Time Series

Time series graphs show how your metrics change over time—watch loss decrease, accuracy improve, and learning rates adjust as training progresses.

This is the default visualization in Valohai. Any numeric metric you log automatically becomes plottable.

---

Quick Start

1. Log Metrics with a Step Counter

Print metrics with a progression counter (epoch, step, iteration):

import json

for epoch in range(num_epochs):
    train_loss = train_epoch(model, train_loader)
    val_loss = validate(model, val_loader)
    
    print(json.dumps({
        "epoch": epoch,
        "train_loss": train_loss,
        "val_loss": val_loss
    }))

---

2. Open Metadata Tab

1. Open your execution

2. Click the Metadata tab

3. Time series graphs appear automatically

---

3. Select Horizontal Axis

Default: Metrics plot against _time (when they were logged)

Better: Use your step counter for clearer progression

1. Click the Horizontal Axis dropdown

2. Select epoch (or step, iteration, etc.)

3. Graph updates immediately

---

Customizing Your Graph

Add or Remove Metrics

To add metrics to the graph:

1. Look for the Vertical Axes section on the right

2. Click Add/remove

3. Select which metrics to visualize

4. Each metric appears as a line on the graph

To remove metrics:

1. Find the metric in the Vertical Axes list

2. Click the × or remove button

3. Metric disappears from the graph

---

Smooth Noisy Data

Training metrics often have noise. Use smoothing to see trends more clearly.

To apply smoothing:

1. Find the metric in the Vertical Axes section

2. Drag the Smoothing slider

3. Move right to increase smoothing

4. "No smoothing" appears when the slider is at zero

How it works: Smoothing applies a moving average to your data. The original data points remain, but the line becomes smoother.

---

Use Multiple Vertical Axes

Plot metrics with different scales on the same graph:

Example: Loss (0-1 range) and learning rate (0.0001-0.01 range)

To use left and right axes:

1. Each metric has a Vertical Axis Side control

2. Click < for left axis (default)

3. Click > for right axis

4. Text shows which axis is active: "Using left vertical axis" or "Using right vertical axis"

This prevents one metric from dwarfing another due to scale differences.

---

Chart Options

At the bottom of the controls, you'll find:

Logarithmic Vertical Scale Use log scale for the vertical axis. Useful when metrics span multiple orders of magnitude (e.g., learning rate decay from 0.1 to 0.0001).

Show Legend Toggle the legend that identifies each line. Uncheck to hide legend for a cleaner view.

One Color per Execution When comparing multiple executions, use the same color for all metrics from one execution. Makes it easier to track which execution each line belongs to.

---

Multiple Graphs in Tabs

Create separate tabs for different metric groups.

To create a new tab:

1. Click the + button next to visualization tabs

2. Type a name (e.g., "Loss Curves", "Accuracy", "Learning Rate")

3. Configure metrics for this tab independently

Use cases:

• Loss tab: train_loss and val_loss together

• Accuracy tab: train_accuracy and val_accuracy

• Learning rate tab: Track LR schedule separately

• Resource usage tab: GPU utilization, memory usage

Each tab can have different:

• Horizontal axis selection

• Smoothing settings

• Vertical axis configurations

• Logarithmic scale settings

---

Interactive Features

All graphs use Plotly, which provides rich interactivity:

Zoom

Click and drag on the graph to zoom into a specific region.

Use cases:

• Examine early training behavior

• Investigate sudden spikes or drops

• Focus on final epochs for fine-grained convergence analysis

---

Pan

Hold Shift and drag to pan around the graph while zoomed in.

---

Reset View

Double-click anywhere on the graph to reset zoom and pan to the original view.

---

Hover for Values

Hover over any point to see:

• Exact value

• Horizontal axis value (epoch/step)

• Metric name

---

Save as Image

Click the camera icon in the graph toolbar (top right) to download the graph as a PNG image.

Perfect for:

• Including in reports

• Sharing with teammates

• Presentations

---

Common Patterns

Monitor Training vs. Validation

Plot train and validation metrics together to detect overfitting:

print(json.dumps({
    "epoch": epoch,
    "train_loss": train_loss,
    "train_accuracy": train_acc,
    "val_loss": val_loss,
    "val_accuracy": val_acc
}))

What to look for:

• Healthy training: Train and val losses both decrease

• Overfitting: Train loss decreases but val loss increases

• Underfitting: Both losses remain high

---

Track Learning Rate Schedule

Log learning rate alongside loss to understand training dynamics:

print(json.dumps({
    "epoch": epoch,
    "learning_rate": optimizer.param_groups[0]['lr'],
    "train_loss": train_loss
}))

Visualization tip: Put learning rate on the right vertical axis (different scale than loss).

---

Multiple Metrics on Same Scale

Compare related metrics with similar scales:

print(json.dumps({
    "epoch": epoch,
    "precision": precision,
    "recall": recall,
    "f1_score": f1
}))

All three metrics range from 0-1, so they plot well together on one axis.

---

Best Practices

Always Include a Step Counter

Graphs need a progression indicator:

# Good: Clear progression
print(json.dumps({
    "epoch": epoch,
    "loss": loss
}))

# Avoid: No progression
print(json.dumps({
    "loss": loss  # Will plot against _time, hard to interpret
}))

---

Use Consistent Naming

Keep metric names identical across experiments for easy comparison:

# Good: Consistent
"train_loss"
"val_loss"
"test_loss"

# Avoid: Inconsistent
"training_loss"
"validation_loss"
"testLoss"

---

Group Related Metrics

Log related metrics in the same JSON object so they share the same timestamp:

# Good: Same timestamp
print(json.dumps({
    "epoch": epoch,
    "train_loss": train_loss,
    "val_loss": val_loss
}))

# ⚠️ Avoid: Separate prints (slightly different timestamps)
print(json.dumps({"epoch": epoch, "train_loss": train_loss}))
print(json.dumps({"epoch": epoch, "val_loss": val_loss}))

---

Don't Over-Smooth

Smoothing hides detail. Use the minimum smoothing needed to see trends:

# For noisy batch-level metrics: More smoothing
# For epoch-level metrics: Less smoothing
# For final results: No smoothing

---

Use Descriptive Names

Choose metric names that are self-explanatory:

# Good: Clear meaning
"train_accuracy"
"val_f1_score"
"learning_rate"

# Avoid: Cryptic abbreviations
"tr_acc"
"v_f1"
"lr"

---

Troubleshooting

Metric Not Appearing in Horizontal Axis Dropdown

Cause: Metric has non-numeric values or missing values

Solution: Only consistent numeric metrics can be used as the horizontal axis. Use a simple counter:

# Works as horizontal axis
{"epoch": [1, 2, 3, 4, ...]}
{"step": [100, 200, 300, ...]}

# Won't work as horizontal axis
{"epoch": ["epoch_1", "epoch_2", ...]}  # String
{"loss": [0.5, None, 0.3, ...]}  # Has None

---

Graph Shows Unexpected Jumps

Cause: Using _time as horizontal axis, but logging happened at irregular intervals

Solution: Switch to a step counter like epoch for consistent spacing:

1. Click Horizontal Axis dropdown

2. Select epoch or step (assuming you have recorded this)

---

Lines Overlap and Hard to Distinguish

Cause: Too many metrics on one graph

Solutions:

• Create multiple tabs for different metric groups

• Use different vertical axes (left/right) for different scales

• Remove less important metrics temporarily

---

Can't See Early Training Behavior

Cause: Graph auto-scales to show all data

Solution:

1. Zoom into the region of interest by clicking and dragging

2. Or create a separate tab with logarithmic scale

---

Download Data for External Analysis

Export your metrics for custom visualizations:

1. Click Download raw data (top right of Metadata tab)

2. Choose CSV or JSON

3. Use in Python, R, Excel, or other tools

---

Next Steps

• Compare multiple executions to see how different settings affect training

• Create confusion matrices for classification analysis

• Compare output images from different runs

• Back to Visualize Metrics overview