lr_scheduler_from_optim_params setting

[BraveBoBo]

    elif lr_scheduler_type == "cosine":
        assert net_optim_params["learning_rate"]["step_every_batch"]
        # source: https://github.com/huggingface/diffusers/blob/ee7e141d805b0d87ad207872060ae1f15ce65943/src/diffusers/optimization.py#L154
        num_warmup_steps = net_optim_params["learning_rate"].get("warmup_steps", 0)
        num_training_steps = num_train_batches * num_epochs
        num_cycles = net_optim_params["learning_rate"].get("num_cycles", 0.5) # number of cosine cycles (0 to 2pi) in the LR schedule, default to half-cycle 
        def lr_lambda(current_step):
            if current_step < num_warmup_steps:
                return float(current_step) / float(max(1, num_warmup_steps))
            progress = float(current_step - num_warmup_steps) / float(max(1, num_training_steps - num_warmup_steps))
            return max(0.0, 0.5 * (1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress)))
        return optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)

Traceback (most recent call last):
File "/home/libo/project/VectorizedCollisionLab/scripts/imitation_learning/robomimic/train.py", line 428, in main
train(config, device, log_dir, ckpt_dir, video_dir)
File "/home/libo/project/VectorizedCollisionLab/scripts/imitation_learning/robomimic/train.py", line 220, in train
model = algo_factory(
^^^^^^^^^^^^^
File "/home/libo/project/VectorizedCollisionLab/third_party/robomimic/robomimic/algo/algo.py", line 78, in algo_factory
return algo_cls(
^^^^^^^^^
File "/home/libo/project/VectorizedCollisionLab/third_party/robomimic/robomimic/algo/algo.py", line 133, in init
self._create_optimizers()
File "/home/libo/project/VectorizedCollisionLab/third_party/robomimic/robomimic/algo/algo.py", line 195, in _create_optimizers
self.lr_schedulers[k] = TorchUtils.lr_scheduler_from_optim_params(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/libo/project/VectorizedCollisionLab/third_party/robomimic/robomimic/utils/torch_utils.py", line 170, in lr_scheduler_from_optim_params
num_training_steps = num_train_batches * num_epochs
~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~
TypeError: unsupported operand type(s) for *: 'Config' and 'Config'

[amandlek]

Can you share a way to reproduce this issue?

[BraveBoBo]

Thanks for your response. I’ve resolved the problem. When training with diffusion_policy, I used the provided template robomimic/exps/templates/diffusion_policy.json, but I forgot to include:

"num_epochs": 2000,
"num_train_batches": 100,

[BraveBoBo]

Thanks for your response. I’ve resolved the problem. When training with diffusion_policy, I used the provided template robomimic/exps/templates/diffusion_policy.json, but I forgot to include:

"num_epochs": 2000,
"num_train_batches": 100,

When using "scheduler_type": "cosine", an issue occurs.