I am using the TF-2.5-vanilla Conda environment. It has tensorflow2.5 installed using the command pip install tensorflow==2.5.0 .
I have installed your bert using pip install bert-for-tf2. I have written the following test code to check the activation of your bert.
Please check this code snippet.
import bertimport numpy as npimport tensorflow as tf#wget --quiet https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-12_H-768_A-12.zip and unzipmodel_dir = "./uncased_L-12_H-768_A-12"bert_params = bert.params_from_pretrained_ckpt(model_dir)l_bert = bert.BertModelLayer.from_params(bert_params, name="bert")#Pass1np.random.seed(171)input_ids = tf.Variable(np.random.randint(low=0, high=30522, size=(32, 384)))np.save('input.npy',input_ids)output = l_bert(input_ids)print(type(output))print()print(output.shape)np.save('file1.npy',output)
I have used this above code snippet twice with the same Conda environment TF-2.5-vanilla.
The first time, the input, is generated with the seed 171, is stored in input.npy. The second time, the input, is generated with the same seed(171), which is stored in input2.npy. and after comparison of these np arrays, input.npy==input2.npy
The first time, the output of l_bert, using input_ids, is stored in file1.npy. The second time, the output of l_bert, using the same input_ids, is stored in file2.npy. and after comparison of these np arrays, file1.npy != file2.npy
Why is my output of l_bert, using the same input_ids, different?
I am using the TF-2.5-vanilla Conda environment. It has tensorflow2.5 installed using the command
pip install tensorflow==2.5.0. I have installed your bert usingpip install bert-for-tf2. I have written the following test code to check the activation of your bert. Please check this code snippet.import bertimport numpy as npimport tensorflow as tf#wget --quiet https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-12_H-768_A-12.zip and unzipmodel_dir = "./uncased_L-12_H-768_A-12"bert_params = bert.params_from_pretrained_ckpt(model_dir)l_bert = bert.BertModelLayer.from_params(bert_params, name="bert")#Pass1np.random.seed(171)input_ids = tf.Variable(np.random.randint(low=0, high=30522, size=(32, 384)))np.save('input.npy',input_ids)output = l_bert(input_ids)print(type(output))print()print(output.shape)np.save('file1.npy',output)I have used this above code snippet twice with the same Conda environment TF-2.5-vanilla. The first time, the input, is generated with the seed 171, is stored in input.npy. The second time, the input, is generated with the same seed(171), which is stored in input2.npy. and after comparison of these np arrays, input.npy==input2.npy
The first time, the output of l_bert, using input_ids, is stored in file1.npy. The second time, the output of l_bert, using the same input_ids, is stored in file2.npy. and after comparison of these np arrays, file1.npy != file2.npy
Why is my output of l_bert, using the same input_ids, different?