tanner/min-dalle-test
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

previous commit broke colab example, so adjusting flax requirement to 0.4.2 for now

Browse Source
This commit is contained in:
Brett Kuprel
2022-06-28 08:04:08 -04:00
parent 38ebe54a38
commit 34df2b97df
2 changed files with 5 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
min_dalle/models/dalle_bart_decoder_flax.py
View File

@@ -35,7 +35,7 @@ class DecoderSelfAttentionFlax(AttentionFlax):
) -> Tuple[jnp.ndarray, Tuple[jnp.ndarray, jnp.ndarray]]:
shape_split = decoder_state.shape[:2] + (self.head_count, -1)
keys_state = lax.dynamic_update_slice(
keys_state,
keys_state,
self.k_proj(decoder_state).reshape(shape_split),
state_index
)
@@ -205,7 +205,6 @@ class DalleBartDecoderFlax(nn.Module):
params: dict
) -> jnp.ndarray:
attention_mask = jnp.not_equal(text_tokens, 1)
encoder_state = encoder_state.astype(jnp.float16)
def sample_next_image_token(
state: SampleState,
@@ -248,8 +247,8 @@ class DalleBartDecoderFlax(nn.Module):
initial_state = SampleState(
prev_token = self.start_token,
prng_key = prng_key,
keys_state = jnp.zeros(state_shape, dtype=jnp.float16),
values_state = jnp.zeros(state_shape, dtype=jnp.float16)
keys_state = jnp.zeros(state_shape),
values_state = jnp.zeros(state_shape)
)
_, image_tokens = lax.scan(