From a867ea5209dfd6e95d4f6b02195bd00ce8de4a9b Mon Sep 17 00:00:00 2001
From: D-X-Y <280835372@qq.com>
Date: Thu, 25 Feb 2021 00:24:56 -0800
Subject: [PATCH] Start prototype

---
 .gitignore                            |   3 +-
 .latent-data/init-configs/vimrc       |   4 +-
 exps/trading/test-qlib.py             |   3 -
 exps/trading/workflow_test.py         |  94 ++++++
 lib/layers/__init__.py                |   2 +
 lib/layers/drop.py                    | 169 ++++++++++
 lib/layers/weight_init.py             |  61 ++++
 lib/trade_models/__init__.py          |   1 +
 lib/trade_models/quant_transformer.py | 461 ++++++++++++++++++++++++++
 9 files changed, 792 insertions(+), 6 deletions(-)
 delete mode 100644 exps/trading/test-qlib.py
 create mode 100644 exps/trading/workflow_test.py
 create mode 100644 lib/layers/__init__.py
 create mode 100644 lib/layers/drop.py
 create mode 100644 lib/layers/weight_init.py
 create mode 100644 lib/trade_models/__init__.py
 create mode 100755 lib/trade_models/quant_transformer.py

diff --git a/.gitignore b/.gitignore
index c705264..8d5cec0 100644
--- a/.gitignore
+++ b/.gitignore
@@ -127,4 +127,5 @@ TEMP-L.sh
 */*.swo
 
 # Visual Studio Code
-.vscode
\ No newline at end of file
+.vscode
+mlruns
diff --git a/.latent-data/init-configs/vimrc b/.latent-data/init-configs/vimrc
index d139c97..c843e86 100644
--- a/.latent-data/init-configs/vimrc
+++ b/.latent-data/init-configs/vimrc
@@ -1,4 +1,4 @@
-color desert 
+color desert
 "设置背景色，每种配色有两种方案，一个light、一个dark 
 "set background=light 
 ""打开语法高亮 
@@ -16,7 +16,7 @@ endif
 "显示行号 
 "set number 
 ""设置缩进有三个取值cindent(c风格)、smartindent(智能模式，其实不觉得有什么智能)、autoindent(简单的与上一行保持一致) 
-set cindent 
+set autoindent 
 "在windows版本中vim的退格键模式默认与vi兼容，与我们的使用习惯不太符合，下边这条可以改过来
 "set backspace=indent,eol,start 
 ""用空格键替换制表符 
diff --git a/exps/trading/test-qlib.py b/exps/trading/test-qlib.py
deleted file mode 100644
index 8ba0624..0000000
--- a/exps/trading/test-qlib.py
+++ /dev/null
@@ -1,3 +0,0 @@
-import os
-
-print('xxx123')
\ No newline at end of file
diff --git a/exps/trading/workflow_test.py b/exps/trading/workflow_test.py
new file mode 100644
index 0000000..e9f98e0
--- /dev/null
+++ b/exps/trading/workflow_test.py
@@ -0,0 +1,94 @@
+#####################################################
+# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2019.01 #
+#####################################################
+# Refer to:
+# - https://github.com/microsoft/qlib/blob/main/examples/workflow_by_code.ipynb
+# - https://github.com/microsoft/qlib/blob/main/examples/workflow_by_code.py
+# python exps/trading/workflow_test.py
+#####################################################
+import sys, site
+from pathlib import Path
+
+lib_dir = (Path(__file__).parent / '..' / '..' / 'lib').resolve()
+if str(lib_dir) not in sys.path: sys.path.insert(0, str(lib_dir))
+
+import qlib
+import pandas as pd
+from qlib.config import REG_CN
+from qlib.contrib.model.gbdt import LGBModel
+from qlib.contrib.data.handler import Alpha158
+from qlib.contrib.strategy.strategy import TopkDropoutStrategy
+from qlib.contrib.evaluate import (
+    backtest as normal_backtest,
+    risk_analysis,
+)
+from qlib.utils import exists_qlib_data, init_instance_by_config
+from qlib.workflow import R
+from qlib.workflow.record_temp import SignalRecord, PortAnaRecord
+from qlib.utils import flatten_dict
+
+
+# use default data
+# NOTE: need to download data from remote: python scripts/get_data.py qlib_data_cn --target_dir ~/.qlib/qlib_data/cn_data
+provider_uri = "~/.qlib/qlib_data/cn_data"  # target_dir
+if not exists_qlib_data(provider_uri):
+    print(f"Qlib data is not found in {provider_uri}")
+    sys.path.append(str(scripts_dir))
+    from get_data import GetData
+    GetData().qlib_data(target_dir=provider_uri, region=REG_CN)
+qlib.init(provider_uri=provider_uri, region=REG_CN)
+
+
+market = "csi300"
+benchmark = "SH000300"
+
+
+###################################
+# train model
+###################################
+data_handler_config = {
+    "start_time": "2008-01-01",
+    "end_time": "2020-08-01",
+    "fit_start_time": "2008-01-01",
+    "fit_end_time": "2014-12-31",
+    "instruments": market,
+}
+
+task = {
+    "model": {
+        "class": "QuantTransformer",
+        "module_path": "trade_models",
+        "kwargs": {
+            "loss": "mse",
+            "GPU": "0",
+            "metric": "loss",
+        },
+    },
+    "dataset": {
+        "class": "DatasetH",
+        "module_path": "qlib.data.dataset",
+        "kwargs": {
+            "handler": {
+                "class": "Alpha158",
+                "module_path": "qlib.contrib.data.handler",
+                "kwargs": data_handler_config,
+            },
+            "segments": {
+                "train": ("2008-01-01", "2014-12-31"),
+                "valid": ("2015-01-01", "2016-12-31"),
+                "test": ("2017-01-01", "2020-08-01"),
+            },
+        },
+    },
+}
+
+# model initiaiton
+model = init_instance_by_config(task["model"])
+dataset = init_instance_by_config(task["dataset"])
+
+# start exp to train model
+with R.start(experiment_name="train_model"):
+    R.log_params(**flatten_dict(task))
+    model.fit(dataset)
+    R.save_objects(trained_model=model)
+    rid = R.get_recorder().id
diff --git a/lib/layers/__init__.py b/lib/layers/__init__.py
new file mode 100644
index 0000000..21bca49
--- /dev/null
+++ b/lib/layers/__init__.py
@@ -0,0 +1,2 @@
+from .drop import DropBlock2d, DropPath
+from .weight_init import trunc_normal_
diff --git a/lib/layers/drop.py b/lib/layers/drop.py
new file mode 100644
index 0000000..f9ebef7
--- /dev/null
+++ b/lib/layers/drop.py
@@ -0,0 +1,169 @@
+""" Borrowed from https://github.com/rwightman/pytorch-image-models
+DropBlock, DropPath
+
+PyTorch implementations of DropBlock and DropPath (Stochastic Depth) regularization layers.
+
+Papers:
+DropBlock: A regularization method for convolutional networks (https://arxiv.org/abs/1810.12890)
+
+Deep Networks with Stochastic Depth (https://arxiv.org/abs/1603.09382)
+
+Code:
+DropBlock impl inspired by two Tensorflow impl that I liked:
+ - https://github.com/tensorflow/tpu/blob/master/models/official/resnet/resnet_model.py#L74
+ - https://github.com/clovaai/assembled-cnn/blob/master/nets/blocks.py
+
+Hacked together by / Copyright 2020 Ross Wightman
+"""
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def drop_block_2d(
+    x, drop_prob: float = 0.1, block_size: int = 7,  gamma_scale: float = 1.0,
+    with_noise: bool = False, inplace: bool = False, batchwise: bool = False):
+  """ DropBlock. See https://arxiv.org/pdf/1810.12890.pdf
+
+  DropBlock with an experimental gaussian noise option. This layer has been tested on a few training
+  runs with success, but needs further validation and possibly optimization for lower runtime impact.
+  """
+  B, C, H, W = x.shape
+  total_size = W * H
+  clipped_block_size = min(block_size, min(W, H))
+  # seed_drop_rate, the gamma parameter
+  gamma = gamma_scale * drop_prob * total_size / clipped_block_size ** 2 / (
+    (W - block_size + 1) * (H - block_size + 1))
+
+  # Forces the block to be inside the feature map.
+  w_i, h_i = torch.meshgrid(torch.arange(W).to(x.device), torch.arange(H).to(x.device))
+  valid_block = ((w_i >= clipped_block_size // 2) & (w_i < W - (clipped_block_size - 1) // 2)) & \
+                ((h_i >= clipped_block_size // 2) & (h_i < H - (clipped_block_size - 1) // 2))
+  valid_block = torch.reshape(valid_block, (1, 1, H, W)).to(dtype=x.dtype)
+
+  if batchwise:
+    # one mask for whole batch, quite a bit faster
+    uniform_noise = torch.rand((1, C, H, W), dtype=x.dtype, device=x.device)
+  else:
+    uniform_noise = torch.rand_like(x)
+  block_mask = ((2 - gamma - valid_block + uniform_noise) >= 1).to(dtype=x.dtype)
+  block_mask = -F.max_pool2d(
+    -block_mask,
+    kernel_size=clipped_block_size,  # block_size,
+    stride=1,
+    padding=clipped_block_size // 2)
+
+  if with_noise:
+    normal_noise = torch.randn((1, C, H, W), dtype=x.dtype, device=x.device) if batchwise else torch.randn_like(x)
+    if inplace:
+      x.mul_(block_mask).add_(normal_noise * (1 - block_mask))
+    else:
+      x = x * block_mask + normal_noise * (1 - block_mask)
+  else:
+    normalize_scale = (block_mask.numel() / block_mask.to(dtype=torch.float32).sum().add(1e-7)).to(x.dtype)
+    if inplace:
+      x.mul_(block_mask * normalize_scale)
+    else:
+      x = x * block_mask * normalize_scale
+  return x
+
+
+def drop_block_fast_2d(
+    x: torch.Tensor, drop_prob: float = 0.1, block_size: int = 7,
+    gamma_scale: float = 1.0, with_noise: bool = False, inplace: bool = False, batchwise: bool = False):
+  """ DropBlock. See https://arxiv.org/pdf/1810.12890.pdf
+
+  DropBlock with an experimental gaussian noise option. Simplied from above without concern for valid
+  block mask at edges.
+  """
+  B, C, H, W = x.shape
+  total_size = W * H
+  clipped_block_size = min(block_size, min(W, H))
+  gamma = gamma_scale * drop_prob * total_size / clipped_block_size ** 2 / (
+      (W - block_size + 1) * (H - block_size + 1))
+
+  if batchwise:
+    # one mask for whole batch, quite a bit faster
+    block_mask = torch.rand((1, C, H, W), dtype=x.dtype, device=x.device) < gamma
+  else:
+    # mask per batch element
+    block_mask = torch.rand_like(x) < gamma
+  block_mask = F.max_pool2d(
+    block_mask.to(x.dtype), kernel_size=clipped_block_size, stride=1, padding=clipped_block_size // 2)
+
+  if with_noise:
+    normal_noise = torch.randn((1, C, H, W), dtype=x.dtype, device=x.device) if batchwise else torch.randn_like(x)
+    if inplace:
+      x.mul_(1. - block_mask).add_(normal_noise * block_mask)
+    else:
+      x = x * (1. - block_mask) + normal_noise * block_mask
+  else:
+    block_mask = 1 - block_mask
+    normalize_scale = (block_mask.numel() / block_mask.to(dtype=torch.float32).sum().add(1e-7)).to(dtype=x.dtype)
+    if inplace:
+      x.mul_(block_mask * normalize_scale)
+    else:
+      x = x * block_mask * normalize_scale
+  return x
+
+
+class DropBlock2d(nn.Module):
+  """ DropBlock. See https://arxiv.org/pdf/1810.12890.pdf
+  """
+  def __init__(self,
+         drop_prob=0.1,
+         block_size=7,
+         gamma_scale=1.0,
+         with_noise=False,
+         inplace=False,
+         batchwise=False,
+         fast=True):
+    super(DropBlock2d, self).__init__()
+    self.drop_prob = drop_prob
+    self.gamma_scale = gamma_scale
+    self.block_size = block_size
+    self.with_noise = with_noise
+    self.inplace = inplace
+    self.batchwise = batchwise
+    self.fast = fast  # FIXME finish comparisons of fast vs not
+
+  def forward(self, x):
+    if not self.training or not self.drop_prob:
+      return x
+    if self.fast:
+      return drop_block_fast_2d(
+        x, self.drop_prob, self.block_size, self.gamma_scale, self.with_noise, self.inplace, self.batchwise)
+    else:
+      return drop_block_2d(
+        x, self.drop_prob, self.block_size, self.gamma_scale, self.with_noise, self.inplace, self.batchwise)
+
+
+def drop_path(x, drop_prob: float = 0., training: bool = False):
+  """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+
+  This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+  the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+  See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+  changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+  'survival rate' as the argument.
+
+  """
+  if drop_prob == 0. or not training:
+    return x
+  keep_prob = 1 - drop_prob
+  shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+  random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
+  random_tensor.floor_()  # binarize
+  output = x.div(keep_prob) * random_tensor
+  return output
+
+
+class DropPath(nn.Module):
+  """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+  """
+  def __init__(self, drop_prob=None):
+    super(DropPath, self).__init__()
+    self.drop_prob = drop_prob
+
+  def forward(self, x):
+    return drop_path(x, self.drop_prob, self.training)
diff --git a/lib/layers/weight_init.py b/lib/layers/weight_init.py
new file mode 100644
index 0000000..a70904b
--- /dev/null
+++ b/lib/layers/weight_init.py
@@ -0,0 +1,61 @@
+# Borrowed from https://github.com/rwightman/pytorch-image-models
+import torch
+import math
+import warnings
+
+
+def _no_grad_trunc_normal_(tensor, mean, std, a, b):
+  # Cut & paste from PyTorch official master until it's in a few official releases - RW
+  # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+  def norm_cdf(x):
+    # Computes standard normal cumulative distribution function
+    return (1. + math.erf(x / math.sqrt(2.))) / 2.
+
+  if (mean < a - 2 * std) or (mean > b + 2 * std):
+    warnings.warn("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+                  "The distribution of values may be incorrect.",
+                  stacklevel=2)
+
+  with torch.no_grad():
+    # Values are generated by using a truncated uniform distribution and
+    # then using the inverse CDF for the normal distribution.
+    # Get upper and lower cdf values
+    l = norm_cdf((a - mean) / std)
+    u = norm_cdf((b - mean) / std)
+
+    # Uniformly fill tensor with values from [l, u], then translate to
+    # [2l-1, 2u-1].
+    tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+    # Use inverse cdf transform for normal distribution to get truncated
+    # standard normal
+    tensor.erfinv_()
+
+    # Transform to proper mean, std
+    tensor.mul_(std * math.sqrt(2.))
+    tensor.add_(mean)
+
+    # Clamp to ensure it's in the proper range
+    tensor.clamp_(min=a, max=b)
+    return tensor
+
+
+def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
+  # type: (Tensor, float, float, float, float) -> Tensor
+  r"""Fills the input Tensor with values drawn from a truncated
+  normal distribution. The values are effectively drawn from the
+  normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
+  with values outside :math:`[a, b]` redrawn until they are within
+  the bounds. The method used for generating the random values works
+  best when :math:`a \leq \text{mean} \leq b`.
+  Args:
+    tensor: an n-dimensional `torch.Tensor`
+    mean: the mean of the normal distribution
+    std: the standard deviation of the normal distribution
+    a: the minimum cutoff value
+    b: the maximum cutoff value
+  Examples:
+    >>> w = torch.empty(3, 5)
+    >>> nn.init.trunc_normal_(w)
+  """
+  return _no_grad_trunc_normal_(tensor, mean, std, a, b)
diff --git a/lib/trade_models/__init__.py b/lib/trade_models/__init__.py
new file mode 100644
index 0000000..ee0721e
--- /dev/null
+++ b/lib/trade_models/__init__.py
@@ -0,0 +1 @@
+from .quant_transformer import QuantTransformer
diff --git a/lib/trade_models/quant_transformer.py b/lib/trade_models/quant_transformer.py
new file mode 100755
index 0000000..fb0c43c
--- /dev/null
+++ b/lib/trade_models/quant_transformer.py
@@ -0,0 +1,461 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+
+from __future__ import division
+from __future__ import print_function
+
+import os
+import numpy as np
+import pandas as pd
+import copy
+from functools import partial
+from sklearn.metrics import roc_auc_score, mean_squared_error
+from typing import Optional
+import logging
+
+from qlib.utils import (
+  unpack_archive_with_buffer,
+  save_multiple_parts_file,
+  create_save_path,
+  drop_nan_by_y_index,
+)
+from qlib.log import get_module_logger, TimeInspector
+
+import torch
+import torch.nn as nn
+import torch.optim as optim
+
+from layers import DropPath, trunc_normal_
+
+from qlib.model.base import Model
+from qlib.data.dataset import DatasetH
+from qlib.data.dataset.handler import DataHandlerLP
+
+
+class QuantTransformer(Model):
+  """Transformer-based Quant Model
+
+  """
+
+  def __init__(
+    self,
+    d_feat=6,
+    hidden_size=64,
+    num_layers=2,
+    dropout=0.0,
+    n_epochs=200,
+    lr=0.001,
+    metric="",
+    batch_size=2000,
+    early_stop=20,
+    loss="mse",
+    optimizer="adam",
+    GPU=0,
+    seed=None,
+    **kwargs
+  ):
+    # Set logger.
+    self.logger = get_module_logger("QuantTransformer")
+    self.logger.info("QuantTransformer pytorch version...")
+
+    # set hyper-parameters.
+    self.d_feat = d_feat
+    self.hidden_size = hidden_size
+    self.num_layers = num_layers
+    self.dropout = dropout
+    self.n_epochs = n_epochs
+    self.lr = lr
+    self.metric = metric
+    self.batch_size = batch_size
+    self.early_stop = early_stop
+    self.optimizer = optimizer.lower()
+    self.loss = loss
+    self.device = torch.device("cuda:{:}".format(GPU) if torch.cuda.is_available() else "cpu")
+    self.use_gpu = torch.cuda.is_available()
+    self.seed = seed
+
+    self.logger.info(
+      "GRU parameters setting:"
+      "\nd_feat : {}"
+      "\nhidden_size : {}"
+      "\nnum_layers : {}"
+      "\ndropout : {}"
+      "\nn_epochs : {}"
+      "\nlr : {}"
+      "\nmetric : {}"
+      "\nbatch_size : {}"
+      "\nearly_stop : {}"
+      "\noptimizer : {}"
+      "\nloss_type : {}"
+      "\nvisible_GPU : {}"
+      "\nuse_GPU : {}"
+      "\nseed : {}".format(
+        d_feat,
+        hidden_size,
+        num_layers,
+        dropout,
+        n_epochs,
+        lr,
+        metric,
+        batch_size,
+        early_stop,
+        optimizer.lower(),
+        loss,
+        GPU,
+        self.use_gpu,
+        seed,
+      )
+    )
+
+    if self.seed is not None:
+      np.random.seed(self.seed)
+      torch.manual_seed(self.seed)
+
+    self.model = TransformerModel(d_feat=self.d_feat)
+    if optimizer.lower() == "adam":
+      self.train_optimizer = optim.Adam(self.model.parameters(), lr=self.lr)
+    elif optimizer.lower() == "gd":
+      self.train_optimizer = optim.SGD(self.model.parameters(), lr=self.lr)
+    else:
+      raise NotImplementedError("optimizer {:} is not supported!".format(optimizer))
+
+    self.fitted = False
+    self.model.to(self.device)
+
+  def mse(self, pred, label):
+    loss = (pred - label) ** 2
+    return torch.mean(loss)
+
+  def loss_fn(self, pred, label):
+    mask = ~torch.isnan(label)
+
+    if self.loss == "mse":
+      return self.mse(pred[mask], label[mask])
+
+    raise ValueError("unknown loss `%s`" % self.loss)
+
+  def metric_fn(self, pred, label):
+
+    mask = torch.isfinite(label)
+
+    if self.metric == "" or self.metric == "loss":
+      return -self.loss_fn(pred[mask], label[mask])
+
+    raise ValueError("unknown metric `%s`" % self.metric)
+
+  def train_epoch(self, x_train, y_train):
+
+    x_train_values = x_train.values
+    y_train_values = np.squeeze(y_train.values)
+
+    self.model.train()
+
+    indices = np.arange(len(x_train_values))
+    np.random.shuffle(indices)
+
+    for i in range(len(indices))[:: self.batch_size]:
+
+      if len(indices) - i < self.batch_size:
+        break
+
+      feature = torch.from_numpy(x_train_values[indices[i : i + self.batch_size]]).float().to(self.device)
+      label = torch.from_numpy(y_train_values[indices[i : i + self.batch_size]]).float().to(self.device)
+
+      pred = self.model(feature)
+      loss = self.loss_fn(pred, label)
+
+      self.train_optimizer.zero_grad()
+      loss.backward()
+      torch.nn.utils.clip_grad_value_(self.model.parameters(), 3.0)
+      self.train_optimizer.step()
+
+  def test_epoch(self, data_x, data_y):
+
+    # prepare training data
+    x_values = data_x.values
+    y_values = np.squeeze(data_y.values)
+
+    self.model.eval()
+
+    scores = []
+    losses = []
+
+    indices = np.arange(len(x_values))
+    import pdb; pdb.set_trace()
+
+    for i in range(len(indices))[:: self.batch_size]:
+
+      if len(indices) - i < self.batch_size:
+        break
+
+      feature = torch.from_numpy(x_values[indices[i : i + self.batch_size]]).float().to(self.device)
+      label = torch.from_numpy(y_values[indices[i : i + self.batch_size]]).float().to(self.device)
+
+      pred = self.model(feature)
+      loss = self.loss_fn(pred, label)
+      losses.append(loss.item())
+
+      score = self.metric_fn(pred, label)
+      scores.append(score.item())
+
+    return np.mean(losses), np.mean(scores)
+
+  def fit(
+    self,
+    dataset: DatasetH,
+    evals_result=dict(),
+    verbose=True,
+    save_path=None,
+  ):
+
+    df_train, df_valid, df_test = dataset.prepare(
+      ["train", "valid", "test"],
+      col_set=["feature", "label"],
+      data_key=DataHandlerLP.DK_L,
+    )
+
+    x_train, y_train = df_train["feature"], df_train["label"]
+    x_valid, y_valid = df_valid["feature"], df_valid["label"]
+
+    if save_path == None:
+      save_path = create_save_path(save_path)
+    stop_steps = 0
+    train_loss = 0
+    best_score = -np.inf
+    best_epoch = 0
+    evals_result["train"] = []
+    evals_result["valid"] = []
+
+    # train
+    self.logger.info("training...")
+    self.fitted = True
+
+    for step in range(self.n_epochs):
+      self.logger.info("Epoch%d:", step)
+      self.logger.info("training...")
+      self.train_epoch(x_train, y_train)
+      self.logger.info("evaluating...")
+      train_loss, train_score = self.test_epoch(x_train, y_train)
+      val_loss, val_score = self.test_epoch(x_valid, y_valid)
+      self.logger.info("train %.6f, valid %.6f" % (train_score, val_score))
+      evals_result["train"].append(train_score)
+      evals_result["valid"].append(val_score)
+
+      if val_score > best_score:
+        best_score = val_score
+        stop_steps = 0
+        best_epoch = step
+        best_param = copy.deepcopy(self.model.state_dict())
+      else:
+        stop_steps += 1
+        if stop_steps >= self.early_stop:
+          self.logger.info("early stop")
+          break
+
+    self.logger.info("best score: %.6lf @ %d" % (best_score, best_epoch))
+    self.model.load_state_dict(best_param)
+    torch.save(best_param, save_path)
+
+    if self.use_gpu:
+      torch.cuda.empty_cache()
+
+  def predict(self, dataset):
+    if not self.fitted:
+      raise ValueError("model is not fitted yet!")
+
+    x_test = dataset.prepare("test", col_set="feature")
+    index = x_test.index
+    self.model.eval()
+    x_values = x_test.values
+    sample_num = x_values.shape[0]
+    preds = []
+
+    for begin in range(sample_num)[:: self.batch_size]:
+
+      if sample_num - begin < self.batch_size:
+        end = sample_num
+      else:
+        end = begin + self.batch_size
+
+      x_batch = torch.from_numpy(x_values[begin:end]).float().to(self.device)
+
+      with torch.no_grad():
+        if self.use_gpu:
+          pred = self.model(x_batch).detach().cpu().numpy()
+        else:
+          pred = self.model(x_batch).detach().numpy()
+
+      preds.append(pred)
+
+    return pd.Series(np.concatenate(preds), index=index)
+
+
+# Real Model
+
+
+class Mlp(nn.Module):
+  def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+    super().__init__()
+    out_features = out_features or in_features
+    hidden_features = hidden_features or in_features
+    self.fc1 = nn.Linear(in_features, hidden_features)
+    self.act = act_layer()
+    self.fc2 = nn.Linear(hidden_features, out_features)
+    self.drop = nn.Dropout(drop)
+
+  def forward(self, x):
+    x = self.fc1(x)
+    x = self.act(x)
+    x = self.drop(x)
+    x = self.fc2(x)
+    x = self.drop(x)
+    return x
+
+
+class Attention(nn.Module):
+  def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
+    super().__init__()
+    self.num_heads = num_heads
+    head_dim = dim // num_heads
+    # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
+    self.scale = qk_scale or head_dim ** -0.5
+
+    self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+    self.attn_drop = nn.Dropout(attn_drop)
+    self.proj = nn.Linear(dim, dim)
+    self.proj_drop = nn.Dropout(proj_drop)
+
+  def forward(self, x):
+    B, N, C = x.shape
+    qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+    q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)
+
+    attn = (q @ k.transpose(-2, -1)) * self.scale
+    attn = attn.softmax(dim=-1)
+    attn = self.attn_drop(attn)
+
+    x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+    x = self.proj(x)
+    x = self.proj_drop(x)
+    return x
+
+
+class Block(nn.Module):
+
+  def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+         drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+    super().__init__()
+    self.norm1 = norm_layer(dim)
+    self.attn = Attention(
+      dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+    # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+    self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+    self.norm2 = norm_layer(dim)
+    mlp_hidden_dim = int(dim * mlp_ratio)
+    self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+  def forward(self, x):
+    x = x + self.drop_path(self.attn(self.norm1(x)))
+    x = x + self.drop_path(self.mlp(self.norm2(x)))
+    return x
+
+
+class SimpleEmbed(nn.Module):
+
+  def __init__(self, d_feat, embed_dim):
+    super(SimpleEmbed, self).__init__()
+    self.proj = nn.Linear(d_feat, embed_dim)
+
+  def forward(self, x):
+    import pdb; pdb.set_trace()
+    B, C, H, W = x.shape
+    # FIXME look at relaxing size constraints
+    assert H == self.img_size[0] and W == self.img_size[1], \
+      f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+    x = self.proj(x).flatten(2).transpose(1, 2)
+    return x
+
+
+class TransformerModel(nn.Module):
+  def __init__(self,
+         d_feat: int,
+         embed_dim: int = 64,
+         depth: int = 4,
+         num_heads: int = 4,
+         mlp_ratio: float = 4.,
+         qkv_bias: bool = True,
+         qk_scale: Optional[float] = None,
+         drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=None):
+    """
+    Args:
+      d_feat (int, tuple): input image size
+      embed_dim (int): embedding dimension
+      depth (int): depth of transformer
+      num_heads (int): number of attention heads
+      mlp_ratio (int): ratio of mlp hidden dim to embedding dim
+      qkv_bias (bool): enable bias for qkv if True
+      qk_scale (float): override default qk scale of head_dim ** -0.5 if set
+      drop_rate (float): dropout rate
+      attn_drop_rate (float): attention dropout rate
+      drop_path_rate (float): stochastic depth rate
+      norm_layer: (nn.Module): normalization layer
+    """
+    super(TransformerModel, self).__init__()
+    self.embed_dim = embed_dim
+    self.num_features = embed_dim
+    norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
+
+    self.input_embed = SimpleEmbed(d_feat, embed_dim=embed_dim)
+
+    """
+    self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+    self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+    self.pos_drop = nn.Dropout(p=drop_rate)
+    """
+
+    dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+    self.blocks = nn.ModuleList([
+      Block(
+        dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+        drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer)
+      for i in range(depth)])
+    self.norm = norm_layer(embed_dim)
+
+    # regression head
+    self.head = nn.Linear(self.num_features, 1)
+
+    """
+    trunc_normal_(self.pos_embed, std=.02)
+    trunc_normal_(self.cls_token, std=.02)
+    """
+    self.apply(self._init_weights)
+
+  def _init_weights(self, m):
+    if isinstance(m, nn.Linear):
+      trunc_normal_(m.weight, std=.02)
+      if isinstance(m, nn.Linear) and m.bias is not None:
+        nn.init.constant_(m.bias, 0)
+    elif isinstance(m, nn.LayerNorm):
+      nn.init.constant_(m.bias, 0)
+      nn.init.constant_(m.weight, 1.0)
+
+  def forward_features(self, x):
+    B = x.shape[0]
+    x = self.input_embed(x)
+
+    cls_tokens = self.cls_token.expand(B, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
+    x = torch.cat((cls_tokens, x), dim=1)
+    x = x + self.pos_embed
+    x = self.pos_drop(x)
+
+    for blk in self.blocks:
+      x = blk(x)
+
+    x = self.norm(x)[:, 0]
+    return x
+
+  def forward(self, x):
+    x = self.forward_features(x)
+    x = self.head(x)
+    return x