f0k

#!/usr/bin/env python
# -*- coding: utf-8 -*-

"""
Outputs some information on CUDA-enabled devices on your computer,
including current memory usage.

It's a port of https://gist.github.com/f0k/0d6431e3faa60bffc788f8b4daa029b1
from C to Python with ctypes, so it can run without compiling anything. Note
that this is a direct translation with no attempt to make the code Pythonic.

goldsborough

#include <cudnn.h>
#include <cassert>
#include <cstdlib>
#include <iostream>
#include <opencv2/opencv.hpp>

#define checkCUDNN(expression)                               \
  {                                                          \
    cudnnStatus_t status = (expression);                     \
    if (status != CUDNN_STATUS_SUCCESS) {                    \

neodelphis

def conv_forward_naive(x, w, b, conv_param):
    """
    A naive implementation of the forward pass for a convolutional layer.

    The input consists of N data points, each with C channels, height H and
    width W. We convolve each input with F different filters, where each filter
    spans all C channels and has height HH and width WW.

    Input:
    - x: Input data of shape (N, C, H, W)

neodelphis

def conv_backward_naive(dout, cache):
    """
    A naive implementation of the backward pass for a convolutional layer.

    Inputs:
    - dout: Upstream derivatives.
    - cache: A tuple of (x, w, b, conv_param) as in conv_forward_naive

    Returns a tuple of:
    - dx: Gradient with respect to x