Skip to content

Roi#1

Open
roisantos wants to merge 33 commits intomasterfrom
roi
Open

Roi#1
roisantos wants to merge 33 commits intomasterfrom
roi

Conversation

@roisantos
Copy link
Owner

No description provided.

vvvlv and others added 30 commits March 5, 2025 18:48
Dynamic Zero-Rate Registration:
All layers in dict_module now get zero-rate, dead neuron rate, and activation stats hooks.
Dead neuron rate is tracked and logged per epoch.
Activation & Weight Histograms:
Activation histograms are logged for all tracked layers (every heavy epoch).
Weight histograms are already logged.
IoU Metric:
IoU is now computed and returned in traverseDataset and logged/printed in the training loop.
CosineAnnealingLR:
Added as the learning rate scheduler in the training loop, stepped every epoch.
…, Merging Inputs, and PCh Outputs

 Add channel_convert to dict_module in SantosNet Variants

The main "correction" to ensure this works robustly, especially with GPU memory, is to move the tensor to the CPU before appending it to self.activations list within the hook. This prevents a large list of GPU tensors from accumulating if clone() still keeps it on the same device.

Importante: Cambio scripts a SimpleResidualBlock y añado  Logging Skip Connections, Merging Inputs, and PCh Outputs
The EpochActivationStats class has now been updated to only store the last seen activations. This should significantly reduce memory usage related to histogram logging and hopefully resolve the OOM error. The histogram tag is now hist_last_batch to reflect this change.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants

Comments