Add pytorch model

config.json

@@ -1,5 +1,5 @@
 {
-  "_name_or_path": "/home/patrick/hugging_face/t5/t5-v1_1-base",
+  "_name_or_path": ".",
   "architectures": [
     "T5ForConditionalGeneration"
   ],
@@ -21,6 +21,7 @@
   "pad_token_id": 0,
   "relative_attention_num_buckets": 32,
   "tie_word_embeddings": false,
+  "torch_dtype": "float32",
   "transformers_version": "4.13.0",
   "use_cache": true,
   "vocab_size": 32103

flax_to_pytorch.py

@@ -0,0 +1,26 @@
+import torch
+import numpy as np
+import jax.numpy as jnp
+from transformers import AutoTokenizer
+from transformers import FlaxT5ForConditionalGeneration
+from transformers import T5ForConditionalGeneration
+tokenizer = AutoTokenizer.from_pretrained(".")
+model_fx = FlaxT5ForConditionalGeneration.from_pretrained(".")
+model_pt = T5ForConditionalGeneration.from_pretrained(".", from_flax=True)
+model_pt.save_pretrained("./")
+text = "Hoe gaat het?"
+e_input_ids_fx = tokenizer(text, return_tensors="np", padding=True, max_length=128, truncation=True)
+d_input_ids_fx = jnp.ones((e_input_ids_fx.input_ids.shape[0], 1), dtype="i4") * model_fx.config.decoder_start_token_id
+e_input_ids_pt = tokenizer(text, return_tensors="pt", padding=True, max_length=128, truncation=True)
+d_input_ids_pt = np.ones((e_input_ids_pt.input_ids.shape[0], 1), dtype="i4") * model_pt.config.decoder_start_token_id
+print(e_input_ids_fx)
+print(d_input_ids_fx)
+print()
+encoder_pt = model_fx.encode(**e_input_ids_pt)
+decoder_pt = model_fx.decode(d_input_ids_pt, encoder_pt)
+logits_pt = decoder_pt.logits
+print(logits_pt)
+encoder_fx = model_fx.encode(**e_input_ids_fx)
+decoder_fx = model_fx.decode(d_input_ids_fx, encoder_fx)
+logits_fx = decoder_fx.logits
+print(logits_fx)

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa8b87f8bb924ddaf9823ed6c9ed8f57adbee415b398049da58ddbe36997cf9a
+size 990280781