M3d-Cam with 3DCNN-RNN

[nasir3843]

Hi,
My designed network is 3D CNN followed by RNN.

I am trying to use M3D-Cam to get attention maps from the last convolution layer of the proposed network however I get the following error. Kindly have a look and guide me how can I achieve it

model = 3DCNN_RNN().to(device)
weights_pth = 'trained_weights'
model.load_state_dict(torch.load(weights_pth))
save_pth = 'model_att_maps/'

model_c = medcam.inject(model, backend='gcam', replace = True, label = 1 ,layer = 'conv4')
model_c.eval()
for (img1, _ , label) in train_dataloader:
img_vol = img1.unsqueeze(1)
img_cuda = img_vol.float().to(device)
output = model_c(img_cuda)
break

I get the following error

<< ======================================================>>

RuntimeError Traceback (most recent call last)
in
54 img_vol = img1.unsqueeze(1)
55 img_cuda = img_vol.float().to(device)
---> 56 output = model_c(img_cuda)
57
58 break

~/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
725 result = self._slow_forward(*input, **kwargs)
726 else:
--> 727 result = self.forward(*input, **kwargs)
728 for hook in itertools.chain(
729 _global_forward_hooks.values(),

~/miniconda3/lib/python3.8/site-packages/medcam/medcam_inject.py in forward(self, batch, label, mask, raw_input)
201 if self.medcam_dict['enabled']:
202 with torch.enable_grad():
--> 203 self.test_run(batch, internal=True)
204 if self.medcam_dict['layer'] == 'full' and not self.medcam_dict['tested']:
205 raise ValueError("Layer mode 'full' requires a test run either during injection or by calling test_run() afterwards")

~/miniconda3/lib/python3.8/site-packages/medcam/medcam_inject.py in test_run(self, batch, internal)
230 if batch is not None and not self.medcam_dict['tested']:
231 with torch.enable_grad():
--> 232 _ = self.medcam_dict['model_backend'].generate_attention_map(batch, None)
233 registered_hooks = self.medcam_dict['model_backend'].get_registered_hooks()
234 self.medcam_dict['tested'] = True

~/miniconda3/lib/python3.8/site-packages/medcam/backends/base.py in generate_attention_map(self, batch, label)
20 """Handles the generation of the attention map from start to finish."""
21 output = self.forward(batch)
---> 22 self.backward(label=label)
23 attention_map = self.generate()
24 return output, attention_map, self.output_batch_size, self.output_channels, self.output_shape

~/miniconda3/lib/python3.8/site-packages/medcam/backends/base.py in backward(self, label)
40
41 self.mask = self._isolate_class(self.logits, label)
---> 42 self.logits.backward(gradient=self.mask, retain_graph=self.retain_graph)
43 self.remove_hook(forward=True, backward=True)
44

~/miniconda3/lib/python3.8/site-packages/torch/tensor.py in backward(self, gradient, retain_graph, create_graph)
219 retain_graph=retain_graph,
220 create_graph=create_graph)
--> 221 torch.autograd.backward(self, gradient, retain_graph, create_graph)
222
223 def register_hook(self, hook):

~/miniconda3/lib/python3.8/site-packages/torch/autograd/init.py in backward(tensors, grad_tensors, retain_graph, create_graph, grad_variables)
128 retain_graph = create_graph
129
--> 130 Variable.execution_engine.run_backward(
131 tensors, grad_tensors, retain_graph, create_graph,
132 allow_unreachable=True) # allow_unreachable flag

RuntimeError: cudnn RNN backward can only be called in training mode