remove compute_speedup call from tutorials

Author: guilhermeleobas

intermediate_source/ensembling.py

@@ -27,7 +27,6 @@
 import torch.nn.functional as F
 from torch._dynamo import config
 config.inline_inbuilt_nn_modules = 1
-import profile_utils
 torch.manual_seed(0)
 
 # Here's a simple MLP
@@ -133,7 +132,6 @@ def compute_predictions1(params, buffers, minibatches):
     return vmap(fmodel)(params, buffers, minibatches)
 
 predictions1_vmap = compute_predictions1(params, buffers, minibatches)
-profile_utils.compute_speedup(compute_predictions1, (params, buffers, minibatches), device)
 
 # verify the ``vmap`` predictions match the
 assert torch.allclose(predictions1_vmap, torch.stack(predictions_diff_minibatch_loop), atol=1e-3, rtol=1e-5)
@@ -150,7 +148,6 @@ def compute_predictions2(params, buffers, minibatch):
     return vmap(fmodel, in_dims=(0, 0, None))(params, buffers, minibatch)
 
 predictions2_vmap = compute_predictions2(params, buffers, minibatch)
-profile_utils.compute_speedup(compute_predictions2, (params, buffers, minibatch), device)
 
 assert torch.allclose(predictions2_vmap, torch.stack(predictions2), atol=1e-3, rtol=1e-5)
 

intermediate_source/neural_tangent_kernels.py

@@ -22,7 +22,6 @@
 """
 
 import torch
-import profile_utils
 import torch.nn as nn
 from torch.func import functional_call, vmap, vjp, jvp, jacrev
 from torch._dynamo import config
@@ -117,7 +116,6 @@ def empirical_ntk_jacobian_contraction(fnet_single, params, x1, x2):
 
 result = empirical_ntk_jacobian_contraction(fnet_single, params, x_train, x_test)
 print(result.shape)
-profile_utils.compute_speedup(empirical_ntk_jacobian_contraction, (fnet_single, params, x_train, x_test), device)
 
 ######################################################################
 # In some cases, you may only want the diagonal or the trace of this quantity,
@@ -154,7 +152,6 @@ def empirical_ntk_jacobian_contraction(fnet_single, params, x1, x2, compute):
 
 result = empirical_ntk_jacobian_contraction(fnet_single, params, x_train, x_test, 'trace')
 print(result.shape)
-profile_utils.compute_speedup(empirical_ntk_jacobian_contraction, (fnet_single, params, x_train, x_test, 'trace'), device)
 
 ######################################################################
 # The asymptotic time complexity of this method is :math:`N O [FP]` (time to
@@ -236,7 +233,6 @@ def get_ntk_slice(vec):
 with torch.backends.cudnn.flags(allow_tf32=False):
     result_from_jacobian_contraction = empirical_ntk_jacobian_contraction(fnet_single, params, x_test, x_train, 'full')
     result_from_ntk_vps = empirical_ntk_ntk_vps(fnet_single, params, x_test, x_train, 'full')
-    profile_utils.compute_speedup(empirical_ntk_ntk_vps, (fnet_single, params, x_train, x_test, 'full'), device)
 
 assert torch.allclose(result_from_jacobian_contraction, result_from_ntk_vps, atol=1e-5)
 

intermediate_source/per_sample_grads.py

@@ -16,7 +16,6 @@
 
 """
 
-import profile_utils
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -171,7 +170,6 @@ def vmap_ft_compute_grad(params, buffers, data, targets):
     return ft_compute_sample_grad_(params, buffers, data, targets)
 
 ft_per_sample_grads = vmap_ft_compute_grad(params, buffers, data, targets)
-profile_utils.compute_speedup(vmap_ft_compute_grad, (params, buffers, data, targets), device)
 
 ######################################################################
 # we can double check that the results using ``grad`` and ``vmap`` match the