Inception V3 gives wrong predictions

[erogol]

I guess there is something wrong about the released network or at least the preprocessing code. I tried to use prediction-with-pretrained example but the results are mistaken.

I also realized that output layer has 1008 nodes where as the label txt has 1001 classes

[antinucleon]

The model is converted from TensorFlow model. Note preprocessing code is different, see preprocessing.py in the zip. Also 1008 is from Google, and 1-1000 is ILVRC2012 label.

[erogol]

I used the provided processing code but still same problem.

[antinucleon]

I am not sure what is your problem, but on my side it works well, and I verified on ILSVRC 2012 validation set and TensorFlow sample image.

[501177639]

I happen the same problem as erogol, could you show the code how you verify on ILSVRC 2012 validation set?

[ghost]

Same here.

[piiswrong]

What does wrong result mean? What is your accuracy on imagenet?

[antinucleon]

First, resize the raw image into 384, then you can do in this way (code I used a month ago)

import mxnet as mx
import numpy as np
val = mx.io.ImageRecordIter(
        path_imgrec = "model/val-384.rec",
        mean_r      = 128,
        mean_g      = 128,
        mean_b      = 128,
        scale       = 0.0078125,
        rand_crop   = False,
        rand_mirror = False,
        data_shape  = (3, 299, 299),
        batch_size  = 128)

symbol, arg_params, aux_params = mx.model.load_checkpoint("model/Inception-7", 1)
model = mx.model.FeedForward(symbol=symbol, ctx=mx.gpu(), arg_params=arg_params, aux_params=aux_params, numpy_batch_size=1)
prob = model.predict(val)


import csv
fi = csv.reader(open("old_synset.txt"), delimiter=' ')
old = {}
idx = 0
for line in fi:
    old[idx] = line[0]
    idx += 1

fi = csv.reader(open("val.lst"), delimiter='\t')
ans = []
for line in fi:
    ans.append(old[int(line[1])])


fi = csv.reader(open("model/synset.txt"), delimiter=' ')
new = {}
idx = 0
for line in fi:
    new[line[0]] = idx
    idx += 1

new_ans = [new[s] for s in ans]

top_1 = 0.
top_5 = 0.
for i in range(len(new_ans)):
    sol = new_ans[i]
    pred_top5 = prob[i, :].argsort()[::-1][:5]
    if pred_top5[0] == sol:
        top_1 += 1
    if sol in pred_top5:
        top_5 += 1
print(top_1 / 50000)
print(top_5 / 50000)

[luxiangju]

i couldn't download the model, who can share the model with me?

[ghost]

synset.txt might wrong in this model. This leads to a wrong mapping.

[erogol]

Accuracy is not important since the results are obviously flawed for couple of obvious class images that my previous net solves successfully. I also believe that the synset.txt is wrong since number of output nodes and the number of lines in the synset is not matching.

[antinucleon]

The synset is correct. Again, in Google's released model, there is only 1008 outputs. There is a mapping in old synset and new synset, which I have provided code above. If it wrong, it can't produce 77% accuracy.

[ghost]

The question is where can we find the old_synset.txt?

[antinucleon]

You can find old_synset.txt in old Inception-BN model.

On Thu, Jan 14, 2016 at 4:07 PM, Shuo Zhang [email protected]
wrote:

The question is where can we find the old_synset.txt?

—
Reply to this email directly or view it on GitHub
#7 (comment)
.

[501177639]

I use the code above to test the validation set again, but still get the very low accuracy. I guess there are something wrong in either the Inception-7-0001.params or the synset.txt. Can you test again and add your code here https://s3.amazonaws.com/dmlc/model/inception-v3.tar.gz ,so we can run it directly to get the right result?

[erogol]

I observe that the model works fine with cpu but not with gpu. All top predictions are skewed in gpu setting.

[u1234x1234]

I've also had some problems with this model: it gaves me wrong predictions with cudnn v3. mxnet without cudnn, with cudnn v4, and cpu version worked fine for me.

[501177639]

something wrong with cudnn v3?

[erogol]

Yeap I update to cudnn 4 and problem barely resolved. Thanks for pointing @u1234x1234 .

But still gpu execution gives different top5 ordering in relation to cpu. At least results make sense for both cases.