Spaces:
Running
on
Zero
Running
on
Zero
arthur-qiu
commited on
Commit
·
cca304f
1
Parent(s):
2cc3d41
clean code
Browse files- pipeline_freescale.py +0 -121
- pipeline_freescale_turbo.py +0 -121
pipeline_freescale.py
CHANGED
@@ -55,127 +55,6 @@ EXAMPLE_DOC_STRING = """
|
|
55 |
```
|
56 |
"""
|
57 |
|
58 |
-
def default(val, d):
|
59 |
-
if exists(val):
|
60 |
-
return val
|
61 |
-
return d() if isfunction(d) else d
|
62 |
-
|
63 |
-
def exists(val):
|
64 |
-
return val is not None
|
65 |
-
|
66 |
-
def extract_into_tensor(a, t, x_shape):
|
67 |
-
b, *_ = t.shape
|
68 |
-
out = a.gather(-1, t)
|
69 |
-
return out.reshape(b, *((1,) * (len(x_shape) - 1)))
|
70 |
-
|
71 |
-
def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
|
72 |
-
if schedule == "linear":
|
73 |
-
betas = (
|
74 |
-
torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
|
75 |
-
)
|
76 |
-
elif schedule == "cosine":
|
77 |
-
timesteps = (
|
78 |
-
torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
|
79 |
-
)
|
80 |
-
alphas = timesteps / (1 + cosine_s) * np.pi / 2
|
81 |
-
alphas = torch.cos(alphas).pow(2)
|
82 |
-
alphas = alphas / alphas[0]
|
83 |
-
betas = 1 - alphas[1:] / alphas[:-1]
|
84 |
-
betas = np.clip(betas, a_min=0, a_max=0.999)
|
85 |
-
elif schedule == "sqrt_linear":
|
86 |
-
betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
|
87 |
-
elif schedule == "sqrt":
|
88 |
-
betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
|
89 |
-
else:
|
90 |
-
raise ValueError(f"schedule '{schedule}' unknown.")
|
91 |
-
return betas.numpy()
|
92 |
-
|
93 |
-
to_torch = partial(torch.tensor, dtype=torch.float16)
|
94 |
-
betas = make_beta_schedule("linear", 1000, linear_start=0.00085, linear_end=0.012)
|
95 |
-
alphas = 1. - betas
|
96 |
-
alphas_cumprod = np.cumprod(alphas, axis=0)
|
97 |
-
sqrt_alphas_cumprod = to_torch(np.sqrt(alphas_cumprod))
|
98 |
-
sqrt_one_minus_alphas_cumprod = to_torch(np.sqrt(1. - alphas_cumprod))
|
99 |
-
|
100 |
-
def q_sample(x_start, t, init_noise_sigma = 1.0, noise=None, device=None):
|
101 |
-
noise = default(noise, lambda: torch.randn_like(x_start)).to(device) * init_noise_sigma
|
102 |
-
return (extract_into_tensor(sqrt_alphas_cumprod.to(device), t, x_start.shape) * x_start +
|
103 |
-
extract_into_tensor(sqrt_one_minus_alphas_cumprod.to(device), t, x_start.shape) * noise)
|
104 |
-
|
105 |
-
def get_views(height, width, h_window_size=128, w_window_size=128, h_window_stride=64, w_window_stride=64, vae_scale_factor=8):
|
106 |
-
height //= vae_scale_factor
|
107 |
-
width //= vae_scale_factor
|
108 |
-
num_blocks_height = int((height - h_window_size) / h_window_stride - 1e-6) + 2 if height > h_window_size else 1
|
109 |
-
num_blocks_width = int((width - w_window_size) / w_window_stride - 1e-6) + 2 if width > w_window_size else 1
|
110 |
-
total_num_blocks = int(num_blocks_height * num_blocks_width)
|
111 |
-
views = []
|
112 |
-
for i in range(total_num_blocks):
|
113 |
-
h_start = int((i // num_blocks_width) * h_window_stride)
|
114 |
-
h_end = h_start + h_window_size
|
115 |
-
w_start = int((i % num_blocks_width) * w_window_stride)
|
116 |
-
w_end = w_start + w_window_size
|
117 |
-
|
118 |
-
if h_end > height:
|
119 |
-
h_start = int(h_start + height - h_end)
|
120 |
-
h_end = int(height)
|
121 |
-
if w_end > width:
|
122 |
-
w_start = int(w_start + width - w_end)
|
123 |
-
w_end = int(width)
|
124 |
-
if h_start < 0:
|
125 |
-
h_end = int(h_end - h_start)
|
126 |
-
h_start = 0
|
127 |
-
if w_start < 0:
|
128 |
-
w_end = int(w_end - w_start)
|
129 |
-
w_start = 0
|
130 |
-
|
131 |
-
random_jitter = True
|
132 |
-
if random_jitter:
|
133 |
-
h_jitter_range = (h_window_size - h_window_stride) // 4
|
134 |
-
w_jitter_range = (w_window_size - w_window_stride) // 4
|
135 |
-
h_jitter = 0
|
136 |
-
w_jitter = 0
|
137 |
-
|
138 |
-
if (w_start != 0) and (w_end != width):
|
139 |
-
w_jitter = random.randint(-w_jitter_range, w_jitter_range)
|
140 |
-
elif (w_start == 0) and (w_end != width):
|
141 |
-
w_jitter = random.randint(-w_jitter_range, 0)
|
142 |
-
elif (w_start != 0) and (w_end == width):
|
143 |
-
w_jitter = random.randint(0, w_jitter_range)
|
144 |
-
if (h_start != 0) and (h_end != height):
|
145 |
-
h_jitter = random.randint(-h_jitter_range, h_jitter_range)
|
146 |
-
elif (h_start == 0) and (h_end != height):
|
147 |
-
h_jitter = random.randint(-h_jitter_range, 0)
|
148 |
-
elif (h_start != 0) and (h_end == height):
|
149 |
-
h_jitter = random.randint(0, h_jitter_range)
|
150 |
-
h_start += (h_jitter + h_jitter_range)
|
151 |
-
h_end += (h_jitter + h_jitter_range)
|
152 |
-
w_start += (w_jitter + w_jitter_range)
|
153 |
-
w_end += (w_jitter + w_jitter_range)
|
154 |
-
|
155 |
-
views.append((h_start, h_end, w_start, w_end))
|
156 |
-
return views
|
157 |
-
|
158 |
-
def gaussian_kernel(kernel_size=3, sigma=1.0, channels=3):
|
159 |
-
x_coord = torch.arange(kernel_size)
|
160 |
-
gaussian_1d = torch.exp(-(x_coord - (kernel_size - 1) / 2) ** 2 / (2 * sigma ** 2))
|
161 |
-
gaussian_1d = gaussian_1d / gaussian_1d.sum()
|
162 |
-
gaussian_2d = gaussian_1d[:, None] * gaussian_1d[None, :]
|
163 |
-
kernel = gaussian_2d[None, None, :, :].repeat(channels, 1, 1, 1)
|
164 |
-
|
165 |
-
return kernel
|
166 |
-
|
167 |
-
def gaussian_filter(latents, kernel_size=3, sigma=1.0):
|
168 |
-
channels = latents.shape[1]
|
169 |
-
kernel = gaussian_kernel(kernel_size, sigma, channels).to(latents.device, latents.dtype)
|
170 |
-
if len(latents.shape) == 5:
|
171 |
-
b = latents.shape[0]
|
172 |
-
latents = rearrange(latents, 'b c t i j -> (b t) c i j')
|
173 |
-
blurred_latents = F.conv2d(latents, kernel, padding=kernel_size//2, groups=channels)
|
174 |
-
blurred_latents = rearrange(blurred_latents, '(b t) c i j -> b c t i j', b=b)
|
175 |
-
else:
|
176 |
-
blurred_latents = F.conv2d(latents, kernel, padding=kernel_size//2, groups=channels)
|
177 |
-
|
178 |
-
return blurred_latents
|
179 |
|
180 |
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
|
181 |
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
|
|
|
55 |
```
|
56 |
"""
|
57 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
|
59 |
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
|
60 |
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
|
pipeline_freescale_turbo.py
CHANGED
@@ -55,127 +55,6 @@ EXAMPLE_DOC_STRING = """
|
|
55 |
```
|
56 |
"""
|
57 |
|
58 |
-
def default(val, d):
|
59 |
-
if exists(val):
|
60 |
-
return val
|
61 |
-
return d() if isfunction(d) else d
|
62 |
-
|
63 |
-
def exists(val):
|
64 |
-
return val is not None
|
65 |
-
|
66 |
-
def extract_into_tensor(a, t, x_shape):
|
67 |
-
b, *_ = t.shape
|
68 |
-
out = a.gather(-1, t)
|
69 |
-
return out.reshape(b, *((1,) * (len(x_shape) - 1)))
|
70 |
-
|
71 |
-
def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
|
72 |
-
if schedule == "linear":
|
73 |
-
betas = (
|
74 |
-
torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
|
75 |
-
)
|
76 |
-
elif schedule == "cosine":
|
77 |
-
timesteps = (
|
78 |
-
torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
|
79 |
-
)
|
80 |
-
alphas = timesteps / (1 + cosine_s) * np.pi / 2
|
81 |
-
alphas = torch.cos(alphas).pow(2)
|
82 |
-
alphas = alphas / alphas[0]
|
83 |
-
betas = 1 - alphas[1:] / alphas[:-1]
|
84 |
-
betas = np.clip(betas, a_min=0, a_max=0.999)
|
85 |
-
elif schedule == "sqrt_linear":
|
86 |
-
betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
|
87 |
-
elif schedule == "sqrt":
|
88 |
-
betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
|
89 |
-
else:
|
90 |
-
raise ValueError(f"schedule '{schedule}' unknown.")
|
91 |
-
return betas.numpy()
|
92 |
-
|
93 |
-
to_torch = partial(torch.tensor, dtype=torch.float16)
|
94 |
-
betas = make_beta_schedule("linear", 1000, linear_start=0.00085, linear_end=0.012)
|
95 |
-
alphas = 1. - betas
|
96 |
-
alphas_cumprod = np.cumprod(alphas, axis=0)
|
97 |
-
sqrt_alphas_cumprod = to_torch(np.sqrt(alphas_cumprod))
|
98 |
-
sqrt_one_minus_alphas_cumprod = to_torch(np.sqrt(1. - alphas_cumprod))
|
99 |
-
|
100 |
-
def q_sample(x_start, t, init_noise_sigma = 1.0, noise=None, device=None):
|
101 |
-
noise = default(noise, lambda: torch.randn_like(x_start)).to(device) * init_noise_sigma
|
102 |
-
return (extract_into_tensor(sqrt_alphas_cumprod.to(device), t, x_start.shape) * x_start +
|
103 |
-
extract_into_tensor(sqrt_one_minus_alphas_cumprod.to(device), t, x_start.shape) * noise)
|
104 |
-
|
105 |
-
def get_views(height, width, h_window_size=128, w_window_size=128, h_window_stride=64, w_window_stride=64, vae_scale_factor=8):
|
106 |
-
height //= vae_scale_factor
|
107 |
-
width //= vae_scale_factor
|
108 |
-
num_blocks_height = int((height - h_window_size) / h_window_stride - 1e-6) + 2 if height > h_window_size else 1
|
109 |
-
num_blocks_width = int((width - w_window_size) / w_window_stride - 1e-6) + 2 if width > w_window_size else 1
|
110 |
-
total_num_blocks = int(num_blocks_height * num_blocks_width)
|
111 |
-
views = []
|
112 |
-
for i in range(total_num_blocks):
|
113 |
-
h_start = int((i // num_blocks_width) * h_window_stride)
|
114 |
-
h_end = h_start + h_window_size
|
115 |
-
w_start = int((i % num_blocks_width) * w_window_stride)
|
116 |
-
w_end = w_start + w_window_size
|
117 |
-
|
118 |
-
if h_end > height:
|
119 |
-
h_start = int(h_start + height - h_end)
|
120 |
-
h_end = int(height)
|
121 |
-
if w_end > width:
|
122 |
-
w_start = int(w_start + width - w_end)
|
123 |
-
w_end = int(width)
|
124 |
-
if h_start < 0:
|
125 |
-
h_end = int(h_end - h_start)
|
126 |
-
h_start = 0
|
127 |
-
if w_start < 0:
|
128 |
-
w_end = int(w_end - w_start)
|
129 |
-
w_start = 0
|
130 |
-
|
131 |
-
random_jitter = True
|
132 |
-
if random_jitter:
|
133 |
-
h_jitter_range = (h_window_size - h_window_stride) // 4
|
134 |
-
w_jitter_range = (w_window_size - w_window_stride) // 4
|
135 |
-
h_jitter = 0
|
136 |
-
w_jitter = 0
|
137 |
-
|
138 |
-
if (w_start != 0) and (w_end != width):
|
139 |
-
w_jitter = random.randint(-w_jitter_range, w_jitter_range)
|
140 |
-
elif (w_start == 0) and (w_end != width):
|
141 |
-
w_jitter = random.randint(-w_jitter_range, 0)
|
142 |
-
elif (w_start != 0) and (w_end == width):
|
143 |
-
w_jitter = random.randint(0, w_jitter_range)
|
144 |
-
if (h_start != 0) and (h_end != height):
|
145 |
-
h_jitter = random.randint(-h_jitter_range, h_jitter_range)
|
146 |
-
elif (h_start == 0) and (h_end != height):
|
147 |
-
h_jitter = random.randint(-h_jitter_range, 0)
|
148 |
-
elif (h_start != 0) and (h_end == height):
|
149 |
-
h_jitter = random.randint(0, h_jitter_range)
|
150 |
-
h_start += (h_jitter + h_jitter_range)
|
151 |
-
h_end += (h_jitter + h_jitter_range)
|
152 |
-
w_start += (w_jitter + w_jitter_range)
|
153 |
-
w_end += (w_jitter + w_jitter_range)
|
154 |
-
|
155 |
-
views.append((h_start, h_end, w_start, w_end))
|
156 |
-
return views
|
157 |
-
|
158 |
-
def gaussian_kernel(kernel_size=3, sigma=1.0, channels=3):
|
159 |
-
x_coord = torch.arange(kernel_size)
|
160 |
-
gaussian_1d = torch.exp(-(x_coord - (kernel_size - 1) / 2) ** 2 / (2 * sigma ** 2))
|
161 |
-
gaussian_1d = gaussian_1d / gaussian_1d.sum()
|
162 |
-
gaussian_2d = gaussian_1d[:, None] * gaussian_1d[None, :]
|
163 |
-
kernel = gaussian_2d[None, None, :, :].repeat(channels, 1, 1, 1)
|
164 |
-
|
165 |
-
return kernel
|
166 |
-
|
167 |
-
def gaussian_filter(latents, kernel_size=3, sigma=1.0):
|
168 |
-
channels = latents.shape[1]
|
169 |
-
kernel = gaussian_kernel(kernel_size, sigma, channels).to(latents.device, latents.dtype)
|
170 |
-
if len(latents.shape) == 5:
|
171 |
-
b = latents.shape[0]
|
172 |
-
latents = rearrange(latents, 'b c t i j -> (b t) c i j')
|
173 |
-
blurred_latents = F.conv2d(latents, kernel, padding=kernel_size//2, groups=channels)
|
174 |
-
blurred_latents = rearrange(blurred_latents, '(b t) c i j -> b c t i j', b=b)
|
175 |
-
else:
|
176 |
-
blurred_latents = F.conv2d(latents, kernel, padding=kernel_size//2, groups=channels)
|
177 |
-
|
178 |
-
return blurred_latents
|
179 |
|
180 |
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
|
181 |
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
|
|
|
55 |
```
|
56 |
"""
|
57 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
|
59 |
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
|
60 |
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
|