Jump of loss when resume training from checkpoint

[BZandi]

I noticed that the loss jumps every time I resume the training (see figure). I use 'resume_from_checkpoint' to continue the training. The parameters seem to load correctly, including the optimisation state. Does anyone know where this behaviour comes from, or is it normal? Thanks in advance for any ideas

# Save model
trainer.save_checkpoint('Path')

# Load model and resume training
Model = FF_Class.FeedForwardShuffle.load_from_checkpoint('Path to Checkpoint')
trainer = pl.Trainer(resume_from_checkpoint=Max_Checkpoint,...)

# Start training
trainer.fit(Model)