df.values is a numpy array, and accessing values that way is always faster than np.array.
df.values
np.array
scipy.sparse.csr_matrix(df.values)
You might need to take the transpose first, like df.values.T. In DataFrames, the columns are axis 0.
df.values.T