NVIDIA-Merlin
[BUG] Inconsistent inference and evaluation results of the XLNET-CLM even on the training set!
Bug description
Hello, I followed the example and successfully trained an XLNet-CLM model on my custom dataset. However, I noticed that while the model performs well on the validation set with trainer.evaluate() (even achieving 90% recall@5), I encountered many errors when using trainer.predict() for inference, falling far short of the expected performance.
So, I conducted an experiment:
I took a portion of the training set data and input it into both functions. I used sequence[:] for evaluate() and sequence[:-1] for predict():
=========data for eval===============
session_id item_id-list
0 1 [26, 26, 26, 26, 4, 4, 4, 4, 4, 4, 4, 4]
1 2 [7, 43, 35, 3, 3, 7, 29, 35, 35, 111, 5, 9]
2 3 [74, 7, 74, 7, 7, 110, 32, 67, 4, 4, 17, 7]
4 5 [56, 25, 25, 25, 25, 25, 25, 23, 34, 4, 19, 43]
5 6 [270, 41, 41, 41, 41, 7, 43, 34, 78, 38, 71, 23]
6 7 [74, 28, 28, 5, 5, 24, 9, 5, 5, 59, 4, 91]
=========data for infer===============
session_id item_id-list
0 1 [26, 26, 26, 26, 4, 4, 4, 4, 4, 4, 4]
1 2 [7, 43, 35, 3, 3, 7, 29, 35, 35, 111, 5]
2 3 [74, 7, 74, 7, 7, 110, 32, 67, 4, 4, 17]
4 5 [56, 25, 25, 25, 25, 25, 25, 23, 34, 4, 19]
5 6 [270, 41, 41, 41, 41, 7, 43, 34, 78, 38, 71]
6 7 [74, 28, 28, 5, 5, 24, 9, 5, 5, 59, 4]
=========labels:===============
0 4
1 9
2 7
4 43
5 23
6 91
Name: item_id-list, dtype: int64
Ideally, the predictions from both functions should be similar, but it seems the inference results is significantly worse:
=========inference===============
PredictionOutput(predictions=(array([[ 4, 19, 7, 5, 11],
[ 5, 15, 23, 22, 7],
[17, 7, 30, 70, 15],
[19, 4, 11, 5, 15],
[ 7, 4, 79, 34, 6],
[ 4, 7, 19, 11, 5]]), array([[7.677282 , 5.3613596, 5.00848 , 4.6888046, 4.319791 ],
[7.15173 , 5.7525525, 5.726646 , 4.9717607, 4.903692 ],
[7.843027 , 7.171323 , 5.413306 , 5.378131 , 5.3367157],
[6.309387 , 5.833148 , 5.6351004, 4.9878273, 4.4770455],
[7.6406755, 5.9271154, 5.91128 , 5.8692527, 5.589204 ],
[7.7423315, 5.4810033, 5.0784097, 4.7432585, 4.603503 ]],
dtype=float32)), label_ids=None, metrics={'test_runtime': 0.792, 'test_samples_per_second': 7.576, 'test_steps_per_second': 3.788})
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 120.82it/s]
=========evaluation===============
PredictionOutput(predictions=(array([[ 4, 12, 19, 7, 36],
[ 9, 12, 10, 33, 49],
[ 7, 34, 4, 30, 79],
[ 43, 110, 129, 6, 126],
[ 23, 34, 15, 56, 42],
[ 91, 59, 102, 121, 83]]), array([[18.594198 , 10.1901245, 10.138629 , 9.487766 , 9.470682 ],
[18.095396 , 11.1290245, 10.906204 , 10.659245 , 10.633735 ],
[16.824923 , 9.672426 , 9.347423 , 8.993546 , 8.8174095],
[15.070214 , 10.285906 , 9.3903475, 8.776609 , 8.753696 ],
[16.2281 , 10.282413 , 10.155615 , 9.619689 , 9.48603 ],
[14.316702 , 10.41172 , 9.787606 , 9.7606325, 9.695224 ]],
dtype=float32)), label_ids=array([ 4, 9, 7, 43, 23, 91]), metrics={'eval_/next-item/ndcg_at_5': 1.0, 'eval_/next-item/ndcg_at_10': 1.0, 'eval_/next-item/recall_at_5': 1.0, 'eval_/next-item/recall_at_10': 1.0, 'eval_/next-item/avg_precision_at_5': 1.0, 'eval_/next-item/avg_precision_at_10': 1.0, 'eval_/loss': 0.02894706465303898, 'eval_runtime': 0.0639, 'eval_samples_per_second': 93.926, 'eval_steps_per_second': 46.963})
Please note that I conducted this experiment on the training set. The outputs from evaluation are expected, but the inference not. I'm curious to know why this is happening. Thanks!
Here is my code that generates the outputs above:
tr_model.load_state_dict(torch.load("tmp/checkpoint-450/pytorch_model.bin"))
tr_model.eval()
args = tr.trainer.T4RecTrainingArguments(
output_dir="tmp",
per_device_eval_batch_size=2,
max_sequence_length=30,
fp16=True,
)
trainer = tr.Trainer(
model=tr_model,
args=args,
schema=schema,
compute_metrics=True,
)
trainer.test_dataset_or_path = 'data/preproc_sessions_by_day_3827/1/train_truncate_for_infer.parquet'
trainer.eval_dataset_or_path = 'data/preproc_sessions_by_day_3827/1/train_truncate_for_eval.parquet'
trainer.args.predict_top_k = 5
prediction = trainer.predict(trainer.test_dataset_or_path)
print("=========inference===============")
print(prediction)
# a small monkey patch to output the predictions, not only the metrics
tr.Trainer.evaluate = evaluate_change_output
prediction = trainer.evaluate()
print("=========evaluation===============")
print(prediction)
Environment details
- Transformers4Rec version: 23.06
- Platform: linux
- Python version: 3.10
- Huggingface Transformers version: 4.33.2
- PyTorch version (GPU?):
- Tensorflow version (GPU?):
I've reviewed the clm masking code, and I'm little confused about this line here: https://github.com/NVIDIA-Merlin/Transformers4Rec/blob/625897c3f55135342dbe53c32b7650a8ecb86c75/transformers4rec/torch/masking.py#L319 I would like to know, what is the purpose of removing the last padding item if the input is padded? Should the last item be removed instead of the feature of the last padding item?
anyway, I changed some code in apply_mask_to_inputs of clm masking:
def apply_mask_to_inputs_CLM(
self,
inputs: torch.Tensor,
mask_schema: torch.Tensor,
training: bool = False,
testing: bool = False,
) -> torch.Tensor:
if not training and not testing:
# Replacing the inputs corresponding to padded items with a trainable embedding
# To mimic training and evaluation masking strategy
inputs = torch.where(
mask_schema.unsqueeze(-1).bool(),
inputs,
self.masked_item_embedding.to(inputs.dtype),
)
return inputs
# # shift sequence of interaction embeddings
# pos_emb_inp = inputs[:, :-1]
# # Adding a masked item in the sequence to return to the initial sequence.
# pos_emb_inp = torch.cat( # type: ignore
# [
# pos_emb_inp,
# torch.zeros(
# (pos_emb_inp.shape[0], 1, pos_emb_inp.shape[2]),
# dtype=pos_emb_inp.dtype,
# ).to(inputs.device),
# ],
# axis=1,
# )
pos_emb_inp = inputs
pos_emb_inp_new = pos_emb_inp.clone()
# Iterate over each row in the boolean tensor
for i in range(mask_schema.shape[0]):
# Find the index of the last True value in the row
# If there's no True value, idx will be -1
idx = (mask_schema[i].nonzero(as_tuple=True)[0]).max().item() if mask_schema[i].any() else -1
# Replace corresponding item in feature tensor with a zero matrix
if idx != -1:
pos_emb_inp_new[i, idx] = torch.zeros(pos_emb_inp.shape[2], dtype=pos_emb_inp.dtype).to(inputs.device)
pos_emb_inp = pos_emb_inp_new
# Replacing the inputs corresponding to padded items with a trainable embedding
pos_emb_inp = torch.where(
mask_schema.unsqueeze(-1).bool(),
pos_emb_inp,
self.masked_item_embedding.to(pos_emb_inp.dtype),
)
return pos_emb_inp
Interestingly, with this modification, the metrics of XLNet in the CLM setting have decreased compared to before, making it more reasonable. I've also noticed that this #719 and #746 mentioned a similar issue. Additionally, I observed that the outputs of the predict and evaluate functions have become similar:
# same inputs as before!
=========inference===============
PredictionOutput(predictions=(array([[ 4, 3, 7, 5, 6],
[ 5, 7, 30, 3, 22],
[17, 7, 6, 26, 11],
[ 3, 4, 25, 7, 18],
[71, 26, 24, 4, 3],
[ 4, 5, 3, 7, 22]]), array([[ 8.456369 , 5.8312187, 5.6498675, 5.1875997, 4.956415 ],
[10.017425 , 6.9653053, 6.9261403, 6.617892 , 6.447962 ],
[ 9.492199 , 8.0066 , 6.5381365, 6.097307 , 5.9652367],
[ 6.56517 , 6.2221594, 5.269926 , 5.1904283, 5.092497 ],
[ 6.0176606, 5.6072693, 5.431714 , 5.273334 , 5.039026 ],
[ 8.518238 , 6.537841 , 5.880717 , 5.6688414, 5.0900187]],
dtype=float32)), label_ids=None, metrics={'test_runtime': 2.0936, 'test_samples_per_second': 2.866, 'test_steps_per_second': 1.433})
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 14.08it/s]
=========evaluation===============
PredictionOutput(predictions=(array([[ 4, 7, 3, 6, 30],
[ 5, 7, 30, 3, 4],
[17, 7, 6, 26, 4],
[ 3, 4, 25, 39, 7],
[ 4, 24, 71, 26, 14],
[ 4, 30, 5, 3, 7]]), array([[7.4311004, 4.9409456, 4.876472 , 4.848051 , 4.7532825],
[7.4449983, 6.3823347, 6.3242016, 6.160487 , 5.6924496],
[7.9830294, 7.4659915, 6.0589595, 5.7209487, 5.57537 ],
[6.721603 , 5.4623756, 5.102635 , 4.9805446, 4.9469085],
[4.9136906, 4.843173 , 4.796401 , 4.7077065, 4.5187464],
[7.7905493, 5.1441355, 5.0995426, 5.0602183, 5.031133 ]],
dtype=float32)), label_ids=array([ 4, 9, 7, 43, 23, 91]), metrics={'eval_/next-item/ndcg_at_5': 0.27182161808013916, 'eval_/next-item/ndcg_at_10': 0.27182161808013916, 'eval_/next-item/recall_at_5': 0.3333333432674408, 'eval_/next-item/recall_at_10': 0.3333333432674408, 'eval_/next-item/avg_precision_at_5': 0.25, 'eval_/next-item/avg_precision_at_10': 0.25, 'eval_/loss': 4.302186489105225, 'eval_runtime': 0.6978, 'eval_samples_per_second': 8.599, 'eval_steps_per_second': 4.299})
I'm not sure if my changes are correct, and I strongly recommend that you pay attention to this issue. Thanks!