Edit model card

mlp-geron-time-series

A multi-layer perceptron (MLP) trained on Aurelien Geron's synthetic time series dataset - from Lesson 8 of the Udacity course "Intro to TensorFlow for Deep Learning". It is a PyTorch adaptation of the TensorFlow model trained in the lesson.

The time series contains a linear trend, a yearly seasonality and some white noise.

The MLP takes windows of 30 time steps of the series as input, and predicts the next time step of the series as output.

Code: https://github.com/sambitmukherjee/handson-ml3-pytorch/blob/main/bonus/udacity/Intro_to_TF_for_DL/Lesson_8_A.ipynb

Experiment tracking: https://wandb.ai/sadhaklal/mlp-geron-time-series

Usage

import numpy as np

# Create synthetic time series:
def trend(time, slope=0):
    return slope * time

def seasonal_pattern(season_time):
    """Just an arbitrary pattern; you can change it if you wish."""
    return np.where(season_time < 0.4, np.cos(season_time * 2 * np.pi), 1 / np.exp(3 * season_time))

def seasonality(time, period, amplitude=1, phase=0):
    """Repeats the same pattern in each period."""
    season_time = ((time + phase) % period) / period
    return amplitude * seasonal_pattern(season_time)

def white_noise(time, noise_level=1, seed=None):
    return np.random.RandomState(seed).randn(len(time)) * noise_level

time = np.arange(4 * 365 + 1) # 1 is added since every four years there is a leap year.
baseline = 10
slope = 0.05
amplitude = 50
series = baseline + trend(time, slope) + seasonality(time, period=365, amplitude=amplitude)
noise_level = 5
noise = white_noise(time, noise_level, seed=42)
series += noise

# Define training & validation periods:
split_time = 1000
x_train = series[:split_time]
x_valid = series[split_time:]

# Create custom `Dataset` class; it'll be used by the `model_forecast` function below:
import torch
from torch.utils.data import Dataset, DataLoader

class WindowDataset(Dataset):
    def __init__(self, series, window_size):
        window_size += 1
        self.windows = []
        for i in range(0, len(series) - window_size + 1, 1):
            self.windows.append(series[i:i + window_size])

    def __len__(self):
        return len(self.windows)

    def __getitem__(self, idx):
        window = self.windows[idx]
        return window[:-1], window[-1]

# Re-create model & load weights:
import torch.nn as nn
from huggingface_hub import PyTorchModelHubMixin

device = torch.device("cpu")

class DenseModel(nn.Module, PyTorchModelHubMixin):
    def __init__(self, window_size):
        super().__init__()
        self.dense = nn.Sequential(
            nn.Linear(window_size, 10),
            nn.ReLU(),
            nn.Linear(10, 10),
            nn.ReLU(),
            nn.Linear(10, 1)
        )

    def forward(self, x):
        return self.dense(x)

window_size = 30
model = DenseModel.from_pretrained("sadhaklal/mlp-geron-time-series", window_size=window_size)
model.to(device)

# Forecast on validation period:
def model_forecast(model, series):
    series = torch.tensor(series, dtype=torch.float32)
    ds = WindowDataset(series, window_size)
    dl = DataLoader(ds, batch_size=32, shuffle=False)
    forecast = []
    for x_batch, y_batch in dl:
        x_batch, y_batch = x_batch.to(device), y_batch.to(device)
        with torch.no_grad():
            preds = model(x_batch)
        forecast.append(preds.squeeze())
    forecast = torch.cat(forecast)
    return forecast.cpu().numpy()

dense_forecast = model_forecast(model, series[split_time - window_size:])

Metric

MAE on the validation period: 5.01


This model has been pushed to the Hub using the PyTorchModelHubMixin integration.

Downloads last month
5
Safetensors
Model size
431 params
Tensor type
F32
·
Inference API (serverless) does not yet support pytorch models for this pipeline type.