Spaces:
Running
Running
feat(wandb-examples): use model file
Browse files
dev/predictions/wandb-examples.py
CHANGED
@@ -4,16 +4,14 @@
|
|
4 |
import random
|
5 |
|
6 |
import jax
|
7 |
-
import flax.linen as nn
|
8 |
from flax.training.common_utils import shard
|
9 |
from flax.jax_utils import replicate, unreplicate
|
10 |
|
11 |
from transformers.models.bart.modeling_flax_bart import *
|
12 |
from transformers import BartTokenizer, FlaxBartForConditionalGeneration
|
13 |
|
14 |
-
import
|
15 |
|
16 |
-
import requests
|
17 |
from PIL import Image
|
18 |
import numpy as np
|
19 |
import matplotlib.pyplot as plt
|
@@ -23,58 +21,24 @@ import torchvision.transforms as T
|
|
23 |
import torchvision.transforms.functional as TF
|
24 |
from torchvision.transforms import InterpolationMode
|
25 |
|
|
|
26 |
from vqgan_jax.modeling_flax_vqgan import VQModel
|
27 |
|
28 |
-
#
|
29 |
-
|
30 |
-
OUTPUT_LENGTH = 256 + 1 # number of encoded tokens + 1 for bos
|
31 |
-
BOS_TOKEN_ID = 16384
|
32 |
-
BASE_MODEL = 'facebook/bart-large-cnn'
|
33 |
-
|
34 |
-
class CustomFlaxBartModule(FlaxBartModule):
|
35 |
-
def setup(self):
|
36 |
-
# we keep shared to easily load pre-trained weights
|
37 |
-
self.shared = nn.Embed(
|
38 |
-
self.config.vocab_size,
|
39 |
-
self.config.d_model,
|
40 |
-
embedding_init=jax.nn.initializers.normal(self.config.init_std, self.dtype),
|
41 |
-
dtype=self.dtype,
|
42 |
-
)
|
43 |
-
# a separate embedding is used for the decoder
|
44 |
-
self.decoder_embed = nn.Embed(
|
45 |
-
OUTPUT_VOCAB_SIZE,
|
46 |
-
self.config.d_model,
|
47 |
-
embedding_init=jax.nn.initializers.normal(self.config.init_std, self.dtype),
|
48 |
-
dtype=self.dtype,
|
49 |
-
)
|
50 |
-
self.encoder = FlaxBartEncoder(self.config, dtype=self.dtype, embed_tokens=self.shared)
|
51 |
-
|
52 |
-
# the decoder has a different config
|
53 |
-
decoder_config = BartConfig(self.config.to_dict())
|
54 |
-
decoder_config.max_position_embeddings = OUTPUT_LENGTH
|
55 |
-
decoder_config.vocab_size = OUTPUT_VOCAB_SIZE
|
56 |
-
self.decoder = FlaxBartDecoder(decoder_config, dtype=self.dtype, embed_tokens=self.decoder_embed)
|
57 |
-
|
58 |
-
class CustomFlaxBartForConditionalGenerationModule(FlaxBartForConditionalGenerationModule):
|
59 |
-
def setup(self):
|
60 |
-
self.model = CustomFlaxBartModule(config=self.config, dtype=self.dtype)
|
61 |
-
self.lm_head = nn.Dense(
|
62 |
-
OUTPUT_VOCAB_SIZE,
|
63 |
-
use_bias=False,
|
64 |
-
dtype=self.dtype,
|
65 |
-
kernel_init=jax.nn.initializers.normal(self.config.init_std, self.dtype),
|
66 |
-
)
|
67 |
-
self.final_logits_bias = self.param("final_logits_bias", self.bias_init, (1, OUTPUT_VOCAB_SIZE))
|
68 |
-
|
69 |
-
class CustomFlaxBartForConditionalGeneration(FlaxBartForConditionalGeneration):
|
70 |
-
module_class = CustomFlaxBartForConditionalGenerationModule
|
71 |
-
|
72 |
|
73 |
import wandb
|
74 |
import os
|
|
|
|
|
|
|
|
|
75 |
os.environ["WANDB_SILENT"] = "true"
|
76 |
os.environ["WANDB_CONSOLE"] = "off"
|
77 |
|
|
|
|
|
|
|
78 |
# set id to None so our latest images don't get overwritten
|
79 |
id = None
|
80 |
run = wandb.init(id=id,
|
@@ -87,8 +51,10 @@ artifact = run.use_artifact('wandb/hf-flax-dalle-mini/model-4oh3u7ca:latest', ty
|
|
87 |
artifact_dir = artifact.download()
|
88 |
|
89 |
# create our model
|
90 |
-
tokenizer = BartTokenizer.from_pretrained(BASE_MODEL)
|
91 |
model = CustomFlaxBartForConditionalGeneration.from_pretrained(artifact_dir)
|
|
|
|
|
|
|
92 |
model.config.force_bos_token_to_be_generated = False
|
93 |
model.config.forced_bos_token_id = None
|
94 |
model.config.forced_eos_token_id = None
|
@@ -143,9 +109,6 @@ p_get_images = jax.pmap(get_images, "batch")
|
|
143 |
bart_params = replicate(model.params)
|
144 |
vqgan_params = replicate(vqgan.params)
|
145 |
|
146 |
-
# ## CLIP Scoring
|
147 |
-
from transformers import CLIPProcessor, FlaxCLIPModel
|
148 |
-
|
149 |
clip = FlaxCLIPModel.from_pretrained("openai/clip-vit-base-patch32")
|
150 |
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
|
151 |
|
@@ -170,16 +133,12 @@ def hallucinate(prompt, num_images=64):
|
|
170 |
|
171 |
def clip_top_k(prompt, images, k=8):
|
172 |
inputs = processor(text=prompt, images=images, return_tensors="np", padding=True)
|
|
|
173 |
outputs = clip(**inputs)
|
174 |
logits = outputs.logits_per_text
|
175 |
scores = np.array(logits[0]).argsort()[-k:][::-1]
|
176 |
return [images[score] for score in scores]
|
177 |
|
178 |
-
|
179 |
-
# ## Log to wandb
|
180 |
-
|
181 |
-
from dalle_mini.helpers import captioned_strip
|
182 |
-
|
183 |
def log_to_wandb(prompts):
|
184 |
strips = []
|
185 |
for prompt in prompts:
|
|
|
4 |
import random
|
5 |
|
6 |
import jax
|
|
|
7 |
from flax.training.common_utils import shard
|
8 |
from flax.jax_utils import replicate, unreplicate
|
9 |
|
10 |
from transformers.models.bart.modeling_flax_bart import *
|
11 |
from transformers import BartTokenizer, FlaxBartForConditionalGeneration
|
12 |
|
13 |
+
import os
|
14 |
|
|
|
15 |
from PIL import Image
|
16 |
import numpy as np
|
17 |
import matplotlib.pyplot as plt
|
|
|
21 |
import torchvision.transforms.functional as TF
|
22 |
from torchvision.transforms import InterpolationMode
|
23 |
|
24 |
+
from dalle_mini.model import CustomFlaxBartForConditionalGeneration
|
25 |
from vqgan_jax.modeling_flax_vqgan import VQModel
|
26 |
|
27 |
+
# ## CLIP Scoring
|
28 |
+
from transformers import CLIPProcessor, FlaxCLIPModel
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
import wandb
|
31 |
import os
|
32 |
+
|
33 |
+
from dalle_mini.helpers import captioned_strip
|
34 |
+
|
35 |
+
|
36 |
os.environ["WANDB_SILENT"] = "true"
|
37 |
os.environ["WANDB_CONSOLE"] = "off"
|
38 |
|
39 |
+
# TODO: used for legacy support
|
40 |
+
BASE_MODEL = 'facebook/bart-large-cnn'
|
41 |
+
|
42 |
# set id to None so our latest images don't get overwritten
|
43 |
id = None
|
44 |
run = wandb.init(id=id,
|
|
|
51 |
artifact_dir = artifact.download()
|
52 |
|
53 |
# create our model
|
|
|
54 |
model = CustomFlaxBartForConditionalGeneration.from_pretrained(artifact_dir)
|
55 |
+
|
56 |
+
# TODO: legacy support (earlier models)
|
57 |
+
tokenizer = BartTokenizer.from_pretrained(BASE_MODEL)
|
58 |
model.config.force_bos_token_to_be_generated = False
|
59 |
model.config.forced_bos_token_id = None
|
60 |
model.config.forced_eos_token_id = None
|
|
|
109 |
bart_params = replicate(model.params)
|
110 |
vqgan_params = replicate(vqgan.params)
|
111 |
|
|
|
|
|
|
|
112 |
clip = FlaxCLIPModel.from_pretrained("openai/clip-vit-base-patch32")
|
113 |
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
|
114 |
|
|
|
133 |
|
134 |
def clip_top_k(prompt, images, k=8):
|
135 |
inputs = processor(text=prompt, images=images, return_tensors="np", padding=True)
|
136 |
+
# FIXME: image should be resized and normalized prior to being processed by CLIP
|
137 |
outputs = clip(**inputs)
|
138 |
logits = outputs.logits_per_text
|
139 |
scores = np.array(logits[0]).argsort()[-k:][::-1]
|
140 |
return [images[score] for score in scores]
|
141 |
|
|
|
|
|
|
|
|
|
|
|
142 |
def log_to_wandb(prompts):
|
143 |
strips = []
|
144 |
for prompt in prompts:
|