fmassa’s gists

fmassa / query_distribution.py

Created June 28, 2020 16:11

Code to reproduce Fig 7 in "End to End Object Detection with Transformers"

	# this file needs to be added to the root folder of detr github repo
	import torch
	import time
	import torchvision

	import numpy as np
	import tqdm

	import matplotlib.pyplot as plt

fmassa / compute_flops.py

Created June 28, 2020 13:34

Utility functions used to compute flops in DETR.

	# this is the main entrypoint
	# as we describe in the paper, we compute the flops over the first 100 images
	# on COCO val2017, and report the average result
	import torch
	import time
	import torchvision

	import numpy as np
	import tqdm

fmassa / visualize_maskrcnn_predictions.py

Created December 2, 2019 10:52

Script to visualize predictions from torchvision Mask R-CNN

	import torch
	import numpy as np
	import matplotlib.pyplot as plt
	import cv2


	def select_top_predictions(predictions, threshold):
	idx = (predictions["scores"] > threshold).nonzero().squeeze(1)
	new_predictions = {}
	for k, v in predictions.items():

fmassa / sparse_adam_jit.py

Created July 31, 2019 17:41

Fused sparse adam with JIT

	import math
	import torch
	from torch.optim.optimizer import Optimizer


	class SparseAdam(Optimizer):
	r"""Implements lazy version of Adam algorithm suitable for sparse tensors.
	In this variant, only moments that show up in the gradient get updated, and
	only those portions of the gradient get applied to the parameters.
	Arguments:

fmassa / recover_optnet.lua

Created November 1, 2016 16:25

Script to recover saved optnet models for which the tensor pointer changed after optimization

	local optnet = require 'optnet'
	local net = torch.load('celeba_24_G.t7')

	local keys
	-- the tensor pointers that were saved in the model
	-- have changed and are not valid anymore.
	-- In order to try to recover, let's suppose that
	-- the mapping is given by the offsets in ascending
	-- order.
	do

fmassa / optnet_test_resnet101.lua

Created August 10, 2016 16:18

Test script to verify that ResNet101 fits in memory with batch size 64 with latest optnet

	local optnet = require 'optnet'
	require 'cudnn'
	require 'cunn'

	local createModel = require 'resnet'
	local opt = {dataset='imagenet', depth=101}

	local model = createModel(opt)

	local input = torch.zeros(4,3,224,224):cuda()

fmassa / differentLearningRates.lua

Last active July 3, 2017 08:25

	-- suppose you have a model called model
	lrs_model = model:clone()
	lrs = lrs_model:getParameters()
	lrs:fill(1) -- setting the base learning rate to 1

	-- now lets set the learning rate factor of the bias of module 5 to 2
	lrs_model:get(5).bias:fill(2)
	-- same thing for the weights of module 2, let's set them to 3
	lrs_model:get(2).weight:fill(3)

fmassa / convertLinear2Conv1x1.lua

Last active May 30, 2018 20:16

Simple example on how to convert a Linear model to a 1x1 convolution

	require 'nn'

	-- you just need to provide the linear module you want to convert,
	-- and the dimensions of the field of view of the linear layer
	function convertLinear2Conv1x1(linmodule,in_size)
	local s_in = linmodule.weight:size(2)/(in_size[1]*in_size[2])
	local s_out = linmodule.weight:size(1)
	local convmodule = nn.SpatialConvolutionMM(s_in,s_out,in_size[1],in_size[2],1,1)
	convmodule.weight:copy(linmodule.weight)
	convmodule.bias:copy(linmodule.bias)

Francisco Massa fmassa