Skip to content

Instantly share code, notes, and snippets.

View wassname2's full-sized avatar
🤖

Michael.Clark2 wassname2

🤖
10 followers · 59 following
View GitHub Profile
@wassname2
wassname2 / upgrad_print_hook.py
Last active September 2, 2025 07:28
when using upgrad of any torchjd loss weight you can print the aggregation weight using a `with` function, and it's tidier than manually applying hooks
import torch
from torch import Tensor
from torch.nn.functional import cosine_similarity
from contextlib import AbstractContextManager
from typing import List, Dict
from functools import partial
class PrintGradWeights(AbstractContextManager):
"""Capture and print UPGrad weights + cosine-similarity during a forward pass.
@wassname2
wassname2 / caching_pattern_for_long_running.py
Last active July 8, 2024 01:34
caching_pattern_for_long_running.py
from pathlib import Path
import pandas as pd
from tqdm.auto import tqdm
experiments = {
'exp1': dict(lr=1e-4),
]
cache_dir = Path('../data/30_processed/results_cache1')
cache_dir.mkdir(exists_ok=True)
@wassname2
wassname2 / pd_util.py
Last active May 23, 2024 02:00
pandas utils
import pandas as pd
import numpy as np
def pandas_groupby_agg_mixed_dtype_df(df, groupby: list, agg='max', str_agg='first'):
"""
resample a df with numeric and strings
"""
numerics_cols = list(df.select_dtypes(include='number').columns)+groupby
other_cols = list(df.select_dtypes(exclude='number').columns)+groupby
@wassname2
wassname2 / time_chunking.py
Last active May 14, 2024 02:29
time_chunking.py a good way to chunk time, in human readable chunks of decreasing size
import pandas as pd
def chunking_by_time(startTime: pd.Timestamp, endTime: pd.Timestamp, freqs=['Y', 'M', 'D']):
"""
How do we cache timespans? We want to take year chunks, then for the remainder months, and so on.
That way old data is cached in big chunks, and new data is rechunked as needed
```py
startTime = pd.to_datetime('2023-01-02 01:01')