# Train a CIFAR-10 model

[Download the notebook](https://ambiqai.github.io/helia-edge/notebooks/train-cifar-model.ipynb)

This page includes the notebook’s saved outputs. It does not run training in your browser.

[Open in Colab](https://colab.research.google.com/github/AmbiqAI/helia-edge/blob/main/docs/guides/train-cifar-model.ipynb) · [View source](https://github.com/AmbiqAI/helia-edge/blob/main/docs/guides/train-cifar-model.ipynb)

## Train a CIFAR-10 Model

This guide demonstrates how to train a CIFAR-10 model using HeliaEdge. Install `helia-edge[tensorflow,plotting]` to include the training and visualization dependencies.

```python
import os

import keras
import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf

import helia_edge as helia
```

Saved output

```text
2024-08-29 18:22:27.004701: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
2024-08-29 18:22:27.011355: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:485] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
2024-08-29 18:22:27.019207: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:8454] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
2024-08-29 18:22:27.021477: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1452] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
2024-08-29 18:22:27.027787: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2024-08-29 18:22:27.419630: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
```

```python
# Constants
AUG_SIZE = 64
CROP_SIZE = 48
NUM_CLASSES = 10
SHUFFLE_SIZE = 10000
BATCH_SIZE = 256
EPOCHS = 50
LR_RATE = 1e-3
LR_CYCLES = 1
VERBOSE = 1
```

```python
(x_train, y_train), (x_test, y_test) = keras.datasets.cifar10.load_data()
y_train = y_train.squeeze()
y_test = y_test.squeeze()
```

```python
y_train[:15]
```

Saved output

```text
array([6, 9, 9, 4, 1, 1, 2, 7, 8, 3, 4, 7, 7, 2, 9], dtype=uint8)
```

```python
train_ds = tf.data.Dataset.from_tensor_slices((x_train, y_train))
val_ds = tf.data.Dataset.from_tensor_slices((x_test, y_test))
```

Saved output

```text
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1724955748.587058  209138 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1724955748.606303  209138 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1724955748.606421  209138 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1724955748.607778  209138 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1724955748.607872  209138 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1724955748.607917  209138 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1724955748.648330  209138 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1724955748.648420  209138 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
2024-08-29 18:22:28.648458: W tensorflow/core/common_runtime/gpu/gpu_bfc_allocator.cc:47] Overriding orig_value setting because the TF_FORCE_GPU_ALLOW_GROWTH environment variable is set. Original config value was 0.
I0000 00:00:1724955748.648484  209138 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
2024-08-29 18:22:28.648530: I tensorflow/core/common_runtime/gpu/gpu_device.cc:2021] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 21858 MB memory:  -> device: 0, name: NVIDIA GeForce RTX 4090, pci bus id: 0000:01:00.0, compute capability: 8.9
```

```python
# # Plot a random image
fig, ax = plt.subplots(figsize=(2, 2))
ax.imshow(x_train[np.random.randint(0, x_train.shape[0])])
_ = ax.axis("off")
```

Saved output

![Saved notebook figure](https://ambiqai.github.io/helia-edge/notebooks/train-cifar-model-7-0.png)

```python
augmenter = helia.layers.preprocessing.AugmentationPipeline(
    layers=[
        helia.layers.preprocessing.Resizing2D(height=AUG_SIZE, width=AUG_SIZE, name="Resizing"),
        helia.layers.preprocessing.RandomFlip2D(horizontal=True, vertical=False, name="RandomFlip"),
        helia.layers.preprocessing.RandomCrop2D(
            height=CROP_SIZE, width=CROP_SIZE, name="RandomCrop", auto_vectorize=True
        ),
        helia.layers.preprocessing.LayerNormalization2D(name="normalize"),
    ],
)
```

```python
y = augmenter(x_train[:1].astype(np.float32))
```

```python
# # Plot a random image
fig, ax = plt.subplots(figsize=(2, 2))
ax.imshow(y[0].numpy(), vmin=-2.6, vmax=4)
_ = ax.axis("off")
```

Saved output

```text
[08/29/24 18:22:28] WARNING  WARNING:matplotlib.image:Clipping input data to the valid range for       ]8;id=450190;file:///workspaces/helia-edge/.venv/lib/python3.11/site-packages/matplotlib/image.py\image.py]8;;\:]8;id=197316;file:///workspaces/helia-edge/.venv/lib/python3.11/site-packages/matplotlib/image.py#705\705]8;;\
                             imshow with RGB data ([0..1] for floats or [0..255] for integers). Got
                             range [-2.9038177..3.3586729].
```

![Saved notebook figure](https://ambiqai.github.io/helia-edge/notebooks/train-cifar-model-10-1.png)

```python
train_ds = (
    train_ds.shuffle(buffer_size=SHUFFLE_SIZE)
    .batch(batch_size=BATCH_SIZE)
    .map(
        lambda data, labels: {
            "data": tf.cast(data, "float32"),
            "labels": tf.one_hot(labels, NUM_CLASSES),
        },
        num_parallel_calls=tf.data.AUTOTUNE,
    )
    .map(augmenter, num_parallel_calls=tf.data.AUTOTUNE)
    .map(
        lambda data: (data["data"], data["labels"]),
        num_parallel_calls=tf.data.AUTOTUNE,
    )
    .prefetch(tf.data.AUTOTUNE)
)
```

```python
val_ds = (
    val_ds.shuffle(buffer_size=SHUFFLE_SIZE)
    .batch(batch_size=BATCH_SIZE)
    .map(
        lambda data, labels: {
            "data": tf.cast(data, "float32"),
            "labels": tf.one_hot(labels, NUM_CLASSES),
        },
        num_parallel_calls=tf.data.AUTOTUNE,
    )
    .map(augmenter, num_parallel_calls=tf.data.AUTOTUNE)
    .map(
        lambda data: (data["data"], data["labels"]),
        num_parallel_calls=tf.data.AUTOTUNE,
    )
    .cache()
    .prefetch(tf.data.AUTOTUNE)
)
```

```python
model = helia.models.build(
    helia.models.ModelSpec(
        params=helia.models.EfficientNetParams(
            input_filters=32,
            input_kernel_size=(3, 3),
            input_strides=(2, 2),
            blocks=[
                helia.layers.MBConvParams(
                    filters=48, depth=1, kernel_size=(3, 3), strides=(1, 1), ex_ratio=1, se_ratio=4
                ),
                helia.layers.MBConvParams(
                    filters=64, depth=1, kernel_size=(3, 3), strides=(2, 2), ex_ratio=1, se_ratio=4
                ),
                helia.layers.MBConvParams(
                    filters=96, depth=1, kernel_size=(3, 3), strides=(2, 2), ex_ratio=1, se_ratio=4
                ),
                helia.layers.MBConvParams(
                    filters=128, depth=1, kernel_size=(3, 3), strides=(2, 2), ex_ratio=1, se_ratio=4
                ),
            ],
            output_filters=0,
            include_top=True,
            num_classes=NUM_CLASSES,
            dropout=0.2,
            drop_connect_rate=0.2,
        ),
        input_shape=(CROP_SIZE, CROP_SIZE, 3),
    )
)
```

```python
flops = helia.metrics.flops.get_flops(model, batch_size=1, fpath=os.devnull)
model.summary(layer_range=("inputs", model.layers[10].name))
print(f"FLOPS: {flops / 1e6:.2f}M")
```

Saved output

```text
WARNING:tensorflow:From /workspaces/helia-edge/.venv/lib/python3.11/site-packages/tensorflow/python/ops/nn_ops.py:5256: tensor_shape_from_node_def_name (from tensorflow.python.framework.graph_util_impl) is deprecated and will be removed in a future version.
Instructions for updating:
This API was designed for TensorFlow v1. See https://www.tensorflow.org/guide/migrate for instructions on how to migrate your code to TensorFlow v2.
```

```text
[08/29/24 18:22:29] WARNING  WARNING:tensorflow:From                                              ]8;id=231804;file:///workspaces/helia-edge/.venv/lib/python3.11/site-packages/tensorflow/python/util/deprecation.py\deprecation.py]8;;\:]8;id=705631;file:///workspaces/helia-edge/.venv/lib/python3.11/site-packages/tensorflow/python/util/deprecation.py#50\50]8;;\
                             /workspaces/helia-edge/.venv/lib/python3.11/site-packages/tenso
                             rflow/python/ops/nn_ops.py:5256: tensor_shape_from_node_def_name
                             (from tensorflow.python.framework.graph_util_impl) is deprecated and
                             will be removed in a future version.
                             Instructions for updating:
                             This API was designed for TensorFlow v1. See
                             https://www.tensorflow.org/guide/migrate for instructions on how to
                             migrate your code to TensorFlow v2.
```

```text
Model: "functional"
```

```text
┏━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┓
┃ Layer (type)        ┃ Output Shape      ┃    Param # ┃ Connected to      ┃
┡━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━┩
│ inputs (InputLayer) │ (None, 48, 48, 3) │          0 │ -                 │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stem.conv (Conv2D)  │ (None, 24, 24,    │        864 │ inputs[0][0]      │
│                     │ 32)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stem.bn             │ (None, 24, 24,    │        128 │ stem.conv[0][0]   │
│ (BatchNormalizatio… │ 32)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stem.act            │ (None, 24, 24,    │          0 │ stem.bn[0][0]     │
│ (Activation)        │ 32)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.dp   │ (None, 24, 24,    │        288 │ stem.act[0][0]    │
│ (DepthwiseConv2D)   │ 32)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.dp.… │ (None, 24, 24,    │        128 │ stage1.mbconv1.d… │
│ (BatchNormalizatio… │ 32)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.dp.… │ (None, 24, 24,    │          0 │ stage1.mbconv1.d… │
│ (Activation)        │ 32)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.se.… │ (None, 1, 1, 32)  │          0 │ stage1.mbconv1.d… │
│ (GlobalAveragePool… │                   │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.se.… │ (None, 1, 1, 8)   │        264 │ stage1.mbconv1.s… │
│ (Conv2D)            │                   │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.se.… │ (None, 1, 1, 8)   │          0 │ stage1.mbconv1.s… │
│ (Activation)        │                   │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.se.… │ (None, 1, 1, 32)  │        288 │ stage1.mbconv1.s… │
│ (Conv2D)            │                   │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.se.… │ (None, 1, 1, 32)  │          0 │ stage1.mbconv1.s… │
│ (Activation)        │                   │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.se.… │ (None, 24, 24,    │          0 │ stage1.mbconv1.d… │
│ (Multiply)          │ 32)               │            │ stage1.mbconv1.s… │
└─────────────────────┴───────────────────┴────────────┴───────────────────┘
```

```text
 Total params: 38,406 (150.02 KB)
```

```text
 Trainable params: 37,190 (145.27 KB)
```

```text
 Non-trainable params: 1,216 (4.75 KB)
```

```text
FLOPS: 5.08M
```

Training saves the best checkpoint to `model.keras` in the working directory when this notebook runs. The repository does not include trained weights.

```python
t_mul = 1
first_steps = (EPOCHS * len(train_ds)) / (np.power(LR_CYCLES, t_mul) - t_mul + 1)
scheduler = keras.optimizers.schedules.CosineDecayRestarts(
    initial_learning_rate=LR_RATE,
    first_decay_steps=np.ceil(first_steps),
    t_mul=t_mul,
    m_mul=0.5,
)

optimizer = keras.optimizers.Adam(scheduler)
loss = keras.losses.CategoricalFocalCrossentropy(from_logits=True)

metrics = [
    keras.metrics.CategoricalAccuracy(name="acc"),
    keras.metrics.F1Score(name="f1", average="weighted"),
]

callbacks = [
    keras.callbacks.EarlyStopping(monitor="val_loss", patience=25, restore_best_weights=True, verbose=VERBOSE),
    keras.callbacks.ModelCheckpoint(monitor="val_loss", verbose=VERBOSE, save_best_only=True, filepath="model.keras"),
]
```

```python
model.compile(
    optimizer=optimizer,
    loss=loss,
    metrics=metrics,
)
```

```python
history = model.fit(
    train_ds,
    validation_data=val_ds,
    epochs=EPOCHS,
    callbacks=callbacks,
    verbose=VERBOSE,
)
```

Saved output

```text
Epoch 1/50
```

```text
2024-08-29 18:22:29.789875: E tensorflow/core/util/util.cc:131] oneDNN supports DT_INT32 only on platforms with AVX-512. Falling back to the default Eigen-based implementation if present.
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1724955751.896582  209302 service.cc:146] XLA service 0x77e9a0002f90 initialized for platform CUDA (this does not guarantee that XLA will be used). Devices:
I0000 00:00:1724955751.896661  209302 service.cc:154]   StreamExecutor device (0): NVIDIA GeForce RTX 4090, Compute Capability 8.9
2024-08-29 18:22:32.062786: I tensorflow/compiler/mlir/tensorflow/utils/dump_mlir_util.cc:268] disabling MLIR crash reproducer, set env var `MLIR_CRASH_REPRODUCER_DIRECTORY` to enable.
2024-08-29 18:22:32.346082: I external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:531] Loaded cuDNN version 8907
```

```text
 27/196 ━━━━━━━━━━━━━━━━━━━━ 1s 6ms/step - acc: 0.1293 - f1: 0.1271 - loss: 0.4749
```

```text
I0000 00:00:1724955756.415054  209302 device_compiler.h:188] Compiled cluster using XLA!  This line is logged at most once for the lifetime of the process.
```

```text
196/196 ━━━━━━━━━━━━━━━━━━━━ 0s 25ms/step - acc: 0.2348 - f1: 0.2300 - loss: 0.3985
Epoch 1: val_loss improved from inf to 0.29081, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 14s 36ms/step - acc: 0.2351 - f1: 0.2303 - loss: 0.3982 - val_acc: 0.4022 - val_f1: 0.3935 - val_loss: 0.2908
Epoch 2/50
195/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.4083 - f1: 0.4059 - loss: 0.2860
Epoch 2: val_loss improved from 0.29081 to 0.25709, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.4085 - f1: 0.4061 - loss: 0.2859 - val_acc: 0.4660 - val_f1: 0.4641 - val_loss: 0.2571
Epoch 3/50
180/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.4687 - f1: 0.4664 - loss: 0.2526
Epoch 3: val_loss improved from 0.25709 to 0.23724, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.4696 - f1: 0.4673 - loss: 0.2521 - val_acc: 0.4950 - val_f1: 0.4859 - val_loss: 0.2372
Epoch 4/50
196/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.5023 - f1: 0.4994 - loss: 0.2330
Epoch 4: val_loss improved from 0.23724 to 0.22668, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.5023 - f1: 0.4994 - loss: 0.2330 - val_acc: 0.5079 - val_f1: 0.5082 - val_loss: 0.2267
Epoch 5/50
189/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.5213 - f1: 0.5201 - loss: 0.2205
Epoch 5: val_loss did not improve from 0.22668
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.5215 - f1: 0.5202 - loss: 0.2204 - val_acc: 0.4902 - val_f1: 0.4825 - val_loss: 0.2376
Epoch 6/50
187/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.5367 - f1: 0.5358 - loss: 0.2120
Epoch 6: val_loss improved from 0.22668 to 0.20614, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 8ms/step - acc: 0.5370 - f1: 0.5361 - loss: 0.2118 - val_acc: 0.5485 - val_f1: 0.5481 - val_loss: 0.2061
Epoch 7/50
193/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.5539 - f1: 0.5523 - loss: 0.2018
Epoch 7: val_loss did not improve from 0.20614
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.5539 - f1: 0.5524 - loss: 0.2018 - val_acc: 0.5300 - val_f1: 0.5267 - val_loss: 0.2139
Epoch 8/50
196/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.5736 - f1: 0.5724 - loss: 0.1903
Epoch 8: val_loss improved from 0.20614 to 0.19365, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.5736 - f1: 0.5724 - loss: 0.1903 - val_acc: 0.5636 - val_f1: 0.5613 - val_loss: 0.1937
Epoch 9/50
184/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.5811 - f1: 0.5801 - loss: 0.1870
Epoch 9: val_loss did not improve from 0.19365
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.5809 - f1: 0.5800 - loss: 0.1870 - val_acc: 0.5645 - val_f1: 0.5631 - val_loss: 0.1941
Epoch 10/50
186/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.5879 - f1: 0.5867 - loss: 0.1816
Epoch 10: val_loss improved from 0.19365 to 0.18366, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.5880 - f1: 0.5869 - loss: 0.1815 - val_acc: 0.5871 - val_f1: 0.5869 - val_loss: 0.1837
Epoch 11/50
190/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.5861 - f1: 0.5856 - loss: 0.1814
Epoch 11: val_loss improved from 0.18366 to 0.18201, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.5862 - f1: 0.5857 - loss: 0.1814 - val_acc: 0.5874 - val_f1: 0.5883 - val_loss: 0.1820
Epoch 12/50
196/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.5939 - f1: 0.5929 - loss: 0.1764
Epoch 12: val_loss improved from 0.18201 to 0.18121, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.5939 - f1: 0.5929 - loss: 0.1764 - val_acc: 0.5855 - val_f1: 0.5868 - val_loss: 0.1812
Epoch 13/50
186/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6016 - f1: 0.6011 - loss: 0.1730
Epoch 13: val_loss improved from 0.18121 to 0.17338, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6018 - f1: 0.6013 - loss: 0.1729 - val_acc: 0.6038 - val_f1: 0.6031 - val_loss: 0.1734
Epoch 14/50
192/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6101 - f1: 0.6094 - loss: 0.1672
Epoch 14: val_loss did not improve from 0.17338
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6101 - f1: 0.6094 - loss: 0.1672 - val_acc: 0.6003 - val_f1: 0.5985 - val_loss: 0.1742
Epoch 15/50
184/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6197 - f1: 0.6188 - loss: 0.1647
Epoch 15: val_loss improved from 0.17338 to 0.16967, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6197 - f1: 0.6188 - loss: 0.1646 - val_acc: 0.6086 - val_f1: 0.6083 - val_loss: 0.1697
Epoch 16/50
193/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6250 - f1: 0.6243 - loss: 0.1591
Epoch 16: val_loss improved from 0.16967 to 0.16697, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6249 - f1: 0.6242 - loss: 0.1592 - val_acc: 0.6137 - val_f1: 0.6116 - val_loss: 0.1670
Epoch 17/50
191/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6273 - f1: 0.6271 - loss: 0.1588
Epoch 17: val_loss did not improve from 0.16697
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6273 - f1: 0.6271 - loss: 0.1588 - val_acc: 0.6038 - val_f1: 0.6034 - val_loss: 0.1747
Epoch 18/50
178/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6334 - f1: 0.6330 - loss: 0.1554
Epoch 18: val_loss improved from 0.16697 to 0.16457, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6331 - f1: 0.6327 - loss: 0.1555 - val_acc: 0.6201 - val_f1: 0.6212 - val_loss: 0.1646
Epoch 19/50
186/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6431 - f1: 0.6422 - loss: 0.1506
Epoch 19: val_loss improved from 0.16457 to 0.16057, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6427 - f1: 0.6418 - loss: 0.1508 - val_acc: 0.6260 - val_f1: 0.6263 - val_loss: 0.1606
Epoch 20/50
189/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6340 - f1: 0.6332 - loss: 0.1530
Epoch 20: val_loss did not improve from 0.16057
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6342 - f1: 0.6334 - loss: 0.1530 - val_acc: 0.6048 - val_f1: 0.6075 - val_loss: 0.1758
Epoch 21/50
195/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6385 - f1: 0.6380 - loss: 0.1514
Epoch 21: val_loss did not improve from 0.16057
196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 7ms/step - acc: 0.6386 - f1: 0.6380 - loss: 0.1514 - val_acc: 0.6234 - val_f1: 0.6246 - val_loss: 0.1609
Epoch 22/50
187/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6433 - f1: 0.6430 - loss: 0.1500
Epoch 22: val_loss did not improve from 0.16057
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6434 - f1: 0.6430 - loss: 0.1500 - val_acc: 0.6203 - val_f1: 0.6190 - val_loss: 0.1665
Epoch 23/50
188/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6422 - f1: 0.6419 - loss: 0.1506
Epoch 23: val_loss improved from 0.16057 to 0.15481, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 8ms/step - acc: 0.6423 - f1: 0.6419 - loss: 0.1506 - val_acc: 0.6368 - val_f1: 0.6358 - val_loss: 0.1548
Epoch 24/50
191/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6509 - f1: 0.6507 - loss: 0.1466
Epoch 24: val_loss did not improve from 0.15481
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6509 - f1: 0.6506 - loss: 0.1466 - val_acc: 0.6351 - val_f1: 0.6361 - val_loss: 0.1563
Epoch 25/50
188/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6570 - f1: 0.6568 - loss: 0.1428
Epoch 25: val_loss improved from 0.15481 to 0.15440, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6569 - f1: 0.6566 - loss: 0.1429 - val_acc: 0.6353 - val_f1: 0.6353 - val_loss: 0.1544
Epoch 26/50
188/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6549 - f1: 0.6548 - loss: 0.1445
Epoch 26: val_loss improved from 0.15440 to 0.15137, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6549 - f1: 0.6548 - loss: 0.1445 - val_acc: 0.6423 - val_f1: 0.6419 - val_loss: 0.1514
Epoch 27/50
182/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6555 - f1: 0.6549 - loss: 0.1435
Epoch 27: val_loss did not improve from 0.15137
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6556 - f1: 0.6550 - loss: 0.1435 - val_acc: 0.6286 - val_f1: 0.6324 - val_loss: 0.1576
Epoch 28/50
188/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6587 - f1: 0.6581 - loss: 0.1426
Epoch 28: val_loss improved from 0.15137 to 0.15019, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6586 - f1: 0.6580 - loss: 0.1427 - val_acc: 0.6439 - val_f1: 0.6456 - val_loss: 0.1502
Epoch 29/50
191/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6651 - f1: 0.6647 - loss: 0.1385
Epoch 29: val_loss improved from 0.15019 to 0.14894, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6650 - f1: 0.6646 - loss: 0.1386 - val_acc: 0.6467 - val_f1: 0.6453 - val_loss: 0.1489
Epoch 30/50
191/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6618 - f1: 0.6612 - loss: 0.1405
Epoch 30: val_loss improved from 0.14894 to 0.14712, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6618 - f1: 0.6612 - loss: 0.1406 - val_acc: 0.6506 - val_f1: 0.6499 - val_loss: 0.1471
Epoch 31/50
192/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6636 - f1: 0.6631 - loss: 0.1408
Epoch 31: val_loss did not improve from 0.14712
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6636 - f1: 0.6631 - loss: 0.1408 - val_acc: 0.6413 - val_f1: 0.6409 - val_loss: 0.1507
Epoch 32/50
195/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6649 - f1: 0.6646 - loss: 0.1396
Epoch 32: val_loss did not improve from 0.14712
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6649 - f1: 0.6646 - loss: 0.1396 - val_acc: 0.6484 - val_f1: 0.6492 - val_loss: 0.1477
Epoch 33/50
190/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6757 - f1: 0.6751 - loss: 0.1349
Epoch 33: val_loss did not improve from 0.14712
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6756 - f1: 0.6750 - loss: 0.1350 - val_acc: 0.6436 - val_f1: 0.6450 - val_loss: 0.1498
Epoch 34/50
180/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6739 - f1: 0.6734 - loss: 0.1336
Epoch 34: val_loss improved from 0.14712 to 0.14656, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6737 - f1: 0.6732 - loss: 0.1338 - val_acc: 0.6483 - val_f1: 0.6482 - val_loss: 0.1466
Epoch 35/50
185/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6757 - f1: 0.6753 - loss: 0.1356
Epoch 35: val_loss did not improve from 0.14656
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6755 - f1: 0.6751 - loss: 0.1357 - val_acc: 0.6488 - val_f1: 0.6499 - val_loss: 0.1467
Epoch 36/50
185/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6683 - f1: 0.6677 - loss: 0.1362
Epoch 36: val_loss improved from 0.14656 to 0.14477, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6683 - f1: 0.6677 - loss: 0.1362 - val_acc: 0.6493 - val_f1: 0.6490 - val_loss: 0.1448
Epoch 37/50
195/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6799 - f1: 0.6800 - loss: 0.1320
Epoch 37: val_loss improved from 0.14477 to 0.14390, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6799 - f1: 0.6799 - loss: 0.1320 - val_acc: 0.6546 - val_f1: 0.6546 - val_loss: 0.1439
Epoch 38/50
196/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6745 - f1: 0.6742 - loss: 0.1344
Epoch 38: val_loss did not improve from 0.14390
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6745 - f1: 0.6742 - loss: 0.1344 - val_acc: 0.6504 - val_f1: 0.6513 - val_loss: 0.1443
Epoch 39/50
193/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6820 - f1: 0.6815 - loss: 0.1304
Epoch 39: val_loss did not improve from 0.14390
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6819 - f1: 0.6814 - loss: 0.1305 - val_acc: 0.6550 - val_f1: 0.6550 - val_loss: 0.1441
Epoch 40/50
190/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6741 - f1: 0.6738 - loss: 0.1341
Epoch 40: val_loss improved from 0.14390 to 0.14272, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 8ms/step - acc: 0.6742 - f1: 0.6738 - loss: 0.1341 - val_acc: 0.6555 - val_f1: 0.6554 - val_loss: 0.1427
Epoch 41/50
187/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6791 - f1: 0.6784 - loss: 0.1309
Epoch 41: val_loss did not improve from 0.14272
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6789 - f1: 0.6782 - loss: 0.1311 - val_acc: 0.6546 - val_f1: 0.6551 - val_loss: 0.1432
Epoch 42/50
180/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6806 - f1: 0.6796 - loss: 0.1323
Epoch 42: val_loss did not improve from 0.14272
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6803 - f1: 0.6793 - loss: 0.1324 - val_acc: 0.6530 - val_f1: 0.6543 - val_loss: 0.1444
Epoch 43/50
190/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6726 - f1: 0.6723 - loss: 0.1356
Epoch 43: val_loss improved from 0.14272 to 0.14239, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 8ms/step - acc: 0.6727 - f1: 0.6724 - loss: 0.1355 - val_acc: 0.6549 - val_f1: 0.6553 - val_loss: 0.1424
Epoch 44/50
189/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6697 - f1: 0.6692 - loss: 0.1366
Epoch 44: val_loss did not improve from 0.14239
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6698 - f1: 0.6693 - loss: 0.1366 - val_acc: 0.6538 - val_f1: 0.6547 - val_loss: 0.1428
Epoch 45/50
178/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6771 - f1: 0.6769 - loss: 0.1322
Epoch 45: val_loss did not improve from 0.14239
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6768 - f1: 0.6766 - loss: 0.1323 - val_acc: 0.6536 - val_f1: 0.6542 - val_loss: 0.1428
Epoch 46/50
195/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6685 - f1: 0.6683 - loss: 0.1386
Epoch 46: val_loss improved from 0.14239 to 0.14178, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 8ms/step - acc: 0.6686 - f1: 0.6683 - loss: 0.1386 - val_acc: 0.6546 - val_f1: 0.6548 - val_loss: 0.1418
Epoch 47/50
181/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6797 - f1: 0.6794 - loss: 0.1315
Epoch 47: val_loss did not improve from 0.14178
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6796 - f1: 0.6793 - loss: 0.1316 - val_acc: 0.6521 - val_f1: 0.6532 - val_loss: 0.1434
Epoch 48/50
182/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6743 - f1: 0.6736 - loss: 0.1347
Epoch 48: val_loss improved from 0.14178 to 0.14157, saving model to model.keras
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6745 - f1: 0.6739 - loss: 0.1346 - val_acc: 0.6551 - val_f1: 0.6553 - val_loss: 0.1416
Epoch 49/50
187/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6798 - f1: 0.6796 - loss: 0.1319
Epoch 49: val_loss did not improve from 0.14157
196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 7ms/step - acc: 0.6797 - f1: 0.6795 - loss: 0.1319 - val_acc: 0.6530 - val_f1: 0.6528 - val_loss: 0.1422
Epoch 50/50
194/196 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - acc: 0.6764 - f1: 0.6762 - loss: 0.1334
Epoch 50: val_loss did not improve from 0.14157
196/196 ━━━━━━━━━━━━━━━━━━━━ 2s 7ms/step - acc: 0.6765 - f1: 0.6762 - loss: 0.1334 - val_acc: 0.6530 - val_f1: 0.6540 - val_loss: 0.1435
Restoring model weights from the end of the best epoch: 48.
```

```python
_ = helia.plotting.plot_history_metrics(
    history=history.history, metrics=["loss", "acc"], title="Training History", stack=True
)
```

Saved output

![Saved notebook figure](https://ambiqai.github.io/helia-edge/notebooks/train-cifar-model-19-0.png)

```python

```