I’m working on Lesson 2, following along with the videos. The current lesson2-download notebook is different than the one covered in the class recording, so I have some questions.
After cleaning up using `ImageCleaner()’ my understanding is that the deletion and re-labeling has not occurred on the original downloaded dataset. Rather, a new .CSV file is created that outlines these changes.
So I went to the docs to find out how to make a data bunch from csv:
cleaned_data = ImageDataBunch.from_csv(path, csv_labels='cleaned.csv', ds_tfms=get_transforms(), size=224).normalize(imagenet_stats)
Then I created a new learner:
learner_cln = cnn_learner(cleaned_data, models.resnet34, metrics=error_rate)
learner_cln.load('stage-1')
learner_cln.unfreeze()
learner_cln.fit_one_cycle(4, max_lr=slice(1e-3, 1e-1))
resulting in very bad error when I began training again:
| epoch | train_loss | valid_loss | error_rate | time |
|---|---|---|---|---|
| 0 | 4.679137 | 1474541.000000 | 0.804688 | 01:14 |
| 1 | 4.246964 | 445326.031250 | 0.890625 | 01:14 |
| 2 | 3.991158 | 2795.947754 | 0.796875 | 01:14 |
| 3 | 3.548282 | 44.752438 | 0.648438 | 01:14 |
So, did I misunderstand how to do this part? The directions are very vague in the new lesson notebook, so I figure I just don’t know what I’m doing.
