Workshop DogsCats

Part 1 (2018)
1

permission.png1344×758 103 KB

I have tried many things but couldn’t figure the problem out. Yesterday when I ran the notebook locally there was no problem. Today I am using paperspace machine and os.mkdirs doc says that some operating systems act diffrently. But still I have no clue. Appreciate your help thanks

2

I’m making changes at the moment - looks like you did a git pull so you have some new stuff. That particular issue requires a new crestle feature that was just added yesterday, so you’ll need to create a new crestle instance to get this working.

(If you’re not using crestle, you don’t need to use any of those lines - you can just link directly to wherever you download the data to.)

1 Like
3

Thanks for the reply. I am using paperspace and running jupyter notebook through that machine. What I did was, I’ve cloned fastai repo, then inside courses/dl1 I’ve downloaded and unzipped data/dogcats.

Since we are first learning from top to bottom I didn’t went through the library and explore each functionality. But I am really curious about the need of /cache/tmp for ConvLearner.pretrained(…,data) to work where my data PATH is data/dogscats which I’ve unzipped after downloading directly from the link : http://files.fast.ai/data/dogscats.zip

It seems like after running the pretrained model it tries to write something to data/dogscats/tmp which I fail to create to begin with. So would it be ok if I manually create an empty /tmp file inside data/dogscats then run the model ?

Edit: Creating an empty /tmp dir seems to work but still don’t understand the need of /cache/tmp.

Thanks

4

Yup the /cache thing is a special requirement only for crestle.

1 Like
5

Hello,

I have a question about ConvLearner.pretrained object and fit() method .

Are the above statements true for correct usage in general:

  • ConvLearner.pretrained(CNNmodel, data, precompute=False): Creates an object as a CNN, which has last layer of FC modified by data(output size) and CONV + FC weights initialized based on this loaded CNNmodel. By default it returns a learn object which we can apply fit only on FC layers.

## Create a learn object with pretrained weights learn = ConvLearner.pretrained(model , data, precompute = True)

## Allow model to train both Conv and FC weights
## Fit model & updated weights stored in learn object learn.fit()

## freeze to just train FC learn.freeze()

## save model learn.save()

After this every time learn.fit is called it keeps training with latest weights but by only updating FC.

Thanks

6

If you go down towards the end of the notebook there’s a step by step description. We’ll be covering it in detail in our first lesson.

1 Like