Notebooks update

[shaydeci]

Segmentation: transfer learning, quickstart and custom dataset connection.
Updated readme links + makefile

[Louis-Dupont]

I didnt manage to comment directly on the notebooks, so here are a few notes

1. This cell doesnt print anything. Is it on purpose (cleared) or unintended ?

from prettyformatter import pprint

print('Dataloader parameters:')
pprint(train_loader.dataloader_params)
print('Dataset parameters')
pprint(train_loader.dataset.dataset_params)

*Note: I just saw that in transfer_learning_semantic_segmentation.ipynb is it shown but in the next cell

2. We are not using infery, so lett's remove it

! pip install -qq infery

[shaydeci]

I didnt manage to comment directly on the notebooks, so here are a few notes

1. This cell doesnt print anything. Is it on purpose (cleared) or unintended ?

from prettyformatter import pprint

print('Dataloader parameters:')
pprint(train_loader.dataloader_params)
print('Dataset parameters')
pprint(train_loader.dataset.dataset_params)

*Note: I just saw that in transfer_learning_semantic_segmentation.ipynb is it shown but in the next cell

2. We are not using infery, so lett's remove it

! pip install -qq infery

Removed infery refs in 5d2c93f

[BloodAxe]

Few notes on my end:

1. prettyformatter is not needed and we can remove it, since there is pprint package with is supported in all python 3.x versions and does the same stuff as prettyformatter

2. Printing to stdout is broken in our notebooks, when using DataLoaders, I have a fix for that which I can do a

3. Quickstart segmentation notebook uses ema setting without explicitly specifying decay_type, which I think we should change to silence the warning

4. Quickstart segmentation notebook installs onnx-simplifier without any version specifier, which is a) bad practice b) unnecessary since we have onnx-simplifier already part of SG dependencies. (See section 7) I suggest removing this install.

5. Quickstart segmentation notebook installs torch 1.12 (See section 7) without saying a word why we ask for this specific version. I suggest removing this install as well since it's probably comes from very very old times.

[BloodAxe]

6. Export call uses opset_version=11 which is pretty old, suggest to remove this explicit opset argument at all.

[BloodAxe]

segmentation_connect_custom_dataset.ipynb

1. Same with prettyformatter
2. Strange naming of sections 2.B Create Torch Dataset & 2.C Create Torch Dataloader. Why not 2.1 / 2.2? My first take was this some kind of typo
3. Notebook emits warning DeprecationWarning: Object name cross_entropyis now deprecated. Please replace it withCrossEntropyLoss.

[BloodAxe]

notebooks/transfer_learning_semantic_segmentation.ipynb

1. prettyformatter

2. SUPERVISELY_DATASET_DOWNLOAD_PATH="/home/data" has two issues. First - is requires Linux/Mac OS, second - it assumes that /home is writable which may not be the case (Since it's home folder for all users and not the current one). Let's avoid using absolute directories and use os.cwd() or simply ./ and build our path from it instead. We want to follow the rule that notebook should be able to be executed "as is" without any changes in cells.

3. supervisely_dataset_dir_path should be constructed using os.path.join or pathlib.Path and not via string concatenation. It's just a bad practice.

4. Notebook emits DeprecationWarning: Object name bce_dice_lossis now deprecated. Please replace it withBCEDiceLoss.

5. Notebook emits WARNING - ema.py - Parameter decayis not specified for EMA params. Please specifydecay parameter explicitly in your config