model--BART--Shubham09 NAN output

[AmosLewis]

We have to revisit it later when we have time.

In the nlp.py, when you change the input to exact the same as huggingface model.

            if self.name == "model--BART--Shubham09":
                tensors[0] = numpy.array([[    0, 40710,  3923, 12346,    16,  2351,    10,  3944,    14, 26232, 7396,  4687,     4,     2]])
                tensors[1] = numpy.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]])
                tensors[2] = numpy.array([[    2,     0, 40710,  3923, 12346]])
                print("Change input of model--BART--Shubham09")

The output NAN for both golden and iree.

matching values with (rtol,atol) = [0.001, 0.001]: 0 of 251325 = 0.0%
Test Result:
TestResult(name='model--BART--Shubham09', input=TestTensors(<class 'numpy.ndarray'>): (array([[    0, 40710,  3923, 12346,    16,  2351,    10,  3944,    14,
        26232,  7396,  4687,     4,     2]]), array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]), array([[    2,     0, 40710,  3923, 12346]])), gold_output=TestTensors(<class 'numpy.ndarray'>): [array([[[nan, nan, nan, ..., nan, nan, nan],
        [nan, nan, nan, ..., nan, nan, nan],
        [nan, nan, nan, ..., nan, nan, nan],
        [nan, nan, nan, ..., nan, nan, nan],
        [nan, nan, nan, ..., nan, nan, nan]]], dtype=float32)], output=TestTensors(<class 'numpy.ndarray'>): (array([[[nan, nan, nan, ..., nan, nan, nan],
        [nan, nan, nan, ..., nan, nan, nan],
        [nan, nan, nan, ..., nan, nan, nan],
        [nan, nan, nan, ..., nan, nan, nan],
        [nan, nan, nan, ..., nan, nan, nan]]], dtype=float32),))

The way to get the huggingface model input.

from transformers import BartTokenizer, BartForConditionalGeneration

# Initialize tokenizer and model
tokenizer = BartTokenizer.from_pretrained('Shubham09/BART')
model = BartForConditionalGeneration.from_pretrained('Shubham09/BART')

# Example input and decoder sequences
input_text = "Hugging Face is creating a tool that democratizes AI."
decoder_text = "</s><s>Hugging Face"

# Tokenize inputs
inputs = tokenizer(input_text, return_tensors='pt')
decoder_inputs = tokenizer(decoder_text, return_tensors='pt', add_special_tokens=False)
print("decoder_input_ids: ")
print(decoder_inputs['input_ids'])

# Generate output
outputs = model.generate(
    input_ids=inputs['input_ids'],
    attention_mask=inputs['attention_mask'],
    decoder_input_ids=decoder_inputs['input_ids']
)

print("input_ids: ")
print(inputs['input_ids']) # numpy.array([[    0, 40710,  3923, 12346,    16,  2351,    10,  3944,    14, 26232, 7396,  4687,     4,     2]])
print("attention_mask: ")
print(inputs['attention_mask']) # array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]])
print("decoder_input_ids: ")
print(decoder_inputs['input_ids']) # array([[    2,     0, 40710,  3923, 12346]])

# Decode the output
decoded_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(decoded_output)