[FIXED] keras Input layer passes a 2 dimensional tensors while its shape is 1 dimensional

Issue

I'm having my first steps with tensorflow and keras, I have created a dataset containing tuples with embeddings of images from VGG16 model as the data and labels which are binary multilabel. for example here is a print of an element from my dataset:

(<tf.Tensor: shape=(4096,), dtype=float32, numpy=
array([0.32185513, 0.14869851, 0.4276036 , ..., 0.        , 0.        ,
       1.7438936 ], dtype=float32)>, <tf.Tensor: shape=(12,), dtype=int32, numpy=array([0, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 0])>)

I built the following model:

inputs = Input(shape = (4096,), name = 'input_1')
dense_1 = Dense(units = 2048, name = "dense_1", activation = 'sigmoid')(inputs)
dense_2 = Dense(units = 2048, name = "dense_2", activation = 'sigmoid')(dense_1)
output = Dense(units = 12, name = "output", activation = 'sigmoid')(dense_2)
model = Model(inputs = inputs, outputs = output)
model.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = 'accuracy')
model.summary()
model.fit(train_ds, validation_data = val_ds)

when I try to fit the model I get the following error

ValueError: Input 0 of layer dense_1 is incompatible with the layer: expected axis -1 of input shape to have value 4096 but received input with shape [4096, 1]

from what I understand the model expects a 1D tensor as input but gets a 2D tensor, the arrays in the data set are 1D, hence I don't understand why they are passed as a 2D tensor from the Input layer. any help figuring out what causes this error and how to fix it?

edit: the code to create the dataset is:

embed_train = np.load("Desktop/DL projects/pawpularity/petfinder-pawpularity-score/VGG16_embed.npy")
print(embed_train.shape)
>>>(9912, 4096)
print(train_labels.shape)
>>>(9912, 12)
train_ds = tf.data.Dataset.from_tensor_slices((embed_train, train_labels))
train_ds = train_ds.shuffle(10000)
test_ds = train_ds.take(1000)
train_ds = train_ds.skip(1000)
val_ds = train_ds.take(2000)
train_ds = train_ds.skip(2000)

both embed_train and train_labels are numpy arrays, as you can see I have a total of 9912 examples, each containing 4096 features, that needs to be classified for each label (12 labels in total) independently.

Solution

I was able to get your code to work by not using the Dataset, instead, passing the training data and labels straight ahead.

import tensorflow as tf
import numpy as np
from tensorflow.keras import Input
from tensorflow.keras.layers import Dense
from tensorflow.keras.models import Model

# random data in your shape - you can split for validation/validation labels if you want. 
# You would use your own embed_train and labels here
embed_train = np.random.rand(9912, 4096) 
train_labels= np.random.randint(0, 12, [9912, 12])

inputs = Input(shape=(4096,), name ='input_1')
dense_1 = Dense(units = 2048, name = "dense_1", activation = 'sigmoid', input_shape=(4096,1))(inputs)
dense_2 = Dense(units = 2048, name = "dense_2", activation = 'sigmoid')(dense_1)
output = Dense(units = 12, name = "output", activation = 'sigmoid')(dense_2)
model = Model(inputs = inputs, outputs = output)
model.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = 'accuracy')
model.summary()
model.fit(embed_train , train_labels)

If you want to shuffle the data, you can use this answer from another post:

indices = tf.range(start=0, limit=tf.shape(embed_train)[0], dtype=tf.int32)
idx = tf.random.shuffle(indices)

x_data = tf.gather(embed_train, idx)
y_data = tf.gather(train_labels, idx)

model.fit(x_data, y_data)

Edit: For Datasets, I believe you need to specify a batch size:

train_ds = train_ds.batch(32)

model.fit(train_ds)

this works on my PC.

Answered By - I M