def compute_variance(res, n_avg):
activations, rate_map, g, pos = compute_ratemaps(model,
data_manager,
options,
res=res,
n_avg=n_avg)
counts = np.zeros([res, res])
variance = np.zeros([res, res])
x_all = (pos[:, 0] + options['box_width'] / 2) / options['box_width'] * res
y_all = (pos[:, 1] +
options['box_height'] / 2) / options['box_height'] * res
for i in tqdm(range(len(g))):
x = int(x_all[i])
y = int(y_all[i])
if x >= 0 and x < res and y >= 0 and y < res:
counts[x, y] += 1
variance[x, y] += np.linalg.norm(
g[i] - activations[:, x, y]) / np.linalg.norm(
g[i]) / np.linalg.norm(activations[:, x, y])
for x in range(res):
for y in range(res):
if counts[x, y] > 0:
variance[x, y] /= counts[x, y]
return variance
def validation_epoch_end(self, outputs):
avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
avg_err = torch.stack([x['val_err'].mean() for x in outputs]).mean()
# avg position error per step in sequence
avg_seq_err = (torch.stack([x['val_err'] for x in outputs]).mean(dim=0)).mean(dim=-1)
save_seq_err(avg_seq_err, self.options)
# maybe only need to do outputs['val_loss']
tensorboard_logs = {'val_loss': avg_loss, 'val_err': avg_err}
# these are the full size of the epoch
pos = torch.cat([x['pos'] for x in outputs], dim=0).to('cpu').detach().numpy()
act = torch.cat([x['act'] for x in outputs], dim=0).to('cpu').detach().numpy()
# TODO: fix these vars
print(pos.shape, act.shape)
# Save a picture of rate maps
# save_ratemaps(self.model, self.trajectory_generator, self.options, step=tot_step)
compute_ratemaps(pos, act, self.options, epoch=self.current_epoch) # TODO: specify range
del pos, act
return {'val_loss': avg_loss, 'val_err': avg_err, 'log': tensorboard_logs}
def validation_epoch_end(self, outputs):
avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
avg_err = torch.stack([x['val_err'].mean() for x in outputs]).mean()
# avg position error per step in sequence
avg_seq_err = (torch.stack([x['val_err'] for x in outputs
]).mean(dim=0)).mean(dim=-1)
save_seq_err(avg_seq_err, self.options)
# maybe only need to do outputs['val_loss']
tensorboard_logs = {'val_loss': avg_loss, 'val_err': avg_err}
# these are the full size of the epoch
pos = torch.cat([x['coord'] for x in outputs], dim=0)
# all_but_last_two_dims = pos.size()[:-2]
# pos = pos.view(*all_but_last_two_dims, -1)
pos = pos.to('cpu').detach().numpy()
act = torch.cat([x['layer_outputs'] for x in outputs], dim=0)
# all_but_last_two_dims = act.size()[:-2]
# act = act.view(*all_but_last_two_dims, -1)
act = act.to('cpu').detach().numpy()
# TODO: fix these vars
# Save a picture of rate maps
# save_ratemaps(self.model, self.trajectory_generator, self.options, step=tot_step)
for i in range(pos.shape[0], act.shape[0] + pos.shape[0],
pos.shape[0]):
tmp_options = self.options
j = i / pos.shape[0]
tmp_options.run_ID + '_L{j}'
ppos = pos.reshape(-1, pos.shape[-1])
aact = act[i - pos.shape[0]:i].reshape(
-1, act[i - pos.shape[0]:i].shape[-1])
compute_ratemaps(ppos, aact, tmp_options, epoch=self.current_epoch)
# TODO: specify range
del pos, act
return {
'val_loss': avg_loss,
'val_err': avg_err,
'log': tensorboard_logs
}