Agfdhyk

I'm trying to train a simple linear model feeding it with a simulated sparse input matrix. I've no errors but the model does not learn. My first debug was to print the logits, which said me that I'm doing badly the operations, so I have an output matrix instead of a vector (I suppose that I'm performing outer product, probably I wrongly defined shapes):

import tensorflow as tf

import numpy as np

import pandas as pd

from scipy.sparse import coo_matrix

from sklearn.datasets import make_blobs

from sklearn.preprocessing import OneHotEncoder

from sklearn.model_selection import train_test_split



samples = 800

# getting datasets

X_values, y_flat = make_blobs(n_features=2, n_samples=samples, centers=3, random_state=500)

y = OneHotEncoder().fit_transform(y_flat.reshape(-1, 1)).todense()

y = np.array(y)    

X_train, X_test, y_train, y_test, y_train_flat, y_test_flat = train_test_split(X_values, y, y_flat)



X_test += np.random.randn(*X_test.shape) * 1.5



n_features = X_values.shape[1]

n_classes = len(set(y_flat))



weights_shape = (n_features, n_classes)

bias_shape = (1, n_classes)



b = tf.Variable(dtype=tf.float32, initial_value=tf.random_normal(bias_shape))

W = tf.Variable(dtype=tf.float32, initial_value=tf.random_normal(weights_shape))

x = tf.sparse.placeholder(tf.float32)

Y_true = tf.placeholder(dtype=tf.float32)



Y_pred = tf.sparse.matmul(x, W) + b



loss_function = tf.losses.softmax_cross_entropy(Y_true, Y_pred)

learner = tf.train.GradientDescentOptimizer(0.1).minimize(loss_function)



with tf.Session() as sess:

  sess.run(tf.global_variables_initializer())

  indices = np.vstack([coo_matrix(X_train).row, coo_matrix(X_train).col]).T

  values = coo_matrix(X_train).data

  shape = np.array(coo_matrix(X_train).shape)



  for i in range(100):

      result = sess.run([learner, Y_pred], feed_dict={

           x: tf.SparseTensorValue(indices, values, shape), Y_true: y_train})  # Will succeed.

      if i % 10 == 0:

          print(result)

I've followed the TF documentation for sparse matrix multiplication https://www.tensorflow.org/api_docs/python/tf/sparse/matmul, but simply I cant solve the problem.