debugging.txt

parse_args(), training seed
dataset_splits (load_dataset(args))
raw_dataset = class DatasetDict(dict(MutableMapping(abc))); train & val ("datasets.arrow_dataset.Dataset", uses Apache Arrows 4 storage, shape (73, 6)&(7, 6) -> all columns of ds)
feature_extractor = Wav2Vec2FeatureExtractor(SequenceFeatureExtractor(FeatureExtractionMixin(PushToHubMixin))): general feature extractor(general FE 4 speech recog(saving/loading 4 FE(push model/tokenizer to hub))); input: raw audiowaves, output: dictionary with "input_values" (preprocessed data(float list / list of batches)) & "attention_mask" (real data=1, padding=0)
raw_dataset.map(args): (general with .map(): input size != output size, but all values in output dict gotta have same numb of elements); remove_columns=raw_datasets["train"].column_names removes all columns (except the ones created with .map())
.map() & .filter() return vectorized_datasets ("train" & "validation" both datasets.arrow_dataset.Dataset type and shape (62,1)&(7,1) -> only coumn is input_values, also dict with lists of floats)
class DataCollatorForWav2Vec2Pretraining: automatic contrastive loss with masking
features: list of dicts (0= {"input_values": [...]}), shape(8, 755) -> 755 variable from batch to batch
__call__ function: takes in features from feature extractor with data from one example each (either as list[int] or torch.tensor), returns dict of torch.tensors (processed & padded batches)
one batch contains: data (input_values (shape [8, 113600]), attention_mask (same shape) and sub_attention_mask ([8, 354] (prop for conv)) each as torch.tensor), mask_time_indices (np.ndarray, torch.long) and sampled_negative_indices (np.ndarray [8, 354, 100]) for contrastive learning 
outputs: torch,tensor
num_losses: torch.tensor
outputs=model(**batch) (after loss (line709)) forward pass
loss is still torch.tensor
accelerator.backward(loss) for backprop and distributed training
def multiply_grads() for stable grads i guess
(with prepare_dataset(batch): "inputs" now contains data(dict) with "input_values"(list with np.ndarrays) and "attention_mask" (also list of np.ndarrays)) still former form in the end