Hi guys,
As part of my adventure using fastai_v2, I’m trying to reimplement some of my v1 models. I’m having trouble with a custom callback. Here’s a prototype:
class TestCB(Callback):
def __init__(self, text:str=' ', temperature:float=1., n_len:int=10, max_mols:int=5):
super().__init__()
self.text = text
self.temperature = temperature
self.n_len = n_len
self.max_mols = max_mols
def after_train(self):
decoder=decode_spec_tokens
self.model.reset()
stop_index = self.dls.train.vocab.index(EOS)
seqs = []
num = self.dls.train_ds.numericalize
xb = num(self.text).to(self.dls.device)
print(xb)
preds,_ = self.get_preds(dl=[(xb[None],)])
for _ in range(self.n_len):
res = preds[0][-1]
if self.temperature != 1.:
res.pow_(1 / self.temperature)
idx = torch.multinomial(res, 1).item()
if idx != stop_index:
seqs.append(idx)
xb = xb.new_tensor([idx])[None]
else:
break
seqs = ''.join([num.vocab[i] for i in seqs if num.vocab[i] not in [BOS, PAD, EOS]])
print(seqs)
This callback is supposed to generate max_mols SMILES strings (an 1D representation of a molecule) of length n_len at the end of every epoch. Starting with a seed token (i.e., text), the model will keep sampling tokens until the EOS is sampled; when that happens the iteration stops.
The callback does work, but only alone! When I tried to use the CSVLoggerCallback, I got this error:
ValueError: I/O operation on closed file.
In addition, when using TestCB, the result table is empty throughout the fit. I can’t see the losses, accuracy or any other metric if TestCB is used. Like this:
epoch train_loss valid_loss accuracy perplexity time
Am I missing something here?