update

zero-cost-nas/foresight/pruners/measures/ntk.py

@@ -0,0 +1,94 @@
+# Copyright 2021 Samsung Electronics Co., Ltd.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# =============================================================================
+
+import torch
+import numpy as np
+
+from . import measure
+
+
+def recal_bn(network, inputs, targets, recalbn, device):
+    for m in network.modules():
+        if isinstance(m, torch.nn.BatchNorm2d):
+            m.running_mean.data.fill_(0)
+            m.running_var.data.fill_(0)
+            m.num_batches_tracked.data.zero_()
+            m.momentum = None
+    network.train()
+    with torch.no_grad():
+        for i, (inputs, targets) in enumerate(zip(inputs, targets)):
+            if i >= recalbn: break
+            inputs = inputs.cuda(device=device, non_blocking=True)
+            _, _ = network(inputs)
+    return network
+
+
+def get_ntk_n(inputs, targets, network, device, recalbn=0, train_mode=False, num_batch=1):
+    device = device
+    # if recalbn > 0:
+    #     network = recal_bn(network, xloader, recalbn, device)
+    #     if network_2 is not None:
+    #         network_2 = recal_bn(network_2, xloader, recalbn, device)
+    network.eval()
+    networks = []
+    networks.append(network)
+    ntks = []
+    # if train_mode:
+    #     networks.train()
+    # else:
+    #     networks.eval()
+    ######
+    grads = [[] for _ in range(len(networks))]
+    for i in range(num_batch):
+        if num_batch > 0 and i >= num_batch: break
+        inputs = inputs.cuda(device=device, non_blocking=True)
+        for net_idx, network in enumerate(networks):
+            network.zero_grad()
+            # print(inputs.size())
+            inputs_ = inputs.clone().cuda(device=device, non_blocking=True)
+            logit = network(inputs_)
+            if isinstance(logit, tuple):
+                logit = logit[1]  # 201 networks: return features and logits
+            for _idx in range(len(inputs_)):
+                logit[_idx:_idx + 1].backward(torch.ones_like(logit[_idx:_idx + 1]), retain_graph=True)
+                grad = []
+                for name, W in network.named_parameters():
+                    if 'weight' in name and W.grad is not None:
+                        grad.append(W.grad.view(-1).detach())
+                grads[net_idx].append(torch.cat(grad, -1))
+                network.zero_grad()
+                torch.cuda.empty_cache()
+    ######
+    grads = [torch.stack(_grads, 0) for _grads in grads]
+    ntks = [torch.einsum('nc,mc->nm', [_grads, _grads]) for _grads in grads]
+    for ntk in ntks:
+        eigenvalues, _ = torch.linalg.eigh(ntk)  # ascending
+        conds = np.nan_to_num((eigenvalues[0] / eigenvalues[-1]).item(), copy=True, nan=100000.0)
+    return conds
+
+@measure('ntk', bn=True)
+def compute_ntk(net, inputs, targets, split_data=1, loss_fn=None):
+    device = inputs.device
+    # Compute gradients (but don't apply them)
+    net.zero_grad()
+
+
+    try:
+        conds = get_ntk_n(inputs, targets, net, device)
+    except Exception as e:
+        print(e)
+        conds= np.nan
+
+    return conds