Weights densenet #1855
Weights densenet #1855
Conversation
| x = keras.layers.Rescaling(scale=1 / 255.0)(image_input) | ||
| x = keras.layers.Normalization( | ||
| axis=channel_axis, | ||
| mean=(0.485, 0.456, 0.406), |
There was a problem hiding this comment.
why do we need a normalization layer in addition to rescaling, and where are these constants coming from?
There was a problem hiding this comment.
These are from the standard Imagenet mean and std, which is also mentioned here https://huggingface.co/timm/densenet201.tv_in1k/blob/main/config.json.
In the Torch model, they don't have the normalization applied, we need to apply this for the channel axis.
Without the normalization to timm model, the difference is 0.6 in 201 variation, after applying normalization, the difference dropped down to 0.3.
There was a problem hiding this comment.
Here is how timm is applying the configs, transforms = timm.data.create_transform(**data_config, is_training=False)
There was a problem hiding this comment.
Interesting. Let me chat with @fchollet about this, but this is making me think we might want to push rescaling and this normalization to our preprocessing layers, and remove include_rescaling everywhere. That seems conceptually cleaner, and it will give us a good way to propagate these as config instead of hardcoding them.
Let me chat with him.
There was a problem hiding this comment.
Okay, we may also have to check the
"interpolation": "bicubic",
"crop_pct": 0.95,
``` which timm is using in many of the models so that the results comes almost closer to the timm/ benchmark result.
There was a problem hiding this comment.
I don't think we need to match crop percent. Better to do simple rescaling of user image than match their crop. Interpolation matching might be worth doing?
|
Also, looks like some legit test failures to look into. |
|
#1859 is in, so this just needs to base over those changes, reupload kaggle presets, and fix these failing tests. |
|
Uploaded the Latest weights and configurations to DenseNet Kaggle https://www.kaggle.com/models/kerashub/densenet |
This reverts commit f67b4db.
Tested Colab: