Upload convert_model.py
Browse files- convert_model.py +30 -10
convert_model.py
CHANGED
@@ -39,9 +39,9 @@ class ImageEncoder(torch.nn.Module):
|
|
39 |
|
40 |
def convert_model(opts):
|
41 |
src_model = uform.get_model(opts.model_name)
|
42 |
-
input_ids = torch.ones(1,
|
43 |
-
attention_mask = torch.ones(1,
|
44 |
-
image = torch.ones(1, 3,
|
45 |
|
46 |
print('Tracing models…')
|
47 |
image_encoder = ImageEncoder(src_model.image_encoder).eval()
|
@@ -51,13 +51,18 @@ def convert_model(opts):
|
|
51 |
|
52 |
print('Converting models…')
|
53 |
|
|
|
|
|
|
|
|
|
|
|
54 |
image_encoder = ct.convert(
|
55 |
image_encoder,
|
56 |
convert_to='mlprogram',
|
57 |
inputs=[
|
58 |
ct.TensorType(
|
59 |
name='image',
|
60 |
-
shape=(
|
61 |
dtype=image.numpy().dtype
|
62 |
)],
|
63 |
outputs=[
|
@@ -71,18 +76,23 @@ def convert_model(opts):
|
|
71 |
compute_precision=ct.precision.FLOAT16 if opts.use_fp16 else ct.precision.FLOAT32
|
72 |
)
|
73 |
|
|
|
|
|
|
|
|
|
|
|
74 |
text_encoder = ct.convert(
|
75 |
text_encoder,
|
76 |
convert_to='mlprogram',
|
77 |
inputs=[
|
78 |
ct.TensorType(
|
79 |
name='input_ids',
|
80 |
-
shape=(
|
81 |
dtype=input_ids.numpy().dtype
|
82 |
),
|
83 |
ct.TensorType(
|
84 |
name='attention_mask',
|
85 |
-
shape=(
|
86 |
dtype=attention_mask.numpy().dtype
|
87 |
)],
|
88 |
outputs=[
|
@@ -110,15 +120,25 @@ if __name__ == '__main__':
|
|
110 |
type=str,
|
111 |
help='UForm model name')
|
112 |
|
113 |
-
opts.add_argument('--
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
114 |
action='store',
|
115 |
type=int,
|
116 |
-
help='lower bound of batch size')
|
117 |
|
118 |
-
opts.add_argument('--
|
119 |
action='store',
|
120 |
type=int,
|
121 |
-
help='upper bound of batch size')
|
122 |
|
123 |
opts.add_argument('-use_fp16',
|
124 |
action='store_true',
|
|
|
39 |
|
40 |
def convert_model(opts):
|
41 |
src_model = uform.get_model(opts.model_name)
|
42 |
+
input_ids = torch.ones(1, src_model.text_encoder.max_position_embeddings, dtype=torch.int32)
|
43 |
+
attention_mask = torch.ones(1, src_model.text_encoder.max_position_embeddings, dtype=torch.int32)
|
44 |
+
image = torch.ones(1, 3, src_model.image_encoder.image_size, src_model.image_encoder.image_size, dtype=torch.float32)
|
45 |
|
46 |
print('Tracing models…')
|
47 |
image_encoder = ImageEncoder(src_model.image_encoder).eval()
|
|
|
51 |
|
52 |
print('Converting models…')
|
53 |
|
54 |
+
if opts.image_batchsize_lb == opts.image_batchsize_ub:
|
55 |
+
image_batch_dim_shape = opts.image_batchsize_lb
|
56 |
+
else:
|
57 |
+
image_batch_dim_shape = ct.RangeDim(lower_bound=opts.image_batchsize_lb, upper_bound=opts.image_batchsize_ub, default=1)
|
58 |
+
|
59 |
image_encoder = ct.convert(
|
60 |
image_encoder,
|
61 |
convert_to='mlprogram',
|
62 |
inputs=[
|
63 |
ct.TensorType(
|
64 |
name='image',
|
65 |
+
shape=(image_batch_dim_shape,) + image.shape[1:],
|
66 |
dtype=image.numpy().dtype
|
67 |
)],
|
68 |
outputs=[
|
|
|
76 |
compute_precision=ct.precision.FLOAT16 if opts.use_fp16 else ct.precision.FLOAT32
|
77 |
)
|
78 |
|
79 |
+
if opts.text_batchsize_lb == opts.text_batchsize_ub:
|
80 |
+
text_batch_dim_shape = opts.text_batchsize_lb
|
81 |
+
else:
|
82 |
+
text_batch_dim_shape = ct.RangeDim(lower_bound=opts.text_batchsize_lb, upper_bound=opts.text_batchsize_ub, default=1)
|
83 |
+
|
84 |
text_encoder = ct.convert(
|
85 |
text_encoder,
|
86 |
convert_to='mlprogram',
|
87 |
inputs=[
|
88 |
ct.TensorType(
|
89 |
name='input_ids',
|
90 |
+
shape=(text_batch_dim_shape,) + input_ids.shape[1:],
|
91 |
dtype=input_ids.numpy().dtype
|
92 |
),
|
93 |
ct.TensorType(
|
94 |
name='attention_mask',
|
95 |
+
shape=(text_batch_dim_shape,) + attention_mask.shape[1:],
|
96 |
dtype=attention_mask.numpy().dtype
|
97 |
)],
|
98 |
outputs=[
|
|
|
120 |
type=str,
|
121 |
help='UForm model name')
|
122 |
|
123 |
+
opts.add_argument('--text_batchsize_lb',
|
124 |
+
action='store',
|
125 |
+
type=int,
|
126 |
+
help='lower bound of batch size for text encoder')
|
127 |
+
|
128 |
+
opts.add_argument('--text_batchsize_ub',
|
129 |
+
action='store',
|
130 |
+
type=int,
|
131 |
+
help='upper bound of batch size for text encoder')
|
132 |
+
|
133 |
+
opts.add_argument('--image_batchsize_lb',
|
134 |
action='store',
|
135 |
type=int,
|
136 |
+
help='lower bound of batch size for image encoder')
|
137 |
|
138 |
+
opts.add_argument('--image_batchsize_ub',
|
139 |
action='store',
|
140 |
type=int,
|
141 |
+
help='upper bound of batch size for image encoder')
|
142 |
|
143 |
opts.add_argument('-use_fp16',
|
144 |
action='store_true',
|