Does Metalhead.classify support pretrained models other than VGG19 ?
#30
Comments
|
On my PC, I'm using Julia1.0.1 with the following libraries |
|
I found handle data as using Flux: onecold
using Metalhead: ResNet, VGG19
using Images
# Take the len-by-len square of pixels at the center of image `im`
function center_crop(im, len)
l2 = div(len,2)
adjust = len % 2 == 0 ? 1 : 0
return im[div(end,2)-l2:div(end,2)+l2-adjust,div(end,2)-l2:div(end,2)+l2-adjust]
end
# Resize an image such that its smallest dimension is the given length
function resize_smallest_dimension(im, len)
reduction_factor = len/minimum(size(im)[1:2])
new_size = size(im)
new_size = (
round(Int, size(im,1)*reduction_factor),
round(Int, size(im,2)*reduction_factor),
)
if reduction_factor < 1.0
# Images.jl's imresize() needs to first lowpass the image, it won't do it for us
im = imfilter(im, KernelFactors.gaussian(0.75/reduction_factor), Inner())
end
return imresize(im, new_size)
end
function preprocess(im::AbstractMatrix{<:AbstractRGB})
# Resize such that smallest edge is 256 pixels long
im = resize_smallest_dimension(im, 256)
# Center-crop to 224x224
im = center_crop(im, 224)
# Convert to channel view and normalize (these coefficients taken
# from PyTorch's ImageNet normalization code)
μ = [0.485, 0.456, 0.406]
σ = [0.229, 0.224, 0.225]
im = (channelview(im) .- μ)./σ
# Convert from CHW (Image.jl's channel ordering) to WHCN (Flux.jl's ordering)
# convert eltype as Float64
return Float64.(permutedims(im, (3, 2, 1))[:,:,:,:].*255)
end
const label_txt = expanduser("~/.julia/packages/Metalhead/rGGAv/datasets/meta/ILSVRC_synset_mappings.txt")
imagenet_labels = String[]
for (idx, line) in enumerate(eachline(label_txt))
synset = line[1:9]
label = line[11:end]
push!(imagenet_labels, label)
end
function main()
model = ResNet()
preprocessed = preprocess(load("elephant.jpeg"))
result = imagenet_labels[onecold(model(preprocessed))][1]
@show result
end
main() |
|
@terasakisatoshi There is some issue with the pretrained weights of Resnet and Densenet that were imported using ONNX. So I am not surprised by the incorrect results. |
|
Sorry to bother you too much today, Why don't you load batchnorm parameters in ResNet example? |
|
IIRC I had discussed this with @ayush1999. The onnx file imported had no batchnorm parameters. That's the reason the resnet doesn't perform as expected. |
|
When I dump the .bson file, there are values such that: I believe: bn_b -> batchnorm bias parameter |
|
Thanks. I will add them as soon as possible. :) |
|
@ekinakyurek I have updated the Resnet Model. U can checkout the |
|
I couldn't see newmodels branch. Is it working ? |
|
Here's the branch |
|
Maybe there is something wrong on my end but ResNet50 is not giving correct results. A picture of an elephant, for example, is classified as ""Persian cat". VGG19 seems to work fine. There might still be an issue with importing the pre-trained weights? Also, could you coordinate merging the whole thing with all the fixes under the default FluxML/Metalhead.jl? Very much appreciated. |
I recently failed to import ResNet with ONNX.jl, and then succeeded with ONNXmutable.jl. |
|
The latest release no longer has the I'm closing this issue for now, so that we only have #72 to track the incorrect results. |
Hi, I'm trying to use pretrained model stored in this repository
VGG,ResNet,GoogleNet,DenseNetand so on.
I wrote the following code to see what happens.
This works fine showing a kind ofelephant as result.
How about the other model e.g. DenseNet ?
also ResNet, GoogleNet show similar error.
Does
Metalhead.classifysupport pretrained models other than VGG19 ?The text was updated successfully, but these errors were encountered: