Example that shows how torch saves more float precision than it prints by default

validate_precision.py

#!/usr/bin/env python3

import torch
from torch.autograd import Variable

import numpy as np

def load_embeddings(count):
  word_embeddings = {}

  with open('2014_tudarmstadt_german_100mincount.vocab', 'r') as f:
    for line in f:
      line_list = line.strip().split(" ")
      word = line_list[0]
      word_embeddings[word] = line_list[1:]

      if len(word_embeddings) >= count:
        return word_embeddings

if __name__ == "__main__":
  embeddings = load_embeddings(100)

  for word, vector_str in embeddings.items():
    print("++ %s ++" % word)
    vector_float = [float(s) for s in vector_str]
    vector_np = np.array(vector_float)
    vector_torch = torch.from_numpy(vector_np)
    vector_torch_list = vector_torch.tolist()
    print("strings", vector_str[:5])
    print("floats ", vector_float[:5])
    print("numpy  ", vector_np[:5])
    print("torch  ", vector_torch[:5])
    print("torch l", vector_torch_list[:5])

    for i in range(100):
      number_str = vector_str[i]
      number_float = vector_float[i]
      number_np = vector_np[i]
      number_torch = vector_torch_list[i]

      assert(number_str == str(number_float))
      assert(number_str == str(number_np))
      assert(number_str == str(number_torch))