Victor Escorcia escorciav

FIle issue onto Qualcomm:SNPE

Hi,

Does QNN supports weight sharing of two independent models/graphs?
I'm trying to export a single binary with two models/graphs who share the same weights, as show in the image.

TL;DR: using syntactic pytorch data loader abstraction to feed your TF model.

WIP. The code should work. Take a look at the benchmark for a particular use case.

Perform PCA of features on Thumos14-val.

A is [n_samples, feat_dim] tall matrix i.e. n_samples >> feat_dim. We compute A^T A out of core to perform dimensionality reduction of feat_dim using in-memory eigen-decomposition.

Wouldn't it be cool to annotate a box inside the IPython notebook and use it to test your algorithm? This notebook will show you how to do that.

Usage: run the first and second cell of the notebook, select a rectangle on the figure in the second cell, run the next cells.

Motivation: I didn't find a simple exampe for doing this without setup a server or using sophisticated connections with IPython via JS. AFAIK this example was the most trivial-pythonic code to annotate boxes but it doesn't explain how to push information to IPython 1.

Acknowledgements: Bokeh for improving the documentation and interface of this package. @jakevdp for share his [pythonic perambulations](https://jakevdp.github.io/blog/2013/06/01/ipython-notebook-javascript-python-commun

This is an small attempt to replicate some features used to describe crowds in images.

LICENSE

MIT

This is the results that I got after ~60000 iterations (> 4 days) on ILSVCR-2014 with 1000 categories.

I'm using the same model proposed by caffe developers.

Training Loss vs Iters

Training Learning Rate vs Iters

Validation Loss vs Iters

	{
	"embed_dim": 512,
	"quick_gelu": true,
	"vision_cfg": {
	"image_size": 224,
	"timm_model_name": "vit_base_patch16_224",
	"timm_model_pretrained": false
	},
	"text_cfg": {
	"context_length": 77,

	name: 'EuclideanExample'
	layer {
	type: 'DummyData'
	name: 'x'
	top: 'x'
	dummy_data_param {
	shape: { dim: 10 dim: 3 dim: 2 }
	data_filler: { type: 'gaussian' }
	}
	}