[ { "function_name": "collect_grads", "file_path": "collect_grad_reps.py", "function_calls": [ { "function_call_name": "torch.random.manual_seed", "functionality": "Sets the seed for generating random numbers in PyTorch for reproducibility.", "input": "seed (int)", "output": "None", "location_or_namespace": "torch" }, { "function_call_name": "torch.stack", "functionality": "Concatenates a sequence of tensors along a new dimension.", "input": "sequence of tensors", "output": "stacked tensor", "location_or_namespace": "torch" }, { "function_call_name": "to", "functionality": "Casts a tensor to the specified type/device.", "input": "dtype, device, or both", "output": "converted tensor", "location_or_namespace": "torch.Tensor" }, { "function_call_name": "projector.project", "functionality": "Projects the input gradient tensor into a lower-dimensional subspace using a random projection matrix.", "input": "gradient tensor, model_id (int)", "output": "projected gradient tensor", "location_or_namespace": "trak.projectors.BasicProjector or CudaProjector" }, { "function_call_name": "cpu", "functionality": "Moves a tensor to CPU memory.", "input": "None", "output": "tensor on CPU", "location_or_namespace": "torch.Tensor" }, { "function_call_name": "torch.cat", "functionality": "Concatenates a sequence of tensors along an existing dimension.", "input": "sequence of tensors", "output": "concatenated tensor", "location_or_namespace": "torch" }, { "function_call_name": "os.path.join", "functionality": "Joins one or more path components intelligently.", "input": "directory path(s), file name(s)", "output": "joined path string", "location_or_namespace": "os" }, { "function_call_name": "torch.save", "functionality": "Saves an object to a disk file.", "input": "object, file path", "output": "None", "location_or_namespace": "torch" }, { "function_call_name": "print", "functionality": "Prints the given object(s) to the console or standard output.", "input": "object(s), optional flush parameter", "output": "None", "location_or_namespace": "builtins" }, { "function_call_name": "next", "functionality": "Retrieves the next item from an iterator.", "input": "iterator", "output": "next item from iterator", "location_or_namespace": "builtins" }, { "function_call_name": "model.parameters", "functionality": "Returns an iterator over model parameters.", "input": "None", "output": "iterator over parameters", "location_or_namespace": "torch.nn.Module" }, { "function_call_name": "assert", "functionality": "Raises an AssertionError if the condition is false.", "input": "condition", "output": "None or exception", "location_or_namespace": "builtins" }, { "function_call_name": "prepare_optimizer_state", "functionality": "", "input": "model, adam_optimizer_state, device", "output": "m, v (momentum statistics)", "location_or_namespace": "collect_grad_reps.py" }, { "function_call_name": "get_trak_projector", "functionality": "", "input": "device", "output": "projector class/factory", "location_or_namespace": "collect_grad_reps.py" }, { "function_call_name": "get_number_of_params", "functionality": "", "input": "model", "output": "number of parameters (int)", "location_or_namespace": "collect_grad_reps.py" }, { "function_call_name": "os.makedirs", "functionality": "Creates a directory recursively; does not raise an error if directory exists.", "input": "directory path, exist_ok (bool)", "output": "None", "location_or_namespace": "os" }, { "function_call_name": "get_max_saved_index", "functionality": "", "input": "output_dir, prefix", "output": "max index (int)", "location_or_namespace": "collect_grad_reps.py" }, { "function_call_name": "tqdm", "functionality": "Wraps an iterable with a progress bar for iteration.", "input": "iterable, total", "output": "iterator with progress bar", "location_or_namespace": "tqdm" }, { "function_call_name": "prepare_batch", "functionality": "", "input": "batch", "output": "", "location_or_namespace": "collect_grad_reps.py" }, { "function_call_name": "obtain_gradients_with_adam", "functionality": "", "input": "model, batch, m, v", "output": "gradient vector", "location_or_namespace": "collect_grad_reps.py" }, { "function_call_name": "obtain_sign_gradients", "functionality": "", "input": "model, batch", "output": "gradient vector", "location_or_namespace": "collect_grad_reps.py" }, { "function_call_name": "obtain_gradients", "functionality": "", "input": "model, batch", "output": "gradient vector", "location_or_namespace": "collect_grad_reps.py" }, { "function_call_name": "model.zero_grad", "functionality": "Sets gradients of all model parameters to zero.", "input": "None", "output": "None", "location_or_namespace": "torch.nn.Module" }, { "function_call_name": "torch.cuda.empty_cache", "functionality": "Releases all unused cached memory from CUDA.", "input": "None", "output": "None", "location_or_namespace": "torch.cuda" }, { "function_call_name": "merge_and_normalize_info", "functionality": "", "input": "output_dir, prefix", "output": "", "location_or_namespace": "collect_grad_reps.py" }, { "function_call_name": "merge_info", "functionality": "", "input": "output_dir, prefix", "output": "", "location_or_namespace": "collect_grad_reps.py" } ] } ]