How to determine input shape in keras?

Question

I am having difficulty finding where my error is while building deep learning models, but I typically have issues when setting the input layer input shape.

This is my model:

model = Sequential([
Dense(32, activation='relu', input_shape=(1461, 75)),
Dense(32, activation='relu'),
Dense(ytrain.size),])

It is returning the following error:

 ValueError: Error when checking input: expected dense_1_input to have 3

 dimensions, but got array with shape (1461, 75)

The array is the training set from the kaggle housing price competition and my dataset has 75 columns and 1461 rows. My array is 2 dimensional, so why are 3 dimensions expected? I have tried adding a redundant 3rd dimension of 1 or flattening the array before the first dense layer but the error simply becomes:

ValueError: Input 0 is incompatible with layer flatten_1: expected 

min_ndim=3, found ndim=2

How do you determine what the input size should be and why do the dimensions it expects seem so arbitrary?

For reference, I attached the rest of my code:

xtrain = pd.read_csv("pricetrain.csv")
test = pd.read_csv("pricetest.csv")
xtrain.fillna(xtrain.mean(), inplace=True)
xtrain.drop(["Alley"], axis=1, inplace=True)
xtrain.drop(["PoolQC"], axis=1, inplace=True)
xtrain.drop(["Fence"], axis=1, inplace=True)
xtrain.drop(["MiscFeature"], axis=1, inplace=True)
xtrain.drop(["PoolArea"], axis=1, inplace=True)
columns = list(xtrain)
for i in columns:
    if xtrain[i].dtypes == 'object':
        xtrain[i] = pd.Categorical(pd.factorize(xtrain[i])[0])
from sklearn import preprocessing

le = preprocessing.LabelEncoder()
for i in columns:
    if xtrain[i].dtypes == 'object':
        xtrain[i] = le.fit_transform(xtrain[i])
ytrain = xtrain["SalePrice"]
xtrain.drop(["SalePrice"], axis=1, inplace=True)
ytrain = ytrain.values
xtrain = xtrain.values
ytrain.astype("float32")

size = xtrain.size
print(ytrain)
model = Sequential(
    [Flatten(),
     Dense(32, activation='relu', input_shape=(109575,)),
     Dense(32, activation='relu'),
     Dense(ytrain.size),
     ])
model.compile(loss='mse', optimizer='adam')
model.fit(xtrain, ytrain, epochs=10, verbose=1)

Any advice would be incredibly helpful!

Thank you.

Answer 1

The 0th dimension (sample-axis) is determined by the batch_size of the training. You omit it when defining the input shape. That makes sense since otherwise your model would be dependent on the number of samples in the dataset.

Same goes for the output. It seems you're only predicting a single value per example ("SalePrice"). So the output layer has shape 1.

model = Sequential([
    Dense(32, activation='relu', input_shape=(75, )),
    Dense(32, activation='relu'),
    Dense(1),
])