Tensorflow : logits and labels must have the same first dimension

Question

I am new in tensoflow and I want to adapt the MNIST tutorial https://www.tensorflow.org/tutorials/layers with my own data (images of 40x40). This is my model function :

Answer 1

I already had this problem in my first time using tensorflow, i figured out that my problem was forgetting to add the attribute class_mode='sparse' / class_mode='binary' to the function that uploads tha training data and validation data :

So try to look after the class_mode option

image_gen_val = ImageDataGenerator(rescale=1./255)
val_data_gen = image_gen_val.flow_from_directory(batch_size=batch_size,
                                                 directory=val_dir,
                                                 target_size=(IMG_SHAPE, IMG_SHAPE),
                                                 class_mode='sparse')

Answer 2

I had a similar problem and it turned out that one pooling layer was not reshaped correctly. I was incorrectly using in my case tf.reshape(pool, shape=[-1, 64 * 7 * 7]) instead of tf.reshape(pool, shape=[-1, 64 * 14 * 14]), which lead to a similar error massage about logits and labels. Altering the factors, e.g. tf.reshape(pool, shape=[-1, 64 * 12 * 12]) lead to a completely different, less misleading error message.

Perhaps this is also the case here. I recommend going through the code checking the shapes of the nodes, just in case.

Answer 3

The problem is in your target shape and is related to the correct choice of an appropriate loss function. you have 2 possibilities:

1. possibility: if you have 1D integer encoded target, you can use sparse_categorical_crossentropy as loss function

n_class = 3
n_features = 100
n_sample = 1000

X = np.random.randint(0,10, (n_sample,n_features))
y = np.random.randint(0,n_class, n_sample)

inp = Input((n_features,))
x = Dense(128, activation='relu')(inp)
out = Dense(n_class, activation='softmax')(x)

model = Model(inp, out)
model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
history = model.fit(X, y, epochs=3)

2. possibility: if you have one-hot encoded your target in order to have 2D shape (n_samples, n_class), you can use categorical_crossentropy

n_class = 3
n_features = 100
n_sample = 1000

X = np.random.randint(0,10, (n_sample,n_features))
y = pd.get_dummies(np.random.randint(0,n_class, n_sample)).values

inp = Input((n_features,))
x = Dense(128, activation='relu')(inp)
out = Dense(n_class, activation='softmax')(x)

model = Model(inp, out)
model.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
history = model.fit(X, y, epochs=3)

Answer 4

Your logits shape looks right, batch size of 3, and output layer of size 2, which is what you defined as your output layer. Your labels should be shape [3, 2] also. Batch of 3, and each batch has 2 [1,0] or [0,1].

Also note that when you have a boolean classification output you shouldn't have 2 neurons on the output/logits layer. You can just output a single value which takes on 0 or 1, you can probably see how 2 outputs of [1,0], and [0,1] is redundant and can be expressed as a simple [0|1] value. Also you tend to get better results when you do it this way.

Thus, your logits should end up being [3,1] and your labels should be an array of 3 values, one for each of the samples in your batch.

Answer 5

I resolved it changing from sparse_categorical_crossentropy to categorical_crossentropy and is now running fine.