Lesson 1 discussion

[jeremy (Jeremy Howard)]

Is this because of the ‘shuffle=true’ in the ‘get_batches’ function?

Exactly right

[jeremy (Jeremy Howard)]

Hi @jeremy, is fine tie something similar to the concept of Transfer Learning

Yes, it is a way to do transfer learning

[jeremy (Jeremy Howard)]

Why is it a better choice to pop the last layer of the network in order to replace it with our own one instead of appending a new layer after the vgg16 layers ?

We will discuss that in lesson 2

[jeremy (Jeremy Howard)]

Hi, within vgg16.py there is a import command
import bcolz. which I don’t have installed

It’s installed automatically in the AWS AMI. If you’re not using that, you’ll need to install it yourself. If you’ve followed our advice and are using anaconda, just type conda install bcolz . If you’re not following our advice and are using anaconda, you may want to reconsider - otherwise you’ll have to figure out how to fix the pip install

[Daniel (深度碎片)]

I prefer iTerm2 instead of the default terminal for everything on my mac

iTerm2 is very useful and seems a lot easier, thanks again

[yad.faeq]

Hey @mclasson What is your setup while running the Model of VGG16?
This might be an issue with not having the supported installations for your machine, either GPU or the drivers related such as cuDnn, are those things in place?

[meftaul (Meftaul Haque)]

Please look at your keras configuration. It might be configured for tensorflow.

I faced the same issue.
To solve the issue I have changed the keras configuration file.
~/.keras/keras.json
{
“image_dim_ordering”: “th”,
“epsilon”: 1e-07,
“floatx”: “float32”,
“backend”: “theano”
}

[mclasson (Marcus Classon)]

Thanks. I had two problems. g++ missing and keras using Tensorflow as you suggested. Now I instead get this one:
Exception: The shape of the input to “Flatten” is not fully defined (got (0, 7, 512). Make sure to pass a complete “input_shape” or “batch_input_shape” argument to the first layer in your model. after printing this

(Subtensor{int64}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Subtensor{int64}.0)
(Subtensor{int64}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Subtensor{int64}.0)
(Subtensor{int64}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Subtensor{int64}.0)
(Subtensor{int64}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Subtensor{int64}.0)
(Subtensor{int64}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Subtensor{int64}.0)
(Subtensor{int64}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Subtensor{int64}.0)
(Subtensor{int64}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Subtensor{int64}.0)
(Subtensor{int64}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Subtensor{int64}.0)
(Subtensor{int64}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Subtensor{int64}.0)
(Subtensor{int64}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Subtensor{int64}.0)
(Subtensor{int64}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Subtensor{int64}.0)
(Subtensor{int64}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Subtensor{int64}.0)
(Subtensor{int64}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Subtensor{int64}.0)

[mclasson (Marcus Classon)]

Beautiful! Helped me a lot. Thanks

[zaiddabaeen]

Is there a way to save the trained model so it can be reused by loading it?

[james_goldfarb (James W Goldfarb)]

You can use model.save(filepath) and model.load_model(filepath) .

See https://keras.io/getting-started/faq/

You can also save the “training”, the model weights, with model.save_weights(filepath) and model.load_weights(filepath).

[martin (Marlon)]

Is the ‘weights’ exactly the same as a model? In another word, does “model_save” function the same as “save_weights”?

[martin (Marlon)]

Training accuracy VS validation accuracy. In lesson 1, I trained on a small portion of the data, a little more than 1000 instances.

Why is the highest training accuracy is under 90%, but the validation accuracy is still over 97%?
1655/1655 [==============================] - 1047s - loss: 0.5973 - acc: 0.8749 - val_loss: 0.1039 - val_acc: 0.9759

I am also using a smaller validation set.

[james_goldfarb (James W Goldfarb)]

model.save(filepath) saves everything.
model.save_weights(filepath) save only the the weights. You need the actual model definition.

[james_goldfarb (James W Goldfarb)]

Why is the highest training accuracy is under 90%, but the validation accuracy is still over 97%?

Keep going with the videos. This is covered in detail in Lesson 3.

[martin (Marlon)]

batch_size = 64
vgg = Vgg16()
path = "data/redux/"
vgg.model.load_model(path+‘results/model.h5’)

I tested with replacing “load_weights” with ‘load_model’, however, I received this error:

Using Theano backend.
Traceback (most recent call last):
File “/home/abigail/workspace/fast-ai/redux_test.py”, line 14, in
vgg.model.load_model(path+‘results/model.h5’)
AttributeError: ‘Sequential’ object has no attribute ‘load_model’

Why can’t it load?

[james_goldfarb (James W Goldfarb)]

I never used it…but…
You probably do not need "vgg = Vgg16()"
and it should be “vgg=”, it should return the model.
vgg=model.load_model(path+‘results/model.h5’)

[martin (Marlon)]

@james_goldfarb, i think i know what’s going on. It would take a while to modify the code to use load and save ‘model’ in lesson one, due to the use of the pretrained model.

[jeremy (Jeremy Howard)]

The shape of the input to “Flatten” is not fully defined

You probably didn’t set image_dim_ordering to th. See the post immediately above yours.

[martin (Marlon)]

Where to enable theano for Keras in Vgg16?

According to Keras documentations, https://keras.io/getting-started/faq/#how-can-i-run-keras-on-gpu. In order to run GPU on theano, it needs to do something like this:

import theano
theano.config.device = 'gpu’
theano.config.floatX = ‘float32’

However, I can’t see where this is set in either Utility.py or vgg16.py. Hi, @jeremy, could you shed some light on this? Just curious.