpy as np
from tqdm import tqdm
import jax.numpy as jnp
from flax import linen as nn
from typing import Any, Dict, List
from flax.jax_utils import replicate, unreplicate
from flax.training import checkpoints, train_state
from flax.core.frozen_dict import freeze, unfreeze
from flax.traverse_util import flatten_dict, unflatten_dict
from flax.training.common_utils import get_metrics, onehot, shard
from transformers.models.vit.modeling_flax_vit import ViTConfig, FlaxViTForImageClassification
from model import OldLMCFlaxViTForImageClassification
from lmc_model import LMCFlaxViTForImageClassification, print_model, print_model_with_prefix
from datasets import build_dataset
import multiprocessing as mp
from pprint import pprint
def imagenet_data_loader(args):
    dataset_train, args.nb_classes = build_dataset(is_train=True, args=args)
    dataset_val, _ = build_dataset(is_train=False, args=args)
    sampler_train = torch.utils.data.RandomSampler(dataset_train)
    sampler_val = torch.utils.data.SequentialSampler(dataset_val)
    data_loader_train = torch.utils.data.DataLoader(
        dataset_train, sampler=sampler_train, batch_size=args.batch_size,
        num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=True
    )
    data_loader_val = torch.utils.data.DataLoader(
        dataset_val, sampler=sampler_val, batch_size=args.batch_size,
        num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=False
    )
    return data_loader_train, data_loader_val
def prepare_image_batch(images:torch.Tensor,labels:torch.Tensor) -> Dict[str, Any]:
    images, labels =  jnp.array(images),jnp.array(labels)
    return {'images': shard(images),'labels': shard(labels)}
def accuracy(logits, labels, topk=(1,)):
    maxk = max(topk)
    batch_size = labels.shape[0]
    topk_preds = jnp.argsort(logits, axis=-1)[:, -maxk:][:, ::-1]  # Top-k predictions
    res = []
    for k in topk:
        correct = (topk_preds[:, :k] == labels[:, None])
        correct = jnp.any(correct, axis=1)
        correct = jnp.sum(correct)
        res.append(100.0 * correct / batch_size)
    return res  # list of [acc@1, acc@5]
def pretrained_new2old(new_params,old_params,config):
    new_params = unfreeze(new_params)
    old_params = unfreeze(old_params)
    print_model(new_params)
    print_model(old_params)
    # 1. Copy top-level params (embeddings, layernorm, classifier)
    new_params["vit"]["embeddings"] = copy.deepcopy(old_params["vit"]["embeddings"])
    new_params["vit"]["layernorm"] = copy.deepcopy(old_params["vit"]["layernorm"])
    new_params["classifier"] = copy.deepcopy(old_params["classifier"])
    # 2. Copy encoder layers
    for i in range(config.num_hidden_layers):
        str_i = str(i)
        ref_layer = old_params["vit"]["encoder"]["layer"][str_i]  # use layer 0 from pretrained
        target_layer = new_params["vit"]["encoder"]["layer"][str_i]
        # Copy shared parts
        target_layer["layernorm_before"] = copy.deepcopy(ref_layer["layernorm_before"])
        target_layer["layernorm_after"] = copy.deepcopy(ref_layer["layernorm_after"])
        target_layer["moe"]['shared_experts']['intermediate'] = copy.deepcopy(ref_layer["mlp"]['intermediate']['dense'])
        target_layer["moe"]['shared_experts']['output'] = copy.deepcopy(ref_layer["mlp"]['output']['dense'])
        #Copy Attention weights 
        target_layer["attention"] = copy.deepcopy(ref_layer["attention"])
    return freeze(new_params)
def main(args: argparse.Namespace):
    train_loader, val_loader = imagenet_data_loader(args)
    save_path = "/mnt/data/vinhbk/weights/imagenet/lr0.0005-rope-epochs300-batch256-shared1-routed0-topk0/"
    old_config = ViTConfig.from_pretrained('/mnt/data/vinhbk/weights/imagenet/lr0.0005-rope-epochs300-batch256/config.json')
    old_model = OldLMCFlaxViTForImageClassification(old_config,dtype=jnp.bfloat16) 
    new_config = copy.deepcopy(old_config)
    new_config.position_embeddings = old_config.position_embeddings
    new_config.rotary_value = old_config.rotary_value
    new_config.q_lora_rank = 8
    new_config.qk_rope_head_dim = 64
    new_config.kv_lora_rank = 8
    new_config.v_head_dim = 64
    new_config.qk_nope_head_dim = 64
    new_config.attention_bias = True
    new_config.routed_scaling_factor = 1.0
    new_config.lmc_layer_indices = []
    new_config.num_shared_experts = 1
    new_config.num_routed_experts = 0
    new_config.topk = 0
    # 1. Recreate LR schedule
    lr_schedule = optax.warmup_cosine_decay_schedule(
        init_value=1e-6,
        peak_value=5e-4,
        warmup_steps=5*5004,
        decay_steps=300*5004,
        end_value=1e-5,
    )
    # 2. Recreate optimizer tx
    tx = optax.adamw(
        learning_rate=lr_schedule,
        b1=0.9,
        b2=0.999,
        eps=1e-8,
        weight_decay=0.01,
    )
    old_state = train_state.TrainState.create(apply_fn=old_model.__call__, params=old_model.params, tx=tx)
    old_params = old_state.params
    new_model = LMCFlaxViTForImageClassification(new_config,input_shape=(1,new_config.image_size, new_config.image_size, new_config.num_channels),seed=args.seed,dtype=jnp.bfloat16) 
    new_model.params = pretrained_new2old(new_params=copy.deepcopy(new_model.params),old_params=copy.deepcopy(old_state.params),config=new_model.config)
    # print(new_model.params['classifier']['bias'])
    # 3. Create new TrainState with new_model
    new_state = train_state.TrainState.create(apply_fn=new_model.__call__,params=new_model.params,tx=tx,)
    print(old_state.params['classifier']['bias'])
    print(new_state.params['classifier']['bias'])
    # 4. Load old values
    os.makedirs(save_path,exist_ok=True)
    new_model.config.save_pretrained(save_path)
    # checkpoints.save_checkpoint(ckpt_dir=save_path,target=new_state,step=old_state.step,prefix="best_",keep=1,overwrite=True)
    def train_step(state, batch, rng):
        dropout_rng, new_dropout_rng = jax.random.split(rng)
        def loss_fn(params):
            outputs = state.apply_fn(params=params,pixel_values=batch["images"],train=True,dropout_rng=dropout_rng,)
            logits = outputs.logits if hasattr(outputs, "logits") else outputs[0]
            loss = optax.softmax_cross_entropy_with_integer_labels(logits, batch["labels"]).mean()
            return loss, logits
        (loss, logits), grads = jax.value_and_grad(loss_fn, has_aux=True)(state.params)
        grads = jax.lax.pmean(grads, axis_name="batch")
        state = state.apply_gradients(grads=grads)
        acc1, acc5 = accuracy(logits, batch["labels"], topk=(1, 5))
        metrics = {"loss": loss,"acc1": acc1,"acc5": acc5,"learning_rate": lr_schedule(state.step),}
        metrics = jax.lax.pmean(metrics, axis_name="batch")
        return state, metrics, new_dropout_rng
    def eval_step(state, batch):
        outputs = state.apply_fn(params=state.params,pixel_values=batch["images"],train=False,)
        logits = outputs.logits if hasattr(outputs, "logits") else outputs[0]
        loss = optax.softmax_cross_entropy_with_integer_labels(logits, batch["labels"]).mean()
        acc1, acc5 = accuracy(logits, batch["labels"], topk=(1, 5))
        metrics = {"loss": loss,"acc1": acc1,"acc5": acc5,}
        metrics = jax.lax.pmean(metrics, axis_name="batch")
        return metrics
    parallel_train_step = jax.pmap(train_step, "batch")
    parallel_eval_step = jax.pmap(eval_step, "batch")
    state = replicate(new_state)
    rng = jax.random.PRNGKey(0)
    train_metrics_stack = []
    train_loss = 0.0
    best_val_acc1 = 0.0
    ###JUST FOR TESTING####
    eval_results = []
    pbar = tqdm(enumerate(val_loader), desc="Evaluating", leave=False)
    for batch_idx, (images, labels) in pbar:
        batch = prepare_image_batch(images, labels)
        # Run eval step
        eval_metric = parallel_eval_step(state, batch)
        eval_results.append(eval_metric)            # Run train step
    # Compute mean metrics across all eval batches
    eval_metrics = get_metrics(eval_results)
    eval_metrics = unreplicate(eval_metrics)
    eval_metrics = jax.tree_util.tree_map(lambda x: x.mean(), eval_metrics)
    val_loss, val_acc1, val_acc5 = float(eval_metrics["loss"]), float(eval_metrics["acc1"]), float(eval_metrics["acc5"])
    print("-" * 100)
    print(f"valid loss {val_loss:5.4f} | valid acc@1 {val_acc1:6.2f}% | valid acc@5 {val_acc5:6.2f}%")
    print("-" * 100)
if __name__ =="__main__":
    parser = argparse.ArgumentParser(description="Fine-tune ViT with MoE on Imagenet")
    parser.add_argument("--batch-size", type=int, default=256)
    parser.add_argument("--data-path", type=str, required=True)
    parser.add_argument('--data-set', default='IMNET', choices=['CIFAR', 'IMNET', 'INAT', 'INAT19'])
    parser.add_argument("--input-size", type=int, default=224)
    parser.add_argument('--num_workers', type=int, default=8)
    parser.add_argument('--pin-mem', action='store_true')
    parser.add_argument('--seed', type=int, default=0)
    parser.add_argument('--color-jitter', type=float, default=0.4)
    parser.add_argument('--aa', type=str, default='rand-m9-mstd0.5-inc1')
    parser.add_argument('--train-interpolation', type=str, default='bicubic')
    parser.add_argument('--reprob', type=float, default=0.25)
    parser.add_argument('--remode', type=str, default='pixel')
    parser.add_argument('--recount', type=int, default=1)
    parser.add_argument("--dtype", choices=["float32", "float16", "bfloat16"], default="bfloat16", help="model datatype")
    args = parser.parse_args()
    main(args)