[Flax] Correct shift labels for seq2seq models in Flax (huggingface#1…

…2720)

* fix_torch_device_generate_test

* remove @

* push

* fix marian

* fix

* up

src/transformers/models/bart/modeling_flax_bart.py

@@ -19,6 +19,8 @@
 from functools import partial
 from typing import Callable, Optional, Tuple
 
+import numpy as np
+
 import flax.linen as nn
 import jax
 import jax.numpy as jnp
@@ -212,15 +214,15 @@
 """
 
 
-def shift_tokens_right(input_ids: jnp.ndarray, pad_token_id: int, decoder_start_token_id: int) -> jnp.ndarray:
+def shift_tokens_right(input_ids: np.array, pad_token_id: int, decoder_start_token_id: int) -> np.ndarray:
     """
     Shift input ids one token to the right.
     """
-    shifted_input_ids = jnp.roll(input_ids, 1, axis=-1)
-    shifted_input_ids = jax.ops.index_update(shifted_input_ids, (..., 0), decoder_start_token_id)
-    # replace possible -100 values in labels by `pad_token_id`
-    shifted_input_ids = jnp.where(shifted_input_ids == -100, pad_token_id, shifted_input_ids)
+    shifted_input_ids = np.zeros_like(input_ids)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1]
+    shifted_input_ids[:, 0] = decoder_start_token_id
 
+    shifted_input_ids = np.where(shifted_input_ids == -100, pad_token_id, shifted_input_ids)
     return shifted_input_ids
 
 

src/transformers/models/marian/modeling_flax_marian.py

@@ -221,11 +221,11 @@ def shift_tokens_right(input_ids: jnp.ndarray, pad_token_id: int, decoder_start_
     """
     Shift input ids one token to the right.
     """
-    shifted_input_ids = jnp.roll(input_ids, 1, axis=-1)
-    shifted_input_ids = jax.ops.index_update(shifted_input_ids, (..., 0), decoder_start_token_id)
-    # replace possible -100 values in labels by `pad_token_id`
-    shifted_input_ids = jnp.where(shifted_input_ids == -100, pad_token_id, shifted_input_ids)
+    shifted_input_ids = np.zeros_like(input_ids)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1]
+    shifted_input_ids[:, 0] = decoder_start_token_id
 
+    shifted_input_ids = np.where(shifted_input_ids == -100, pad_token_id, shifted_input_ids)
     return shifted_input_ids
 
 

src/transformers/models/mbart/modeling_flax_mbart.py

@@ -19,6 +19,8 @@
 from functools import partial
 from typing import Callable, Optional, Tuple
 
+import numpy as np
+
 import flax.linen as nn
 import jax
 import jax.numpy as jnp
@@ -217,20 +219,19 @@ def shift_tokens_right(input_ids: jnp.ndarray, pad_token_id: int) -> jnp.ndarray
     Shift input ids one token to the right, and wrap the last non pad token (the <LID> token) Note that MBart does not
     have a single `decoder_start_token_id` in contrast to other Bart-like models.
     """
-    prev_output_tokens = jnp.array(input_ids).clone()
+    prev_output_tokens = np.array(input_ids).copy()
 
     assert pad_token_id is not None, "self.model.config.pad_token_id has to be defined."
 
     # replace possible -100 values in labels by `pad_token_id`
-    prev_output_tokens = jnp.where(prev_output_tokens == -100, pad_token_id, input_ids)
-    index_of_eos = (jnp.where(prev_output_tokens != pad_token_id, 1, 0).sum(axis=-1) - 1).reshape(-1, 1)
-    decoder_start_tokens = jnp.array(
-        [prev_output_tokens[i, eos_idx] for i, eos_idx in enumerate(index_of_eos)]
+    prev_output_tokens = np.where(prev_output_tokens == -100, pad_token_id, input_ids)
+    index_of_eos = (np.where(prev_output_tokens != pad_token_id, 1, 0).sum(axis=-1) - 1).reshape(-1, 1)
+    decoder_start_tokens = np.array(
+        [prev_output_tokens[i, eos_idx] for i, eos_idx in enumerate(index_of_eos)], dtype=np.int32
     ).squeeze()
-    # for loop basically does jax-compatible version of prev_output_tokens[:, 1:] = prev_output_tokens[:, :-1].clone()
-    for i in range(prev_output_tokens.shape[1], 0, -1):
-        prev_output_tokens = jax.ops.index_update(prev_output_tokens, (..., i), prev_output_tokens[:, i - 1])
-    prev_output_tokens = jax.ops.index_update(prev_output_tokens, (..., 0), decoder_start_tokens)
+
+    prev_output_tokens[:, 1:] = prev_output_tokens[:, :-1].copy()
+    prev_output_tokens[:, 0] = decoder_start_tokens
 
     return prev_output_tokens
 

src/transformers/models/t5/modeling_flax_t5.py

@@ -47,15 +47,16 @@
 _TOKENIZER_FOR_DOC = "T5Tokenizer"
 
 
-def shift_tokens_right(input_ids: jnp.ndarray, pad_token_id: int, decoder_start_token_id: int) -> jnp.ndarray:
+# Copied from transformers.models.bart.modeling_flax_bart.shift_tokens_right
+def shift_tokens_right(input_ids: np.array, pad_token_id: int, decoder_start_token_id: int) -> np.ndarray:
     """
     Shift input ids one token to the right.
     """
-    shifted_input_ids = jnp.roll(input_ids, 1, axis=-1)
-    shifted_input_ids = jax.ops.index_update(shifted_input_ids, (..., 0), decoder_start_token_id)
-    # replace possible -100 values in labels by `pad_token_id`
-    shifted_input_ids = jnp.where(shifted_input_ids == -100, pad_token_id, shifted_input_ids)
+    shifted_input_ids = np.zeros_like(input_ids)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1]
+    shifted_input_ids[:, 0] = decoder_start_token_id
 
+    shifted_input_ids = np.where(shifted_input_ids == -100, pad_token_id, shifted_input_ids)
     return shifted_input_ids