"Error(s) in loading state_dict for Model:" when I try to export a model edited with a recipe about sparseml HOT 10 CLOSED

Hi @antimo22, this recipe was created originally for the COCO dataset. Other datasets can require more or fewer epochs and this is a hyperparameter that you should feel free to edit. We're also going to be updating some new flows to make the editing of these hyperparams in recipes easier.

If you have a chance, could you try and run one of the transfer recipes such as this one: https://github.com/neuralmagic/sparseml/blob/main/integrations/ultralytics-yolov5/recipes/yolov5.transfer_learn_pruned_quantized.md This will run in a shorter amount of time. If you'd like to run for fewer epochs then we'd recommend editing the recipe and changing the num_epochs and the quantization_start_epoch to smaller values that meet your requirements.

from sparseml.

Hi @antimo22 could you share the recipe you applied as well as the number of epochs the model was trained for?

from sparseml.

Hi @bfineran, thank you for your reply. I applied this recipe: yolov5s.pruned_quantized.md (https://github.com/neuralmagic/sparseml/blob/main/integrations/ultralytics-yolov5/tutorials/sparsifying_yolov5_using_recipes.md)

I tried to use the weights of the epoch1, but also of next epochs (I've noticed that the weights size are different, they go from 86MB to 114MB).

BTW, I had another problem, related to this error: "cannot import name 'container_abcs' from 'torch._six'"
That I solved importing container_abcs from collections.abc
But I think this is related to some python version incompatibility.

from sparseml.

Hi @antimo22 it looks like you are trying to export a model with a quantization recipe with non quantized weights (ie weights obtained from epoch 1). This will cause issues because the export script will expect the checkpoint to have quantized weights.

When you notice the weight file size increase, this is likely when quantization begins as extra parameters are injected to track the quantization (such as "model.0.conv.conv.quant.activation_post_process.scale"). Have you tried exporting from one of these checkpoints?

from sparseml.

Yes, I tried to export the model from the epoch 6, that is around 114 MB. Should I have to continue the training for more epochs?

from sparseml.

stopping early depends on the amount of recovery you are looking to get. Additionally, to get a quantized checkpoint, you must stop after quantization has been applied in the recipe. Could you point me to the exact recipe you are using? If quantization is applied before epoch 6, then you should be fine to export, recovery aside.

from sparseml.

This is the recipe I've applied: https://github.com/neuralmagic/sparseml/blob/main/integrations/ultralytics-yolov5/recipes/yolov5s.pruned_quantized.md

I see something like pruning_end_epoch 100, so I have to train for 100 epochs to end the pruning process?
BTW, I used also the epoch 10 with weights around 140MB, still the same error.

from sparseml.

Hi @antimo22
Checking in here to see if you were able to try out one of our recommended recipes? We're happy to help.
Thanks!
Jeannie / Neural Magic

from sparseml.

Hello @antimo22
As there have been no further comments on this, we will close out this issue. Feel free to re-open it however if you would like to continue the conversation. Thank you!
Jeannie / Neural Magic

from sparseml.

To anyone facing the same issue - I had the same problem as @antimo22 and confirm that the solution by @markurtz works!

from sparseml.