diff --git a/exps/basic/xmain.py b/exps/basic/xmain.py
new file mode 100644
index 0000000..87a0cfa
--- /dev/null
+++ b/exps/basic/xmain.py
@@ -0,0 +1,294 @@
+#####################################################
+# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2019.01 #
+#####################################################
+import sys, time, torch, random, argparse
+from PIL import ImageFile
+
+ImageFile.LOAD_TRUNCATED_IMAGES = True
+from copy import deepcopy
+from pathlib import Path
+
+from xautodl.datasets import get_datasets
+from xautodl.config_utils import load_config, obtain_basic_args as obtain_args
+from xautodl.procedures import (
+    prepare_seed,
+    prepare_logger,
+    save_checkpoint,
+    copy_checkpoint,
+)
+from xautodl.procedures import get_optim_scheduler, get_procedures
+from xautodl.models import obtain_model
+from xautodl.xmodels import obtain_model as obtain_xmodel
+from xautodl.nas_infer_model import obtain_nas_infer_model
+from xautodl.utils import get_model_infos
+from xautodl.log_utils import AverageMeter, time_string, convert_secs2time
+
+
+def main(args):
+    assert torch.cuda.is_available(), "CUDA is not available."
+    torch.backends.cudnn.enabled = True
+    torch.backends.cudnn.benchmark = True
+    # torch.backends.cudnn.deterministic = True
+    # torch.set_num_threads(args.workers)
+
+    prepare_seed(args.rand_seed)
+    logger = prepare_logger(args)
+
+    train_data, valid_data, xshape, class_num = get_datasets(
+        args.dataset, args.data_path, args.cutout_length
+    )
+    train_loader = torch.utils.data.DataLoader(
+        train_data,
+        batch_size=args.batch_size,
+        shuffle=True,
+        num_workers=args.workers,
+        pin_memory=True,
+    )
+    valid_loader = torch.utils.data.DataLoader(
+        valid_data,
+        batch_size=args.batch_size,
+        shuffle=False,
+        num_workers=args.workers,
+        pin_memory=True,
+    )
+    # get configures
+    model_config = load_config(args.model_config, {"class_num": class_num}, logger)
+    optim_config = load_config(args.optim_config, {"class_num": class_num}, logger)
+
+    if args.model_source == "normal":
+        base_model = obtain_model(model_config)
+    elif args.model_source == "nas":
+        base_model = obtain_nas_infer_model(model_config, args.extra_model_path)
+    elif args.model_source == "autodl-searched":
+        base_model = obtain_model(model_config, args.extra_model_path)
+    elif args.model_source in ("x", "xmodel"):
+        base_model = obtain_xmodel(model_config)
+    else:
+        raise ValueError("invalid model-source : {:}".format(args.model_source))
+    flop, param = get_model_infos(base_model, xshape)
+    logger.log("model ====>>>>:\n{:}".format(base_model))
+    logger.log("model information : {:}".format(base_model.get_message()))
+    logger.log("-" * 50)
+    logger.log(
+        "Params={:.2f} MB, FLOPs={:.2f} M ... = {:.2f} G".format(
+            param, flop, flop / 1e3
+        )
+    )
+    logger.log("-" * 50)
+    logger.log("train_data : {:}".format(train_data))
+    logger.log("valid_data : {:}".format(valid_data))
+    optimizer, scheduler, criterion = get_optim_scheduler(
+        base_model.parameters(), optim_config
+    )
+    logger.log("optimizer  : {:}".format(optimizer))
+    logger.log("scheduler  : {:}".format(scheduler))
+    logger.log("criterion  : {:}".format(criterion))
+
+    last_info, model_base_path, model_best_path = (
+        logger.path("info"),
+        logger.path("model"),
+        logger.path("best"),
+    )
+    network, criterion = torch.nn.DataParallel(base_model).cuda(), criterion.cuda()
+
+    if last_info.exists():  # automatically resume from previous checkpoint
+        logger.log(
+            "=> loading checkpoint of the last-info '{:}' start".format(last_info)
+        )
+        last_infox = torch.load(last_info)
+        start_epoch = last_infox["epoch"] + 1
+        last_checkpoint_path = last_infox["last_checkpoint"]
+        if not last_checkpoint_path.exists():
+            logger.log(
+                "Does not find {:}, try another path".format(last_checkpoint_path)
+            )
+            last_checkpoint_path = (
+                last_info.parent
+                / last_checkpoint_path.parent.name
+                / last_checkpoint_path.name
+            )
+        checkpoint = torch.load(last_checkpoint_path)
+        base_model.load_state_dict(checkpoint["base-model"])
+        scheduler.load_state_dict(checkpoint["scheduler"])
+        optimizer.load_state_dict(checkpoint["optimizer"])
+        valid_accuracies = checkpoint["valid_accuracies"]
+        max_bytes = checkpoint["max_bytes"]
+        logger.log(
+            "=> loading checkpoint of the last-info '{:}' start with {:}-th epoch.".format(
+                last_info, start_epoch
+            )
+        )
+    elif args.resume is not None:
+        assert Path(args.resume).exists(), "Can not find the resume file : {:}".format(
+            args.resume
+        )
+        checkpoint = torch.load(args.resume)
+        start_epoch = checkpoint["epoch"] + 1
+        base_model.load_state_dict(checkpoint["base-model"])
+        scheduler.load_state_dict(checkpoint["scheduler"])
+        optimizer.load_state_dict(checkpoint["optimizer"])
+        valid_accuracies = checkpoint["valid_accuracies"]
+        max_bytes = checkpoint["max_bytes"]
+        logger.log(
+            "=> loading checkpoint from '{:}' start with {:}-th epoch.".format(
+                args.resume, start_epoch
+            )
+        )
+    elif args.init_model is not None:
+        assert Path(
+            args.init_model
+        ).exists(), "Can not find the initialization file : {:}".format(args.init_model)
+        checkpoint = torch.load(args.init_model)
+        base_model.load_state_dict(checkpoint["base-model"])
+        start_epoch, valid_accuracies, max_bytes = 0, {"best": -1}, {}
+        logger.log("=> initialize the model from {:}".format(args.init_model))
+    else:
+        logger.log("=> do not find the last-info file : {:}".format(last_info))
+        start_epoch, valid_accuracies, max_bytes = 0, {"best": -1}, {}
+
+    train_func, valid_func = get_procedures(args.procedure)
+
+    total_epoch = optim_config.epochs + optim_config.warmup
+    # Main Training and Evaluation Loop
+    start_time = time.time()
+    epoch_time = AverageMeter()
+    for epoch in range(start_epoch, total_epoch):
+        scheduler.update(epoch, 0.0)
+        need_time = "Time Left: {:}".format(
+            convert_secs2time(epoch_time.avg * (total_epoch - epoch), True)
+        )
+        epoch_str = "epoch={:03d}/{:03d}".format(epoch, total_epoch)
+        LRs = scheduler.get_lr()
+        find_best = False
+        # set-up drop-out ratio
+        if hasattr(base_model, "update_drop_path"):
+            base_model.update_drop_path(
+                model_config.drop_path_prob * epoch / total_epoch
+            )
+        logger.log(
+            "\n***{:s}*** start {:s} {:s}, LR=[{:.6f} ~ {:.6f}], scheduler={:}".format(
+                time_string(), epoch_str, need_time, min(LRs), max(LRs), scheduler
+            )
+        )
+
+        # train for one epoch
+        train_loss, train_acc1, train_acc5 = train_func(
+            train_loader,
+            network,
+            criterion,
+            scheduler,
+            optimizer,
+            optim_config,
+            epoch_str,
+            args.print_freq,
+            logger,
+        )
+        # log the results
+        logger.log(
+            "***{:s}*** TRAIN [{:}] loss = {:.6f}, accuracy-1 = {:.2f}, accuracy-5 = {:.2f}".format(
+                time_string(), epoch_str, train_loss, train_acc1, train_acc5
+            )
+        )
+
+        # evaluate the performance
+        if (epoch % args.eval_frequency == 0) or (epoch + 1 == total_epoch):
+            logger.log("-" * 150)
+            valid_loss, valid_acc1, valid_acc5 = valid_func(
+                valid_loader,
+                network,
+                criterion,
+                optim_config,
+                epoch_str,
+                args.print_freq_eval,
+                logger,
+            )
+            valid_accuracies[epoch] = valid_acc1
+            logger.log(
+                "***{:s}*** VALID [{:}] loss = {:.6f}, accuracy@1 = {:.2f}, accuracy@5 = {:.2f} | Best-Valid-Acc@1={:.2f}, Error@1={:.2f}".format(
+                    time_string(),
+                    epoch_str,
+                    valid_loss,
+                    valid_acc1,
+                    valid_acc5,
+                    valid_accuracies["best"],
+                    100 - valid_accuracies["best"],
+                )
+            )
+            if valid_acc1 > valid_accuracies["best"]:
+                valid_accuracies["best"] = valid_acc1
+                find_best = True
+                logger.log(
+                    "Currently, the best validation accuracy found at {:03d}-epoch :: acc@1={:.2f}, acc@5={:.2f}, error@1={:.2f}, error@5={:.2f}, save into {:}.".format(
+                        epoch,
+                        valid_acc1,
+                        valid_acc5,
+                        100 - valid_acc1,
+                        100 - valid_acc5,
+                        model_best_path,
+                    )
+                )
+            num_bytes = (
+                torch.cuda.max_memory_cached(next(network.parameters()).device) * 1.0
+            )
+            logger.log(
+                "[GPU-Memory-Usage on {:} is {:} bytes, {:.2f} KB, {:.2f} MB, {:.2f} GB.]".format(
+                    next(network.parameters()).device,
+                    int(num_bytes),
+                    num_bytes / 1e3,
+                    num_bytes / 1e6,
+                    num_bytes / 1e9,
+                )
+            )
+            max_bytes[epoch] = num_bytes
+        if epoch % 10 == 0:
+            torch.cuda.empty_cache()
+
+        # save checkpoint
+        save_path = save_checkpoint(
+            {
+                "epoch": epoch,
+                "args": deepcopy(args),
+                "max_bytes": deepcopy(max_bytes),
+                "FLOP": flop,
+                "PARAM": param,
+                "valid_accuracies": deepcopy(valid_accuracies),
+                "model-config": model_config._asdict(),
+                "optim-config": optim_config._asdict(),
+                "base-model": base_model.state_dict(),
+                "scheduler": scheduler.state_dict(),
+                "optimizer": optimizer.state_dict(),
+            },
+            model_base_path,
+            logger,
+        )
+        if find_best:
+            copy_checkpoint(model_base_path, model_best_path, logger)
+        last_info = save_checkpoint(
+            {
+                "epoch": epoch,
+                "args": deepcopy(args),
+                "last_checkpoint": save_path,
+            },
+            logger.path("info"),
+            logger,
+        )
+
+        # measure elapsed time
+        epoch_time.update(time.time() - start_time)
+        start_time = time.time()
+
+    logger.log("\n" + "-" * 200)
+    logger.log(
+        "Finish training/validation in {:} with Max-GPU-Memory of {:.2f} MB, and save final checkpoint into {:}".format(
+            convert_secs2time(epoch_time.sum, True),
+            max(v for k, v in max_bytes.items()) / 1e6,
+            logger.path("info"),
+        )
+    )
+    logger.log("-" * 200 + "\n")
+    logger.close()
+
+
+if __name__ == "__main__":
+    args = obtain_args()
+    main(args)
diff --git a/notebooks/spaces/test-transformer-encoder.ipynb b/notebooks/spaces/test-transformer-encoder.ipynb
new file mode 100644
index 0000000..89d4452
--- /dev/null
+++ b/notebooks/spaces/test-transformer-encoder.ipynb
@@ -0,0 +1,277 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "3f754c96",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "from xautodl import spaces\n",
+    "from xautodl.xlayers import super_core\n",
+    "\n",
+    "def _create_stel(input_dim, output_dim, order):\n",
+    "    return super_core.SuperSequential(\n",
+    "        super_core.SuperLinear(input_dim, output_dim),\n",
+    "        super_core.SuperTransformerEncoderLayer(\n",
+    "            output_dim,\n",
+    "            num_heads=spaces.Categorical(2, 4, 6),\n",
+    "            mlp_hidden_multiplier=spaces.Categorical(1, 2, 4),\n",
+    "            order=order,\n",
+    "        ),\n",
+    "    )"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "81d42f4b",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "batch, seq_dim, input_dim = 1, 4, 6\n",
+    "order = super_core.LayerOrder.PreNorm"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "id": "8056b37c",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "SuperSequential(\n",
+      "  (0): SuperSequential(\n",
+      "    (0): SuperLinear(in_features=6, out_features=Categorical(candidates=[12, 24, 36], default_index=None), bias=True)\n",
+      "    (1): SuperTransformerEncoderLayer(\n",
+      "      (norm1): SuperLayerNorm1D(shape=Categorical(candidates=[12, 24, 36], default_index=None), eps=1e-06, elementwise_affine=True)\n",
+      "      (mha): SuperSelfAttention(\n",
+      "        input_dim=Categorical(candidates=[12, 24, 36], default_index=None), proj_dim=Categorical(candidates=[12, 24, 36], default_index=None), num_heads=Categorical(candidates=[2, 4, 6], default_index=None), mask=False, infinity=1000000000.0\n",
+      "        (q_fc): SuperLinear(in_features=Categorical(candidates=[12, 24, 36], default_index=None), out_features=Categorical(candidates=[12, 24, 36], default_index=None), bias=False)\n",
+      "        (k_fc): SuperLinear(in_features=Categorical(candidates=[12, 24, 36], default_index=None), out_features=Categorical(candidates=[12, 24, 36], default_index=None), bias=False)\n",
+      "        (v_fc): SuperLinear(in_features=Categorical(candidates=[12, 24, 36], default_index=None), out_features=Categorical(candidates=[12, 24, 36], default_index=None), bias=False)\n",
+      "        (attn_drop): SuperDrop(p=0.0, dims=[-1, -1, -1, -1], recover=True)\n",
+      "      )\n",
+      "      (drop): Dropout(p=0.0, inplace=False)\n",
+      "      (norm2): SuperLayerNorm1D(shape=Categorical(candidates=[12, 24, 36], default_index=None), eps=1e-06, elementwise_affine=True)\n",
+      "      (mlp): SuperMLPv2(\n",
+      "        in_features=Categorical(candidates=[12, 24, 36], default_index=None), hidden_multiplier=Categorical(candidates=[1, 2, 4], default_index=None), out_features=Categorical(candidates=[12, 24, 36], default_index=None), drop=None, fc1 -> act -> drop -> fc2 -> drop,\n",
+      "        (_params): ParameterDict(\n",
+      "            (fc1_super_weight): Parameter containing: [torch.FloatTensor of size 144x36]\n",
+      "            (fc1_super_bias): Parameter containing: [torch.FloatTensor of size 144]\n",
+      "            (fc2_super_weight): Parameter containing: [torch.FloatTensor of size 36x144]\n",
+      "            (fc2_super_bias): Parameter containing: [torch.FloatTensor of size 36]\n",
+      "        )\n",
+      "        (act): GELU()\n",
+      "        (drop): Dropout(p=0.0, inplace=False)\n",
+      "      )\n",
+      "    )\n",
+      "  )\n",
+      "  (1): SuperSequential(\n",
+      "    (0): SuperLinear(in_features=Categorical(candidates=[12, 24, 36], default_index=None), out_features=Categorical(candidates=[24, 36, 48], default_index=None), bias=True)\n",
+      "    (1): SuperTransformerEncoderLayer(\n",
+      "      (norm1): SuperLayerNorm1D(shape=Categorical(candidates=[24, 36, 48], default_index=None), eps=1e-06, elementwise_affine=True)\n",
+      "      (mha): SuperSelfAttention(\n",
+      "        input_dim=Categorical(candidates=[24, 36, 48], default_index=None), proj_dim=Categorical(candidates=[24, 36, 48], default_index=None), num_heads=Categorical(candidates=[2, 4, 6], default_index=None), mask=False, infinity=1000000000.0\n",
+      "        (q_fc): SuperLinear(in_features=Categorical(candidates=[24, 36, 48], default_index=None), out_features=Categorical(candidates=[24, 36, 48], default_index=None), bias=False)\n",
+      "        (k_fc): SuperLinear(in_features=Categorical(candidates=[24, 36, 48], default_index=None), out_features=Categorical(candidates=[24, 36, 48], default_index=None), bias=False)\n",
+      "        (v_fc): SuperLinear(in_features=Categorical(candidates=[24, 36, 48], default_index=None), out_features=Categorical(candidates=[24, 36, 48], default_index=None), bias=False)\n",
+      "        (attn_drop): SuperDrop(p=0.0, dims=[-1, -1, -1, -1], recover=True)\n",
+      "      )\n",
+      "      (drop): Dropout(p=0.0, inplace=False)\n",
+      "      (norm2): SuperLayerNorm1D(shape=Categorical(candidates=[24, 36, 48], default_index=None), eps=1e-06, elementwise_affine=True)\n",
+      "      (mlp): SuperMLPv2(\n",
+      "        in_features=Categorical(candidates=[24, 36, 48], default_index=None), hidden_multiplier=Categorical(candidates=[1, 2, 4], default_index=None), out_features=Categorical(candidates=[24, 36, 48], default_index=None), drop=None, fc1 -> act -> drop -> fc2 -> drop,\n",
+      "        (_params): ParameterDict(\n",
+      "            (fc1_super_weight): Parameter containing: [torch.FloatTensor of size 192x48]\n",
+      "            (fc1_super_bias): Parameter containing: [torch.FloatTensor of size 192]\n",
+      "            (fc2_super_weight): Parameter containing: [torch.FloatTensor of size 48x192]\n",
+      "            (fc2_super_bias): Parameter containing: [torch.FloatTensor of size 48]\n",
+      "        )\n",
+      "        (act): GELU()\n",
+      "        (drop): Dropout(p=0.0, inplace=False)\n",
+      "      )\n",
+      "    )\n",
+      "  )\n",
+      "  (2): SuperSequential(\n",
+      "    (0): SuperLinear(in_features=Categorical(candidates=[24, 36, 48], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=True)\n",
+      "    (1): SuperTransformerEncoderLayer(\n",
+      "      (norm1): SuperLayerNorm1D(shape=Categorical(candidates=[36, 72, 100], default_index=None), eps=1e-06, elementwise_affine=True)\n",
+      "      (mha): SuperSelfAttention(\n",
+      "        input_dim=Categorical(candidates=[36, 72, 100], default_index=None), proj_dim=Categorical(candidates=[36, 72, 100], default_index=None), num_heads=Categorical(candidates=[2, 4, 6], default_index=None), mask=False, infinity=1000000000.0\n",
+      "        (q_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
+      "        (k_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
+      "        (v_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
+      "        (attn_drop): SuperDrop(p=0.0, dims=[-1, -1, -1, -1], recover=True)\n",
+      "      )\n",
+      "      (drop): Dropout(p=0.0, inplace=False)\n",
+      "      (norm2): SuperLayerNorm1D(shape=Categorical(candidates=[36, 72, 100], default_index=None), eps=1e-06, elementwise_affine=True)\n",
+      "      (mlp): SuperMLPv2(\n",
+      "        in_features=Categorical(candidates=[36, 72, 100], default_index=None), hidden_multiplier=Categorical(candidates=[1, 2, 4], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), drop=None, fc1 -> act -> drop -> fc2 -> drop,\n",
+      "        (_params): ParameterDict(\n",
+      "            (fc1_super_weight): Parameter containing: [torch.FloatTensor of size 400x100]\n",
+      "            (fc1_super_bias): Parameter containing: [torch.FloatTensor of size 400]\n",
+      "            (fc2_super_weight): Parameter containing: [torch.FloatTensor of size 100x400]\n",
+      "            (fc2_super_bias): Parameter containing: [torch.FloatTensor of size 100]\n",
+      "        )\n",
+      "        (act): GELU()\n",
+      "        (drop): Dropout(p=0.0, inplace=False)\n",
+      "      )\n",
+      "    )\n",
+      "  )\n",
+      ")\n"
+     ]
+    }
+   ],
+   "source": [
+    "out1_dim = spaces.Categorical(12, 24, 36)\n",
+    "out2_dim = spaces.Categorical(24, 36, 48)\n",
+    "out3_dim = spaces.Categorical(36, 72, 100)\n",
+    "layer1 = _create_stel(input_dim, out1_dim, order)\n",
+    "layer2 = _create_stel(out1_dim, out2_dim, order)\n",
+    "layer3 = _create_stel(out2_dim, out3_dim, order)\n",
+    "model = super_core.SuperSequential(layer1, layer2, layer3)\n",
+    "print(model)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4fd53a7c",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "> \u001b[0;32m/Users/xuanyidong/anaconda3/lib/python3.8/site-packages/xautodl-0.9.9-py3.8.egg/xautodl/xlayers/super_transformer.py\u001b[0m(116)\u001b[0;36mforward_raw\u001b[0;34m()\u001b[0m\n",
+      "\u001b[0;32m    114 \u001b[0;31m              \u001b[0;32mimport\u001b[0m \u001b[0mpdb\u001b[0m\u001b[0;34m;\u001b[0m \u001b[0mpdb\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mset_trace\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0m\u001b[0;32m    115 \u001b[0;31m            \u001b[0;31m# feed-forward layer -- MLP\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0m\u001b[0;32m--> 116 \u001b[0;31m            \u001b[0my\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mnorm2\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0m\u001b[0;32m    117 \u001b[0;31m            \u001b[0mouts\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mx\u001b[0m \u001b[0;34m+\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mmlp\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0my\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0m\u001b[0;32m    118 \u001b[0;31m        \u001b[0;32melif\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_order\u001b[0m \u001b[0;32mis\u001b[0m \u001b[0mLayerOrder\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mPostNorm\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0m\n",
+      "ipdb> print(self)\n",
+      "SuperTransformerEncoderLayer(\n",
+      "  (norm1): SuperLayerNorm1D(shape=Categorical(candidates=[36, 72, 100], default_index=None), eps=1e-06, elementwise_affine=True)\n",
+      "  (mha): SuperSelfAttention(\n",
+      "    input_dim=Categorical(candidates=[36, 72, 100], default_index=None), proj_dim=Categorical(candidates=[36, 72, 100], default_index=None), num_heads=Categorical(candidates=[2, 4, 6], default_index=None), mask=False, infinity=1000000000.0\n",
+      "    (q_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
+      "    (k_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
+      "    (v_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
+      "    (attn_drop): SuperDrop(p=0.0, dims=[-1, -1, -1, -1], recover=True)\n",
+      "  )\n",
+      "  (drop): Dropout(p=0.0, inplace=False)\n",
+      "  (norm2): SuperLayerNorm1D(shape=Categorical(candidates=[36, 72, 100], default_index=None), eps=1e-06, elementwise_affine=True)\n",
+      "  (mlp): SuperMLPv2(\n",
+      "    in_features=Categorical(candidates=[36, 72, 100], default_index=None), hidden_multiplier=Categorical(candidates=[1, 2, 4], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), drop=None, fc1 -> act -> drop -> fc2 -> drop,\n",
+      "    (_params): ParameterDict(\n",
+      "        (fc1_super_weight): Parameter containing: [torch.FloatTensor of size 400x100]\n",
+      "        (fc1_super_bias): Parameter containing: [torch.FloatTensor of size 400]\n",
+      "        (fc2_super_weight): Parameter containing: [torch.FloatTensor of size 100x400]\n",
+      "        (fc2_super_bias): Parameter containing: [torch.FloatTensor of size 100]\n",
+      "    )\n",
+      "    (act): GELU()\n",
+      "    (drop): Dropout(p=0.0, inplace=False)\n",
+      "  )\n",
+      ")\n",
+      "ipdb> print(inputs.shape)\n",
+      "torch.Size([1, 4, 100])\n",
+      "ipdb> print(x.shape)\n",
+      "torch.Size([1, 4, 96])\n",
+      "ipdb> print(self.mha)\n",
+      "SuperSelfAttention(\n",
+      "  input_dim=Categorical(candidates=[36, 72, 100], default_index=None), proj_dim=Categorical(candidates=[36, 72, 100], default_index=None), num_heads=Categorical(candidates=[2, 4, 6], default_index=None), mask=False, infinity=1000000000.0\n",
+      "  (q_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
+      "  (k_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
+      "  (v_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
+      "  (attn_drop): SuperDrop(p=0.0, dims=[-1, -1, -1, -1], recover=True)\n",
+      ")\n",
+      "ipdb> print(self.mha.candidate)\n",
+      "*** AttributeError: 'SuperSelfAttention' object has no attribute 'candidate'\n",
+      "ipdb> print(self.mha.abstract_candidate)\n",
+      "*** AttributeError: 'SuperSelfAttention' object has no attribute 'abstract_candidate'\n",
+      "ipdb> print(self.mha._abstract_child)\n",
+      "None\n",
+      "ipdb> print(self.abstract_child)\n",
+      "None\n",
+      "ipdb> print(self.abstract_child.abstract_child)\n",
+      "*** AttributeError: 'NoneType' object has no attribute 'abstract_child'\n",
+      "ipdb> print(self.mha)\n",
+      "SuperSelfAttention(\n",
+      "  input_dim=Categorical(candidates=[36, 72, 100], default_index=None), proj_dim=Categorical(candidates=[36, 72, 100], default_index=None), num_heads=Categorical(candidates=[2, 4, 6], default_index=None), mask=False, infinity=1000000000.0\n",
+      "  (q_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
+      "  (k_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
+      "  (v_fc): SuperLinear(in_features=Categorical(candidates=[36, 72, 100], default_index=None), out_features=Categorical(candidates=[36, 72, 100], default_index=None), bias=False)\n",
+      "  (attn_drop): SuperDrop(p=0.0, dims=[-1, -1, -1, -1], recover=True)\n",
+      ")\n"
+     ]
+    }
+   ],
+   "source": [
+    "inputs = torch.rand(batch, seq_dim, input_dim)\n",
+    "outputs = model(inputs)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "05332b98",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "abstract_space = model.abstract_search_space\n",
+    "abstract_space.clean_last()\n",
+    "abstract_child = abstract_space.random(reuse_last=True)\n",
+    "# print(\"The abstract child program is:\\n{:}\".format(abstract_child))\n",
+    "model.enable_candidate()\n",
+    "model.set_super_run_type(super_core.SuperRunMode.Candidate)\n",
+    "model.apply_candidate(abstract_child)\n",
+    "outputs = model(inputs)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "3289f938",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "print(outputs.shape)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "36951cdf",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.8.8"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}
diff --git a/xautodl/xlayers/super_attention.py b/xautodl/xlayers/super_attention.py
index c70ea38..c141b53 100644
--- a/xautodl/xlayers/super_attention.py
+++ b/xautodl/xlayers/super_attention.py
@@ -38,12 +38,15 @@ class SuperSelfAttention(SuperModule):
         self._use_mask = use_mask
         self._infinity = 1e9
 
-        mul_head_dim = (input_dim // num_heads) * num_heads
-        self.q_fc = SuperLinear(input_dim, mul_head_dim, bias=qkv_bias)
-        self.k_fc = SuperLinear(input_dim, mul_head_dim, bias=qkv_bias)
-        self.v_fc = SuperLinear(input_dim, mul_head_dim, bias=qkv_bias)
+        mul_head_dim = (
+            spaces.get_max(input_dim) // spaces.get_min(num_heads)
+        ) * spaces.get_min(num_heads)
+        assert mul_head_dim == spaces.get_max(input_dim)
+        self.q_fc = SuperLinear(input_dim, input_dim, bias=qkv_bias)
+        self.k_fc = SuperLinear(input_dim, input_dim, bias=qkv_bias)
+        self.v_fc = SuperLinear(input_dim, input_dim, bias=qkv_bias)
 
-        self.attn_drop = SuperDrop(attn_drop, [-1, -1, -1, -1], recover=True)
+        self.attn_drop = SuperDrop(attn_drop or 0.0, [-1, -1, -1, -1], recover=True)
         if proj_dim is None:
             self.proj = SuperLinear(input_dim, proj_dim)
             self.proj_drop = SuperDropout(proj_drop or 0.0)
@@ -127,7 +130,18 @@ class SuperSelfAttention(SuperModule):
         attn_v1 = attn_v1.softmax(dim=-1)  # B * #head * N * N
         attn_v1 = self.attn_drop(attn_v1)
         feats_v1 = (attn_v1 @ v_v1).permute(0, 2, 1, 3).reshape(B, N, -1)
-        return feats_v1
+        if C == head_dim * num_head:
+            feats = feats_v1
+        else:  # The channels can not be divided by num_head, the remainder forms an additional head
+            q_v2 = q[:, :, num_head * head_dim :]
+            k_v2 = k[:, :, num_head * head_dim :]
+            v_v2 = v[:, :, num_head * head_dim :]
+            attn_v2 = (q_v2 @ k_v2.transpose(-2, -1)) * math.sqrt(q_v2.shape[-1])
+            attn_v2 = attn_v2.softmax(dim=-1)
+            attn_v2 = self.attn_drop(attn_v2)
+            feats_v2 = attn_v2 @ v_v2
+            feats = torch.cat([feats_v1, feats_v2], dim=-1)
+        return feats
 
     def forward_candidate(self, input: torch.Tensor) -> torch.Tensor:
         # check the num_heads:
diff --git a/xautodl/xmodels/__init__.py b/xautodl/xmodels/__init__.py
index 04f21fe..e44777d 100644
--- a/xautodl/xmodels/__init__.py
+++ b/xautodl/xmodels/__init__.py
@@ -5,3 +5,6 @@
 #####################################################
 
 from .transformers import get_transformer
+
+def obtain_model(config):
+  raise NotImplementedError