Using fine-tuned model in Kaggle Dog competition #226
Conversation
| for aug in auglist: | ||
| im = aug(im) | ||
| # 将数据格式从"高*宽*通道"改为"通道*高*宽"。 | ||
| im = nd.transpose(im, (2,0,1)) | ||
| return (im, nd.array([label]).asscalar().astype('float32')) | ||
|
|
||
| def transform_test(data, label): | ||
| im = image.imresize(data.astype('float32') / 255, 96, 96) | ||
| im = image.imresize(data.astype('float32') / 255, 256, 256) |
There was a problem hiding this comment.
Why do we need to resize to 256, 256 for test? Can we do the following:
im = data.astype('float32') / 255
before invoking image.imresize?
There was a problem hiding this comment.
It's a ImageNet benchmarking standard, we can omit it here for simplicity.
| 我们这里使用了[ResNet-18](resnet-gluon.md)模型。我们使用[hybridizing](../chapter_gluon-advances/hybridize.md)来提升执行效率。 | ||
| 我们这里借助[迁移学习](fine-tuning.md)的思想，调用`Gluon`中经过预训练的[ResNet-34](resnet-gluon.md)模型，并通过微调在新数据集上进行训练。 | ||
|
|
||
| 这里，我们使用与前述教程略微不同的迁移学习方法。在新的训练数据与预训练数据相似的情况下，我们认为原有特征是可重用的。基于这个原因，在一个预训练好的新模型上，我们可以不去改变原已训练好的权重，而是在原网络结构上新加一个小的输出网络。这样能够节省后向传播的时间，也能避免了在特征层上储存梯度所需要的内存空间。 |
There was a problem hiding this comment.
The audience are mostly users without any background on deep learning. Thus, we need to make it even clearer. For instance, ResNet-34 is pre-trained on ImageNet. Our problem is a subset of ImageNet. So we can apply this pre-trained model.
For "而是在原网络结构上新加一个小的输出网络", we should explicitly say that parameters of this layer need to be trained.
|
|
||
| 注意，我们在之前定义的数据预处理函数用了ImageNet数据集上的均值和标准差做标准化，这样才能保证预训练模型能够捕捉正确的数据特征。 | ||
|
|
||
|  |
There was a problem hiding this comment.
In images, pretrained -> pre-trained
| print('Block %d output: %s'%(i+1, out.shape)) | ||
| return out | ||
|
|
||
| from mxnet.gluon.model_zoo import vision as models | ||
|
|
||
| def get_net(ctx): |
There was a problem hiding this comment.
Within get_net, we need to add comments to explain, such as line by line (if necessary).
|
@astonzhang PR is updated accordingly. |
|
LGTM. Just let me know when you got the performance after 30 epochs. |
|
The full training process uses roughly 4GB GPU ram with batch=128. The accuracy is at 0.446 on the leaderboard. |
|
Good. Thanks! |
By setting
epoch=30, one can achieve score ~0.49 on the leaderboard.@astonzhang Please review.