Lesson 4 IMDB: can't run the training code

Part 1 (2018)
1

I’m running the notebook for lesson 4, without any changes: https://github.com/fastai/fastai/blob/master/courses/dl1/lesson4-imdb.ipynb

But this code:
learner.fit(3e-3, 4, wds=1e-6, cycle_len=1, cycle_mult=2)
from the section on training the model produces a RuntimeError:

RuntimeError: shape ‘[1000000, 1]’ is invalid for input of size 2000

How can I fix this ?

2

Apparently there can be some problems related to the torch and fastai versions: Lesson 4 IMDB Test Part Fails - #2 by EinAeffchen

I just used the installation commands on fastai’s github.
Conda says that the following are my installed packages:

fastai 1.0.28 py_1 fastai
pytorch-nightly 1.0.0.dev20181121 py3.6_cuda9.0.176_cudnn7.4.1_0 pytorch
torchvision-nightly 0.2.1 py_0 fastai
torchtext 0.3.1

3

Yup, it’s the packages.

On the fastai github, you’re instructed to install nightly versions. But that brings torch 1.
Fastai expects torch 0.4

As far as I can see, the correct way of installing fastai is now:
conda install pytorch torchvision -c pytorch
conda install -c fastai
pip install torchtext

This will give you
pytorch 0.4.1
torchvision 0.2.1
torchtext 0.3.1
fastai 1.0.28

I’ll test my setup a bit more (by running more sample code).
But as far as I can see, this is the working combination.

4

Le me start saying I have the same problems as you: dimensions don’t match.

Fastai v1 expects pytorch v1. The lesson about IMDB is taught with fastai v1 / torch 1.0

Besides, I find the same problem with fastai 0.7, pytorch 0.4. Note that nobody who asked on the forums about that issued received an answer whatsoever.

5

Really hope somebody can give an answer!

6

Looks like beginners just got to fend for themselves! Here’s the solution I’ve got: create an additional environment so that both versions of fast.ai and both versions of dl1 can run on the same instance. To switch, I just need to select from Jupyter’s pull-down menu, no need to quit an instance and start another. Step-by-step instructions I’m pasting here:
http://www.marychin.org/machinelearning/2018/12/13/my-ultimate-fast-ai-instance/
in case fellow newbies find themselves caught between the two versions: (fast.ai 0.7.x, course/dl1) and (fast.ai 1.0.?, courses-v3/nbs/dl1). Worth the trouble, as there are goodies in the earlier course which cannot be found in the newer one. And, obviously, there’s a video series accompanying the earlier but not the later.

7

same problem occur when i run:
fit(lr=lr, n_cycle=n_cycle,wds=wd,cycle_len=cycle_len,use_clr=(32,10),cycle_mult=cycle_mult,_save_name = ‘language_model’)

Traceback (most recent call last):
File “E:/Mindtree_IDP/neural_language_model.py”, line 41, in
wd=1e-6)
File “E:\Mindtree_IDP\lang_model_utils.py”, line 244, in train_lang_model
best_save_name = ‘language_model’
File “C:\Users\GK\AppData\Roaming\Python\Python36\site-packages\fastai-0.7.0-py3.6.egg\fastai\learner.py”, line 287, in fit
return self.fit_gen(self.model, self.data, layer_opt, n_cycle, **kwargs)
File “C:\Users\GK\AppData\Roaming\Python\Python36\site-packages\fastai-0.7.0-py3.6.egg\fastai\learner.py”, line 234, in fit_gen
swa_eval_freq=swa_eval_freq, **kwargs)
File “C:\Users\GK\AppData\Roaming\Python\Python36\site-packages\fastai-0.7.0-py3.6.egg\fastai\model.py”, line 132, in fit
loss = model_stepper.step(V(x),V(y), epoch)
File “C:\Users\GK\AppData\Roaming\Python\Python36\site-packages\fastai-0.7.0-py3.6.egg\fastai\model.py”, line 50, in step
output = self.m(*xs)
File “c:\users\gk\appdata\local\continuum\anaconda3\envs\mindtree\lib\site-packages\torch\nn\modules\module.py”, line 489, in call
result = self.forward(*input, **kwargs)
File “c:\users\gk\appdata\local\continuum\anaconda3\envs\mindtree\lib\site-packages\torch\nn\modules\container.py”, line 92, in forward
input = module(input)
File “c:\users\gk\appdata\local\continuum\anaconda3\envs\mindtree\lib\site-packages\torch\nn\modules\module.py”, line 489, in call
result = self.forward(*input, **kwargs)
File “C:\Users\GK\AppData\Roaming\Python\Python36\site-packages\fastai-0.7.0-py3.6.egg\fastai\lm_rnn.py”, line 97, in forward
raw_output, new_h = rnn(raw_output, self.hidden[l])
File “c:\users\gk\appdata\local\continuum\anaconda3\envs\mindtree\lib\site-packages\torch\nn\modules\module.py”, line 489, in call
result = self.forward(*input, **kwargs)
File “C:\Users\GK\AppData\Roaming\Python\Python36\site-packages\fastai-0.7.0-py3.6.egg\fastai\rnn_reg.py”, line 122, in forward
return self.module.forward(*args)
File “c:\users\gk\appdata\local\continuum\anaconda3\envs\mindtree\lib\site-packages\torch\nn\modules\rnn.py”, line 179, in forward
self.dropout, self.training, self.bidirectional, self.batch_first)
RuntimeError: shape ‘[1000000, 1]’ is invalid for input of size 2000