mirror of
https://github.com/wassname/seq2seq-time.git
synced 2026-08-18 12:30:31 +08:00
5.0 MiB
5.0 MiB
Sequence to Sequence Models for Timeseries Regression
In this notebook we are going to tackle a harder problem:
- predicting the future on a timeseries
- using an LSTM
- with rough uncertainty (uncalibrated)
- outputing sequence of predictions
In [1]:
# OPTIONAL: Load the "autoreload" extension so that code can change. But blacklist large modules
%load_ext autoreload
%autoreload 2
%aimport -pandas
%aimport -torch
%aimport -numpy
%aimport -matplotlib
%aimport -dask
%aimport -tqdm
%matplotlib inlineIn [2]:
# Imports
import torch
from torch import nn, optim
from torch.nn import functional as F
from torch.autograd import Variable
import torch
import torch.utils.data
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
plt.rcParams['figure.figsize'] = (12.0, 3.0)
plt.style.use('ggplot')
from pathlib import Path
from tqdm.auto import tqdm
import pytorch_lightning as plIn [3]:
import warnings
warnings.simplefilter('once')In [4]:
from seq2seq_time.data.dataset import Seq2SeqDataSet, Seq2SeqDataSets
from seq2seq_time.predict import predict, predict_multi/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel/ipkernel.py:287: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above. and should_run_async(code)
In [5]:
import logging, sys
# logging.basicConfig(stream=sys.stdout, level=logging.INFO)/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel/ipkernel.py:287: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above. and should_run_async(code)
In [6]:
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f'using {device}')
columns_target=['energy(kWh/hh)']
window_past = 48*2
window_future = 48*2
batch_size = 256
num_workers = 5
freq = '30T'
max_rows = 5e5using cuda
In [7]:
def get_smartmeter_df(indir=Path('../data/raw/smart-meters-in-london'), max_files=8):
"""
Data loading and cleanding is always messy, so understand this code is optional.
"""
# Load csv files
csv_files = sorted((indir/'halfhourly_dataset').glob('*.csv'))[:max_files]
dfs = []
for f in csv_files:
df = (pd.read_csv(f, parse_dates=[1], na_values=['Null'])
.groupby('tstp')
.sum()
.sort_index()
)
df['block'] = f.stem
# Drop nan and 0's
df = df[df['energy(kWh/hh)']!=0]
df = df.dropna()
# Add time features
time = df.index.to_series()
df["month"] = time.dt.month
df['day'] = time.dt.day
df['week'] = time.dt.week
df['hour'] = time.dt.hour
df['minute'] = time.dt.minute
df['dayofweek'] = time.dt.dayofweek
# Load weather data
df_weather = pd.read_csv(indir/'weather_hourly_darksky.csv', parse_dates=[3])
use_cols = ['visibility', 'windBearing', 'temperature', 'time', 'dewPoint',
'pressure', 'apparentTemperature', 'windSpeed',
'humidity']
df_weather = df_weather[use_cols].set_index('time')
# Resample to match energy data
# Use first, since we have bearing, and you can't take mean
df_weather = df_weather.resample(freq).first().ffill()
# Join weather and energy data
df = pd.merge(df, df_weather, how='inner', left_index=True, right_index=True, sort=True)
# Holidays
df_hols = pd.read_csv(indir/'uk_bank_holidays.csv', parse_dates=[0])
holidays = set(df_hols['Bank holidays'].dt.round('D'))
def is_holiday(dt):
return dt in holidays
days = df.index.floor('D')
holiday_mapping = days.unique().to_series().apply(is_holiday).astype(int).to_dict()
df['holiday'] = days.to_series().map(holiday_mapping).values
# sort
df.index.name = 'Date'
df = df.loc['2012-09':] # Weird value before this
dfs.append(df)
return pd.concat(dfs)In [ ]:
In [8]:
df = get_smartmeter_df(max_files=12)
# # Just get the first one for now
# dfs = list(dfs)
# # df = df.resample(freq).first().dropna() # Where empty we will backfill, this will respect causality, and mostly maintain the mean
df = df.tail(int(max_rows)).copy() # Just use last X rows
# df = pd.concat(dfs[:6], 0)
# # df = dfs[0]
print(df.block.value_counts())
dfOut [8]:
/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel_launcher.py:26: FutureWarning: Series.dt.weekofyear and Series.dt.week have been deprecated. Please use Series.dt.isocalendar().week instead. /home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel_launcher.py:26: FutureWarning: Series.dt.weekofyear and Series.dt.week have been deprecated. Please use Series.dt.isocalendar().week instead. /home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel_launcher.py:26: FutureWarning: Series.dt.weekofyear and Series.dt.week have been deprecated. Please use Series.dt.isocalendar().week instead. /home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel_launcher.py:26: FutureWarning: Series.dt.weekofyear and Series.dt.week have been deprecated. Please use Series.dt.isocalendar().week instead. /home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel_launcher.py:26: FutureWarning: Series.dt.weekofyear and Series.dt.week have been deprecated. Please use Series.dt.isocalendar().week instead. /home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel_launcher.py:26: FutureWarning: Series.dt.weekofyear and Series.dt.week have been deprecated. Please use Series.dt.isocalendar().week instead. /home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel_launcher.py:26: FutureWarning: Series.dt.weekofyear and Series.dt.week have been deprecated. Please use Series.dt.isocalendar().week instead. /home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel_launcher.py:26: FutureWarning: Series.dt.weekofyear and Series.dt.week have been deprecated. Please use Series.dt.isocalendar().week instead. /home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel_launcher.py:26: FutureWarning: Series.dt.weekofyear and Series.dt.week have been deprecated. Please use Series.dt.isocalendar().week instead. /home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel_launcher.py:26: FutureWarning: Series.dt.weekofyear and Series.dt.week have been deprecated. Please use Series.dt.isocalendar().week instead. /home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel_launcher.py:26: FutureWarning: Series.dt.weekofyear and Series.dt.week have been deprecated. Please use Series.dt.isocalendar().week instead. /home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel_launcher.py:26: FutureWarning: Series.dt.weekofyear and Series.dt.week have been deprecated. Please use Series.dt.isocalendar().week instead.
block_107 26161 block_100 26161 block_10 26161 block_1 26161 block_105 26161 block_0 26161 block_102 26161 block_103 26161 block_108 26161 block_106 26161 block_101 26161 block_104 26161 Name: block, dtype: int64
| energy(kWh/hh) | block | month | day | week | hour | minute | dayofweek | visibility | windBearing | temperature | dewPoint | pressure | apparentTemperature | windSpeed | humidity | holiday | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Date | |||||||||||||||||
| 2012-09-01 00:00:00 | 5.013 | block_0 | 9 | 1 | 35 | 0 | 0 | 5 | 13.36 | 302.0 | 14.08 | 9.74 | 1028.27 | 14.08 | 1.89 | 0.75 | 0 |
| 2012-09-01 00:30:00 | 5.157 | block_0 | 9 | 1 | 35 | 0 | 30 | 5 | 13.36 | 302.0 | 14.08 | 9.74 | 1028.27 | 14.08 | 1.89 | 0.75 | 0 |
| 2012-09-01 01:00:00 | 6.360 | block_0 | 9 | 1 | 35 | 1 | 0 | 5 | 13.50 | 298.0 | 13.93 | 9.81 | 1027.96 | 13.93 | 1.59 | 0.76 | 0 |
| 2012-09-01 01:30:00 | 5.511 | block_0 | 9 | 1 | 35 | 1 | 30 | 5 | 13.50 | 298.0 | 13.93 | 9.81 | 1027.96 | 13.93 | 1.59 | 0.76 | 0 |
| 2012-09-01 02:00:00 | 4.922 | block_0 | 9 | 1 | 35 | 2 | 0 | 5 | 13.21 | 274.0 | 13.52 | 9.94 | 1028.04 | 13.52 | 0.82 | 0.79 | 0 |
| ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... |
| 2014-02-27 22:00:00 | 9.819 | block_108 | 2 | 27 | 9 | 22 | 0 | 3 | 14.00 | 216.0 | 4.10 | 1.64 | 1005.67 | 1.41 | 3.02 | 0.84 | 0 |
| 2014-02-27 22:30:00 | 8.792 | block_108 | 2 | 27 | 9 | 22 | 30 | 3 | 14.00 | 216.0 | 4.10 | 1.64 | 1005.67 | 1.41 | 3.02 | 0.84 | 0 |
| 2014-02-27 23:00:00 | 8.087 | block_108 | 2 | 27 | 9 | 23 | 0 | 3 | 14.03 | 200.0 | 3.93 | 1.61 | 1004.62 | 1.42 | 2.75 | 0.85 | 0 |
| 2014-02-27 23:30:00 | 7.114 | block_108 | 2 | 27 | 9 | 23 | 30 | 3 | 14.03 | 200.0 | 3.93 | 1.61 | 1004.62 | 1.42 | 2.75 | 0.85 | 0 |
| 2014-02-28 00:00:00 | 7.287 | block_108 | 2 | 28 | 9 | 0 | 0 | 4 | 12.63 | 190.0 | 3.81 | 1.53 | 1003.57 | 1.47 | 2.53 | 0.85 | 0 |
313932 rows × 17 columns
In [ ]:
In [ ]:
In [ ]:
In [9]:
import holoviews as hv
from holoviews import opts
from holoviews.plotting.links import RangeToolLink
import datashader as ds
from holoviews.operation.datashader import datashade, shade, dynspread, rasterize
from holoviews.operation import decimate
hv.extension('bokeh')
# def house_curve(Name=None):
# if isinstance(Name, int):
# name = df.block.unique()[Name]
# d = df[df.block == Name]
# d_curve = hv.Curve(d, 'Date', 'energy(kWh/hh)', label=Name).opts(framewise=True)
# return d_curve
# dmap = hv.DynamicMap(house_curve, kdims=['Name'])
# dmap = dmap.redim.values(Name=list(df.block.unique()))
# dynspread(datashade(dmap).opts(width=800,
# height=300,
# tools=['xwheel_zoom', 'pan'],
# active_tools=['xwheel_zoom', 'pan'],
# default_tools=['reset', 'save', 'hover']
# ))/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel/ipkernel.py:287: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above. and should_run_async(code) /home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/holoviews/operation/datashader.py:5: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3,and in 3.9 it will stop working from collections import Callable
[JavaScript output - execution disabled for security]
[JavaScript output - execution disabled for security]
In [ ]:
In [10]:
# from pandas_profiling import ProfileReport
# profile = ProfileReport(df, title="Pandas Profiling Report", minimal=True)
# profile/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel/ipkernel.py:287: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above. and should_run_async(code)
In [11]:
df.describe()Out [11]:
| energy(kWh/hh) | month | day | week | hour | minute | dayofweek | visibility | windBearing | temperature | dewPoint | pressure | apparentTemperature | windSpeed | humidity | holiday | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| count | 313932.000000 | 313932.000000 | 313932.000000 | 313932.000000 | 313932.00000 | 313932.000000 | 313932.000000 | 313932.000000 | 313932.000000 | 313932.000000 | 313932.000000 | 313932.000000 | 313932.000000 | 313932.000000 | 313932.000000 | 313932.000000 |
| mean | 10.304546 | 6.878713 | 15.659187 | 27.962578 | 11.49956 | 14.999427 | 2.998203 | 11.203636 | 194.973128 | 10.196924 | 6.316654 | 1012.613350 | 8.845714 | 3.968334 | 0.784663 | 0.022018 |
| std | 6.425374 | 3.797205 | 8.768608 | 16.528737 | 6.92243 | 15.000024 | 2.001349 | 3.087941 | 91.343087 | 5.891891 | 5.118656 | 11.278742 | 7.116397 | 2.073511 | 0.140025 | 0.146741 |
| min | 1.634000 | 1.000000 | 1.000000 | 1.000000 | 0.00000 | 0.000000 | 0.000000 | 0.270000 | 0.000000 | -3.860000 | -8.920000 | 975.740000 | -8.880000 | 0.040000 | 0.230000 | 0.000000 |
| 25% | 6.121000 | 3.000000 | 8.000000 | 11.000000 | 5.00000 | 0.000000 | 1.000000 | 10.120000 | 119.000000 | 6.050000 | 2.560000 | 1005.900000 | 3.360000 | 2.450000 | 0.710000 | 0.000000 |
| 50% | 8.502000 | 8.000000 | 16.000000 | 31.000000 | 11.00000 | 0.000000 | 3.000000 | 12.230000 | 216.000000 | 9.580000 | 6.460000 | 1013.600000 | 7.990000 | 3.720000 | 0.810000 | 0.000000 |
| 75% | 12.126000 | 10.000000 | 23.000000 | 43.000000 | 17.00000 | 30.000000 | 5.000000 | 13.080000 | 255.000000 | 14.070000 | 10.060000 | 1020.550000 | 14.070000 | 5.160000 | 0.890000 | 0.000000 |
| max | 53.965000 | 12.000000 | 31.000000 | 52.000000 | 23.00000 | 30.000000 | 6.000000 | 16.090000 | 359.000000 | 32.400000 | 18.950000 | 1040.130000 | 32.420000 | 14.800000 | 1.000000 | 1.000000 |
In [12]:
import sklearn
from sklearn.preprocessing import StandardScaler, OrdinalEncoder
from sklearn_pandas import DataFrameMapper
columns_input_numeric = list(df.drop(columns=columns_target)._get_numeric_data().columns)
columns_categorical = list(set(df.columns)-set(columns_input_numeric)-set(columns_target))
output_scalers = [([n], StandardScaler()) for n in columns_target]
transformers=output_scalers + \
[([n], StandardScaler()) for n in columns_input_numeric] + \
[([n], OrdinalEncoder()) for n in columns_categorical]
scaler = DataFrameMapper(transformers, df_out=True)
df_norm = scaler.fit_transform(df)
df_normOut [12]:
/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel/ipkernel.py:287: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above. and should_run_async(code)
| energy(kWh/hh) | month | day | week | hour | minute | dayofweek | visibility | windBearing | temperature | dewPoint | pressure | apparentTemperature | windSpeed | humidity | holiday | block | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Date | |||||||||||||||||
| 2012-09-01 00:00:00 | -0.823540 | 0.558645 | -1.671783 | 0.425770 | -1.661205 | -0.999962 | 1.000225 | 0.698319 | 1.171704 | 0.659055 | 0.668799 | 1.388158 | 0.735526 | -1.002328 | -0.247549 | -0.150044 | 0.0 |
| 2012-09-01 00:30:00 | -0.801129 | 0.558645 | -1.671783 | 0.425770 | -1.661205 | 1.000038 | 1.000225 | 0.698319 | 1.171704 | 0.659055 | 0.668799 | 1.388158 | 0.735526 | -1.002328 | -0.247549 | -0.150044 | 0.0 |
| 2012-09-01 01:00:00 | -0.613902 | 0.558645 | -1.671783 | 0.425770 | -1.516747 | -0.999962 | 1.000225 | 0.743657 | 1.127913 | 0.633597 | 0.682474 | 1.360673 | 0.714448 | -1.147010 | -0.176133 | -0.150044 | 0.0 |
| 2012-09-01 01:30:00 | -0.746035 | 0.558645 | -1.671783 | 0.425770 | -1.516747 | 1.000038 | 1.000225 | 0.743657 | 1.127913 | 0.633597 | 0.682474 | 1.360673 | 0.714448 | -1.147010 | -0.176133 | -0.150044 | 0.0 |
| 2012-09-01 02:00:00 | -0.837703 | 0.558645 | -1.671783 | 0.425770 | -1.372289 | -0.999962 | 1.000225 | 0.649743 | 0.865167 | 0.564009 | 0.707872 | 1.367766 | 0.656834 | -1.518362 | 0.038114 | -0.150044 | 0.0 |
| ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... |
| 2014-02-27 22:00:00 | -0.075567 | -1.284819 | 1.293344 | -1.147251 | 1.516874 | -0.999962 | 0.000898 | 0.905577 | 0.230197 | -1.034801 | -0.913650 | -0.615615 | -1.044872 | -0.457357 | 0.395193 | -0.150044 | 11.0 |
| 2014-02-27 22:30:00 | -0.235402 | -1.284819 | 1.293344 | -1.147251 | 1.516874 | 1.000038 | 0.000898 | 0.905577 | 0.230197 | -1.034801 | -0.913650 | -0.615615 | -1.044872 | -0.457357 | 0.395193 | -0.150044 | 11.0 |
| 2014-02-27 23:00:00 | -0.345124 | -1.284819 | 1.293344 | -1.147251 | 1.661332 | -0.999962 | 0.000898 | 0.915292 | 0.055033 | -1.063654 | -0.919511 | -0.708710 | -1.043467 | -0.587572 | 0.466609 | -0.150044 | 11.0 |
| 2014-02-27 23:30:00 | -0.496555 | -1.284819 | 1.293344 | -1.147251 | 1.661332 | 1.000038 | 0.000898 | 0.915292 | 0.055033 | -1.063654 | -0.919511 | -0.708710 | -1.043467 | -0.587572 | 0.466609 | -0.150044 | 11.0 |
| 2014-02-28 00:00:00 | -0.469630 | -1.284819 | 1.407388 | -1.147251 | -1.661205 | -0.999962 | 0.500561 | 0.461915 | -0.054445 | -1.084021 | -0.935140 | -0.801806 | -1.036441 | -0.693672 | 0.466609 | -0.150044 | 11.0 |
313932 rows × 17 columns
In [13]:
output_scaler = next(filter(lambda r:r[0][0] in columns_target, scaler.features))[-1]
output_scalerOut [13]:
/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel/ipkernel.py:287: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above. and should_run_async(code)
StandardScaler()
In [14]:
# split data, with the test in the future
d0 =df_norm.index.min()
d1 = df_norm.index.max()
split_time = d0+(d1-d0)*0.8
split_time = split_time.round('1D')
print(split_time)
df_train = df_norm.groupby('block').apply(lambda d:d.loc[:split_time]).reset_index(level=0, drop=True)
df_test = df_norm.groupby('block').apply(lambda d:d.loc[split_time:]).reset_index(level=0, drop=True)
# df_test2013-11-11 00:00:00
In [15]:
# # Show split
# df_train['energy(kWh/hh)'].plot(label='train')
# df_test['energy(kWh/hh)'].plot(label='test')
# plt.ylabel('energy(kWh/hh)')
# plt.legend()In [16]:
# # Show split
scatter = dynspread(datashade(hv.Curve(df_train, kdims=['Date'], vdims=['energy(kWh/hh)', 'block']).groupby('block'), cmap='blue'))
scatter *= dynspread(datashade(hv.Curve(df_test, kdims=['Date'], vdims=['energy(kWh/hh)', 'block']).groupby('block'), cmap='red'))
scatter = scatter.opts(plot=dict(width=800))
scatterOut [16]:
/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/bokeh/core/property/bases.py:241: DeprecationWarning: elementwise comparison failed; this will raise an error in the future. return new == old
In [17]:
# ### Dataset
# These are the columns that we wont know in the future
# We need to blank them out in x_future
columns_blank=['visibility',
'windBearing', 'temperature', 'dewPoint', 'pressure',
'apparentTemperature', 'windSpeed', 'humidity']
df_trains = [d.resample(freq).first().ffill().dropna() for _,d in df_train.groupby('block')]
df_tests = [d.resample(freq).first().ffill().dropna() for _,d in df_test.groupby('block')]
ds_train = Seq2SeqDataSets(df_trains,
window_past=window_past,
window_future=window_future,
columns_blank=columns_blank)
ds_test = Seq2SeqDataSets(df_tests,
window_past=window_past,
window_future=window_future,
columns_blank=columns_blank)
print(ds_train)
print(ds_test)/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel/ipkernel.py:287: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above. and should_run_async(code)
<Seq2SeqDataSets([<Seq2SeqDataSet(shape=(20929, 17), times=2012-09-01 00:00:00 to 2012-09-01 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(20929, 17), times=2012-09-01 00:00:00 to 2012-09-01 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(20929, 17), times=2012-09-01 00:00:00 to 2012-09-01 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(20929, 17), times=2012-09-01 00:00:00 to 2012-09-01 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(20929, 17), times=2012-09-01 00:00:00 to 2012-09-01 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(20929, 17), times=2012-09-01 00:00:00 to 2012-09-01 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(20929, 17), times=2012-09-01 00:00:00 to 2012-09-01 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(20929, 17), times=2012-09-01 00:00:00 to 2012-09-01 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(20929, 17), times=2012-09-01 00:00:00 to 2012-09-01 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(20929, 17), times=2012-09-01 00:00:00 to 2012-09-01 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(20929, 17), times=2012-09-01 00:00:00 to 2012-09-01 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(20929, 17), times=2012-09-01 00:00:00 to 2012-09-01 00:30:00 at 30T)>])> <Seq2SeqDataSets([<Seq2SeqDataSet(shape=(5233, 17), times=2013-11-11 00:00:00 to 2013-11-11 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(5233, 17), times=2013-11-11 00:00:00 to 2013-11-11 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(5233, 17), times=2013-11-11 00:00:00 to 2013-11-11 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(5233, 17), times=2013-11-11 00:00:00 to 2013-11-11 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(5233, 17), times=2013-11-11 00:00:00 to 2013-11-11 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(5233, 17), times=2013-11-11 00:00:00 to 2013-11-11 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(5233, 17), times=2013-11-11 00:00:00 to 2013-11-11 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(5233, 17), times=2013-11-11 00:00:00 to 2013-11-11 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(5233, 17), times=2013-11-11 00:00:00 to 2013-11-11 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(5233, 17), times=2013-11-11 00:00:00 to 2013-11-11 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(5233, 17), times=2013-11-11 00:00:00 to 2013-11-11 00:30:00 at 30T)>, <Seq2SeqDataSet(shape=(5233, 17), times=2013-11-11 00:00:00 to 2013-11-11 00:30:00 at 30T)>])>
In [18]:
# we can treat it like an array
ds_train[0]
len(ds_train)
ds_train[-1]Out [18]:
/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel/ipkernel.py:287: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above. and should_run_async(code)
[array([[ 1.0853493 , -0.9875229 , 1.0307775 , ..., 0. ,
-2. , 1. ],
[ 1.0853493 , -0.9875229 , 1.0307775 , ..., 0. ,
-1.9791666 , 1. ],
[ 1.0853493 , -0.9875229 , 1.0307775 , ..., 0. ,
-1.9583334 , 1. ],
...,
[ 1.0853493 , -0.87347955, 1.0307775 , ..., 0. ,
-0.0625 , 1. ],
[ 1.0853493 , -0.75943613, 1.0307775 , ..., 0. ,
-0.04166667, 1. ],
[ 1.0853493 , -0.75943613, 1.0307775 , ..., 0. ,
-0.02083333, 1. ]], dtype=float32),
array([[ 0.21577835],
[ 0.15010113],
[ 0.13095824],
[ 0.02481639],
[ 0.02030303],
[-0.0716762 ],
[ 0.15803841],
[ 0.02917414],
[-0.0674741 ],
[ 0.20052628],
[ 0.19009887],
[ 0.6498394 ],
[ 1.2491828 ],
[ 1.628461 ],
[ 1.5388163 ],
[ 1.9378599 ],
[ 1.5101798 ],
[ 1.8061942 ],
[ 1.569943 ],
[ 1.4147766 ],
[ 1.5297896 ],
[ 0.9147271 ],
[ 0.5668869 ],
[ 0.6613563 ],
[ 0.8767526 ],
[ 1.2315964 ],
[ 1.4297174 ],
[ 1.279687 ],
[ 1.4004583 ],
[ 1.1357263 ],
[ 0.991143 ],
[ 1.7546796 ],
[ 1.8547517 ],
[ 2.5337794 ],
[ 2.6925254 ],
[ 3.1499314 ],
[ 3.0293155 ],
[ 3.4837644 ],
[ 3.7531657 ],
[ 3.8403203 ],
[ 3.7007172 ],
[ 2.7031083 ],
[ 3.0036361 ],
[ 2.3289661 ],
[ 1.6462032 ],
[ 1.4122865 ],
[ 1.1221862 ],
[ 0.47366259],
[ 0.43179724],
[ 0.1378061 ],
[ 0.10450058],
[ 0.05890007],
[-0.01922781],
[ 0.01361078],
[ 0.25001764],
[ 0.20379458],
[ 0.15912783],
[ 0.03804521],
[ 0.0112763 ],
[ 0.40020368],
[ 0.9962789 ],
[ 1.5427071 ],
[ 1.4130647 ],
[ 1.1237426 ],
[ 1.6356201 ],
[ 2.0203454 ],
[ 1.6363983 ],
[ 1.2048274 ],
[ 1.2031155 ],
[ 1.6057385 ],
[ 1.4756292 ],
[ 0.93324745],
[ 1.3648183 ],
[ 2.0363758 ],
[ 1.6476039 ],
[ 2.2405665 ],
[ 2.4416447 ],
[ 2.1408055 ],
[ 1.6493158 ],
[ 2.2284272 ],
[ 2.4018025 ],
[ 2.8567183 ],
[ 2.898895 ],
[ 3.9316769 ],
[ 3.660097 ],
[ 3.2076712 ],
[ 3.409839 ],
[ 3.2286818 ],
[ 3.0336733 ],
[ 2.8353965 ],
[ 2.9273758 ],
[ 2.1327126 ],
[ 1.6311067 ],
[ 1.6121196 ],
[ 0.8972962 ],
[ 0.5648636 ]], dtype=float32),
array([[ 1.0853493 , -0.75943613, 1.0307775 , ..., 0. ,
0. , 0. ],
[ 1.0853493 , -0.75943613, 1.0307775 , ..., 0. ,
0.02083333, 0. ],
[ 1.0853493 , -0.75943613, 1.0307775 , ..., 0. ,
0.04166667, 0. ],
...,
[ 1.0853493 , -0.6453928 , 1.0307775 , ..., 0. ,
1.9166666 , 0. ],
[ 1.0853493 , -0.6453928 , 1.0307775 , ..., 0. ,
1.9375 , 0. ],
[ 1.0853493 , -0.5313494 , 1.0912782 , ..., 0. ,
1.9583334 , 0. ]], dtype=float32),
array([[ 1.9492349e-01],
[ 7.5397186e-02],
[ 2.3336489e-01],
[ 1.2053080e-01],
[ 2.2355998e-01],
[ 2.9982027e-01],
[ 1.5866096e-01],
[ 1.1866322e-01],
[ 1.0512313e-01],
[ 3.0122095e-01],
[ 2.1468890e-01],
[ 5.3949541e-01],
[ 8.2181406e-01],
[ 1.5501775e+00],
[ 1.8349863e+00],
[ 1.6644123e+00],
[ 2.1366036e+00],
[ 2.1674187e+00],
[ 1.9316344e+00],
[ 1.8812094e+00],
[ 2.2492819e+00],
[ 2.3697419e+00],
[ 1.8332744e+00],
[ 1.7370930e+00],
[ 2.1202619e+00],
[ 2.6394544e+00],
[ 2.0871119e+00],
[ 1.7694647e+00],
[ 2.0368426e+00],
[ 2.1297555e+00],
[ 2.0472701e+00],
[ 3.0395873e+00],
[ 3.2934251e+00],
[ 3.4733372e+00],
[ 3.8921461e+00],
[ 3.5867937e+00],
[ 3.4696019e+00],
[ 3.4269586e+00],
[ 2.8092501e+00],
[ 2.9440286e+00],
[ 2.4144087e+00],
[ 2.0331073e+00],
[ 2.2208011e+00],
[ 1.8822988e+00],
[ 1.5588930e+00],
[ 1.4409230e+00],
[ 1.1388389e+00],
[ 1.0885694e+00],
[ 5.5381370e-01],
[ 3.4310017e-02],
[ 1.3157757e-03],
[-1.2521403e-01],
[-7.3855065e-02],
[-1.2101193e-01],
[ 3.9912798e-02],
[ 1.2270970e-01],
[ 4.5204341e-02],
[-1.2677038e-01],
[-1.6115144e-02],
[ 2.1842413e-01],
[ 5.3249198e-01],
[ 4.2136979e-01],
[ 7.1224833e-01],
[ 1.4854342e+00],
[ 1.8214462e+00],
[ 1.6490046e+00],
[ 2.0604987e+00],
[ 2.0366869e+00],
[ 1.6323519e+00],
[ 1.3979683e+00],
[ 1.3878522e+00],
[ 1.4852785e+00],
[ 1.5033319e+00],
[ 1.9745893e+00],
[ 2.0606544e+00],
[ 1.8254926e+00],
[ 1.8941269e+00],
[ 2.4310615e+00],
[ 2.7108901e+00],
[ 2.6917472e+00],
[ 2.9974108e+00],
[ 3.7825804e+00],
[ 3.2772393e+00],
[ 3.5678065e+00],
[ 3.8865433e+00],
[ 3.7761993e+00],
[ 3.8535490e+00],
[ 3.9933076e+00],
[ 3.0651112e+00],
[ 2.7614708e+00],
[ 2.6290269e+00],
[ 2.4046037e+00],
[ 1.4166442e+00],
[ 1.4624003e+00],
[ 9.4772130e-01]], dtype=float32)]In [19]:
# We can get rows
x_past, y_past, x_future, y_future = ds_train.get_rows(10)
# Plot one instance, this is what the model sees
y_past['energy(kWh/hh)'].plot(label='past')
y_future['energy(kWh/hh)'].plot(ax=plt.gca(), label='future')
plt.legend()
plt.ylabel('energy(kWh/hh)')
# Notice we've added on two new columns tsp (time since present) and is_past
x_past.tail()Out [19]:
| month | day | week | hour | minute | dayofweek | visibility | windBearing | temperature | dewPoint | pressure | apparentTemperature | windSpeed | humidity | holiday | block | tsp_days | is_past | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Date | ||||||||||||||||||
| 2013-11-08 17:00:00 | 1.085349 | -0.87348 | 1.030777 | 0.794583 | -0.999962 | 0.500561 | -0.205197 | 0.700950 | -0.350809 | 0.160852 | -0.928594 | -0.347889 | -0.539344 | 1.037935 | -0.150044 | 0.0 | -0.104167 | 1.0 |
| 2013-11-08 17:30:00 | 1.085349 | -0.87348 | 1.030777 | 0.794583 | 1.000038 | 0.500561 | -0.205197 | 0.700950 | -0.350809 | 0.160852 | -0.928594 | -0.347889 | -0.539344 | 1.037935 | -0.150044 | 0.0 | -0.083333 | 1.0 |
| 2013-11-08 18:00:00 | 1.085349 | -0.87348 | 1.030777 | 0.939042 | -0.999962 | 0.500561 | 0.176935 | 0.919905 | -0.349112 | 0.139362 | -0.896675 | -0.442038 | 0.034563 | 1.037935 | -0.150044 | 0.0 | -0.062500 | 1.0 |
| 2013-11-08 18:30:00 | 1.085349 | -0.87348 | 1.030777 | 0.939042 | 1.000038 | 0.500561 | 0.176935 | 0.919905 | -0.349112 | 0.139362 | -0.896675 | -0.442038 | 0.034563 | 1.037935 | -0.150044 | 0.0 | -0.041667 | 1.0 |
| 2013-11-08 19:00:00 | 1.085349 | -0.87348 | 1.030777 | 1.083500 | -0.999962 | 0.500561 | 0.118644 | 0.952749 | -0.391543 | -0.048187 | -0.821312 | -0.517919 | 0.270877 | 0.680856 | -0.150044 | 0.0 | -0.020833 | 1.0 |
In [20]:
# Notice we've hidden some future columns to prevent cheating
x_future.tail()Out [20]:
/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel/ipkernel.py:287: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above. and should_run_async(code)
| month | day | week | hour | minute | dayofweek | visibility | windBearing | temperature | dewPoint | pressure | apparentTemperature | windSpeed | humidity | holiday | block | tsp_days | is_past | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Date | ||||||||||||||||||
| 2013-11-10 17:00:00 | 1.085349 | -0.645393 | 1.030777 | 0.794583 | -0.999962 | 1.499889 | -0.328257 | 0.36157 | 0.77956 | 1.225586 | -1.236252 | 0.835296 | 2.180685 | 0.609441 | -0.150044 | 0.0 | 1.895833 | 0.0 |
| 2013-11-10 17:30:00 | 1.085349 | -0.645393 | 1.030777 | 0.794583 | 1.000038 | 1.499889 | -0.328257 | 0.36157 | 0.77956 | 1.225586 | -1.236252 | 0.835296 | 2.180685 | 0.609441 | -0.150044 | 0.0 | 1.916667 | 0.0 |
| 2013-11-10 18:00:00 | 1.085349 | -0.645393 | 1.030777 | 0.939042 | -0.999962 | 1.499889 | -0.328257 | 0.36157 | 0.77956 | 1.225586 | -1.236252 | 0.835296 | 2.180685 | 0.609441 | -0.150044 | 0.0 | 1.937500 | 0.0 |
| 2013-11-10 18:30:00 | 1.085349 | -0.645393 | 1.030777 | 0.939042 | 1.000038 | 1.499889 | -0.328257 | 0.36157 | 0.77956 | 1.225586 | -1.236252 | 0.835296 | 2.180685 | 0.609441 | -0.150044 | 0.0 | 1.958333 | 0.0 |
| 2013-11-10 19:00:00 | 1.085349 | -0.645393 | 1.030777 | 1.083500 | -0.999962 | 1.499889 | -0.328257 | 0.36157 | 0.77956 | 1.225586 | -1.236252 | 0.835296 | 2.180685 | 0.609441 | -0.150044 | 0.0 | 1.979167 | 0.0 |
In [21]:
def plot_prediction(ds_preds, i):
"""Plot a prediction into the future, at a single point in time."""
d = ds_preds.isel(t_source=i)
# Get arrays
xf = d.t_target
yp = d.y_pred
s = d.y_pred_std
yt = d.y_true
now = d.t_source.squeeze()
plt.figure(figsize=(12, 4))
plt.scatter(xf, yt, label='true', c='k', s=6)
ylim = plt.ylim()
# plot prediction
plt.fill_between(xf, yp-2*s, yp+2*s, alpha=0.25,
facecolor="b",
interpolate=True,
label="2 std",)
plt.plot(xf, yp, label='pred', c='b')
# plot true
plt.scatter(
d.t_past,
d.y_past,
c='k',
s=6
)
# plot a red line for now
plt.vlines(x=now, ymin=0, ymax=1, label='now', color='r')
plt.ylim(*ylim)
now=pd.Timestamp(now.values)
plt.title(f'Prediction NLL={d.nll.mean().item():2.2g}')
plt.xlabel(f'{now.date()}')
plt.ylabel('energy(kWh/hh)')
plt.legend()
plt.xticks(rotation=45)
plt.show()
def plot_performance(ds_preds, full=False):
"""Multiple plots using xr_preds"""
plot_prediction(ds_preds, 24)
ds_preds.mean('t_source').plot.scatter('t_ahead_hours', 'nll') # Mean over all predictions
n = len(ds_preds.t_source)
plt.ylabel('Negative Log Likelihood (lower is better)')
plt.xlabel('Hours ahead')
plt.title(f'NLL vs time ahead (no. samples={n})')
plt.show()
# Make a plot of the NLL over time. Does this solution get worse with time?
if full:
d = ds_preds.mean('t_ahead').groupby('t_source').mean().plot.scatter('t_source', 'nll')
plt.xticks(rotation=45)
plt.title('NLL over source time (lower is better)')
plt.show()
# A scatter plot is easy with xarray
if full:
plt.figure(figsize=(5, 5))
ds_preds.plot.scatter('y_true', 'y_pred', s=.01)
plt.show()
/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel/ipkernel.py:287: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above. and should_run_async(code)
In [ ]:
In [22]:
def plot_hist(trainer):
try:
df_hist = pd.read_csv(trainer.logger.experiment.metrics_file_path)
df_hist['epoch'] = df_hist['epoch'].ffill()
df_histe = df_hist.set_index('epoch').groupby('epoch').mean()
if len(df_histe)>1:
df_histe[['loss/train', 'loss/val']].plot(title='history')
return df_histe
except Exception:
passIn [23]:
import pytorch_lightning as pl
class PL_MODEL(pl.LightningModule):
def __init__(self, model, lr=3e-4, patience=2):
super().__init__()
self._model = model
self.lr = lr
self.patience = patience
def forward(self, x_past, y_past, x_future, y_future=None):
"""Eval/Predict"""
y_dist, extra = self._model(x_past, y_past, x_future, y_future)
return y_dist, extra
def training_step(self, batch, batch_idx, phase='train'):
x_past, y_past, x_future, y_future = batch
y_dist, extra = self.forward(*batch)
loss = -y_dist.log_prob(y_future).mean()
self.log_dict({f'loss/{phase}':loss})
if ('loss' in extra) and (phase=='train'):
# some models have a special loss
loss = extra['loss']
self.log_dict({f'model_loss/{phase}':loss})
return loss
def validation_step(self, batch, batch_idx):
return self.training_step(batch, batch_idx, phase='val')
def configure_optimizers(self):
optim = torch.optim.Adam(self.parameters(), lr=self.lr)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optim,
patience=self.patience,
verbose=True,
min_lr=1e-7,
)
return {'optimizer': optim, 'lr_scheduler': scheduler, 'monitor': 'loss/val'}In [24]:
# # Run
from torch.utils.data import DataLoader
from pytorch_lightning.loggers import CSVLogger
from pytorch_lightning.callbacks.early_stopping import EarlyStoppingIn [25]:
# Init data
x_past, y_past, x_future, y_future = ds_train.get_rows(10)
input_size = x_past.shape[-1]
output_size = y_future.shape[-1]
dl_train = DataLoader(ds_train,
batch_size=batch_size,
shuffle=True,
pin_memory=num_workers==0,
num_workers=num_workers)
dl_test = DataLoader(ds_test, batch_size=batch_size, num_workers=num_workers)In [30]:
from seq2seq_time.models.lstm_seq2seq import LSTMSeq2Seq
from seq2seq_time.models.lstm_seq import LSTMSeq
from seq2seq_time.models.lstm import LSTM
from seq2seq_time.models.baseline import BaselineLast
from seq2seq_time.models.transformer import Transformer
from seq2seq_time.models.transformer_seq2seq import TransformerSeq2Seq
from seq2seq_time.models.transformer_seq import TransformerSeq
from seq2seq_time.models.neural_process import RANP
# ## Plots/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel/ipkernel.py:287: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above. and should_run_async(code)
In [27]:
models = [
RANP(input_size,
output_size),
LSTM(input_size,
output_size,
hidden_size=80,
lstm_layers=3,
lstm_dropout=0.3),
LSTMSeq2Seq(input_size,
output_size,
hidden_size=64,
lstm_layers=2,
lstm_dropout=0.25),
TransformerSeq2Seq(input_size,
output_size,
hidden_size=64,
nhead=8,
nlayers=4,
attention_dropout=0.3),
Transformer(input_size,
output_size,
attention_dropout=0.3,
nhead=8,
nlayers=6,
hidden_size=64),
TransformerSeq(input_size,
output_size),
LSTMSeq(input_size,
output_size),
]In [28]:
# Baseline model
pt_model = BaselineLast()
model = PL_MODEL(pt_model).to(device)
trainer = pl.Trainer(gpus=1,
max_epochs=1,
limit_train_batches=0.01,
logger=CSVLogger("logs",
name=type(pt_model).__name__),
)
trainer.fit(model, dl_train, dl_test)
print(plot_hist(trainer))
ds_predss = predict_multi(model.to(device),
ds_test.datasets,
batch_size*8,
device=device,
scaler=output_scaler)
print(f'baseline nll: {ds_preds.nll.mean().item():2.2g}')GPU available: True, used: True TPU available: False, using: 0 TPU cores LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0] | Name | Type | Params ---------------------------------------- 0 | _model | BaselineLast | 1
HBox(children=(HTML(value='Validation sanity check'), FloatProgress(value=1.0, bar_style='info', layout=Layout…
HBox(children=(HTML(value='Training'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), max…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
None
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=20.0), HTML(value='')))
baseline nll: 2.3
In [29]:
for pt_model in models:
name = type(pt_model).__name__
print(name)
# Wrap in lightning
patience = 2
model = PL_MODEL(pt_model, patience=patience, lr=3e-4).to(device)
# Trainer
trainer = pl.Trainer(gpus=1,
min_epochs=2,
max_epochs=10,
amp_level='O1',
precision=16,
gradient_clip_val=1,
logger=CSVLogger("logs",
name=type(pt_model).__name__),
callbacks=[
EarlyStopping(monitor='loss/val', patience=patience*2),
# PrintTableMetricsCallback2()
],
)
# Train
trainer.fit(model, dl_train, dl_test)
ds_predss = predict_multi(model.to(device),
ds_test.datasets,
batch_size*8,
device=device,
scaler=output_scaler)
print(name)
print(f'mean_NLL {ds_predss.nll.mean().item():2.2f}')
# Performance
ds_preds = ds_predss.isel(block=0)
print(plot_hist(trainer))
plot_performance(ds_preds)/home/wassname/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/ipykernel/ipkernel.py:287: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above. and should_run_async(code) GPU available: True, used: True TPU available: False, using: 0 TPU cores LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0] Using native 16bit precision. | Name | Type | Params -------------------------------- 0 | _model | RANP | 58 K
RANP
HBox(children=(HTML(value='Validation sanity check'), FloatProgress(value=1.0, bar_style='info', layout=Layout…
HBox(children=(HTML(value='Training'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), max…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
Epoch 8: reducing learning rate of group 0 to 3.0000e-05.
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value=''), FloatProgress(value=0.0, max=12.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
RANP
mean_NLL -0.03
loss/train model_loss/train step loss/val
epoch
0.0 0.156459 0.156764 522.650000 0.053744
1.0 -0.129233 -0.128922 1473.800000 -0.003657
2.0 -0.164016 -0.163826 2447.523810 -0.012628
3.0 -0.174735 -0.174592 3423.600000 -0.031114
4.0 -0.197001 -0.196890 4397.333333 -0.043655
5.0 -0.207953 -0.207857 5373.400000 -0.041250
6.0 -0.225731 -0.225642 6347.142857 -0.038535
7.0 -0.233150 -0.233063 7323.200000 -0.024135
8.0 -0.247012 -0.246924 8296.952381 -0.033753
mean_NLL 0.55
GPU available: True, used: True TPU available: False, using: 0 TPU cores LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0] Using native 16bit precision. | Name | Type | Params -------------------------------- 0 | _model | LSTM | 136 K
LSTM
HBox(children=(HTML(value='Validation sanity check'), FloatProgress(value=1.0, bar_style='info', layout=Layout…
HBox(children=(HTML(value='Training'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), max…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
Epoch 6: reducing learning rate of group 0 to 3.0000e-05.
HBox(children=(HTML(value='Validating'), FloatProgress(value=1.0, bar_style='info', layout=Layout(flex='2'), m…
HBox(children=(HTML(value=''), FloatProgress(value=0.0, max=12.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
HBox(children=(HTML(value='predict'), FloatProgress(value=0.0, max=3.0), HTML(value='')))
LSTM
mean_NLL 0.35
loss/train step loss/val
epoch
0.0 0.225341 522.650000 0.327770
1.0 -0.098921 1473.800000 0.319602
2.0 -0.172654 2447.523810 0.228442
3.0 -0.213030 3423.600000 0.382066
4.0 -0.225637 4397.333333 0.311360
5.0 -0.257630 5373.400000 0.379025
6.0 -0.264004 6347.142857 0.347051
mean_NLL 1.38
GPU available: True, used: True TPU available: False, using: 0 TPU cores LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0] Using native 16bit precision. | Name | Type | Params --------------------------------------- 0 | _model | LSTMSeq2Seq | 109 K
LSTMSeq2Seq
HBox(children=(HTML(value='Validation sanity check'), FloatProgress(value=1.0, bar_style='info', layout=Layout…
[0;31m---------------------------------------------------------------------------[0m
[0;31mTypeError[0m Traceback (most recent call last)
[0;32m<ipython-input-29-74f32e8df344>[0m in [0;36m<module>[0;34m[0m
[1;32m 23[0m [0;34m[0m[0m
[1;32m 24[0m [0;31m# Train[0m[0;34m[0m[0;34m[0m[0;34m[0m[0m
[0;32m---> 25[0;31m [0mtrainer[0m[0;34m.[0m[0mfit[0m[0;34m([0m[0mmodel[0m[0;34m,[0m [0mdl_train[0m[0;34m,[0m [0mdl_test[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m 26[0m [0;34m[0m[0m
[1;32m 27[0m [0;34m[0m[0m
[0;32m~/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py[0m in [0;36mfit[0;34m(self, model, train_dataloader, val_dataloaders, datamodule)[0m
[1;32m 438[0m [0mself[0m[0;34m.[0m[0mcall_hook[0m[0;34m([0m[0;34m'on_fit_start'[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[1;32m 439[0m [0;34m[0m[0m
[0;32m--> 440[0;31m [0mresults[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0maccelerator_backend[0m[0;34m.[0m[0mtrain[0m[0;34m([0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m 441[0m [0mself[0m[0;34m.[0m[0maccelerator_backend[0m[0;34m.[0m[0mteardown[0m[0;34m([0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[1;32m 442[0m [0;34m[0m[0m
[0;32m~/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py[0m in [0;36mtrain[0;34m(self)[0m
[1;32m 52[0m [0;34m[0m[0m
[1;32m 53[0m [0;31m# train or test[0m[0;34m[0m[0;34m[0m[0;34m[0m[0m
[0;32m---> 54[0;31m [0mresults[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0mtrain_or_test[0m[0;34m([0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m 55[0m [0;32mreturn[0m [0mresults[0m[0;34m[0m[0;34m[0m[0m
[1;32m 56[0m [0;34m[0m[0m
[0;32m~/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/pytorch_lightning/accelerators/accelerator.py[0m in [0;36mtrain_or_test[0;34m(self)[0m
[1;32m 64[0m [0mresults[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0mtrainer[0m[0;34m.[0m[0mrun_test[0m[0;34m([0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[1;32m 65[0m [0;32melse[0m[0;34m:[0m[0;34m[0m[0;34m[0m[0m
[0;32m---> 66[0;31m [0mresults[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0mtrainer[0m[0;34m.[0m[0mtrain[0m[0;34m([0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m 67[0m [0;32mreturn[0m [0mresults[0m[0;34m[0m[0;34m[0m[0m
[1;32m 68[0m [0;34m[0m[0m
[0;32m~/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py[0m in [0;36mtrain[0;34m(self)[0m
[1;32m 460[0m [0;34m[0m[0m
[1;32m 461[0m [0;32mdef[0m [0mtrain[0m[0;34m([0m[0mself[0m[0;34m)[0m[0;34m:[0m[0;34m[0m[0;34m[0m[0m
[0;32m--> 462[0;31m [0mself[0m[0;34m.[0m[0mrun_sanity_check[0m[0;34m([0m[0mself[0m[0;34m.[0m[0mget_model[0m[0;34m([0m[0;34m)[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m 463[0m [0;34m[0m[0m
[1;32m 464[0m [0;31m# enable train mode[0m[0;34m[0m[0;34m[0m[0;34m[0m[0m
[0;32m~/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py[0m in [0;36mrun_sanity_check[0;34m(self, ref_model)[0m
[1;32m 646[0m [0;34m[0m[0m
[1;32m 647[0m [0;31m# run eval step[0m[0;34m[0m[0;34m[0m[0;34m[0m[0m
[0;32m--> 648[0;31m [0m_[0m[0;34m,[0m [0meval_results[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0mrun_evaluation[0m[0;34m([0m[0mtest_mode[0m[0;34m=[0m[0;32mFalse[0m[0;34m,[0m [0mmax_batches[0m[0;34m=[0m[0mself[0m[0;34m.[0m[0mnum_sanity_val_batches[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m 649[0m [0;34m[0m[0m
[1;32m 650[0m [0;31m# allow no returns from eval[0m[0;34m[0m[0;34m[0m[0;34m[0m[0m
[0;32m~/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py[0m in [0;36mrun_evaluation[0;34m(self, test_mode, max_batches)[0m
[1;32m 566[0m [0;34m[0m[0m
[1;32m 567[0m [0;31m# lightning module methods[0m[0;34m[0m[0;34m[0m[0;34m[0m[0m
[0;32m--> 568[0;31m [0moutput[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0mevaluation_loop[0m[0;34m.[0m[0mevaluation_step[0m[0;34m([0m[0mtest_mode[0m[0;34m,[0m [0mbatch[0m[0;34m,[0m [0mbatch_idx[0m[0;34m,[0m [0mdataloader_idx[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m 569[0m [0moutput[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0mevaluation_loop[0m[0;34m.[0m[0mevaluation_step_end[0m[0;34m([0m[0moutput[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[1;32m 570[0m [0;34m[0m[0m
[0;32m~/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/pytorch_lightning/trainer/evaluation_loop.py[0m in [0;36mevaluation_step[0;34m(self, test_mode, batch, batch_idx, dataloader_idx)[0m
[1;32m 169[0m [0moutput[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0mtrainer[0m[0;34m.[0m[0maccelerator_backend[0m[0;34m.[0m[0mtest_step[0m[0;34m([0m[0margs[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[1;32m 170[0m [0;32melse[0m[0;34m:[0m[0;34m[0m[0;34m[0m[0m
[0;32m--> 171[0;31m [0moutput[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0mtrainer[0m[0;34m.[0m[0maccelerator_backend[0m[0;34m.[0m[0mvalidation_step[0m[0;34m([0m[0margs[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m 172[0m [0;34m[0m[0m
[1;32m 173[0m [0;31m# track batch size for weighted average[0m[0;34m[0m[0;34m[0m[0;34m[0m[0m
[0;32m~/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py[0m in [0;36mvalidation_step[0;34m(self, args)[0m
[1;32m 74[0m [0;32mif[0m [0mself[0m[0;34m.[0m[0mtrainer[0m[0;34m.[0m[0mamp_backend[0m [0;34m==[0m [0mAMPType[0m[0;34m.[0m[0mNATIVE[0m[0;34m:[0m[0;34m[0m[0;34m[0m[0m
[1;32m 75[0m [0;32mwith[0m [0mtorch[0m[0;34m.[0m[0mcuda[0m[0;34m.[0m[0mamp[0m[0;34m.[0m[0mautocast[0m[0;34m([0m[0;34m)[0m[0;34m:[0m[0;34m[0m[0;34m[0m[0m
[0;32m---> 76[0;31m [0moutput[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0m__validation_step[0m[0;34m([0m[0margs[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m 77[0m [0;32melse[0m[0;34m:[0m[0;34m[0m[0;34m[0m[0m
[1;32m 78[0m [0moutput[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0m__validation_step[0m[0;34m([0m[0margs[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0;32m~/anaconda/envs/seq2seq-time/lib/python3.7/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py[0m in [0;36m__validation_step[0;34m(self, args)[0m
[1;32m 84[0m [0mbatch[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0mto_device[0m[0;34m([0m[0mbatch[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[1;32m 85[0m [0margs[0m[0;34m[[0m[0;36m0[0m[0;34m][0m [0;34m=[0m [0mbatch[0m[0;34m[0m[0;34m[0m[0m
[0;32m---> 86[0;31m [0moutput[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0mtrainer[0m[0;34m.[0m[0mmodel[0m[0;34m.[0m[0mvalidation_step[0m[0;34m([0m[0;34m*[0m[0margs[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m 87[0m [0;32mreturn[0m [0moutput[0m[0;34m[0m[0;34m[0m[0m
[1;32m 88[0m [0;34m[0m[0m
[0;32m<ipython-input-23-a397a13f8c11>[0m in [0;36mvalidation_step[0;34m(self, batch, batch_idx)[0m
[1;32m 25[0m [0;34m[0m[0m
[1;32m 26[0m [0;32mdef[0m [0mvalidation_step[0m[0;34m([0m[0mself[0m[0;34m,[0m [0mbatch[0m[0;34m,[0m [0mbatch_idx[0m[0;34m)[0m[0;34m:[0m[0;34m[0m[0;34m[0m[0m
[0;32m---> 27[0;31m [0;32mreturn[0m [0mself[0m[0;34m.[0m[0mtraining_step[0m[0;34m([0m[0mbatch[0m[0;34m,[0m [0mbatch_idx[0m[0;34m,[0m [0mphase[0m[0;34m=[0m[0;34m'val'[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m 28[0m [0;34m[0m[0m
[1;32m 29[0m [0;32mdef[0m [0mconfigure_optimizers[0m[0;34m([0m[0mself[0m[0;34m)[0m[0;34m:[0m[0;34m[0m[0;34m[0m[0m
[0;32m<ipython-input-23-a397a13f8c11>[0m in [0;36mtraining_step[0;34m(self, batch, batch_idx, phase)[0m
[1;32m 15[0m [0;32mdef[0m [0mtraining_step[0m[0;34m([0m[0mself[0m[0;34m,[0m [0mbatch[0m[0;34m,[0m [0mbatch_idx[0m[0;34m,[0m [0mphase[0m[0;34m=[0m[0;34m'train'[0m[0;34m)[0m[0;34m:[0m[0;34m[0m[0;34m[0m[0m
[1;32m 16[0m [0mx_past[0m[0;34m,[0m [0my_past[0m[0;34m,[0m [0mx_future[0m[0;34m,[0m [0my_future[0m [0;34m=[0m [0mbatch[0m[0;34m[0m[0;34m[0m[0m
[0;32m---> 17[0;31m [0my_dist[0m[0;34m,[0m [0mextra[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0mforward[0m[0;34m([0m[0;34m*[0m[0mbatch[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m 18[0m [0mloss[0m [0;34m=[0m [0;34m-[0m[0my_dist[0m[0;34m.[0m[0mlog_prob[0m[0;34m([0m[0my_future[0m[0;34m)[0m[0;34m.[0m[0mmean[0m[0;34m([0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[1;32m 19[0m [0mself[0m[0;34m.[0m[0mlog_dict[0m[0;34m([0m[0;34m{[0m[0;34mf'loss/{phase}'[0m[0;34m:[0m[0mloss[0m[0;34m}[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0;32m<ipython-input-23-a397a13f8c11>[0m in [0;36mforward[0;34m(self, x_past, y_past, x_future, y_future)[0m
[1;32m 10[0m [0;32mdef[0m [0mforward[0m[0;34m([0m[0mself[0m[0;34m,[0m [0mx_past[0m[0;34m,[0m [0my_past[0m[0;34m,[0m [0mx_future[0m[0;34m,[0m [0my_future[0m[0;34m=[0m[0;32mNone[0m[0;34m)[0m[0;34m:[0m[0;34m[0m[0;34m[0m[0m
[1;32m 11[0m [0;34m"""Eval/Predict"""[0m[0;34m[0m[0;34m[0m[0m
[0;32m---> 12[0;31m [0my_dist[0m[0;34m,[0m [0mextra[0m [0;34m=[0m [0mself[0m[0;34m.[0m[0m_model[0m[0;34m([0m[0mx_past[0m[0;34m,[0m [0my_past[0m[0;34m,[0m [0mx_future[0m[0;34m,[0m [0my_future[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m 13[0m [0;32mreturn[0m [0my_dist[0m[0;34m,[0m [0mextra[0m[0;34m[0m[0;34m[0m[0m
[1;32m 14[0m [0;34m[0m[0m
[0;31mTypeError[0m: cannot unpack non-iterable Normal objectIn [ ]:
# ds_preds = predict(model.to(device),
# ds_test.datasets[0],
# batch_size*8,
# device=device,
# scaler=output_scaler)In [ ]:
ds_predss = predict_multi(model.to(device),
ds_test.datasets,
batch_size*8,
device=device,
scaler=output_scaler)In [ ]:
ds_pred_block = ds_predss.isel(block=1)In [ ]:
import holoviews as hv
from holoviews import optsIn [ ]:
def plot_prediction_now(t_source):
"""Plot predictions with holoviews"""
# Let us pass in an int
if isinstance(t_source, int):
t_source = ds_pred_block.t_source[t_source].to_pandas()
d = ds_pred_block.sel(t_source=t_source)
# Sometimes there are duplicate times, take the first
if len(d.t_source.shape) and d.t_source.shape[0] > 0:
d = d.isel(t_source=0)
if len(d.t_source.shape) and d.t_source.shape[0] == 0:
return None
now = d.t_source.to_pandas()
# Plot true
x = np.concatenate([d.t_past, d.t_target])
yt = np.concatenate([d.y_past, d.y_true])
p = hv.Scatter({
'x': x,
'y': yt
}, label='true').opts(color='black')
# Get arrays
xf = d.t_target.values
yp = d.y_pred
s = d.y_pred_std
p *= hv.Curve({
'x': xf,
'y': yp
}, label='pred').opts(color='blue')
p *= hv.Area((xf, yp - 2 * s, yp + 2 * s),
vdims=['y', 'y2'],
label='2*std').opts(alpha=0.5, line_width=0)
# plot now line
p *= hv.VLine(now, label='now').opts(color='red', framewise=True)
return p.opts(title=f'Prediction at {now}. NLL={d.nll.mean().item():2.2f}')
dmap_pred = (hv.DynamicMap(plot_prediction_now, kdims=['t_source'])
.redim.values(t_source=ds_pred_block.t_source.to_pandas())
.opts(width=800,
height=300,
))
dmap_predIn [ ]:
d = ds_preds.mean(['t_source', 'block'])['nll'].groupby('t_ahead_hours').mean()
nll_vs_tahead = hv.Curve((d.t_ahead_hours, d)).redim(x='hours ahead', y='nll').opts(width=800)
nll_vs_taheadIn [ ]:
# def plot_predictions_vs_time(it_ahead):
# """Plot predictions vs time with holoviews"""
# d = ds_pred_block.isel(t_ahead=it_ahead).groupby('t_source').first()
# # print(d)
# p = hv.Scatter({
# 'x': d.t_source,
# 'y': d.y_true
# }, label='true').opts(color='black')
# # Get arrays
# xf = d.t_source.values
# yp = d.y_pred
# s = d.y_pred_std
# p *= hv.Curve({
# 'x': xf,
# 'y': yp
# }, label='pred').opts(color='blue')
# p *= hv.Area((xf, yp - 2 * s, yp + 2 * s),
# vdims=['y', 'y2'],
# label='2*std').opts(alpha=0.5, line_width=0)
# return p.opts(title=f'Prediction at {it_ahead * pd.Timedelta(freq)} ahead. NLL={d.nll.mean().item():2.2f}')
# dmap_preds = (hv.DynamicMap(plot_predictions_vs_time, kdims=['it_ahead'])
# .redim.values(it_ahead=range(ds_pred_block.t_ahead.shape[0]))
# .opts(width=800,
# height=300,
# ))
# dmap_preds
# # TODO fixmeIn [ ]:
In [ ]:
# d = ds_preds.mean(['t_ahead', 'block'])['nll'].groupby('t_source').mean()
# nll_vs_time = hv.Curve(d).opts(width=800)
# nll_vs_timeIn [ ]:
# true_vs_pred = hv.Scatter((ds_preds.y_true, ds_preds.y_pred))
# dynspread(datashade(true_vs_pred))In [ ]:
# # Run learning rate finder
# lr_finder = trainer.tuner.lr_find(model)
# # Results can be found in
# lr_finder.results
# # Plot with
# fig = lr_finder.plot(suggest=True)
# fig.show()
# # Pick point based on plot, or get suggestion
# new_lr = lr_finder.suggestion()In [ ]: