如何在TensorFlow的NumpyReader中插入多变量字典？

Got it, let's walk through the multivariate case since you already have the univariate implementation working. The core idea stays similar—you just need to adjust how you structure your VALUES tensor to hold multiple features per time step, while keeping TIMES as a 1D vector like before.

First, let's recap your univariate code to set the context (I filled in the missing timestamp piece for clarity):

import numpy as np
from tensorflow.contrib.timeseries.python.timeseries import NumpyReader

status = []
time = []
# sensorstatus为列表提供数据
for ss in sensorstatus:
    status.append(int(ss.status))
    time.append(ss.time.timestamp())  # Assuming this is how you fetch timestamps

# Univariate reader initialization
reader = NumpyReader({
    "times": np.array(time),  # Shape: [sequence_length]
    "values": np.array(status)  # Shape: [sequence_length]
})

Adapting for Multivariate Data #

Say you have multiple features to track (e.g., sensor status, temperature, humidity) for each time step. Here's how to structure your data:

1. Collect features per time step: For each entry in sensorstatus, gather all your features into a single list/array instead of just one value.
2. Shape your arrays: Convert your collected data into numpy arrays where:
   • TIMES remains a 1D array with shape [sequence_length] (one timestamp per time step)
   • VALUES becomes a 2D array with shape [sequence_length, num_features] (each row is a time step, each column is a feature)

Here's a complete example:

import numpy as np
from tensorflow.contrib.timeseries.python.timeseries import NumpyReader

time_steps = []
multivariate_features = []

# Iterate through your sensor data
for ss in sensorstatus:
    # Capture the timestamp for this time step
    time_steps.append(ss.time.timestamp())
    # Pack all features for this time step into a single list
    current_step_features = [
        int(ss.status),       # Feature 1: sensor status
        float(ss.temperature),# Feature 2: temperature reading
        float(ss.humidity)    # Feature 3: humidity reading
    ]
    multivariate_features.append(current_step_features)

# Convert to numpy arrays with correct shapes
times_array = np.array(time_steps)          # Shape: (sequence_length,)
values_array = np.array(multivariate_features)  # Shape: (sequence_length, 3)

# Initialize the NumpyReader for multivariate data
multivariate_reader = NumpyReader({
    "times": times_array,
    "values": values_array
})

# Optional: Verify shapes to make sure everything is correct
print(f"TIMES shape: {times_array.shape}")       # Should look like (X,) where X is your sequence length
print(f"VALUES shape: {values_array.shape}")     # Should look like (X, 3) for 3 features

Key Notes to Avoid Issues #

• Consistent feature count: Every time step must have the same number of features—don't mix entries with 2 features and others with 3, or you'll get shape errors.
• Data type consistency: Try to keep feature values in the same data type (e.g., all floats) to avoid unexpected behavior during model training.
• TensorFlow compatibility: The NumpyReader handles the 2D VALUES tensor natively, so downstream time series models (like ARIMA or LSTM-based models) will accept this input format without extra conversion.

内容的提问来源于stack exchange，提问作者Mikko P