CODEWITHSUNDEEP
pythontensorflowmachine-learningkerasdeep-learning
Rain
Rain

Reputation: 43

Keras ValueError: Shapes (None, 365) and (None, 1) are incompatible, Image Dataset from Directory

I'm attempting to train a network on the Places2 dataset and have arranged all the classes into subfolders. When the training and validation datasets are loaded via:

from tensorflow.keras.preprocessing import image_dataset_from_directory

train_ds = image_dataset_from_directory(
    "S:/Places2",
    image_size=(224, 224),
    batch_size=128,
)

validation_ds = image_dataset_from_directory(
    "S:/Places2Val",
    image_size=(224, 224),
    batch_size=128,
)

the console reports that all the images have been found in the correct number of classes:

Found 1803460 files belonging to 365 classes.

Found 36501 files belonging to 365 classes.

However, when trying to train the following network:

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import mixed_precision
from tensorflow.keras.applications import EfficientNetB2

start = keras.Input((224, 224, 3))

# main network
base = EfficientNetB2(
    include_top=False,
    weights="imagenet",
    input_shape=(224, 224, 3)
)
base.trainable = False
base = base(start)

# model head
top = keras.layers.AveragePooling2D()(base)
top = keras.layers.Flatten()(top)
top = keras.layers.Dense(128, activation="relu")(top)
top = keras.layers.Dropout(0.5)(top)
top = keras.layers.Dense(128, activation="relu")(top)
top = keras.layers.Dropout(0.5)(top)
top = keras.layers.Dense(365, activation="softmax")(top)

# build model and print summary
model = keras.Model(inputs=start, outputs=top)
model.summary()

# optimiser
opt = keras.optimizers.SGD(lr=0.1, momentum=0.9, nesterov=True)

# Assemble model with appropriate loss function
model.compile(loss="sparse_categorical_crossentropy", optimizer=opt, metrics=['Accuracy'])

# Train and save model
model.fit(
    train_ds,
    validation_data=validation_ds,
    epochs=1,
    batch_size=128,
    verbose=1
)

model.save("places.tf")

It throws the incompatible shapes error: ValueError: Shapes (None, 365) and (None, 1) are incompatible

This is despite image_dataset_from_directory returning (by default) inferred labels with integer labelling, explicitly to work with sparse categorical crossentropy. If the model has the correct number of outputs, and the data is loading the correct number of image categories, then why is one of the output shapes incorrect?

Particularly confusing is how changing the loss to categorical_crossentropy rearranges the error to: ValueError: Shapes (None, 1) and (None, 365) are incompatible

Printing the labels of the first batch with

for images, labels in train_ds.take(1):
    print(labels)

shows that the labels are formatted as expected--i.e. a length 128 tensor of integer labels. This should be compatible with sparse categorical crossentropy.

tf.Tensor(
[ 17 226 130 186 177  34 342  33 277 284 333 358 245 263  33  72  50 139
 298 331 250 241  50  48 264 276 218 236 303 355   3 185 107 329 277 299
  10 314  62 141 221 200   9  64 227 288 253 234  77 174 358  69 277 345
 361 205   8 197 194 217 114 135 296 305 278  82 355 134 300 129  76 321
 167 296  90 299 291 344  29 291 202 333 168 257 354  79 142  77 280   5
 261 234  78  90 250 245 302 189  97 194 347 272  54 256 160  55 131 206
 284  51 347 163 313 354 263  63 190 150 220  22 102  33   8  35  97  13
  16 277], shape=(128,), dtype=int32)

Upvotes: 1

Views: 413

Answers (1)

Gerry P
Gerry P

Reputation: 8092

well after a few hours of head scratching I figured it out! All you have to do is in model.compile change metrics=['Accuracy'] to metrics=['accuracy']. I went back to an old network I built a few years back that used sparse_categorical_crossentropy and went through it line by line.

Upvotes: 1

Related Questions