Remove unused code

.gitignore

@@ -1,33 +1,31 @@
 # Byte-compiled / optimized / DLL files
 __pycache__/
+models/__pycache__/
 *.py[cod]
 *$py.class
 
 # added manually
 .data
 .embeddings
-__pycache__/
-models/__pycache__/
-multiprocess/__pycache__/
 results/
-tests/embeddings/*
 local_files
 local_scripts/*
 dataset/*
 generated/*
+*events.out.tfevents*
 
 .DS_Store
 .idea/
 .vscode/
 models/.DS_Store
-multiprocess/.DS_Store
-text/.DS_Store
 src/
 workdir/
 *save*/
 
 lightning_logs/
 pytorch_model.bin
+*.pth
+tests/embeddings/*
 tests/dataset-*
 tests/dataset/*
 tests/database/*

genienlp/arguments.py

@@ -77,21 +77,21 @@ def parse_argv(parser):
                         'If sentence_batching is used, this will be interpreted as number of examples.')
     parser.add_argument('--jump_start', default=0, type=int, help='number of iterations to give jump started tasks')
     parser.add_argument('--n_jump_start', default=0, type=int, help='how many tasks to jump start (presented in order)')
-    parser.add_argument('--num_print', default=15, type=int,
+    parser.add_argument('--num_print', default=10, type=int,
                         help='how many validation examples with greedy output to print to std out')
 
     parser.add_argument('--no_tensorboard', action='store_false', dest='tensorboard',
                         help='Turn off tensorboard logging')
     parser.add_argument('--tensorboard_dir', default=None,
                         help='Directory where to save Tensorboard logs (defaults to --save)')
-    parser.add_argument('--max_to_keep', default=5, type=int, help='number of checkpoints to keep')
-    parser.add_argument('--log_every', default=int(1e2), type=int, help='how often to log results in # of iterations')
-    parser.add_argument('--save_every', default=int(1e3), type=int,
+    parser.add_argument('--max_to_keep', default=3, type=int, help='number of checkpoints to keep')
+    parser.add_argument('--log_every', default=100, type=int, help='how often to log results in # of iterations')
+    parser.add_argument('--save_every', default=1000, type=int,
                         help='how often to save a checkpoint in # of iterations')
 
     parser.add_argument('--val_tasks', nargs='+', type=str, dest='val_task_names',
                         help='tasks to collect evaluation metrics for')
-    parser.add_argument('--val_every', default=int(1e3), type=int,
+    parser.add_argument('--val_every', default=1000, type=int,
                         help='how often to run validation in # of iterations')
     parser.add_argument('--val_batch_size', nargs='+', default=[3000], type=int,
                         help='Number of tokens in each batch for validation, corresponding to tasks in --val_tasks')

genienlp/data_utils/example.py

@@ -49,8 +49,6 @@ class Example(NamedTuple):
     answer: str
     context_plus_question: List[str]
 
-    vocab_fields = ['context', 'question', 'answer']
-
     @staticmethod
     def from_raw(example_id: str, context: str, question: str, answer: str, preprocess = identity, lower=False):
         args = [example_id]

genienlp/models/mqan_decoder.py

@@ -68,8 +68,8 @@ class MQANDecoder(nn.Module):
     def forward(self, batch, final_context, context_rnn_state, encoder_loss,
                 current_token_id=None, decoder_wrapper=None, expansion_factor=1, generation_dict=None):
 
-        context, context_lengths, context_limited = batch.context.value, batch.context.length, batch.context.limited
-        answer, answer_lengths, answer_limited = batch.answer.value, batch.answer.length, batch.answer.limited
+        context, context_limited = batch.context.value, batch.context.limited
+        answer, answer_limited = batch.answer.value, batch.answer.limited
         decoder_vocab = self.numericalizer.decoder_vocab
         self.map_to_full = decoder_vocab.decode
         context_padding = context.data == self.pad_idx

genienlp/train.py

@@ -138,7 +138,6 @@ def train_step(model, batch, iteration, opt, devices, lr_scheduler=None, grad_cl
         for p in model.parameters():
             if p.grad is None:
                 continue
-            # print('p.grad = ', p.grad)
             p.grad /= accumulated_batch_lengths
         accumulated_batch_lengths = 0
         if grad_clip > 0.0:
@@ -367,8 +366,7 @@ def train(args, devices, model, opt, lr_scheduler, train_sets, train_iterations,
                                          grad_clip=args.grad_clip,
                                          gradient_accumulation_steps=args.gradient_accumulation_steps)
             if loss is None:
-                logger.info(
-                    'Encountered NAN loss during training... Continue training ignoring the current batch')
+                logger.info('Encountered NAN loss during training... Continue training ignoring the current batch')
                 continue
             if loss < 1e-6:
                 zero_loss += 1

genienlp/util.py

@@ -435,19 +435,6 @@ def make_data_loader(dataset, numericalizer, batch_size, device=None, train=Fals
         return data_loader
 
 
-def pad(x, new_channel, dim, val=None):
-    if x.size(dim) > new_channel:
-        x = x.narrow(dim, 0, new_channel)
-    channels = x.size()
-    assert (new_channel >= channels[dim])
-    if new_channel == channels[dim]:
-        return x
-    size = list(channels)
-    size[dim] = new_channel - size[dim]
-    padding = x.new(*size).fill_(val)
-    return torch.cat([x, padding], dim)
-
-
 def have_multilingual(task_names):
     return any(['multilingual' in name for name in task_names])