extend corr_stop function to return a weighted mean of correlations #16
Conversation
nnfabrik/training/trainers.py
Outdated
| target, output = np.array([]), np.array([]) | ||
| target, output = torch.empty(0), torch.empty(0) | ||
| for images, responses in loader[data_key]: | ||
| output = torch.cat((output, (model(images.to(device), data_key).detach().cpu())), dim=0) |
There was a problem hiding this comment.
when you pass in data_key into the model, be sure to pass it as keyword argument (e.g. data_key=data_key) rather than a positional argument.
| """ | ||
| computes model predictions for a given dataloader and a model | ||
| Returns: | ||
| target: ground truth, i.e. neuronal firing rates of the neurons | ||
| output: responses as predicted by the network | ||
| """ | ||
| target, output = np.array([]), np.array([]) | ||
| target, output = torch.empty(0), torch.empty(0) | ||
| for images, responses in loader[data_key]: |
There was a problem hiding this comment.
This assumes that the loader only return inputs and responses and will break if applied on a dataloader that has more parts (e.g. eye movements). Rather, use something like the following:
for data_batch in loader[data_key]:
# put into a dictionary and also move to target device already
data_batch = {f: getattr(data_batch, f).to(device) for f in data_batch._fields}
# extract inputs and targets that are supposed to always exist
inputs, targets = data_batch.pop('inputs'), data_batch.pop('targets')
output = torch.cat((output, (model(inputs, data_key=data_key, **data_batch).detach().cpu())), dim=0)| target, output = torch.empty(0), torch.empty(0) | ||
| for images, responses in loader[data_key]: | ||
| output = torch.cat((output, (model(images.to(device), data_key).detach().cpu())), dim=0) | ||
| target = torch.cat((target, responses.detach().cpu()), dim=0) |
There was a problem hiding this comment.
if you are going to use my codesnippet above, I called it targets instead of responses already.
| all_correlations = np.append(all_correlations, ret) | ||
| else: | ||
| n_neurons[0,i] = output.shape[1] | ||
| correlations[i,0] = ret.mean() |
There was a problem hiding this comment.
wouldn't it make more sense too just get sum of correlations here and later just divide by the total number of neurons?
| desc='Epoch {}'.format(epoch)): | ||
|
|
||
| loss = full_objective(model, data_key, **data) | ||
| loss = full_objective(model, data_key, *data) |
There was a problem hiding this comment.
This is a dangerous assumption about the order of elements returned from the dataloader. You are assuming that it always comes back as inputs, targets, other1, other2,... but this is not guaranteed. Rather we should construct a dictionary to properly deal with this. Refer to my comments above on how to properly pass a dictionary constructed from the namedtuple.
|
|
||
| model, epoch = run(model=model, | ||
| full_objective=full_objective, | ||
| optimizer=optimizer, | ||
| scheduler=scheduler, | ||
| stop_closure=stop_closure, | ||
| train_loader=train_loader, | ||
| train=train, |
There was a problem hiding this comment.
sorry I think this argument was better off staying as train_loader. I just didn't want the dictiionary of data loaders' key to be train_loader.
No description provided.