nagataka

# Clone the repo and replace these three files #

### training/rank_lstm.py ###
"""Rank_LSTM baseline, PyTorch port (Python 3.14 / PyTorch).

Ported from the original TensorFlow 1.x implementation of
"Temporal Relational Ranking for Stock Prediction" (Feng et al., 2019).

In addition to training the sequential LSTM ranking model, this script can
save the per-day sequential embeddings to ``<data>/pretrain/<name>.npy`` so that

nagataka

def construct_index(docs):
    term_dict = {}
    postings = {}
    for id, doc in docs.items():

        terms = list(set(doc.lower().split()))
        for term in terms:
            term_count = term_dict.get(term, 0)
            term_posting = postings.get(term, None)

nagataka

Object detection で得られた結果を利用する際に、「少し（例えば5%）実際のbboxよりも大きな領域をcropしてcontext情報を含めてやると良い」いうことを慣例としてやっていたのだけど、それってどこかで示されているのだっけ？それとも経験的にやっているだけなのだっけ？とふと疑問に思ったので調べてみた。これとかが参考になるか？ An Empirical Study of Context in Object Detection

AWS

aws login コマンドを初めて使ってみたが、これであればアクセスキーを作らなくていいので良さそうだなと思った。 https://dev.classmethod.jp/articles/aws-cli-aws-login/

SAM (Segment Anything Model)

nagataka

たまに色々と忘れるので便利な資料をメモ

• How to Read Figures, Mathematical Expressions and Equations, and Glossary
• 英語による数学表現

nagataka

• ブログなど
  • 日本語なら 今更聞けないLSTMの基本 がとても参考になる。
  • In English: Understanding LSTM Networks
• Implementation
  • Docs > torch.nn > LSTM
  • SEQUENCE MODELS AND LONG-SHORT TERM MEMORY NETWORKS

その他、未読だけれど参考になりそうなリンク (随時読んだら上へ)

• The Unreasonable Effectiveness of Recurrent Neural Networks

nagataka

# OpenAI gym custom environment mimicking Blocking Maze
# See Sutton and Barto "Reinforcement Learning an Introduction"
#     Example 8.2: Blocking Maze
from enum import Enum
import sys
import copy

import gym
from gym import error, spaces, utils
from gym.utils import seeding

nagataka

{
    "python.formatting.provider": "black",
    "python.linting.pylintEnabled": false,
    "python.linting.flake8Enabled": true,
    "python.linting.flake8Args": [
        "--ignore=E501,W503"
    ],
    "python.sortImports.args": [
        "-m 3"
    ],

nagataka

import random
import numpy as np
np.random.seed(0)

def kerri(p, b):
    """https://en.wikipedia.org/wiki/Kelly_criterion
    """
    return (p*(b+1)-1 )/b

N = 300

nagataka

import gym
import ray
from ray.rllib.agents.ppo import PPOTrainer, DEFAULT_CONFIG

import pprint as pp

#tune.run(PPOTrainer, config={"env": "Breakout-v0", "use_pytorch": True})
ray.init(num_gpus=1, ignore_reinit_error=True, log_to_driver=False)

# https://github.com/ray-project/ray/blob/master/rllib/agents/ppo/ppo.py#L15

nagataka

#!/bin/bash

set -eu

### Incoming WebHooks URL
WEBHOOKURL="https://hooks.slack.com/services/FILL_YOUR_WEBHOOKURL"

### channel
CHANNEL=${CHANNEL:-"#notifications"}