"Input 0 of layer sequential_1 is incompatible with the layer" with tf.data.Dataset

Question

I am getting

ValueError: Input 0 of layer sequential_1 is incompatible with the layer: : expected min_ndim=4, found ndim=3. Full shape received: [300, 300, 3]. Model Summary shows 4 inputs but it says 3.

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import tensorflow_datasets as tfds
import logging
logger = tf.get_logger()
logger.setLevel(logging.ERROR)
from tensorflow.keras.preprocessing.image import ImageDataGenerator

(train_dataset, test_dataset), metadata = tfds.load('rock_paper_scissors',split=['train', 'test'], as_supervised=True, with_info=True)

class_names = ['rock','paper','scissors']

num_train_examples = metadata.splits['train'].num_examples
num_test_examples = metadata.splits['test'].num_examples
print("Number of training examples: {}".format(num_train_examples))
print("Number of test examples:     {}".format(num_test_examples))

get_label_name = metadata.features['label'].int2str
print(get_label_name(0))
print(get_label_name(1))
print(get_label_name(2))

def format_example(image, label):
    # Make image color values to be float.
    image = tf.cast(image, tf.float32)
    # Make image color values to be in [0..1] range.
    image = image / 255.
    return image, label

dataset_train = train_dataset.map(format_example)
dataset_test =test_dataset.map(format_example)

l1 = tf.keras.layers.Conv2D(32, (3,3), activation = 'relu', input_shape=(300,300,3))
l2 = tf.keras.layers.MaxPooling2D(2,2)

l3 = tf.keras.layers.Conv2D(64, (3,3), activation='relu')
l4 = tf.keras.layers.MaxPooling2D(2,2)

l5 = tf.keras.layers.Conv2D(128, (3,3), activation='relu')
l6 = tf.keras.layers.MaxPooling2D(2,2)

l7 = tf.keras.layers.Flatten()
l8 = tf.keras.layers.Dense(512, activation='relu')
l9 = tf.keras.layers.Dense(3, activation='softmax')

model = tf.keras.Sequential([l1,l2,l3,l4,l5,l6,l7,l8,l9])

model.compile(optimizer='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy,
              metrics=['accuracy'])

model.summary()

epochs = 10
batch_size = 32

history = model.fit(
    dataset_train,
    validation_data = dataset_test,
    steps_per_epoch = int(np.ceil(num_train_examples/float(batch_size))),
    validation_steps = int(np.ceil(num_test_examples/float(batch_size))),
    epochs = epochs
)

Where am I going wrong? I have not encountered this error with cats and dogs dataset.

Answer 1

You need to batch your dataset in order to get the right shape:

dataset_train = train_dataset.map(format_example).batch(1)
dataset_test =test_dataset.map(format_example).batch(1)

When you don't batch,a picture with shape (h, w, c) will be returned. If you batch, you will get (n, h, w, c), which is what Keras expects. Do the test yourself:

import tensorflow as tf

images = tf.random.uniform(shape=(10, 224, 224, 3), maxval=256, dtype=tf.int32)

ds = tf.data.Dataset.from_tensor_slices(images)

for pic in ds:
    print(pic.shape)
    break

(224, 224, 3)

With batching:

for pic in ds.batch(4):
    print(pic.shape)
    break

(4, 224, 224, 3)